<\/a>Newly Added Models<\/strong><\/h3>\n\n\n\n\n- NVIDIA Cosmos3-Edge<\/li>\n\n\n\n
- NVIDIA Cosmos-Reason2-2B<\/li>\n\n\n\n
- Gemma-4-E2B-it<\/li>\n\n\n\n
- Qwen3-VL-2B-Instruct<\/li>\n<\/ul>\n\n\n\n
![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/07\/Screenshot-2026-07-16-224332-1024x540.png\")
<\/figure>\n\n\n\nFigure 1. Public leaderboard highlighting the newly added models.<\/em><\/p>\n\n\n\n
\n\n\n\n<\/a>Why Edge Models Need Their Own Evaluation<\/strong><\/h2>\n\n\n\nEdge models aren’t simply smaller versions of frontier-scale systems\u2014they’re designed around a fundamentally different engineering objective.<\/p>\n\n\n\n
Rather than maximizing reasoning capability with virtually unlimited compute, edge models must maximize reasoning within a fixed deployment budget<\/strong>. Latency, memory footprint, power consumption, and hardware availability all become part of the optimization problem. Whether deployed on warehouse robots, manufacturing inspection systems, intelligent traffic cameras, or autonomous platforms, these models often need to process information where it is generated instead of relying on cloud-scale infrastructure.<\/p>\n\n\n\nThat changes what a benchmark should measure. It’s no longer just how well<\/strong> a model reasons, but how effectively<\/strong> it reasons under realistic operational constraints. Evaluating compact reasoning models alongside larger systems provides a more complete understanding of the trade-offs between capability, efficiency, and deployability\u2014an increasingly important consideration as physical AI moves from research to production.<\/p>\n\n\n\n<\/a>Growing with the Field<\/strong><\/p>\n\n\n\nThis update represents more than the addition of four reasoning models\u2014it reflects how rapidly the multimodal AI ecosystem is evolving and the need for benchmarks to evolve alongside it.<\/p>\n\n\n\n
Among the newly evaluated models, Cosmos3 Edge as a reasoning vision language model (VLM) achieved the highest overall performance within the evaluated ~2B parameter class<\/strong>, establishing a strong baseline for this emerging generation of compact vision-language models. More importantly, evaluating these models under the same operational tasks and methodology enables meaningful comparisons within similar deployment budgets, helping researchers better understand the trade-offs between model capability and deployment efficiency. Full results for all four models are available on the VANTAGE-Bench leaderboard.<\/p>\n\n\n\nAt Clemson, we view VANTAGE-Bench<\/strong><\/a> as research infrastructure rather than a static benchmark<\/strong>. Maintaining an operational benchmark is an ongoing effort\u2014one that requires continuously integrating new model families, architectures, and deployment paradigms as the field advances. As Physical AI continues to evolve, VANTAGE-Bench will continue evolving with it, providing the research community with a consistent, reproducible framework for evaluating the next generation of vision-language models.<\/p>\n","protected":false},"excerpt":{"rendered":"The rapid advancement of multimodal AI is reshaping how intelligent systems interact with the physical world. Vision language models (VLMs) are no longer limited to cloud-based assistants\u2014they are increasingly being deployed in warehouses, manufacturing facilities, traffic monitoring systems, robotics, and other physical AI applications where decisions must be made in real time. As these deployments […]<\/p>\n","protected":false},"author":339,"featured_media":295,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-292","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news-item"],"fimg_url":"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/07\/Students-150x150.jpg","_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/292","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=292"}],"version-history":[{"count":1,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/292\/revisions"}],"predecessor-version":[{"id":296,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/292\/revisions\/296"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media\/295"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=292"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=292"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=292"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=292"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":287,"date":"2026-06-01T01:30:00","date_gmt":"2026-06-01T01:30:00","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=287"},"modified":"2026-05-31T16:22:57","modified_gmt":"2026-05-31T16:22:57","slug":"vantage-bench-evaluating-vision-language-models-in-real-world-operational-environments","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/vantage-bench-evaluating-vision-language-models-in-real-world-operational-environments\/","title":{"rendered":"VANTAGE-Bench: Evaluating Vision-Language Models in Real-World Operational Environments"},"content":{"rendered":"\n
Vision-Language Models (VLMs) are rapidly becoming central to the next generation of artificial intelligence systems, powering technologies that combine visual understanding with natural language reasoning. As these systems move beyond research demonstrations and into applications involving transportation systems, logistics networks, robotics platforms, industrial facilities, and public infrastructure, questions about how they should be evaluated are becoming increasingly important.<\/p>\n\n\n\n
Many existing VLM benchmarks rely on consumer-centric source material \u2013 egocentric recordings, broadcast media, web-curated video \u2013 while reducing evaluation to multiple-choice questions that effectively provide the model with the correct answer as one of the options. But operational environments do not arrive pre-framed, and the decisions they demand are rarely multiple choice. They are continuous, dynamic, and often safety- or mission-critical.<\/p>\n\n\n\n
VANTAGE-Bench (Video Analysis Tasks Across Generalized Environments) is a new public benchmark suite and leaderboard developed to evaluate how Vision-Language Models perform across the range of capabilities required for fixed-camera, infrastructure-scale deployment. This includes a previously unmeasured capability: continuous single-object tracking through pure vision-language reasoning. The benchmark is designed to help researchers and model developers better understand how these systems perform in dynamic operational environments where reliability and continuity matter.<\/p>\n\n\n\n
The Infrastructure AI Gap<\/strong><\/p>\n\n\n\nFor more than a decade, progress in vision-language modeling has largely been measured against internet-scale datasets built from social media imagery, curated video, consumer photography, and other highly photographed digital environments. Models have become remarkably capable on this material. But the environments where these systems are increasingly expected to function \u2013 distribution centers, intersections, transit corridors, ports, public facilities, and industrial floors \u2013 often look very different from the datasets used to train and evaluate them.<\/p>\n\n\n\n
These are fixed-camera worlds. They are dense, continuous, and operationally consequential. A missed forklift is not a missed caption. A misread pedestrian is not a misread meme.<\/p>\n\n\n\n
When VANTAGE-Bench’s authors measured frontier models on infrastructure-oriented tasks against their reported scores on standard consumer-centric benchmarks, the degradation was substantial. A representative open-weight model dropped nearly 30 points on causal event verification compared to its score on MLVU, and roughly 24 points on 2D spatial pointing compared to BLINK.<\/p>\n\n\n\n
The cause appears structural: internet-trained models rely on a “cinematic prior” characterized by human-centric framing, high-velocity motion, and edited temporal structure \u2013 assumptions that do not exist in elevated, wide-angle, fixed-camera footage. Without that prior, models must reason from pure spatial-temporal logic, and many struggle to do so reliably.<\/p>\n\n\n\n
The research community has increasingly recognized this disconnect as part of the broader Infrastructure AI Gap<\/em>: the growing distance between benchmark performance and operational deployment. VANTAGE-Bench was created to help measure that gap more directly by evaluating how VLMs perform in dynamic operational scenarios rather than only in familiar benchmark environments.<\/p>\n\n\n\nIn doing so, it shifts the central evaluation question from \u201cCan this model perform well on familiar benchmark data?\u201d<\/em> to a more operationally meaningful one: \u201cCan this model be trusted in the environments where it will actually be used?\u201d<\/em><\/p>\n\n\n\nBenchmark Architecture and Evaluation Design<\/strong><\/p>\n\n\n\nVANTAGE-Bench is organized around four reasoning pillars, each targeting a distinct capability required for reliable Infrastructure AI. Across these pillars, the suite evaluates eight task formulations spanning both image and video modalities, moving deliberately beyond the multiple-choice format that dominates many existing benchmarks.<\/p>\n\n\n\n
Pillar I \u2013 Semantic Understanding (What and Why)<\/strong><\/p>\n\n\n\nTwo tasks evaluate high-level causal and operational reasoning.<\/p>\n\n\n\n
\n- Event Verification<\/strong> asks the model to confirm or reject operational hypotheses against visual evidence, such as determining whether a tailgating event occurred at an entrance.<\/li>\n\n\n\n
- Video Question Answering<\/strong> evaluates multi-step logical reasoning over untrimmed video sequences.<\/li>\n<\/ol>\n\n\n\n
Pillar II \u2013 Spatial Understanding (Where)<\/strong><\/p>\n\n\n\nThree tasks evaluate precise geometric grounding in dense, multi-instance scenes.<\/p>\n\n\n\n
\n- Referring Expressions<\/strong> requires the model to localize a target described in natural language, for example, “the black motorcycles near the bottom right”, by outputting bounding box coordinates.<\/li>\n\n\n\n
- Spatial Pointing<\/strong> tests coordinate selection from positional prompts.<\/li>\n\n\n\n
- 2D Object Localization<\/strong> establishes a baseline for class-level spatial awareness by requiring the model to identify every instance of a given category and report results in structured JSON format.<\/li>\n<\/ol>\n\n\n\n
Pillar III \u2013 Temporal Understanding (When)<\/strong><\/p>\n\n\n\nTwo tasks evaluate the perception of action duration and event boundaries.<\/p>\n\n\n\n
\n- Temporal Localization<\/strong> requires the model to predict precise start and end timestamps for a queried event.<\/li>\n\n\n\n
- Dense Video Captioning<\/strong> requires the model to autonomously segment a video into constituent events and describe each one. This generative task helps expose whether models can localize what they describe or whether their descriptions remain temporally ungrounded.<\/li>\n<\/ol>\n\n\n\n
Pillar IV \u2013 Spatio-Temporal Understanding (Dynamics)<\/strong><\/p>\n\n\n\nThe benchmark introduces a quantitative Single-Object Tracking evaluation for VLMs.<\/p>\n\n\n\n
\n- The model is shown a target bounding box overlaid on the first frame and, in a single inference pass, must output a continuous coordinate trajectory across the remaining frames.<\/li>\n\n\n\n
- Unlike classical tracking pipelines, no rolling memory, motion model, or frame-by-frame state update is permitted. This isolates a model’s emergent ability to maintain spatial context purely through vision-language reasoning. Standardized evaluation for this capability has remained limited, despite growing interest in deploying multimodal AI systems into physical operational environments.<\/li>\n<\/ul>\n\n\n\n
Together, these eight tasks span 3,346 media assets and 35,027 expert annotations, drawn from three operational domains: Warehouse and Logistics, Transportation, and Smart Spaces. All footage originates from fixed-infrastructure cameras that are static, elevated, and wide-angle \u2013 categorically distinct from the egocentric and broadcast-style sources that dominate existing benchmarks.<\/p>\n\n\n\n
Dataset Construction and Annotation Pipeline<\/strong><\/p>\n\n\n\nThe benchmark is primarily human-labeled by trained professional annotators rather than crowdsourced contributors, supported by a multi-tier quality pipeline in which 100% of initial annotations are reviewed by a secondary expert and a random 10% subset is audited by a third-tier QA reviewer.<\/p>\n\n\n\n
Programmatic methods are used only where they can be anchored to human-verified data, allowing the dataset to scale without introducing substantial annotation noise.<\/p>\n\n\n\n
To responsibly evaluate safety-critical events such as collisions and industrial accidents, scenarios that are both rare and heavily restricted by privacy regulations, approximately 20% of the Video QA and Temporal splits use high-fidelity synthetic footage generated through NVIDIA Omniverse<\/a> DRIVE Sim. This allows the benchmark to cover severe long-tail anomalies without ethical compromise.<\/p>\n\n\n\nAll real-world footage was sourced under explicit informed consent, with personally identifiable information removed through automated obfuscation followed by human-in-the-loop verification.<\/p>\n\n\n\n
Public Infrastructure and Benchmark Operations<\/strong><\/p>\n\n\n\nVANTAGE-Bench is built for the researchers advancing multimodal AI, the model developers preparing systems for deployment, and the organizations whose facilities, infrastructure, and public spaces may ultimately depend on those systems working as promised.<\/p>\n\n\n\n
The platform accepts model submissions, evaluates them through a standardized testing pipeline built on an extended version of the open-source VLMEvalKit harness, and maintains a public leaderboard reporting benchmark results.<\/p>\n\n\n\n
Every submission is evaluated against a versioned dataset and evaluation framework designed to support transparent and reproducible comparison over time. As the dataset and evaluation suite evolve, benchmark versions remain clearly identified to preserve scientific consistency and longitudinal comparability.<\/p>\n\n\n\n![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/05\/release2-1024x875.png\")
<\/figure>\n\n\n\nThe open-source benchmark codebase is available at https:\/\/github.com\/Clemson-Capstone\/VANTAGE-Bench<\/a>, and the public leaderboard is hosted at https:\/\/huggingface.co\/spaces\/clemson-computing\/VANTAGE-Bench-Leaderboard<\/a>.<\/p>\n\n\n\nBuilding an Open Evaluation Community<\/strong><\/p>\n\n\n\nThe dataset behind VANTAGE-Bench began through an initial collaboration with NVIDIA and is designed to continue expanding over time. Clemson University\u2019s School of Computing now independently hosts and maintains the benchmark, including the public evaluation infrastructure, leaderboard, and future dataset development.<\/p>\n\n\n\n
Researchers, practitioners, and organizations working in operational environments are invited to contribute new data that expands the diversity of scenes, domains, and environmental conditions represented within the benchmark. Model developers across academia and industry are also encouraged to submit Vision-Language Models to the public leaderboard and participate in broader efforts to establish shared evaluation standards for operational AI systems.<\/p>\n\n\n\n
The dataset is publicly available at https:\/\/huggingface.co\/datasets\/nvidia\/PhysicalAI-VANTAGE-Bench<\/a>, and model submissions can be made through the public submission portal at https:\/\/vantage-bench.org<\/a>.<\/p>\n\n\n\nWhy Clemson, and Why Now?<\/strong><\/p>\n\n\n\nClemson University has long emphasized research connected to real-world systems and societal impact. VANTAGE-Bench reflects that broader approach to artificial intelligence \u2013 not only advancing AI capabilities, but also helping build the infrastructure needed to evaluate those capabilities responsibly.<\/p>\n\n\n\n
The School of Computing views open evaluation as an important part of the scientific process. Public benchmarks maintained by independent research institutions can help provide transparency, reproducibility, and continuity in a rapidly evolving field where evaluation standards are still actively developing.<\/p>\n\n\n\n
This effort also aligns with Clemson\u2019s broader investment in artificial intelligence research, operational AI systems, and interdisciplinary collaboration across areas including computer vision, machine learning, robotics, intelligent systems, and computational imaging.<\/p>\n\n\n\n
Defining the Next Phase of AI Evaluation<\/strong><\/p>\n\n\n\nThe future of AI will not be defined solely by who builds the largest models. Increasingly, it will also be shaped by who develops rigorous evaluation standards capable of measuring whether those systems can operate reliably in the physical world.<\/p>\n\n\n\n
VANTAGE-Bench is one contribution toward that goal, and an invitation to the broader research and industry community to help define how operational AI systems should be evaluated in the years ahead.<\/p>\n\n\n\n
The future of AI will not be decided only by what models can say. It will be shaped by what they can reliably understand in the places where understanding matters most. VANTAGE-Bench is built to measure that, and Clemson is building it to last.<\/p>\n","protected":false},"excerpt":{"rendered":"
Vision-Language Models (VLMs) are rapidly becoming central to the next generation of artificial intelligence systems, powering technologies that combine visual understanding with natural language reasoning. As these systems move beyond research demonstrations and into applications involving transportation systems, logistics networks, robotics platforms, industrial facilities, and public infrastructure, questions about how they should be evaluated are […]<\/p>\n","protected":false},"author":339,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-287","post","type-post","status-publish","format-standard","hentry","category-news-item"],"fimg_url":false,"_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/287","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=287"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/287\/revisions"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=287"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=287"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=287"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=287"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":280,"date":"2026-03-02T17:00:24","date_gmt":"2026-03-02T17:00:24","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=280"},"modified":"2026-03-02T17:00:24","modified_gmt":"2026-03-02T17:00:24","slug":"duke-energy-science-nights-expand-stem-engagement-across-south-carolina","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/duke-energy-science-nights-expand-stem-engagement-across-south-carolina\/","title":{"rendered":"Duke Energy Science Nights Expand STEM Engagement Across South Carolina"},"content":{"rendered":"\n![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/03\/Picture1.png\")
<\/figure>\n\n\n\n<\/p>\n\n\n\n
Excitement is building across South Carolina as 60 schools prepare to host Duke Energy Science Nights during SC STEM Education Month 2026. Through a continued partnership among Duke Energy, South Carolina\u2019s Coalition for Mathematics & Science (SCCMS), and the Morehead Planetarium and Science Center, these hands-on family science events will engage students and caregivers in interactive experiences designed to spark curiosity and inspire future STEM careers.<\/p>\n\n\n\n
Now in its fifth year, Duke Energy Science Nights have continued to grow in reach and impact since launching in South Carolina in 2022. This year\u2019s program marks an expansion from 50 to 60 schools being awarded kits thanks to increased support from the Duke Energy Foundation. This program growth is in response to the increased demand from school applications and a shared commitment from program partners to provide equitable access to high-quality STEM learning opportunities across the state.<\/p>\n\n\n\n
STEM\u2014science, technology, engineering and mathematics\u2014plays a critical role in preparing students for future careers, particularly in South Carolina, where advanced manufacturing, energy, and technology-driven industries are central to the state\u2019s economy. Duke Energy Science Nights help students see the relevance of STEM learning beyond the classroom while encouraging family engagement and community connection.<\/p>\n\n\n\n
The science night kits are funded by a grant from the Duke Energy Foundation and are provided at no cost to participating schools. Each kit includes materials to support up to 200 participants in 10 hands-on activities, along with activity guides, promotional materials, downloadable digital resources (including Spanish-language versions), and virtual support. The kits are produced by the Morehead Planetarium and Science Center\u2019s NCSciFest team in collaboration with SCCMS staff, while SCCMS facilitates the application process and provides ongoing planning and communication support.<\/p>\n\n\n\n
\u201cPrograms like this reinforce our company\u2019s commitment to create a more vibrant economy by investing in our future workforce early on,\u201d said Amanda Dow, Foundation Manager for Duke Energy in South Carolina. \u201cWe know firsthand that STEM is the backbone to many career opportunities, including those here at Duke Energy and we\u2019re proud to support these students and schools.”<\/p>\n\n\n\n
Participating Schools for 2026:<\/p>\n\n\n\n
\n- Westwood Elementary, Abbeville<\/li>\n\n\n\n
- Flat Rock Elementary, Anderson<\/li>\n\n\n\n
- Varennes Elementary, Anderson<\/li>\n\n\n\n
- Nevitt Forest Elementary, Anderson<\/li>\n\n\n\n
- Marshall Primay, Belton<\/li>\n\n\n\n
- Bennettsville Primary School, Bennettsville<\/li>\n\n\n\n
- Lee Central Elementary School, Bishopville<\/li>\n\n\n\n
- John C. Calhoun Elementary, Calhoun Falls<\/li>\n\n\n\n
- Pine Tree Hill Elementary, Camden<\/li>\n\n\n\n
- Great Falls Elementary, Chester<\/li>\n\n\n\n
- Clinton Elementary, Clinton<\/li>\n\n\n\n
- Clio Elementary, Clio<\/li>\n\n\n\n
- Mayo Elementary, Cowpens<\/li>\n\n\n\n
- J.L. Cain Elementary, Darlington<\/li>\n\n\n\n
- Gordon Elementary, Dillon<\/li>\n\n\n\n
- Crosswell Elementary, Easley<\/li>\n\n\n\n
- Dewey L. Carter Elementary, Effingham<\/li>\n\n\n\n
- Henry Timrod Elementary, Florence<\/li>\n\n\n\n
- North Vista Elementary School, Florence<\/li>\n\n\n\n
- B.D. Lee Elementary, Gaffney<\/li>\n\n\n\n
- Midland Elementary, Galivants Ferry<\/li>\n\n\n\n
- Sampit Elementary, Georgetown<\/li>\n\n\n\n
- Greeleyville STEAM Academy, Greeleyville<\/li>\n\n\n\n
- Robert E. Cashion Elementary, Greenville<\/li>\n\n\n\n
- Eleanor S. Rice Elementary, Greenwood<\/li>\n\n\n\n
- Lakeview Elementary, Greenwood<\/li>\n\n\n\n
- Heath Springs Elementary, Heath Springs<\/li>\n\n\n\n
- Johnsonville Elementary, Johnsonville<\/li>\n\n\n\n
- Kenneth Gardner Leadership Academy, Kingstree<\/li>\n\n\n\n
- Main Street Elementary School of Arts and Leadership, Lake City<\/li>\n\n\n\n
- E.B. Morse Elementary, Laurens<\/li>\n\n\n\n
- Ford Elementary, Laurens<\/li>\n\n\n\n
- Lugoff Elementary School, Lugoff<\/li>\n\n\n\n
- Lyman Elementary, Lyman<\/li>\n\n\n\n
- McBee Elementary, McBee<\/li>\n\n\n\n
- McCorminck Elementary, Mullins<\/li>\n\n\n\n
- Petersburg Primary, Pageland<\/li>\n\n\n\n
- West Pelzer Elementary, Pelzer<\/li>\n\n\n\n
- Hagood Elementary, Pickens<\/li>\n\n\n\n
- Prosperity Rikard Elementary, Prosperity<\/li>\n\n\n\n
- Lewisville Elementary, Richburg<\/li>\n\n\n\n
- Roebuck Elementary, Roebuck<\/li>\n\n\n\n
- Saluda Elementary, Saluda<\/li>\n\n\n\n
- Scranton Elementary STEAM Academy, Scranton<\/li>\n\n\n\n
- Blue Ridge elementary, Seneca<\/li>\n\n\n\n
- Meeting Street Academy, Spartanburg<\/li>\n\n\n\n
- High Point Academy, Spartanburg<\/li>\n\n\n\n
- Dr. Rose H. Wilder Elementary, Summerton<\/li>\n\n\n\n
- Crosswell Drive Elementary, Sumter<\/li>\n\n\n\n
- R.E. Davis College Preparatory Academy, Sumter<\/li>\n\n\n\n
- Willow Drive Elementary, Sumter<\/li>\n\n\n\n
- Townville Elementary, Townville<\/li>\n\n\n\n
- Walker Gamble Elementary, Turbeville<\/li>\n\n\n\n
- Foster Park Elementary, Union<\/li>\n\n\n\n
- James M. Brown Elementary, Walhalla<\/li>\n\n\n\n
- Ware Shoals Primary, Ware Shoals<\/li>\n\n\n\n
- Waterloo Elementary, Waterloo<\/li>\n\n\n\n
- Wellford Academy of Science and Technology, Wellford<\/li>\n\n\n\n
- Fairfield Elementary, Winnsboro<\/li>\n\n\n\n
- Cotton Belt Elementary, York<\/li>\n<\/ul>\n\n\n\n
Over the coming weeks, participating schools will finalize event details and prepare to welcome families for evenings centered on exploration, discovery, and shared learning. Reflecting on the impact of past events, Willette Sheard, a teacher at Dr. Rose H. Wilder Elementary shared, \u201cIt\u2019s a chance to spark curiosity, encourage problem-solving, and make learning exciting outside of the classroom. This event also helps families feel connected to their child\u2019s education while building community. Most of all, it inspires kids to see themselves as future scientists, engineers, and problem-solvers.\u201d
Katherine Mulholland, Executive Director of SCCMS, states, \u201cDuke Energy Science Nights give students and families engaging, hands-on STEM experiences that spark curiosity and connect learning to real-world careers. As South Carolina continues to grow and evolve, these events play an important role in inspiring the next generation of problem-solvers and preparing our future workforce. By exposing students early to STEM pathways, we are helping build a strong pipeline of talent that will power our state\u2019s economy for years to come\u201d<\/p>\n\n\n\n
SC STEM Education Month launches annually on Pi Day, March 14, and is sponsored by South Carolina\u2019s Coalition for Mathematics & Science, a statewide organization housed within Clemson University\u2019s College of Engineering, Computing and Applied Sciences.<\/p>\n\n\n\n
Learn more at scstemmonth.org.<\/a><\/p>\n\n\n\nContacts:
Katherine Mulholland,<\/strong> kmulholland@sccoalition.org<\/a>
Elena Stout,<\/strong> estout@s2temsc.org<\/a> <\/p>\n","protected":false},"excerpt":{"rendered":"Excitement is building across South Carolina as 60 schools prepare to host Duke Energy Science Nights during SC STEM Education Month 2026. Through a continued partnership among Duke Energy, South Carolina\u2019s Coalition for Mathematics & Science (SCCMS), and the Morehead Planetarium and Science Center, these hands-on family science events will engage students and caregivers in […]<\/p>\n","protected":false},"author":339,"featured_media":281,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-280","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news-item"],"fimg_url":"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/03\/Picture1-150x135.png","_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/280","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=280"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/280\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media\/281"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=280"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=280"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=280"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=280"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":270,"date":"2026-01-15T00:28:35","date_gmt":"2026-01-15T00:28:35","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=270"},"modified":"2026-01-15T16:28:57","modified_gmt":"2026-01-15T16:28:57","slug":"small-models-big-capabilities","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/small-models-big-capabilities\/","title":{"rendered":"Small Models, Big Capabilities: Clemson and NVIDIA\u2019s Work on Video Tool Calling and Long-Horizon Reasoning"},"content":{"rendered":"\n
\n\n\n\nIn July 2025, Capgemini published a global report projecting that AI agents will deliver more than $450B in economic value by 2028<\/strong>, with 96% of surveyed organizations already experimenting with agentic systems. Industry momentum is unmistakable: enterprises are rapidly shifting from passive large language model (LLM) chatbots to active, tool-using AI agents that can retrieve data, execute workflows, and make decisions.<\/p>\n\n\n\nNVIDIA\u2019s position aligns with this shift. In its paper Small Language Models Are the Future of Agentic AI<\/em>, NVIDIA argues that specialized Small Language Models (SLMs) – not monolithic generalist LLMs – will form the backbone of enterprise agents.<\/p>\n\n\n\nBut the most valuable agents are not limited to text. Real-world deployments increasingly require multimodal understanding, particularly in robotics and autonomous systems where perception drives action. Here, NVIDIA\u2019s Cosmos Stack provides the foundation:<\/p>\n\n\n\n
\n- Cosmos Reason 1<\/strong><\/a>[GU1]<\/a> (7B)<\/strong> for physical, common-sense reasoning across time and space, and<\/li>\n\n\n\n
- Cosmos-Embed1<\/strong> for efficient video and multimodal retrieval.<\/li>\n<\/ul>\n\n\n\n
Yet both SLMs and small vision language models (VLMs) for physical applications carry known limitations.<\/p>\n\n\n\n
\n- SLMs, despite their efficiency, often show reduced tool-calling accuracy, especially for vision-heavy tasks.<\/li>\n\n\n\n
- Cosmos Reason 1, while strong on short-form video and physical reasoning, struggles with long-horizon temporal understanding due to its training distribution.<\/li>\n<\/ul>\n\n\n\n
Our project, a collaboration between NVIDIA and Clemson University, set out to evaluate and patch these weaknesses. We pursued a two-part investigation:<\/p>\n\n\n\n
\n- Finetune Qwen2.5-7B-Instruct<\/strong> to improve multimodal tool calling for enterprise video retrieval, demonstrating that small models can serve as reliable agent controllers.<\/li>\n\n\n\n
- Extend Cosmos Reason 1<\/strong>\u2019s short-form training bias using Cosmos-Embed 1<\/strong> inside a RAG + agentic architecture, enabling long-form video understanding without retraining the underlying model.<\/li>\n<\/ul>\n\n\n\n
Together, these efforts demonstrate a central thesis: Small Language Models, when paired with retrieval and targeted finetuning, can deliver competitive agentic performance across both enterprise and physical AI domains.<\/p>\n\n\n\n
This post details how we built, evaluated, and validated these improvements.<\/p>\n\n\n\n
Approach<\/strong><\/p>\n\n\n\n
\n\n\n\nTo evaluate and extend the capabilities of small multimodal models, we built our system with NVIDIA Metropolis VSS and NVIDIA NeMo Agent Toolkit (NAT)<\/strong> – an agent coordination framework designed to orchestrate tool-using LLMs with minimal overhead. <\/a>[GU2]<\/a> NAT integrates cleanly with existing agent ecosystems such as LangChain, while providing several advantages critical to this project: built-in agent evaluation, prompt optimization utilities, structured tool-trajectory visualization, and a lightweight UI for rapid iteration.<\/p>\n\n\n\nWe hosted all experimentation on NVIDIA Brev<\/strong>, deploying multiple A100 <\/strong>instances to support both model fine tuning and long-video RAG indexing workloads. From this shared infrastructure, the collaboration proceeded along two parallel technical paths:<\/p>\n\n\n\n\n- Finetuning Path: Improving SLM Tool Calling<\/em><\/li>\n<\/ol>\n\n\n\n
This path focused on enhancing the reliability of small language models as agent orchestrators. We investigated whether targeted optimization and reinforcement-based finetuning could improve structured tool invocation, positioning small models as viable controllers for enterprise agent workflows.<\/p>\n\n\n\n
\n- Physical AI Path: Patching Short-Form Video Bias<\/em><\/li>\n<\/ol>\n\n\n\n
The physical AI path focused on extending short-form reasoning models to better understand long-form videos. Instead of scaling model size or increasing context length, we approached temporal reasoning as a coordination problem. NAT provided the structure we needed to treat long-video understanding as an agentic workflow, where retrieval and reasoning are composed iteratively rather than executed in a single pass.<\/p>\n\n\n\n
Within NAT, long videos are broken down into more manageable chunks that can be indexed, retrieved, and reasoned over using a ReAct agent to orchestrate these steps. This design allows Cosmos Reason 1 to keep its strengths in short-term physics and spatial reasoning while still being effective over an extended period of time.<\/p>\n\n\n\n
Methods<\/strong><\/p>\n\n\n\n
\n\n\n\nKeeping consistent with this two track approach, each experiment retained its own system architecture within each respective Brev instance.<\/p>\n\n\n\n
\n- Finetuning Path: Improving SLM Tool Calling<\/em><\/li>\n<\/ol>\n\n\n\n
We implemented a ReAct-based agent configured to interact with a set of Metropolis VSS-style tools designed to emulate warehouse monitoring analytics APIs. The tool suite included functions such as get_all_sensor_ids, list_incidents, and get_fov_object_counts, each requiring structured arguments and returning deterministic outputs. Tool schemas and response formats were fixed across all experiments to ensure reproducibility.<\/p>\n\n\n\n
The primary model evaluated was Qwen2.5-7B-Instruct. For reference, Nemotron-Super-49B and GPT-4.1 were evaluated under the same agent configurations. All models used identical prompts, tool definitions, inference parameters, and agent logic.<\/p>\n\n\n\n
An evaluation dataset consisting of natural-language queries paired with true tool-call traces, in which each trace specifies the expected sequence of tool calls and corresponding arguments, was used to benchmark the full agent trajectories from the aforementioned models.<\/p>\n\n\n\n
The improvement procedure proceeded in two stages. First, NAT\u2019s optimizer feature was used to tune agent-level hyperparameters and system prompts. This stage targeted improvements in tool selection consistency, argument formatting, and call sequencing.<\/p>\n\n\n\n
Second, weight-level adaptation was performed using Agent Reinforcement Trainer (ART). ART applies reinforcement learning over complete agent trajectories, with rewards computed based on alignment between model-generated tool calls and true traces. Training rollouts were generated by using a synthetic task generator that produced diverse natural-language queries paired with known tool-call patterns.<\/p>\n\n\n\n
Evaluation metrics included exact-match accuracy over full tool-call trajectories, as well as an efficiency-weighted accuracy metric. For each evaluation query, the dataset defined a set of acceptable tool-call traces with associated efficiency scores in the range [0,1], where a score of 1.0 corresponds to the most efficient valid trajectory. Model outputs that matched a correct but suboptimal trace were counted as correct under exact-match accuracy, while receiving a proportionally reduced score under the efficiency-weighted metric. All evaluations were conducted using the same query set and agent configuration to isolate the effects of optimization and finetuning.<\/p>\n\n\n\n
\n- Physical AI Path: Patching Short-Form Video Bias<\/em><\/li>\n<\/ol>\n\n\n\n
Over the course of 16 weeks, we primarily worked with two models:<\/p>\n\n\n\n
\n- Cosmos Reason 1 (7B)<\/strong> is a model that specializes in understanding time, space, and physics, due to training on short form cause-and-effect video content<\/li>\n\n\n\n
- Cosmos Embed 1 (448p)<\/strong> is a multimodal embedding model designed to convert video into dense vectors, curated for physical understanding<\/li>\n<\/ul>\n\n\n\n
Cosmos Reason 1 serves as a foundational model within NVIDIA’s Metropolis vision for physical AI – artificial intelligence systems capable of comprehending fundamental physics, spatial relationships, and temporal dynamics to enable common-sense reasoning. The model achieves this capability through fine-tuning on short-form video data, enabling it to infer physical cause-and-effect relationships within brief temporal windows. However, this training methodology introduces a temporal bias that limits performance on long-form video understanding tasks.<\/p>\n\n\n\n
To mitigate this limitation, we developed a ReAct-based agentic architecture within NVIDIA AI Toolkit (NAT) that orchestrates three specialized retrieval tools, inspired by Microsoft\u2019s Deep Video Discovery, and reformed for our use:<\/p>\n\n\n\n
\n- Global Tool: <\/strong>During video ingestion, Cosmos Reason 1 performs object tracking across temporal segments to construct a subject registry – a JSON-formatted index recording entry and exit timestamps for tracked entities. At query time, the Global Tool augments this registry with an event registry, leveraging physical reasoning capabilities to identify and temporally localize events involving tracked subjects.<\/li>\n\n\n\n
- Clip Tool:<\/strong> Videos are preprocessed into 5-second clips sampled at 2 FPS. These clips are encoded using Cosmos Embed 1 and stored in a FAISS HNSW index for efficient similarity search, while corresponding metadata is maintained in a PostgreSQL database. At inference, the system performs cosine similarity retrieval biased toward identified subjects, extracts relevant clip metadata from PostgreSQL to determine the relevant timestamp of each clip, and generates candidate answers with associated temporal ranges.<\/li>\n\n\n\n
- Frame Tool: <\/strong>When the Clip Tool produces insufficient results, the Frame Tool leverages the returned temporal range to extract a uniformly sampled set of 50 frames. These frames are processed by Cosmos Reason 1 for direct visual question answering.<\/li>\n<\/ul>\n\n\n\n
A ReAct agent orchestrates these tools by evaluating query alignment, tool execution history, candidate answers, and tool-reported confidence scores to determine response completeness. Upon convergence, the agent invokes a “finish” tool to return the final answer to the user.<\/p>\n\n\n\n
We evaluated this agentic workflow using LongVideoBench, a benchmark designed to assess both long-form video comprehension and multi-faceted reasoning capabilities including spatial, temporal, and event-based understanding. Results are presented below.<\/p>\n\n\n\n
Results<\/strong><\/a>[GU3]<\/a> <\/p>\n\n\n\n
\n\n\n\n![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/01\/Picture1-1024x552.png\")
<\/figure>\n\n\n\nOur baseline configuration, Qwen2.5-7B-Instruct with NAT\u2019s default react agent prompt, achieved a raw accuracy of 45.70%. This gap illustrates a common challenge in agentic systems: even when models produce correct outputs, inefficient or unnecessary tool calls reduce their practical utility in enterprise settings.<\/p>\n\n\n\n
Introducing optimizations, values found by the optimizer for hyper parameters and the prompt with manually added tool call examples yielded small gains with a 2.87% increase in raw accuracy, suggesting that prompt-level interventions alone are insufficient to meaningfully improve agent reliability.<\/p>\n\n\n\n
The most significant improvement within the model came from fine tuning the model. Qwen2.5-7B-Instruct with finetuning achieved 65.70% raw accuracy, representing an almost 20% improvement over the default baseline. <\/p>\n\n\n\n![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/01\/Picture2-1024x364.png\")
<\/figure>\n\n\n\nOur agentic RAG architecture demonstrated substantial improvements over the baseline Cosmos Reason 1 model, achieving an overall accuracy increase of 10.23%<\/strong> across the LongVideoBench evaluation suite. <\/p>\n\n\n\nPerformance gains were particularly pronounced in reasoning tasks that require long-range temporal understanding and cross-modal association. The system achieved its strongest results in:<\/p>\n\n\n\n
\n- Temporal-to-Action<\/strong> reasoning (+25.83% accuracy)<\/li>\n\n\n\n
- Spatial-to-Event <\/strong>reasoning (+26.86% accuracy)<\/li>\n\n\n\n
- Single-Scene Spatial<\/strong> retrieval & reasoning (+26.14% accuracy, showcasing the potential of Cosmos Embed 1)<\/li>\n\n\n\n
- Temporal-to-Event <\/strong>reasoning (+18.51% accuracy)<\/li>\n<\/ul>\n\n\n\n
These results indicate that our multi-tool architecture effectively compensates for the temporal limitations of short-form video fine-tuning, enabling the system to maintain physical reasoning capabilities while scaling to long-form video understanding tasks. The performance distribution across reasoning categories further suggests that semantic retrieval combined with explicit subject and event tracking creates complementary pathways for temporal understanding in physical AI systems.<\/p>\n\n\n\n
Overall, these results validate our hypothesis that through specialization methods such as agent workflows and finetuning, you can create small, computationally inexpensive models which are competitive with their larger counterparts.<\/p>\n\n\n\n
Impact<\/strong><\/p>\n\n\n\n
\n\n\n\n\n- Physical AI & Robotics Impact<\/em><\/li>\n<\/ol>\n\n\n\n
In our work, we show how agentic retrieval can extend long video understanding in smaller reasoning models. Our architecture enables Cosmos Reason 1 to keep its strengths in short-term physics and spatial reasoning while improving its temporal limitations. This approach mimics real-world robotic perception pipelines, where agents need to reason over an extended period of time.<\/p>\n\n\n\n
We also contributed a video upload and storage pipeline integration to NVIDIA\u2019s open-source Nemo Agent Toolkit. <\/p>\n\n\n\n
This extension addresses a key infrastructure requirement for physical AI and Metropolis-style workflows: persistent access to video assets throughout the agent lifecycle. <\/p>\n\n\n\n
The contribution introduces native support for video uploading and storage within NAT and NAT-UI, including:<\/p>\n\n\n\n
\n- A new video library component in NAT UI\u2019s sidebar for uploading and managing video assets<\/li>\n\n\n\n
- Storage of uploaded videos in a local S3-compatible object store<\/li>\n<\/ul>\n\n\n\n
This functionality enables video-based agent testing and evaluation, allowing agents to be validated against fixed video datasets. These additions directly benefit the Metropolis and robotics teams at NVIDIA, where developing and validating video-centric agents is critical.<\/p>\n\n\n\n
Together, the long-video agent architecture and open-source contributions to NAT provide meaningful impact for physical AI applications.<\/p>\n\n\n\n
\n- Broader Implication for Agent Applications<\/em><\/li>\n<\/ol>\n\n\n\n
You don\u2019t need a giant model if you combine the right specialized SLM, the right retrieval system, and the right agent tools. <\/p>\n\n\n\n
Our results demonstrate that small, finetuned models can reliably function as enterprise-grade agents. By improving tool-calling accuracy rather than scaling model size, we achieved substantial gains in both correctness and efficiency.<\/p>\n\n\n\n
For enterprises, this translates to:<\/p>\n\n\n\n
\n- Lower inference and infrastructure costs<\/li>\n\n\n\n
- Reduced latency for real-time analytics<\/li>\n\n\n\n
- Improved flexibility for agent hosting.<\/li>\n<\/ul>\n\n\n\n
In enterprise agentic systems, specialization beats scale. A fine-tuned small model that knows how to use tools reliably is often more valuable than a larger model that tries to reason end-to-end, taking more time, money and compute.<\/p>\n\n\n\n
Our work with Cosmos Reason demonstrates a fundamental shift in how we approach scaling AI capabilities for domain-specific tasks. Rather than relying solely on parameter count and massive foundation models, we achieved competitive performance on long-form video understanding by orchestrating a specialized small language model with purpose-built retrieval infrastructure and task-specific tooling.<\/p>\n\n\n\n
This research represents a scalable pathway for deploying AI across Metropolis applications – from vision tool calling to smart city infrastructure, where efficiency, interpretability, and task-specific performance are keystones.<\/p>\n\n\n\n
Clemson x NVIDIA Partnership – Conclusion<\/strong><\/p>\n\n\n\n
\n\n\n\nAs Clemson students, getting to be a part of the partnership between our university and NVIDIA has been nothing short of transformative. Due to the hard work of Carrie Russell from Clemson, as well as Karthick Iyer, Roopa Prabhu, Sujit Biswas, and Zac Wang from NVIDIA, we were given a platform to stand on as a means of chasing our ambitions. For students like us, passionate about pushing the boundaries of artificial intelligence and high-performance computing, the Clemson-NVIDIA collaboration hasn\u2019t just been an opportunity – it’s been a launchpad, and it means the world that we\u2019ve been able to learn and grow while proudly representing the Tiger spirit in the global tech landscape.<\/p>\n\n\n\n
A special thanks to everyone involved! We really appreciated all of your support along the way. <\/p>\n\n\n\n
\n\n\n\n<\/p>\n","protected":false},"excerpt":{"rendered":"
Capgemini published a global report projecting that AI agents will deliver more than $450B in economic value by 2028. Industry momentum is unmistakable: enterprises are rapidly shifting from passive large language model (LLM) chatbots to active, tool-using AI agents that can retrieve data, execute workflows, and make decisions. NVIDIA\u2019s position aligns with this shift.<\/p>\n","protected":false},"author":339,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-270","post","type-post","status-publish","format-standard","hentry","category-news-item"],"fimg_url":false,"_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/270","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=270"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/270\/revisions"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=270"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=270"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=270"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=270"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":257,"date":"2025-05-29T15:50:11","date_gmt":"2025-05-29T15:50:11","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=257"},"modified":"2025-06-23T15:12:01","modified_gmt":"2025-06-23T15:12:01","slug":"next-engineers-celebrates-2025-graduates","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/next-engineers-celebrates-2025-graduates\/","title":{"rendered":"Next Engineers Celebrates 2025 Graduates"},"content":{"rendered":"\n![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2025\/05\/NEXTEngineerGrad_2025-1024x675.jpg\")
John Intile (left), Vice President, Engineering GE Vernova and Serita Acker (right), Executive Director of PEER & WISE, smile for a photo with one of this year’s Next Engineers graduates.<\/figcaption><\/figure>\n\n\n\n
\n\n\n\nThe Next Engineers: Engineering Academy Greenville<\/strong> proudly announces the graduation of its 2025 cohort, marking a significant milestone in the program’s mission to inspire the next generation of engineers.<\/p>\n\n\n\nThis year, 33 students from 16 high schools across the Upstate completed the rigorous three-year program, which combines hands-on design challenges, career coaching, and college-readiness workshops. Each graduate is eligible for a $20,000 scholarship upon enrollment in a qualified engineering or engineering-related degree program.<\/p>\n\n\n\n
This is the second graduating class of Next Engineers: Engineering Academy students in Greenville, South Carolina. Since its inception in 2021, this college- and career-readiness program has been a collaborative effort between Clemson University’s PEER & WISE program and GE Vernova, creating opportunities for young people to become engineers. To date, the Next Engineers program in Greenville has reached nearly 5,400 students and awarded over $900,000 in scholarships to qualifying graduates.<\/p>\n\n\n\n
Students in the program engaged in hours of hands-on educational activities, including building water filtration systems, testing prototype helmets, and presenting their designs to peers and professional engineers. These experiences have equipped them with essential skills such as teamwork, problem-solving, and communication.<\/p>\n\n\n
\n![\"\"](\"https:\/\/news.clemson.edu\/wp-content\/uploads\/2024\/06\/thumbnail_Facetune_06-06-2024-09-22-05-190x215.jpg\")
Serita Acker<\/figcaption><\/figure>\n<\/div>\n\n\nSerita Acker<\/strong>, executive director of PEER & WISE, expressed her pride in the program’s success:<\/p>\n\n\n\n“Now in our fourth year, Next Engineers continues to transform lives by investing in the future of STEM and empowering the next generation of diverse innovators. I am especially grateful for the Clemson University faculty, staff, students, and graduate students for their continued engagement, as well as for the partnership and support from Upstate high schools who help make this initiative possible.”<\/em><\/p>\n\n\n\nThe Next Engineers: Engineering Academy is a transformative learning experience designed for students aged 15 to 18. Through a rigorous curriculum, immersive design challenges, and career coaching, participants learn to think and act like engineers.<\/p>\n\n\n\n
The graduates and their high schools are:<\/strong><\/p>\n\n\n\nDavid Bellino<\/strong> \u2014 Greenville Technical Charter High School
John Buckingham<\/strong> \u2014 Greenville Technical Charter High School
Joy Dodd<\/strong> \u2014 Travelers Rest High School
Michael Driscoll<\/strong> \u2014 Greenville Technical Charter High School
Ryan Ellis<\/strong> \u2014 Westside High School
DeAngelo Fuller<\/strong> \u2014 Westside High School
Tristan Greenleaf<\/strong> \u2014 Pickens High School
Christian Harling<\/strong> \u2014 Southside High School
Brock Hinson<\/strong> \u2014 J.L. Mann High School
Abdurrahman Housari<\/strong> \u2014 Greenville Technical Charter High School
Omar Housari<\/strong> \u2014 Greenville Technical Charter High School
Advaith Joshi<\/strong> \u2014 J.L. Mann High School
Thomas Kezman<\/strong> \u2014 Greer Middle College
Noah King<\/strong> \u2014 D.W. Daniel High School
Riya Kot<\/strong> \u2014 Southside High School
Adam Le<\/strong> \u2014 Riverside High School
Rami Makhtoub<\/strong> \u2014 Greenville Technical Charter High School
Noah Miller<\/strong> \u2014 Easley High School
Hayley Morton<\/strong> \u2014 Greenville Technical Charter High School
Levi Naylor<\/strong> \u2014 Easley High School
Tron Paul<\/strong> \u2014 Crescent High School
Hayden Ramsey<\/strong> \u2014 Travelers Rest High School
Harm Ravenhorst<\/strong> \u2014 Greenville High School
Hallie Ray<\/strong> \u2014 D.W. Daniel High School
Addison Reid<\/strong> \u2014 Blue Ridge High School
James Robinson<\/strong> \u2014 Mauldin High School
Alejandra Rodriguez<\/strong> \u2014 Easley High School
Josh Sparks<\/strong> \u2014 Legacy Early College High School
Jeffrey Sweeney<\/strong> \u2014 Pendleton High School
Sai Praneetha Thatavarthi<\/strong> \u2014 J.L Mann High School
Dimitri Tsirkas<\/strong> \u2014 Easley High School
Jade Williams<\/strong> \u2014 Legacy Early College High School
Kylie Willis<\/strong> \u2014 Pendleton High School<\/p>\n\n\n\nRead the GE Vernova press release below:<\/strong>
https:\/\/www.gevernova.com\/news\/press-releases\/ge-vernova-clemson-university-celebrate-next-engineers-academy-graduation<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"The Next Engineers: Engineering Academy Greenville proudly announces the graduation of its 2025 cohort, marking a significant milestone in the program’s mission to inspire the next generation of engineers. This year, 33 students from 16 high schools across the Upstate completed the rigorous three-year program, which combines hands-on design challenges, career coaching, and college-readiness workshops. […]<\/p>\n","protected":false},"author":3922,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14094],"class_list":["post-257","post","type-post","status-publish","format-standard","hentry","category-news-item"],"fimg_url":false,"_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/257","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/3922"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=257"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/257\/revisions"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=257"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=257"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=257"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=257"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":247,"date":"2025-05-16T18:00:08","date_gmt":"2025-05-16T18:00:08","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=247"},"modified":"2025-06-23T15:12:07","modified_gmt":"2025-06-23T15:12:07","slug":"building-resilience-building-leaders-cedc-hosts-2025-spring-summit","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/building-resilience-building-leaders-cedc-hosts-2025-spring-summit\/","title":{"rendered":"Building Resilience, Building Leaders: CEDC Hosts 2025 Spring Summit"},"content":{"rendered":"\n![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2025\/05\/cedc-spring25-a-1024x616.jpg\")
<\/figure>\n\n\n\n<\/p>\n\n\n\n
Clemson Engineers for Developing Communities (CEDC) hosted its 2025 Spring Summit at the end of the semester, showcasing the innovative, student-led projects that are making a difference in South Carolina and beyond. Held at the Hendrix Student Center, the event brought together students, faculty, alumni, and community partners to celebrate another year of impactful work and leadership development.<\/p>\n\n\n\n
Founded in 2009, CEDC is a service-learning program within the College of Engineering, Computing and Applied Sciences (CECAS) that empowers students to develop sustainable solutions for low-capacity communities while cultivating critical leadership, project management, and community engagement skills.<\/p>\n\n\n\n
This year\u2019s Summit highlighted projects across South Carolina, Haiti, Ecuador, and Colombia. From water system designs in the Amazon to infrastructure resilience planning across South Carolina, students tackled real-world challenges with creativity, compassion, and a focus on strengthening community capabilities.<\/p>\n\n\n
\n![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2025\/05\/cedc-spring25-c-1024x695.jpg\")
<\/figure>\n<\/div>\n\n\nNew for 2025, CEDC introduced the Challenge Coin of Leadership, a tradition that encourages students to recognize leadership excellence among their peers. Each coin represents a commitment to passing leadership forward, building a culture of service and accountability that extends beyond the classroom.<\/p>\n\n\n\n
The Summit also spotlighted major initiatives like ResilientSC, a growing effort to address two critical challenges facing South Carolina:<\/p>\n\n\n\n
\n- Brain Drain: <\/strong>With 37% of college graduates leaving the state, ResilientSC is launching programs to retain young talent and support local communities.<\/li>\n\n\n\n
- Infrastructure Vulnerability:<\/strong> South Carolina faces more than $375 million in disaster recovery costs each year, highlighting the urgent need for proactive resilience planning. Through ResilientSC, students are helping communities strengthen critical infrastructure and prepare for future challenges.<\/li>\n<\/ul>\n\n\n\n
![\"\"](\"https:\/\/blogs.clemson.edu\/cecas\/files\/2025\/05\/cedc-spring25-b-1024x719.jpg\")
<\/figure>\n<\/div>\n\n\nIn addition to project presentations, students hosted a cultural awareness reception featuring Mediterranean cuisine, emphasizing the importance of global engagement through food, conversation, and shared experiences.<\/p>\n\n\n\n
Several students were recognized with awards for outstanding service and leadership:<\/p>\n\n\n\n
\n- Distinguished Service Award: Nate Polakowski, Ian Krinock, and William Pautler<\/li>\n\n\n\n
- Excellence in Leadership Award: Ayden Fournier<\/li>\n<\/ul>\n\n\n\n
As CEDC continues to grow, its mission remains clear: Serving the developing world, developing those who serve.<\/em><\/p>\n\n\n\n\u201cThe true measure of our success is not what we build, but who we lift up,\u201d <\/em>said David Vaughn, Director of CEDC and Professor of Practice in CECAS. \u201cThrough CEDC, our students are not just becoming engineers\u2014they\u2019re becoming leaders, change-makers, and community builders.\u201d<\/em><\/p>\n\n\n\nFor more information about CEDC and how to get involved, visit cecas.clemson.edu\/cedc.<\/p>\n","protected":false},"excerpt":{"rendered":"
Clemson Engineers for Developing Communities (CEDC) hosted its 2025 Spring Summit at the end of the semester, showcasing the innovative, student-led projects that are making a difference in South Carolina and beyond. Held at the Hendrix Student Center, the event brought together students, faculty, alumni, and community partners to celebrate another year of impactful work […]<\/p>\n","protected":false},"author":339,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-247","post","type-post","status-publish","format-standard","hentry","category-news-item"],"fimg_url":false,"_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/247","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=247"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/247\/revisions"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=247"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=247"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=247"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=247"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":249,"date":"2025-05-13T18:37:33","date_gmt":"2025-05-13T18:37:33","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=249"},"modified":"2025-06-23T15:12:12","modified_gmt":"2025-06-23T15:12:12","slug":"powered-by-nvidia-clemson-students-create-ai-teaching-assistant-for-smarter-scalable-learning","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/powered-by-nvidia-clemson-students-create-ai-teaching-assistant-for-smarter-scalable-learning\/","title":{"rendered":"Powered by NVIDIA: Clemson Students Create AI Teaching Assistant for Smarter, Scalable Learning"},"content":{"rendered":"\n
Clemson University computer science students teamed up with global AI leader NVIDIA to develop an AI-powered virtual teaching assistant designed to enhance learning and uphold academic integrity.<\/p>\n\n\n\n
They built the tool using NVIDIA\u2019s cutting-edge AI Blueprint for Retrieval-Augmented Generation (RAG), creating a system that helps students strengthen their understanding of core concepts while offering scalable, around-the-clock support. The assistant addresses growing demand for tutoring in high-enrollment STEM courses and demonstrates how AI can responsibly support education.<\/p>\n\n\n\n
\u201cThis project shows what\u2019s possible when students work directly with advanced AI technology from industry leaders like NVIDIA,\u201d said Carrie Russell, professor of practice in Clemson\u2019s School of Computing. \u201cOur team tackled a pressing challenge in education and delivered a solution that\u2019s both innovative and practical. We\u2019re proud to collaborate with NVIDIA on work that has the potential to transform learning at scale.\u201d<\/p>\n\n\n\n
Karthick Iyer, Vice President, Engineering at NVIDIA, said he was impressed with the students\u2019 ability to turn cutting-edge AI tools into a practical solution for real educational challenges.<\/p>\n\n\n\n
\u201cIt\u2019s inspiring to see Clemson students apply NVIDIA\u2019s AI Blueprint for RAG to create a tool that directly tackles real-world issues in education,\u201d Iyer said. \u201cThis collaboration shows the power of pairing AI innovation with hands-on student learning, and we\u2019re thrilled to support projects that advance both technology and education.\u201d<\/p>\n\n\n\n
Students Camden Spehl, Kyle Zheng, Jaylen Goddard, and Cordarro Higgins Redman collaborated with Vinay Raman, a senior deep learning scientist at NVIDIA, to bring the project to life.<\/p>\n\n\n\n
<\/figure>\n\n\n\nFigure 1. Public leaderboard highlighting the newly added models.<\/em><\/p>\n\n\n\n Edge models aren’t simply smaller versions of frontier-scale systems\u2014they’re designed around a fundamentally different engineering objective.<\/p>\n\n\n\n Rather than maximizing reasoning capability with virtually unlimited compute, edge models must maximize reasoning within a fixed deployment budget<\/strong>. Latency, memory footprint, power consumption, and hardware availability all become part of the optimization problem. Whether deployed on warehouse robots, manufacturing inspection systems, intelligent traffic cameras, or autonomous platforms, these models often need to process information where it is generated instead of relying on cloud-scale infrastructure.<\/p>\n\n\n\n That changes what a benchmark should measure. It’s no longer just how well<\/strong> a model reasons, but how effectively<\/strong> it reasons under realistic operational constraints. Evaluating compact reasoning models alongside larger systems provides a more complete understanding of the trade-offs between capability, efficiency, and deployability\u2014an increasingly important consideration as physical AI moves from research to production.<\/p>\n\n\n\n <\/a>Growing with the Field<\/strong><\/p>\n\n\n\n This update represents more than the addition of four reasoning models\u2014it reflects how rapidly the multimodal AI ecosystem is evolving and the need for benchmarks to evolve alongside it.<\/p>\n\n\n\n Among the newly evaluated models, Cosmos3 Edge as a reasoning vision language model (VLM) achieved the highest overall performance within the evaluated ~2B parameter class<\/strong>, establishing a strong baseline for this emerging generation of compact vision-language models. More importantly, evaluating these models under the same operational tasks and methodology enables meaningful comparisons within similar deployment budgets, helping researchers better understand the trade-offs between model capability and deployment efficiency. Full results for all four models are available on the VANTAGE-Bench leaderboard.<\/p>\n\n\n\n At Clemson, we view VANTAGE-Bench<\/strong><\/a> as research infrastructure rather than a static benchmark<\/strong>. Maintaining an operational benchmark is an ongoing effort\u2014one that requires continuously integrating new model families, architectures, and deployment paradigms as the field advances. As Physical AI continues to evolve, VANTAGE-Bench will continue evolving with it, providing the research community with a consistent, reproducible framework for evaluating the next generation of vision-language models.<\/p>\n","protected":false},"excerpt":{"rendered":" The rapid advancement of multimodal AI is reshaping how intelligent systems interact with the physical world. Vision language models (VLMs) are no longer limited to cloud-based assistants\u2014they are increasingly being deployed in warehouses, manufacturing facilities, traffic monitoring systems, robotics, and other physical AI applications where decisions must be made in real time. As these deployments […]<\/p>\n","protected":false},"author":339,"featured_media":295,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-292","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news-item"],"fimg_url":"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/07\/Students-150x150.jpg","_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/292","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=292"}],"version-history":[{"count":1,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/292\/revisions"}],"predecessor-version":[{"id":296,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/292\/revisions\/296"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media\/295"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=292"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=292"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=292"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=292"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":287,"date":"2026-06-01T01:30:00","date_gmt":"2026-06-01T01:30:00","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=287"},"modified":"2026-05-31T16:22:57","modified_gmt":"2026-05-31T16:22:57","slug":"vantage-bench-evaluating-vision-language-models-in-real-world-operational-environments","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/vantage-bench-evaluating-vision-language-models-in-real-world-operational-environments\/","title":{"rendered":"VANTAGE-Bench: Evaluating Vision-Language Models in Real-World Operational Environments"},"content":{"rendered":"\n Vision-Language Models (VLMs) are rapidly becoming central to the next generation of artificial intelligence systems, powering technologies that combine visual understanding with natural language reasoning. As these systems move beyond research demonstrations and into applications involving transportation systems, logistics networks, robotics platforms, industrial facilities, and public infrastructure, questions about how they should be evaluated are becoming increasingly important.<\/p>\n\n\n\n Many existing VLM benchmarks rely on consumer-centric source material \u2013 egocentric recordings, broadcast media, web-curated video \u2013 while reducing evaluation to multiple-choice questions that effectively provide the model with the correct answer as one of the options. But operational environments do not arrive pre-framed, and the decisions they demand are rarely multiple choice. They are continuous, dynamic, and often safety- or mission-critical.<\/p>\n\n\n\n VANTAGE-Bench (Video Analysis Tasks Across Generalized Environments) is a new public benchmark suite and leaderboard developed to evaluate how Vision-Language Models perform across the range of capabilities required for fixed-camera, infrastructure-scale deployment. This includes a previously unmeasured capability: continuous single-object tracking through pure vision-language reasoning. The benchmark is designed to help researchers and model developers better understand how these systems perform in dynamic operational environments where reliability and continuity matter.<\/p>\n\n\n\n The Infrastructure AI Gap<\/strong><\/p>\n\n\n\n For more than a decade, progress in vision-language modeling has largely been measured against internet-scale datasets built from social media imagery, curated video, consumer photography, and other highly photographed digital environments. Models have become remarkably capable on this material. But the environments where these systems are increasingly expected to function \u2013 distribution centers, intersections, transit corridors, ports, public facilities, and industrial floors \u2013 often look very different from the datasets used to train and evaluate them.<\/p>\n\n\n\n These are fixed-camera worlds. They are dense, continuous, and operationally consequential. A missed forklift is not a missed caption. A misread pedestrian is not a misread meme.<\/p>\n\n\n\n When VANTAGE-Bench’s authors measured frontier models on infrastructure-oriented tasks against their reported scores on standard consumer-centric benchmarks, the degradation was substantial. A representative open-weight model dropped nearly 30 points on causal event verification compared to its score on MLVU, and roughly 24 points on 2D spatial pointing compared to BLINK.<\/p>\n\n\n\n The cause appears structural: internet-trained models rely on a “cinematic prior” characterized by human-centric framing, high-velocity motion, and edited temporal structure \u2013 assumptions that do not exist in elevated, wide-angle, fixed-camera footage. Without that prior, models must reason from pure spatial-temporal logic, and many struggle to do so reliably.<\/p>\n\n\n\n The research community has increasingly recognized this disconnect as part of the broader Infrastructure AI Gap<\/em>: the growing distance between benchmark performance and operational deployment. VANTAGE-Bench was created to help measure that gap more directly by evaluating how VLMs perform in dynamic operational scenarios rather than only in familiar benchmark environments.<\/p>\n\n\n\n In doing so, it shifts the central evaluation question from \u201cCan this model perform well on familiar benchmark data?\u201d<\/em> to a more operationally meaningful one: \u201cCan this model be trusted in the environments where it will actually be used?\u201d<\/em><\/p>\n\n\n\n Benchmark Architecture and Evaluation Design<\/strong><\/p>\n\n\n\n VANTAGE-Bench is organized around four reasoning pillars, each targeting a distinct capability required for reliable Infrastructure AI. Across these pillars, the suite evaluates eight task formulations spanning both image and video modalities, moving deliberately beyond the multiple-choice format that dominates many existing benchmarks.<\/p>\n\n\n\n Pillar I \u2013 Semantic Understanding (What and Why)<\/strong><\/p>\n\n\n\n Two tasks evaluate high-level causal and operational reasoning.<\/p>\n\n\n\n Pillar II \u2013 Spatial Understanding (Where)<\/strong><\/p>\n\n\n\n Three tasks evaluate precise geometric grounding in dense, multi-instance scenes.<\/p>\n\n\n\n Pillar III \u2013 Temporal Understanding (When)<\/strong><\/p>\n\n\n\n Two tasks evaluate the perception of action duration and event boundaries.<\/p>\n\n\n\n Pillar IV \u2013 Spatio-Temporal Understanding (Dynamics)<\/strong><\/p>\n\n\n\n The benchmark introduces a quantitative Single-Object Tracking evaluation for VLMs.<\/p>\n\n\n\n Together, these eight tasks span 3,346 media assets and 35,027 expert annotations, drawn from three operational domains: Warehouse and Logistics, Transportation, and Smart Spaces. All footage originates from fixed-infrastructure cameras that are static, elevated, and wide-angle \u2013 categorically distinct from the egocentric and broadcast-style sources that dominate existing benchmarks.<\/p>\n\n\n\n Dataset Construction and Annotation Pipeline<\/strong><\/p>\n\n\n\n The benchmark is primarily human-labeled by trained professional annotators rather than crowdsourced contributors, supported by a multi-tier quality pipeline in which 100% of initial annotations are reviewed by a secondary expert and a random 10% subset is audited by a third-tier QA reviewer.<\/p>\n\n\n\n Programmatic methods are used only where they can be anchored to human-verified data, allowing the dataset to scale without introducing substantial annotation noise.<\/p>\n\n\n\n To responsibly evaluate safety-critical events such as collisions and industrial accidents, scenarios that are both rare and heavily restricted by privacy regulations, approximately 20% of the Video QA and Temporal splits use high-fidelity synthetic footage generated through NVIDIA Omniverse<\/a> DRIVE Sim. This allows the benchmark to cover severe long-tail anomalies without ethical compromise.<\/p>\n\n\n\n All real-world footage was sourced under explicit informed consent, with personally identifiable information removed through automated obfuscation followed by human-in-the-loop verification.<\/p>\n\n\n\n Public Infrastructure and Benchmark Operations<\/strong><\/p>\n\n\n\n VANTAGE-Bench is built for the researchers advancing multimodal AI, the model developers preparing systems for deployment, and the organizations whose facilities, infrastructure, and public spaces may ultimately depend on those systems working as promised.<\/p>\n\n\n\n The platform accepts model submissions, evaluates them through a standardized testing pipeline built on an extended version of the open-source VLMEvalKit harness, and maintains a public leaderboard reporting benchmark results.<\/p>\n\n\n\n Every submission is evaluated against a versioned dataset and evaluation framework designed to support transparent and reproducible comparison over time. As the dataset and evaluation suite evolve, benchmark versions remain clearly identified to preserve scientific consistency and longitudinal comparability.<\/p>\n\n\n\n The open-source benchmark codebase is available at https:\/\/github.com\/Clemson-Capstone\/VANTAGE-Bench<\/a>, and the public leaderboard is hosted at https:\/\/huggingface.co\/spaces\/clemson-computing\/VANTAGE-Bench-Leaderboard<\/a>.<\/p>\n\n\n\n Building an Open Evaluation Community<\/strong><\/p>\n\n\n\n The dataset behind VANTAGE-Bench began through an initial collaboration with NVIDIA and is designed to continue expanding over time. Clemson University\u2019s School of Computing now independently hosts and maintains the benchmark, including the public evaluation infrastructure, leaderboard, and future dataset development.<\/p>\n\n\n\n Researchers, practitioners, and organizations working in operational environments are invited to contribute new data that expands the diversity of scenes, domains, and environmental conditions represented within the benchmark. Model developers across academia and industry are also encouraged to submit Vision-Language Models to the public leaderboard and participate in broader efforts to establish shared evaluation standards for operational AI systems.<\/p>\n\n\n\n The dataset is publicly available at https:\/\/huggingface.co\/datasets\/nvidia\/PhysicalAI-VANTAGE-Bench<\/a>, and model submissions can be made through the public submission portal at https:\/\/vantage-bench.org<\/a>.<\/p>\n\n\n\n Why Clemson, and Why Now?<\/strong><\/p>\n\n\n\n Clemson University has long emphasized research connected to real-world systems and societal impact. VANTAGE-Bench reflects that broader approach to artificial intelligence \u2013 not only advancing AI capabilities, but also helping build the infrastructure needed to evaluate those capabilities responsibly.<\/p>\n\n\n\n The School of Computing views open evaluation as an important part of the scientific process. Public benchmarks maintained by independent research institutions can help provide transparency, reproducibility, and continuity in a rapidly evolving field where evaluation standards are still actively developing.<\/p>\n\n\n\n This effort also aligns with Clemson\u2019s broader investment in artificial intelligence research, operational AI systems, and interdisciplinary collaboration across areas including computer vision, machine learning, robotics, intelligent systems, and computational imaging.<\/p>\n\n\n\n Defining the Next Phase of AI Evaluation<\/strong><\/p>\n\n\n\n The future of AI will not be defined solely by who builds the largest models. Increasingly, it will also be shaped by who develops rigorous evaluation standards capable of measuring whether those systems can operate reliably in the physical world.<\/p>\n\n\n\n VANTAGE-Bench is one contribution toward that goal, and an invitation to the broader research and industry community to help define how operational AI systems should be evaluated in the years ahead.<\/p>\n\n\n\n The future of AI will not be decided only by what models can say. It will be shaped by what they can reliably understand in the places where understanding matters most. VANTAGE-Bench is built to measure that, and Clemson is building it to last.<\/p>\n","protected":false},"excerpt":{"rendered":" Vision-Language Models (VLMs) are rapidly becoming central to the next generation of artificial intelligence systems, powering technologies that combine visual understanding with natural language reasoning. As these systems move beyond research demonstrations and into applications involving transportation systems, logistics networks, robotics platforms, industrial facilities, and public infrastructure, questions about how they should be evaluated are […]<\/p>\n","protected":false},"author":339,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-287","post","type-post","status-publish","format-standard","hentry","category-news-item"],"fimg_url":false,"_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/287","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=287"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/287\/revisions"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=287"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=287"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=287"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=287"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":280,"date":"2026-03-02T17:00:24","date_gmt":"2026-03-02T17:00:24","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=280"},"modified":"2026-03-02T17:00:24","modified_gmt":"2026-03-02T17:00:24","slug":"duke-energy-science-nights-expand-stem-engagement-across-south-carolina","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/duke-energy-science-nights-expand-stem-engagement-across-south-carolina\/","title":{"rendered":"Duke Energy Science Nights Expand STEM Engagement Across South Carolina"},"content":{"rendered":"\n <\/p>\n\n\n\n Excitement is building across South Carolina as 60 schools prepare to host Duke Energy Science Nights during SC STEM Education Month 2026. Through a continued partnership among Duke Energy, South Carolina\u2019s Coalition for Mathematics & Science (SCCMS), and the Morehead Planetarium and Science Center, these hands-on family science events will engage students and caregivers in interactive experiences designed to spark curiosity and inspire future STEM careers.<\/p>\n\n\n\n Now in its fifth year, Duke Energy Science Nights have continued to grow in reach and impact since launching in South Carolina in 2022. This year\u2019s program marks an expansion from 50 to 60 schools being awarded kits thanks to increased support from the Duke Energy Foundation. This program growth is in response to the increased demand from school applications and a shared commitment from program partners to provide equitable access to high-quality STEM learning opportunities across the state.<\/p>\n\n\n\n STEM\u2014science, technology, engineering and mathematics\u2014plays a critical role in preparing students for future careers, particularly in South Carolina, where advanced manufacturing, energy, and technology-driven industries are central to the state\u2019s economy. Duke Energy Science Nights help students see the relevance of STEM learning beyond the classroom while encouraging family engagement and community connection.<\/p>\n\n\n\n The science night kits are funded by a grant from the Duke Energy Foundation and are provided at no cost to participating schools. Each kit includes materials to support up to 200 participants in 10 hands-on activities, along with activity guides, promotional materials, downloadable digital resources (including Spanish-language versions), and virtual support. The kits are produced by the Morehead Planetarium and Science Center\u2019s NCSciFest team in collaboration with SCCMS staff, while SCCMS facilitates the application process and provides ongoing planning and communication support.<\/p>\n\n\n\n \u201cPrograms like this reinforce our company\u2019s commitment to create a more vibrant economy by investing in our future workforce early on,\u201d said Amanda Dow, Foundation Manager for Duke Energy in South Carolina. \u201cWe know firsthand that STEM is the backbone to many career opportunities, including those here at Duke Energy and we\u2019re proud to support these students and schools.”<\/p>\n\n\n\n Participating Schools for 2026:<\/p>\n\n\n\n Over the coming weeks, participating schools will finalize event details and prepare to welcome families for evenings centered on exploration, discovery, and shared learning. Reflecting on the impact of past events, Willette Sheard, a teacher at Dr. Rose H. Wilder Elementary shared, \u201cIt\u2019s a chance to spark curiosity, encourage problem-solving, and make learning exciting outside of the classroom. This event also helps families feel connected to their child\u2019s education while building community. Most of all, it inspires kids to see themselves as future scientists, engineers, and problem-solvers.\u201d SC STEM Education Month launches annually on Pi Day, March 14, and is sponsored by South Carolina\u2019s Coalition for Mathematics & Science, a statewide organization housed within Clemson University\u2019s College of Engineering, Computing and Applied Sciences.<\/p>\n\n\n\n Learn more at scstemmonth.org.<\/a><\/p>\n\n\n\n Contacts: Excitement is building across South Carolina as 60 schools prepare to host Duke Energy Science Nights during SC STEM Education Month 2026. Through a continued partnership among Duke Energy, South Carolina\u2019s Coalition for Mathematics & Science (SCCMS), and the Morehead Planetarium and Science Center, these hands-on family science events will engage students and caregivers in […]<\/p>\n","protected":false},"author":339,"featured_media":281,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-280","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news-item"],"fimg_url":"https:\/\/blogs.clemson.edu\/cecas\/files\/2026\/03\/Picture1-150x135.png","_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/280","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=280"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/280\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media\/281"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=280"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=280"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=280"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=280"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":270,"date":"2026-01-15T00:28:35","date_gmt":"2026-01-15T00:28:35","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=270"},"modified":"2026-01-15T16:28:57","modified_gmt":"2026-01-15T16:28:57","slug":"small-models-big-capabilities","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/small-models-big-capabilities\/","title":{"rendered":"Small Models, Big Capabilities: Clemson and NVIDIA\u2019s Work on Video Tool Calling and Long-Horizon Reasoning"},"content":{"rendered":"\n In July 2025, Capgemini published a global report projecting that AI agents will deliver more than $450B in economic value by 2028<\/strong>, with 96% of surveyed organizations already experimenting with agentic systems. Industry momentum is unmistakable: enterprises are rapidly shifting from passive large language model (LLM) chatbots to active, tool-using AI agents that can retrieve data, execute workflows, and make decisions.<\/p>\n\n\n\n NVIDIA\u2019s position aligns with this shift. In its paper Small Language Models Are the Future of Agentic AI<\/em>, NVIDIA argues that specialized Small Language Models (SLMs) – not monolithic generalist LLMs – will form the backbone of enterprise agents.<\/p>\n\n\n\n But the most valuable agents are not limited to text. Real-world deployments increasingly require multimodal understanding, particularly in robotics and autonomous systems where perception drives action. Here, NVIDIA\u2019s Cosmos Stack provides the foundation:<\/p>\n\n\n\n Yet both SLMs and small vision language models (VLMs) for physical applications carry known limitations.<\/p>\n\n\n\n Our project, a collaboration between NVIDIA and Clemson University, set out to evaluate and patch these weaknesses. We pursued a two-part investigation:<\/p>\n\n\n\n Together, these efforts demonstrate a central thesis: Small Language Models, when paired with retrieval and targeted finetuning, can deliver competitive agentic performance across both enterprise and physical AI domains.<\/p>\n\n\n\n This post details how we built, evaluated, and validated these improvements.<\/p>\n\n\n\n Approach<\/strong><\/p>\n\n\n\n To evaluate and extend the capabilities of small multimodal models, we built our system with NVIDIA Metropolis VSS and NVIDIA NeMo Agent Toolkit (NAT)<\/strong> – an agent coordination framework designed to orchestrate tool-using LLMs with minimal overhead. <\/a>[GU2]<\/a> NAT integrates cleanly with existing agent ecosystems such as LangChain, while providing several advantages critical to this project: built-in agent evaluation, prompt optimization utilities, structured tool-trajectory visualization, and a lightweight UI for rapid iteration.<\/p>\n\n\n\n We hosted all experimentation on NVIDIA Brev<\/strong>, deploying multiple A100 <\/strong>instances to support both model fine tuning and long-video RAG indexing workloads. From this shared infrastructure, the collaboration proceeded along two parallel technical paths:<\/p>\n\n\n\n This path focused on enhancing the reliability of small language models as agent orchestrators. We investigated whether targeted optimization and reinforcement-based finetuning could improve structured tool invocation, positioning small models as viable controllers for enterprise agent workflows.<\/p>\n\n\n\n The physical AI path focused on extending short-form reasoning models to better understand long-form videos. Instead of scaling model size or increasing context length, we approached temporal reasoning as a coordination problem. NAT provided the structure we needed to treat long-video understanding as an agentic workflow, where retrieval and reasoning are composed iteratively rather than executed in a single pass.<\/p>\n\n\n\n Within NAT, long videos are broken down into more manageable chunks that can be indexed, retrieved, and reasoned over using a ReAct agent to orchestrate these steps. This design allows Cosmos Reason 1 to keep its strengths in short-term physics and spatial reasoning while still being effective over an extended period of time.<\/p>\n\n\n\n Methods<\/strong><\/p>\n\n\n\n Keeping consistent with this two track approach, each experiment retained its own system architecture within each respective Brev instance.<\/p>\n\n\n\n We implemented a ReAct-based agent configured to interact with a set of Metropolis VSS-style tools designed to emulate warehouse monitoring analytics APIs. The tool suite included functions such as get_all_sensor_ids, list_incidents, and get_fov_object_counts, each requiring structured arguments and returning deterministic outputs. Tool schemas and response formats were fixed across all experiments to ensure reproducibility.<\/p>\n\n\n\n The primary model evaluated was Qwen2.5-7B-Instruct. For reference, Nemotron-Super-49B and GPT-4.1 were evaluated under the same agent configurations. All models used identical prompts, tool definitions, inference parameters, and agent logic.<\/p>\n\n\n\n An evaluation dataset consisting of natural-language queries paired with true tool-call traces, in which each trace specifies the expected sequence of tool calls and corresponding arguments, was used to benchmark the full agent trajectories from the aforementioned models.<\/p>\n\n\n\n The improvement procedure proceeded in two stages. First, NAT\u2019s optimizer feature was used to tune agent-level hyperparameters and system prompts. This stage targeted improvements in tool selection consistency, argument formatting, and call sequencing.<\/p>\n\n\n\n Second, weight-level adaptation was performed using Agent Reinforcement Trainer (ART). ART applies reinforcement learning over complete agent trajectories, with rewards computed based on alignment between model-generated tool calls and true traces. Training rollouts were generated by using a synthetic task generator that produced diverse natural-language queries paired with known tool-call patterns.<\/p>\n\n\n\n Evaluation metrics included exact-match accuracy over full tool-call trajectories, as well as an efficiency-weighted accuracy metric. For each evaluation query, the dataset defined a set of acceptable tool-call traces with associated efficiency scores in the range [0,1], where a score of 1.0 corresponds to the most efficient valid trajectory. Model outputs that matched a correct but suboptimal trace were counted as correct under exact-match accuracy, while receiving a proportionally reduced score under the efficiency-weighted metric. All evaluations were conducted using the same query set and agent configuration to isolate the effects of optimization and finetuning.<\/p>\n\n\n\n Over the course of 16 weeks, we primarily worked with two models:<\/p>\n\n\n\n Cosmos Reason 1 serves as a foundational model within NVIDIA’s Metropolis vision for physical AI – artificial intelligence systems capable of comprehending fundamental physics, spatial relationships, and temporal dynamics to enable common-sense reasoning. The model achieves this capability through fine-tuning on short-form video data, enabling it to infer physical cause-and-effect relationships within brief temporal windows. However, this training methodology introduces a temporal bias that limits performance on long-form video understanding tasks.<\/p>\n\n\n\n To mitigate this limitation, we developed a ReAct-based agentic architecture within NVIDIA AI Toolkit (NAT) that orchestrates three specialized retrieval tools, inspired by Microsoft\u2019s Deep Video Discovery, and reformed for our use:<\/p>\n\n\n\n A ReAct agent orchestrates these tools by evaluating query alignment, tool execution history, candidate answers, and tool-reported confidence scores to determine response completeness. Upon convergence, the agent invokes a “finish” tool to return the final answer to the user.<\/p>\n\n\n\n We evaluated this agentic workflow using LongVideoBench, a benchmark designed to assess both long-form video comprehension and multi-faceted reasoning capabilities including spatial, temporal, and event-based understanding. Results are presented below.<\/p>\n\n\n\n Results<\/strong><\/a>[GU3]<\/a> <\/p>\n\n\n\n Our baseline configuration, Qwen2.5-7B-Instruct with NAT\u2019s default react agent prompt, achieved a raw accuracy of 45.70%. This gap illustrates a common challenge in agentic systems: even when models produce correct outputs, inefficient or unnecessary tool calls reduce their practical utility in enterprise settings.<\/p>\n\n\n\n Introducing optimizations, values found by the optimizer for hyper parameters and the prompt with manually added tool call examples yielded small gains with a 2.87% increase in raw accuracy, suggesting that prompt-level interventions alone are insufficient to meaningfully improve agent reliability.<\/p>\n\n\n\n The most significant improvement within the model came from fine tuning the model. Qwen2.5-7B-Instruct with finetuning achieved 65.70% raw accuracy, representing an almost 20% improvement over the default baseline. <\/p>\n\n\n\n Our agentic RAG architecture demonstrated substantial improvements over the baseline Cosmos Reason 1 model, achieving an overall accuracy increase of 10.23%<\/strong> across the LongVideoBench evaluation suite. <\/p>\n\n\n\n Performance gains were particularly pronounced in reasoning tasks that require long-range temporal understanding and cross-modal association. The system achieved its strongest results in:<\/p>\n\n\n\n These results indicate that our multi-tool architecture effectively compensates for the temporal limitations of short-form video fine-tuning, enabling the system to maintain physical reasoning capabilities while scaling to long-form video understanding tasks. The performance distribution across reasoning categories further suggests that semantic retrieval combined with explicit subject and event tracking creates complementary pathways for temporal understanding in physical AI systems.<\/p>\n\n\n\n Overall, these results validate our hypothesis that through specialization methods such as agent workflows and finetuning, you can create small, computationally inexpensive models which are competitive with their larger counterparts.<\/p>\n\n\n\n Impact<\/strong><\/p>\n\n\n\n In our work, we show how agentic retrieval can extend long video understanding in smaller reasoning models. Our architecture enables Cosmos Reason 1 to keep its strengths in short-term physics and spatial reasoning while improving its temporal limitations. This approach mimics real-world robotic perception pipelines, where agents need to reason over an extended period of time.<\/p>\n\n\n\n We also contributed a video upload and storage pipeline integration to NVIDIA\u2019s open-source Nemo Agent Toolkit. <\/p>\n\n\n\n This extension addresses a key infrastructure requirement for physical AI and Metropolis-style workflows: persistent access to video assets throughout the agent lifecycle. <\/p>\n\n\n\n The contribution introduces native support for video uploading and storage within NAT and NAT-UI, including:<\/p>\n\n\n\n This functionality enables video-based agent testing and evaluation, allowing agents to be validated against fixed video datasets. These additions directly benefit the Metropolis and robotics teams at NVIDIA, where developing and validating video-centric agents is critical.<\/p>\n\n\n\n Together, the long-video agent architecture and open-source contributions to NAT provide meaningful impact for physical AI applications.<\/p>\n\n\n\n You don\u2019t need a giant model if you combine the right specialized SLM, the right retrieval system, and the right agent tools. <\/p>\n\n\n\n Our results demonstrate that small, finetuned models can reliably function as enterprise-grade agents. By improving tool-calling accuracy rather than scaling model size, we achieved substantial gains in both correctness and efficiency.<\/p>\n\n\n\n For enterprises, this translates to:<\/p>\n\n\n\n In enterprise agentic systems, specialization beats scale. A fine-tuned small model that knows how to use tools reliably is often more valuable than a larger model that tries to reason end-to-end, taking more time, money and compute.<\/p>\n\n\n\n Our work with Cosmos Reason demonstrates a fundamental shift in how we approach scaling AI capabilities for domain-specific tasks. Rather than relying solely on parameter count and massive foundation models, we achieved competitive performance on long-form video understanding by orchestrating a specialized small language model with purpose-built retrieval infrastructure and task-specific tooling.<\/p>\n\n\n\n This research represents a scalable pathway for deploying AI across Metropolis applications – from vision tool calling to smart city infrastructure, where efficiency, interpretability, and task-specific performance are keystones.<\/p>\n\n\n\n Clemson x NVIDIA Partnership – Conclusion<\/strong><\/p>\n\n\n\n As Clemson students, getting to be a part of the partnership between our university and NVIDIA has been nothing short of transformative. Due to the hard work of Carrie Russell from Clemson, as well as Karthick Iyer, Roopa Prabhu, Sujit Biswas, and Zac Wang from NVIDIA, we were given a platform to stand on as a means of chasing our ambitions. For students like us, passionate about pushing the boundaries of artificial intelligence and high-performance computing, the Clemson-NVIDIA collaboration hasn\u2019t just been an opportunity – it’s been a launchpad, and it means the world that we\u2019ve been able to learn and grow while proudly representing the Tiger spirit in the global tech landscape.<\/p>\n\n\n\n A special thanks to everyone involved! We really appreciated all of your support along the way. <\/p>\n\n\n\n <\/p>\n","protected":false},"excerpt":{"rendered":" Capgemini published a global report projecting that AI agents will deliver more than $450B in economic value by 2028. Industry momentum is unmistakable: enterprises are rapidly shifting from passive large language model (LLM) chatbots to active, tool-using AI agents that can retrieve data, execute workflows, and make decisions. NVIDIA\u2019s position aligns with this shift.<\/p>\n","protected":false},"author":339,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-270","post","type-post","status-publish","format-standard","hentry","category-news-item"],"fimg_url":false,"_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/270","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=270"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/270\/revisions"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=270"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=270"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=270"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=270"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":257,"date":"2025-05-29T15:50:11","date_gmt":"2025-05-29T15:50:11","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=257"},"modified":"2025-06-23T15:12:01","modified_gmt":"2025-06-23T15:12:01","slug":"next-engineers-celebrates-2025-graduates","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/next-engineers-celebrates-2025-graduates\/","title":{"rendered":"Next Engineers Celebrates 2025 Graduates"},"content":{"rendered":"\n The Next Engineers: Engineering Academy Greenville<\/strong> proudly announces the graduation of its 2025 cohort, marking a significant milestone in the program’s mission to inspire the next generation of engineers.<\/p>\n\n\n\n This year, 33 students from 16 high schools across the Upstate completed the rigorous three-year program, which combines hands-on design challenges, career coaching, and college-readiness workshops. Each graduate is eligible for a $20,000 scholarship upon enrollment in a qualified engineering or engineering-related degree program.<\/p>\n\n\n\n This is the second graduating class of Next Engineers: Engineering Academy students in Greenville, South Carolina. Since its inception in 2021, this college- and career-readiness program has been a collaborative effort between Clemson University’s PEER & WISE program and GE Vernova, creating opportunities for young people to become engineers. To date, the Next Engineers program in Greenville has reached nearly 5,400 students and awarded over $900,000 in scholarships to qualifying graduates.<\/p>\n\n\n\n Students in the program engaged in hours of hands-on educational activities, including building water filtration systems, testing prototype helmets, and presenting their designs to peers and professional engineers. These experiences have equipped them with essential skills such as teamwork, problem-solving, and communication.<\/p>\n\n\n Serita Acker<\/strong>, executive director of PEER & WISE, expressed her pride in the program’s success:<\/p>\n\n\n\n “Now in our fourth year, Next Engineers continues to transform lives by investing in the future of STEM and empowering the next generation of diverse innovators. I am especially grateful for the Clemson University faculty, staff, students, and graduate students for their continued engagement, as well as for the partnership and support from Upstate high schools who help make this initiative possible.”<\/em><\/p>\n\n\n\n The Next Engineers: Engineering Academy is a transformative learning experience designed for students aged 15 to 18. Through a rigorous curriculum, immersive design challenges, and career coaching, participants learn to think and act like engineers.<\/p>\n\n\n\n The graduates and their high schools are:<\/strong><\/p>\n\n\n\n David Bellino<\/strong> \u2014 Greenville Technical Charter High School Read the GE Vernova press release below:<\/strong> The Next Engineers: Engineering Academy Greenville proudly announces the graduation of its 2025 cohort, marking a significant milestone in the program’s mission to inspire the next generation of engineers. This year, 33 students from 16 high schools across the Upstate completed the rigorous three-year program, which combines hands-on design challenges, career coaching, and college-readiness workshops. […]<\/p>\n","protected":false},"author":3922,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14094],"class_list":["post-257","post","type-post","status-publish","format-standard","hentry","category-news-item"],"fimg_url":false,"_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/257","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/3922"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=257"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/257\/revisions"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=257"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=257"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=257"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=257"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":247,"date":"2025-05-16T18:00:08","date_gmt":"2025-05-16T18:00:08","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=247"},"modified":"2025-06-23T15:12:07","modified_gmt":"2025-06-23T15:12:07","slug":"building-resilience-building-leaders-cedc-hosts-2025-spring-summit","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/building-resilience-building-leaders-cedc-hosts-2025-spring-summit\/","title":{"rendered":"Building Resilience, Building Leaders: CEDC Hosts 2025 Spring Summit"},"content":{"rendered":"\n <\/p>\n\n\n\n Clemson Engineers for Developing Communities (CEDC) hosted its 2025 Spring Summit at the end of the semester, showcasing the innovative, student-led projects that are making a difference in South Carolina and beyond. Held at the Hendrix Student Center, the event brought together students, faculty, alumni, and community partners to celebrate another year of impactful work and leadership development.<\/p>\n\n\n\n Founded in 2009, CEDC is a service-learning program within the College of Engineering, Computing and Applied Sciences (CECAS) that empowers students to develop sustainable solutions for low-capacity communities while cultivating critical leadership, project management, and community engagement skills.<\/p>\n\n\n\n This year\u2019s Summit highlighted projects across South Carolina, Haiti, Ecuador, and Colombia. From water system designs in the Amazon to infrastructure resilience planning across South Carolina, students tackled real-world challenges with creativity, compassion, and a focus on strengthening community capabilities.<\/p>\n\n\n New for 2025, CEDC introduced the Challenge Coin of Leadership, a tradition that encourages students to recognize leadership excellence among their peers. Each coin represents a commitment to passing leadership forward, building a culture of service and accountability that extends beyond the classroom.<\/p>\n\n\n\n The Summit also spotlighted major initiatives like ResilientSC, a growing effort to address two critical challenges facing South Carolina:<\/p>\n\n\n\n In addition to project presentations, students hosted a cultural awareness reception featuring Mediterranean cuisine, emphasizing the importance of global engagement through food, conversation, and shared experiences.<\/p>\n\n\n\n Several students were recognized with awards for outstanding service and leadership:<\/p>\n\n\n\n As CEDC continues to grow, its mission remains clear: Serving the developing world, developing those who serve.<\/em><\/p>\n\n\n\n \u201cThe true measure of our success is not what we build, but who we lift up,\u201d <\/em>said David Vaughn, Director of CEDC and Professor of Practice in CECAS. \u201cThrough CEDC, our students are not just becoming engineers\u2014they\u2019re becoming leaders, change-makers, and community builders.\u201d<\/em><\/p>\n\n\n\n For more information about CEDC and how to get involved, visit cecas.clemson.edu\/cedc.<\/p>\n","protected":false},"excerpt":{"rendered":" Clemson Engineers for Developing Communities (CEDC) hosted its 2025 Spring Summit at the end of the semester, showcasing the innovative, student-led projects that are making a difference in South Carolina and beyond. Held at the Hendrix Student Center, the event brought together students, faculty, alumni, and community partners to celebrate another year of impactful work […]<\/p>\n","protected":false},"author":339,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[14092],"tags":[],"coauthors":[14093],"class_list":["post-247","post","type-post","status-publish","format-standard","hentry","category-news-item"],"fimg_url":false,"_links":{"self":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/247","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/users\/339"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/comments?post=247"}],"version-history":[{"count":0,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/posts\/247\/revisions"}],"wp:attachment":[{"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/media?parent=247"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/categories?post=247"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/tags?post=247"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/blogs.clemson.edu\/cecas\/wp-json\/wp\/v2\/coauthors?post=247"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}},{"id":249,"date":"2025-05-13T18:37:33","date_gmt":"2025-05-13T18:37:33","guid":{"rendered":"https:\/\/blogs.clemson.edu\/cecas\/?p=249"},"modified":"2025-06-23T15:12:12","modified_gmt":"2025-06-23T15:12:12","slug":"powered-by-nvidia-clemson-students-create-ai-teaching-assistant-for-smarter-scalable-learning","status":"publish","type":"post","link":"https:\/\/blogs.clemson.edu\/cecas\/powered-by-nvidia-clemson-students-create-ai-teaching-assistant-for-smarter-scalable-learning\/","title":{"rendered":"Powered by NVIDIA: Clemson Students Create AI Teaching Assistant for Smarter, Scalable Learning"},"content":{"rendered":"\n Clemson University computer science students teamed up with global AI leader NVIDIA to develop an AI-powered virtual teaching assistant designed to enhance learning and uphold academic integrity.<\/p>\n\n\n\n They built the tool using NVIDIA\u2019s cutting-edge AI Blueprint for Retrieval-Augmented Generation (RAG), creating a system that helps students strengthen their understanding of core concepts while offering scalable, around-the-clock support. The assistant addresses growing demand for tutoring in high-enrollment STEM courses and demonstrates how AI can responsibly support education.<\/p>\n\n\n\n \u201cThis project shows what\u2019s possible when students work directly with advanced AI technology from industry leaders like NVIDIA,\u201d said Carrie Russell, professor of practice in Clemson\u2019s School of Computing. \u201cOur team tackled a pressing challenge in education and delivered a solution that\u2019s both innovative and practical. We\u2019re proud to collaborate with NVIDIA on work that has the potential to transform learning at scale.\u201d<\/p>\n\n\n\n Karthick Iyer, Vice President, Engineering at NVIDIA, said he was impressed with the students\u2019 ability to turn cutting-edge AI tools into a practical solution for real educational challenges.<\/p>\n\n\n\n \u201cIt\u2019s inspiring to see Clemson students apply NVIDIA\u2019s AI Blueprint for RAG to create a tool that directly tackles real-world issues in education,\u201d Iyer said. \u201cThis collaboration shows the power of pairing AI innovation with hands-on student learning, and we\u2019re thrilled to support projects that advance both technology and education.\u201d<\/p>\n\n\n\n Students Camden Spehl, Kyle Zheng, Jaylen Goddard, and Cordarro Higgins Redman collaborated with Vinay Raman, a senior deep learning scientist at NVIDIA, to bring the project to life.<\/p>\n\n\n\n
\n\n\n\n<\/a>Why Edge Models Need Their Own Evaluation<\/strong><\/h2>\n\n\n\n
\n
\n
\n
\n
<\/figure>\n\n\n\n<\/figure>\n\n\n\n\n
Katherine Mulholland, Executive Director of SCCMS, states, \u201cDuke Energy Science Nights give students and families engaging, hands-on STEM experiences that spark curiosity and connect learning to real-world careers. As South Carolina continues to grow and evolve, these events play an important role in inspiring the next generation of problem-solvers and preparing our future workforce. By exposing students early to STEM pathways, we are helping build a strong pipeline of talent that will power our state\u2019s economy for years to come\u201d<\/p>\n\n\n\n
Katherine Mulholland,<\/strong> kmulholland@sccoalition.org<\/a>
Elena Stout,<\/strong> estout@s2temsc.org<\/a> <\/p>\n","protected":false},"excerpt":{"rendered":"
\n\n\n\n\n
\n
\n
\n\n\n\n\n
\n
\n\n\n\n\n
\n
\n
\n
\n\n\n\n<\/figure>\n\n\n\n<\/figure>\n\n\n\n\n
\n\n\n\n\n
\n
\n
\n
\n\n\n\n
\n\n\n\n
\n\n\n\n
John Buckingham<\/strong> \u2014 Greenville Technical Charter High School
Joy Dodd<\/strong> \u2014 Travelers Rest High School
Michael Driscoll<\/strong> \u2014 Greenville Technical Charter High School
Ryan Ellis<\/strong> \u2014 Westside High School
DeAngelo Fuller<\/strong> \u2014 Westside High School
Tristan Greenleaf<\/strong> \u2014 Pickens High School
Christian Harling<\/strong> \u2014 Southside High School
Brock Hinson<\/strong> \u2014 J.L. Mann High School
Abdurrahman Housari<\/strong> \u2014 Greenville Technical Charter High School
Omar Housari<\/strong> \u2014 Greenville Technical Charter High School
Advaith Joshi<\/strong> \u2014 J.L. Mann High School
Thomas Kezman<\/strong> \u2014 Greer Middle College
Noah King<\/strong> \u2014 D.W. Daniel High School
Riya Kot<\/strong> \u2014 Southside High School
Adam Le<\/strong> \u2014 Riverside High School
Rami Makhtoub<\/strong> \u2014 Greenville Technical Charter High School
Noah Miller<\/strong> \u2014 Easley High School
Hayley Morton<\/strong> \u2014 Greenville Technical Charter High School
Levi Naylor<\/strong> \u2014 Easley High School
Tron Paul<\/strong> \u2014 Crescent High School
Hayden Ramsey<\/strong> \u2014 Travelers Rest High School
Harm Ravenhorst<\/strong> \u2014 Greenville High School
Hallie Ray<\/strong> \u2014 D.W. Daniel High School
Addison Reid<\/strong> \u2014 Blue Ridge High School
James Robinson<\/strong> \u2014 Mauldin High School
Alejandra Rodriguez<\/strong> \u2014 Easley High School
Josh Sparks<\/strong> \u2014 Legacy Early College High School
Jeffrey Sweeney<\/strong> \u2014 Pendleton High School
Sai Praneetha Thatavarthi<\/strong> \u2014 J.L Mann High School
Dimitri Tsirkas<\/strong> \u2014 Easley High School
Jade Williams<\/strong> \u2014 Legacy Early College High School
Kylie Willis<\/strong> \u2014 Pendleton High School<\/p>\n\n\n\n
https:\/\/www.gevernova.com\/news\/press-releases\/ge-vernova-clemson-university-celebrate-next-engineers-academy-graduation<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<\/figure>\n\n\n\n<\/figure>\n<\/div>\n\n\n\n
<\/figure>\n<\/div>\n\n\n\n