Overview
COLM is an academic venue focused on the study of language modeling, broadly defined, with the goal of creating a community of researchers with expertise in different disciplines, focused on understanding, improving, and critiquing the development of LM technology.
Booth schedule
Presentations:
11:30 - 12:00 PM Talk: Super Weights in LLMs and the Failure of Selective Training - Akarsha Sehwag
12:00 - 12:30 PM Talk: BubbleScale: Predictive Environment Prewarming for Agentic RL Rollout - Sathishkumar Sivashanmugam
1:00 - 1:30 PM Talk: L3M: A Lean Harness for Multi-Agent Work - Carson Eisenach
1:30 - 2:00 PM Talk: Trie Automata for Constrained Decoding over Large Finite Sets - Xingzi Xu
2:00 - 2:30 PM Talk: Human-Centered Edge AI: From Robotics to AR Glasses - Yelin Kim
Meet the scientist:
11:00 AM - 12:00 PM
Domain/research |
Team |
Amazonian |
Reinforcement learning, inference |
eCF |
|
Generative AI, Inference-time compute, Multi-agent systems |
AWS |
|
Generative AI, Large Language Models, Agentic AI |
AWS Generative AI Innovation Center |
12:00 - 1:00 PM
Domain/research |
Team |
Amazonian |
LLMs, Agents, Evaluation, Reasoning, Text Detection and Authorship Attribution |
Worldwide Amazon Stores |
|
AI infra, KV Cache, vLLM etc |
AWS |
Sathishkumar Sivashanmugam |
1:00 - 2:00 PM
Domain/research |
Team |
Amazonian |
Multi-lingual LLMs, Post-Training, Alignment |
Alexa International |
|
Human-Centered Edge AI, Robotics, AR Glasses |
Amazon |
4:30 - 5:30 PM
Domain/research |
Team |
Amazonian |
Post-training for agentic coding models |
AWS - Kiro Science |
|
LLM, Post-Training, Agent |
AGI |
|
Multimodal LLM, NeuroSymbolic and Multi-Agent Reasoning |
AWS Generative AI |
|
NLP, Recommendation System, LLM, Agent, RAG, Responsible AI |
CXBT |
Thanh Tran |
5:00 - 6:00 PM
Domain/research |
Team |
Amazonian |
NLP, Recommendation System, LLM, Agent, RAG, Responsible AI |
CXBT |
Thanh Tran |
Presentations:
11:30 - 12:00 PM Talk: OrchOpt: Validation-Gated Architecture Search for Self-Organizing Multi-Agent Orchestration - Amit Dhanda
12:00 - 12:30 PM Talk: Augmented Hypothesis Testing with Persona-Based LLM Simulations - Saab Mansour
12:30 -1:00 PM Talk: The guide to Amazon AI chips - Armin Agha-Ebrahim, Marisa Klee
1:00 - 1:30 PM Talk: Chart-RL: Policy Optimization Reinforcement Learning for Enhanced Visual Reasoning in Chart Question Answering with Vision Language Models - Amit Dhanda
1:30 - 2:00 PM Talk: Inside Amazon University Talent Acquisition - Ankita Goyal
2:00 - 2:30 PM Talk: Beyond Task Success: An Eight-Metric Tiered Evaluation Protocol for LLM Agents over Fragmented Operational Data - Neeraj Baji
4:30 - 5:00 PM Talk: Brick-DICL: Dynamic In-Context Learning for Automated Brick Schema Classification - Yiyue Qian
Meet the scientist:
11:00 AM - 12:00 PM
Domain/research |
Team |
Amazonian |
Multimodal LLM, NeuroSymbolic and Multi-Agent Reasoning |
AWS Generative AI |
|
Post-training for agentic coding models |
AWS - Kiro Science |
|
Agentic AI Evaluation |
AWS - AMPS |
|
RL, LLMs, Multi-agent |
SCOT |
12:00 - 1:00 PM
Domain/research |
Team |
Amazonian |
Reinforcement Learning, LLMs, Agentic Systems |
AWS Marketplace |
|
Video Understanding, Multi-Modal Learning, World Models, LLM post-training |
CS Stores |
|
Medical NLP and Imaging, LLMs, Representation Learning, Deep Learning, Radiology |
Healthcare AI |
1:00 - 2:00 PM
Domain/research |
Team |
Amazonian |
AF-Tech AI |
Sina Ghotbi |
|
Reasoning using LLMs, RL training LLMs and Agent Harnesses |
WW Ops |
Sreekar Aditya |
Multi-agent systems, low latency AI infra, Domain specific agent evals, Agents for mission critical systems |
Continuous Infrastructure Automation |
4:30 - 5:30 PM
Domain/research |
Team |
Amazonian |
Reinforcement Learning, LLM post-training for system kernel generation |
Annapurna Labs Neuron Science Team |
Jiin Woo |
Multi-lingual LLMs, Post-Training, Alignment |
Alexa International |
|
model post training, SFT, RFT, OPD, Agentic RL, Multi-agents,LLM-as-a-Judge |
AWS Generative AI Innovation Center |
Presentations:
11:30 - 12:00 PM Talk: AI rap battle demo: Can you out-rhyme the agent? - Nikita Kozodoi, Jack Butler
12:30 - 1:00 PM Talk: L3M: A Lean Harness for Multi-Agent Work - Carson Eisenach
2:00 - 2:30 PM Talk: RELISH: LLM REgression with a Latent Iterative State Head - Matt Lease (Amazon Scholar), Yiheng Su (Mentee)
Meet the scientist:
11:00 AM - 12:00 PM
Domain/research |
Team |
Amazonian |
Multimodal LLM, NeuroSymbolic and Multi-Agent Reasoning |
AWS Generative AI |
|
DS Agent, Agent Evaluation, Data Synthesis, Agent Optimization |
AWS |
Chi Zhang |
LLMs, Agents, Simulation, NLP, Multilinguality |
AWS |
Saab Mansour |
12:00 - 1:00 PM
Domain/research |
Team |
Amazonian |
Multi-lingual LLMs, Post-Training, Alignment |
Alexa International |
1:00 - 2:00 PM
Domain/research |
Team |
Amazonian |
Episodic Long-Term Memory for Language Agents |
Devices and Services |
|
Video Understanding, Multi-Modal Learning, World Models, LLM post-training |
CS Stores |
4:30 - 5:30 PM
Domain/research |
Team |
Amazonian |
NLP, Agentic AI |
Stores |
Arjun Mishra |
Reinforcement Learning, Coding Agents, LLMs |
AWS - Kiro Science |
Expo Talk
Speaker: Matt Lease, The University of Texas at Austin, Amazon Scholar
Abstract: How can we best harness rapidly advancing AI capabilities to accelerate scientific discovery? Because the use of AI can bring both benefits & risks, responsible AI governance advocates for assessing risks as well as benefits that could stem from AI adoption. Moreover, rather than react to such harms after they have already occurred, we should anticipate, prevent, and mitigate such harms whenever possible. In general, potential benefits should be assessed to outweigh the risks before pursuing AI adoption, similar to how institutional review boards (IRBs) similarly require proposed human subjects research to justify potential benefits outweigh risks prior to study approval. After briefly introducing foundations of responsible AI governance, I will highlight potential benefits vs. risks posed by greater adoption of AI in scientific practice. My talk bridges ongoing work in UT Austin’s campus-wide responsible AI initiative, Good Systems, with that of our NSF-Simons AI Institute for accelerating discovery, CosmicAI.
