Short Industry Talks pt.2
summary
New decoding rule for discrete diffusion language models claims to beat confidence-based parallel sampling by explore-then-exploit selection of high-entropy tokens, with a provable information-theoretic argument that confidence-only schemes require ~n rounds and experiments on GSM8K showing a better accuracy vs rounds Pareto frontier using exploration batch size one. A browser agent pipeline built for Comet uses consolidated tools (form-fill), compares HTML vs screenshots as views, and runs site-specific instructions plus a prompt-improver loop and a trajectory classifier to block prompt-injection.
AI-written summary · about
chapters (14)
- 0:00Diffusion language models intro
- 3:06Autoregressive vs masked diffusion
- 9:52Confidence-based decoding limits
- 11:47Explore-exploit decoding
- 16:13Training alignment with decoding
- 18:47Summary and future
- 20:23Browser agents intro
- 23:08Generalizing across domains
- 25:37Agent architecture design
- 29:32Learning from user interactions
- 32:05Browser environment challenges
- 32:59Domain-specific LLM expertise
- 34:33Prompt injection risks
- 35:38Closing remarks
related talks
01h 08mFormal Methods for Frontier AI Systems
46mArchitecting Efficient Agents with Semantic Routing and Mixture-of-Tools
38mCoding Agents & Language Evolution: Navigating Uncharted Waters
36mTalk by Kamalika Chaudhuri (Meta)
35mUsing coding agents with open models
40mSurviving the Agentic Hype with Small Language Models
45mDiffusion to Prod: vLLM-Omni Unifies Multimodal Inference
43mAdvanced Prompt Engineering and Beyond