Stable RL Alignment with Unified Pair-GRPO Preference Constraints
Discover how the unified Pair-GRPO framework improves stable and general RL alignment using implicit to explicit preference constraints for better LLM trai...
Grounded Correspondence: Enhancing Temporal Consistency in Video Learning
Discover how Grounded Correspondence improves temporal consistency in video object-centric learning using deterministic matching without prediction models.
AI in Number Theory: LLMs for Algorithms & Verification
Explore how AI and LLMs enhance number theory by generating algorithms and verifying conjectures with machine learning ensemble methods.
Shepherd: Fast Runtime for Meta-Agents with Formal Traces
Discover Shepherd, a runtime substrate that formalizes execution traces and boosts meta-agent efficiency with rapid forking and replay capabilities.
Decision-Centric Memory Framework for AI Agents
Discover a novel rate-distortion framework that optimizes AI agent memory by focusing on decisions over descriptions for better performance.