Alignment Gaps, Agent Governance, and Greener LLMs
Three new papers expose a hidden flaw in DPO training, propose policy-as-code governance for enterprise agents, and cut LLM serving energy use by 26% via GPU power control.
Articles Tagged "Reinforcement Learning"
Self-Correcting Models, Smarter Monitors, AI Designs Itself
Three new papers tackle critique dependency in LLMs, ensemble monitoring for AI control, and agents that autonomously discover better neural architectures.
SU-01
SU-01 is a 30B-A3B MoE reasoning model from Shanghai AI Lab that achieves gold-medal performance on IMO 2025, USAMO 2026, and IPhO 2024/2025 using a three-stage training recipe and test-time scaling.
Olympiad Gold, Broken Memories, and Attention Loss
A 30B model earns IMO gold, memory consolidation silently corrupts agents, and a new metric predicts when LLMs lose track of their instructions.
Agent Memory in 2026: Circuits, Tiers, Evolution
Three new papers reveal how agent memory silently breaks, how a tiered architecture recovers it, and how models can self-improve without human labels.
Async RL Speedups, Unsafe Robots, and Routing Math
Three papers: 2-4x async RL training speedup, alarming 54.4% safety violation rate in medical robots, and a training-free routing trick that lifts math accuracy 3-7%.
David Silver Raises $1.1B to Build AI Without Human Data
David Silver, creator of AlphaGo and AlphaZero, closed a $1.1B seed round for Ineffable Intelligence - a London lab building AI that learns without human data.
Leaner Reasoning, Fragile Agents, and Model Self-Audit
Three new papers tackle reasoning token waste, orchestration failures across 22 agent frameworks, and a method for teaching LLMs to describe their own learned behaviors.
LeCun's JEPA World Model Plans 47x Faster on One GPU
LeWorldModel from Yann LeCun's group strips JEPA world models down to two loss terms, trains 15M parameters on a single GPU in hours, and plans roughly 47x faster than DINO-WM.
Physical Intelligence Launches π0.7 for Untrained Tasks
Physical Intelligence's π0.7 robot model can generalize to tasks it was never explicitly trained on, matching fine-tuned specialist models through compositional skill recombination.
Claude Beat Human Alignment Researchers - Then Failed
Nine Claude Opus 4.6 agents outperformed human researchers on a core alignment benchmark, hitting 97% vs 23% in five days - then showed no statistically significant improvement in production.
Autonomous Research, Broken Reasoning, Smarter Agents
Three new papers: AlphaLab runs autonomous GPU research campaigns, open-weight reasoning models collapse under text reformatting, and HiL-Bench reveals agents can't decide when to ask for help.
