Agents Fail Safety, Probes Miss Fanatics, Better RLHF
Three new papers expose gaps in agent safety evaluation, challenge activation-probe reliability for detecting misaligned models, and fix reward hacking in RLHF training.
Articles Tagged "Reinforcement Learning"
Hyperagents, Milestone Rewards, and the 19x Efficiency Win
Three arXiv papers push AI agents further: metacognitive self-modification, milestone-based RL lifting Gemma3-12B from 6% to 43% on WebArena-Lite, and hybrid workflows cutting inference costs 19x.
MiniMax M2.7 Claims to Automate Its Own Training
MiniMax's new 2,300B MoE model tops the Artificial Analysis Intelligence Index and claims to run 30-50% of its own RL research workflow autonomously.
Cursor Ships Composer 2 - Its First In-House Coding Model
Cursor launches Composer 2, its first in-house coding model trained via RL on long-horizon tasks, scoring 73.7 on SWE-bench Multilingual at $0.50/M input tokens.
Multi-Agent Constitution, Sleeper Defense, Skill RL
Three new arXiv papers tackle constitutional AI rule learning, sleeper agent defense for multi-agent pipelines, and skill-evolving reinforcement learning for math reasoning.
MiniMax M2.7
MiniMax M2.7 is a 230B MoE coding agent that handles 30-50% of MiniMax's own RL research workflow, scoring 56.22% on SWE-Pro and 78% on SWE-bench Verified at $0.30/M input tokens.
Enterprise Agents Stall, Safety Gates, Smarter Tool Use
New research shows enterprise AI agents top out at 37.4% success, a deterministic safety gate beats commercial solutions, and an ICLR 2026 paper cuts RL compute by 81%.
16 Open-Source RL Libraries, One Shared GPU Bottleneck
A Hugging Face survey of 16 open-source reinforcement learning libraries finds the entire ecosystem has converged on async disaggregated training to fix a single brutal bottleneck: GPU idle time during long rollouts.
LeRobot v0.5.0: First Humanoid, 10x Training Speed
Hugging Face ships its largest LeRobot update yet: Unitree G1 humanoid support, Pi0-FAST VLA, Real-Time Chunking, 10x faster image training, and PEFT/LoRA fine-tuning for large robot policies.
VLMs Fail Physics Tests, RL Quits Bad Paths, Agents Lie
Three new papers expose systematic VLM failures on basic physics, introduce RL that learns to abandon bad reasoning paths, and reveal that AI agents deceive primarily through misdirection rather than fabrication.
OpenAI's Head of Post-Training Max Schwarzer Joins Anthropic
Max Schwarzer, VP of Research and Head of Post-Training at OpenAI, leaves after a year leading the team that shipped GPT-5, 5.1, 5.2, and 5.3-Codex to return to RL research at Anthropic.
ByteDance Trained an AI Agent That Writes Faster CUDA Kernels Than You
CUDA Agent uses reinforcement learning trained on actual GPU profiling data to generate optimized CUDA kernels. It beats torch.compile by 2.11x overall and outperforms Claude Opus 4.5 and Gemini 3 Pro by 40 points on the hardest kernels.
