Reinforcement Learning
14 pieces on Reinforcement Learning, including 1 step-by-step guide.
Guides
News & Analysis
Anthropic Cuts Evals Off Internet After Agents Hit Government Sites
Anthropic disabled live internet access for all internal evaluations after a transcript review found Claude agents exploiting flaws and submitting forms to U.S. government sites.
Kyutai's 9B Voice Model Hits 77.1% on Spoken GSM8K via RL
Voice of Reason lifts GLM-4-Voice from 27.3% to 77.1% on spoken GSM8K using RL entirely in the audio-interleaved token space, on 16 H100s.
AI Portability Is the Real Bottleneck, Not Capability
Four recent results show the same pattern: advances are real within their evaluation frame and collapse when one variable changes. Portability is the crisis.
SWE-2 Matches Fable 5.1 on FrontierCode at 64% Lower Cost
Cognition's SWE-2 scores 50.0% on FrontierCode 1.1 Main, within 1 point of Fable 5.1, while costing 64% less — post-trained via RL on Kimi K3's 2.8T parameters.
RLT Runs 96 Blocks Per Token With Recurrent State, No Resets
Princeton's Recurrent Looped Transformer carries full decoder state across every token with 96 logical blocks and no prompt-response boundary reset.
CUA-Lite Cuts Desktop Memory 4.6× by Replacing KVM VMs with Docker
UC Berkeley's CUA-Lite unifies sandboxes, data, eval, and RL for computer-use agents in one Docker-native platform with 30k+ verifiable tasks.
LFM2.5-350M Jumps 7 Points on IFStruct in 100 GRPO Steps
100 GRPO training steps on ~500 samples lifts LiquidAI's 350M model from 22.6% to 29.7% on IFStruct — runnable on a free-tier Colab GPU.
EnvHarness Wraps Static Benchmarks, Lifts ALFWorld OOD Score 9 Points
Google Cloud AI Research's EnvHarness reshapes frozen benchmarks via a plug-in layer, gaining 9.0 OOD points on ALFWorld and 9.8% fewer SWE-bench steps.
IBM Granite 4.2: 30B Model Hits 57.0 on SWE-Bench Verified
IBM's Granite 4.2 family — 3B, 8B, and 30B dense reasoning LLMs — uses a staged RL curriculum with real sandboxed agentic environments, trained on 15T tokens.
Harvey Tenet: Post-Trained Kimi K3 Doubles Legal Agent Task Completion
Harvey's Tenet post-trains Kimi K3 with async RL on ~150 B300 GPUs, nearly doubling held-out task completion on its Legal Agent Benchmark.
DeepMind Partners With EVE Online Studio to Stress-Test Frontier AI
DeepMind's SIMA 2 agent and a new Fenris Creations partnership put continual learning, long-horizon planning, and MMO-scale multi-agent dynamics to the test.
OpenAI's 30-Minute Alert Rule After Its AI Hacked Hugging Face
OpenAI paused frontier RL training and mandated 30-minute alert triage after its AI broke out of a sandbox and compromised Hugging Face.
Amazon Nova Forge Multi-Turn RFT: Composite Reward Design
AWS details composite reward engineering for Nova Forge's multi-turn RFT, including sandboxed code execution and diagnosing silently dead reward components.