Tag: Post-Training

All the articles with the tag "Post-Training".

• From Tool Calls to Policy Updates: A Reproducible Agent RL Stack

  14 Aug, 2026 · 45 min read

  A reproducible blueprint for Agent RL: environment snapshots, verifier contracts, credit assignment, three-policy semantics, asynchronous generation and training, partial rollouts, and runnable smoke tests.

• A Repair Can Save a Trajectory Without Teaching the Agent

  14 Aug, 2026 · 18 min read

  An eight-seed counterfactual-replay experiment found that high-value one-action repairs reliably rescue failed agent trajectories—but a repair-value-dominant SFT selector did not beat simpler rules and recovered only 25% of matched-RL improvement.

• Do Verifier Errors Grow Superlinearly with Horizon? A Three-Stage Experiment

  14 Aug, 2026 · 11 min read

  A controlled long-horizon experiment found a clear horizon effect but no preregistered evidence of superlinear verifier error—and exposed why token measurement and false-negative costs matter before scaling RL evaluations.

• Reproducing CompactRL: What Worked, What Failed, and Why We Did Not Scale

  Updated: 24 Jul, 2026 · 16 min read

  An auditable CompactRL reproduction spanning the public algorithm, a 96-step long-horizon simulation, integration with slime, real Qwen actor-critic training, value-function fixes, 17 experimental phases, and the evidence that stopped us from scaling.

• From Long CoT to Agent Swarms: The Documented Evolution of Kimi's Reinforcement Learning

  18 Jul, 2026 · 15 min read

  A source-grounded history of Kimi's reinforcement-learning stack, from Kimi k1.5's long-context outcome RL and partial rollouts to K2's general RL and K2.5's multimodal GRMs and Parallel-Agent RL.

• Training the Critic Without Crashing the Reward: A Practical Guide to Agentic RL

  12 Jul, 2026 · 20 min read

  A practical framework for critic training and credit assignment in long-horizon LLM agents: IQL, pairwise advantage, hindsight and counterfactual critics, privileged information, turn-level MDPs, chain-of-thought monitoring, and reward-crash diagnosis.