post-training
43 posts tagged post-training.
- Agentic-OPD says your best RL checkpoint is not the whole model
- Token-Level Reward Hacking: Where Preference Optimization and Verifiable RL Actually Stand
- Arete-27B tests values as a post-training layer
- Scientist World Turns Research Papers Into Runnable Environments
- An Independent Replication of Direct Nash Optimization, and What Its Ablations Actually Prove
- What It Takes to Make a Bad Model Behavior Actually Stay Gone
- When the Chain of Thought Doesn't Actually Drive the Answer
- Reading Evolution Strategies as a Center of Mass, Not a Gradient
- The Teacher That Scores Worse Than It Plays: Fixing Agentic Distillation
- When Fine-Tuning for One Behavior Drags Another Along
- MLExpert and the Case for Training Agents on Whole Experiments, Not Single Turns
- When Renaming a Tool Breaks Your Agent: What CIPR Fixes
- Latent Thinking, Text Acting: What HybridAgent Changes About Agent Cost
- Alignment generalization is a behavior problem, not a values checklist
- JOLT makes self-distillation follow the student’s path
- Multi-teacher distillation should copy the delta, not the whole teacher
- The Explorer and the Student: How ExpDis Splits RL Into Two Jobs
- The First Word a Model Writes Decides How Well It Reasons
- Pivot-SD trains the denoising choices that steer diffusion LMs
- When the Prompt's Phrasing Changes the Answer: SCAPO and Token-Level Credit in RLVR
- Not All Quantization Errors Are Equal: STEPQuant and the Memory Cost of Linear Attention
- KV-streams make long-agent training cheaper by not throwing memory away
- Self-Improving Agents Without RL: What Retrospection, Creativity, and Token Forecasting Actually Show
- Teaching a Reasoning Model to Know When It's Sure Cuts Its Token Bill
- MISVO steers frozen language models at inference time
- The World Model That Edits the Agent's Mind, Not the Environment
- onPanda turns alignment feedback into token-level steering
- Jev is a decision model, not another chatbot
- When the Teacher Knows to Quit: RetireOPD and Distillation for Agents
- ComPO and the Case Against Optimizing the Loss You Wrote Down
- RetroThinker lets speech models correct themselves while you are still talking
- Layer-selective unlearning aims at the parts of a model that remember
- An AI Just Outscored the Top Human at IOI 2026. Here's What That Actually Means
- The SFT-RL Split Has a Wide Safe Zone, and You Can Find It Cheap
- Small dialogue agents need repair loops, not just bigger training
- IAR makes retrieval-free document QA less brittle
- The token-budget bug hiding in multi-teacher distillation
- When the Teacher and the Verifier Disagree: Fixing On-Policy Distillation for Long Context
- Post-training is now the behavior layer
- Output Reset makes PPO smoother, not automatically better
- Muon’s agent RL win is real, narrow, and stack-dependent
- Reusing RL Gains Across Model Sizes: The Case for Direct-OPD
- Post-training logprobs as a step-level critic for agents