reinforcement-learning
97 posts tagged reinforcement-learning.
- AutoScaffold trains models with hints, then removes the hints
- Agentic-OPD says your best RL checkpoint is not the whole model
- Dense rewards made clinical agents ask better questions
- MaxIM treats agent memory as a trained compression problem
- Token-Level Reward Hacking: Where Preference Optimization and Verifiable RL Actually Stand
- Grasp-TAGS picks the grasp before the robot commits
- 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
- WebHarbor Puts 153 Fake Real Websites in a Box for Agents to Train On
- Why VLM robot reward models hallucinate progress, and a prompt fix that works
- LARC moves image reasoning out of token space
- When the Chain of Thought Doesn't Actually Drive the Answer
- Verifying Computer-Use Agents by Letting the Verifier Go Look
- Reading Evolution Strategies as a Center of Mass, Not a Gradient
- RobustRLlib maps the failure boundary for real-world RL
- METTA splits agent memory into moments and episodes
- ASTEROID teaches agents to explore before they act
- LongZero trains long-context reasoning without outside documents
- Sparse RLHF makes private preference data less impossible
- Why the Best Solver Model Is Rarely the Best Environment Author
- The Teacher That Scores Worse Than It Plays: Fixing Agentic Distillation
- CoopUQ turns uncertainty into a navigation cost for robot teams
- WeAct Turns LoRA Adapters Into Something an Agent Can Pick Mid-Task
- MLExpert and the Case for Training Agents on Whole Experiments, Not Single Turns
- CoPE trains agents to forecast their own future actions
- Diffusion Models Are Starting to Generate Robot Policies, Not Just Robot Videos
- HiRe-MoMa fine-tunes robots by correcting policies instead of replacing them
- PrismaCoder's real trick: a rubric that both filters data and rewards training
- The Case for Making Research Agents Forget Their History
- When Renaming a Tool Breaks Your Agent: What CIPR Fixes
- Latent Thinking, Text Acting: What HybridAgent Changes About Agent Cost
- Tev's Dense Progress Rewards and the Sparse-Reward Problem in Long-Horizon Agents
- ViSkill and the case for keeping agent memory visual
- JOLT makes self-distillation follow the student’s path
- The Explorer and the Student: How ExpDis Splits RL Into Two Jobs
- Co-Evolving the Attacker: How AdvSim2Real Hardens Web Agents Against Prompt Injection
- Recommendation systems need a throttle, not just a better ranker
- Sherpa trains tutors on learning outcomes, not lesson vibes
- CLIFT turns web-agent judging into reusable verification
- MemPilot and the Case for Deciding Memory at Query Time, Not Ingest Time
- The First Word a Model Writes Decides How Well It Reasons
- RL Environments on Hugging Face Make Agent Work Less Hand-Wavy
- Stratego Is a Better Agent Test Than Another Chess Win
- FERPO avoids action-gradient trust in continuous-control RL
- KaliBench Measures Whether LLMs Can Actually Type the Right Security Command
- When the Prompt's Phrasing Changes the Answer: SCAPO and Token-Level Credit in RLVR
- AdviSD trains a small model to steer bigger frozen models
- Robots Need Correction Loops, Not Just Better Planners
- KV-streams make long-agent training cheaper by not throwing memory away
- Reinforcement learning weakens reasoning distillation defenses
- Self-Improving Agents Without RL: What Retrospection, Creativity, and Token Forecasting Actually Show
- X-Reset turns human hand demos into better robot starting points
- UECR-GRPO treats the teacher as evidence, not an oracle
- A Small Model That Steers a Bigger One's Reasoning
- Speaker-Centered Memory: Why Group Chats Break Your AI Agent
- Critical-State RL trains the one agent call that actually matters
- Where Harness Self-Improvement Actually Stands in Late 2026
- Xiaomi's MiMo V2.6 Ships in Three Flavors, and the Split Matters
- When the Teacher Knows to Quit: RetireOPD and Distillation for Agents
- ScienceBuddy and the case for training the harness before the model
- Bellman Policy Optimization cuts one moving part from RLVR
- OptiFlow treats offline RL policy learning as sample matching
- Recursive Self-Improvement Has a Roadmap Now, and Most of It Isn't Built
- SAGE Uses a Big Model as a Coach, Not a Crutch
- The SFT-RL Split Has a Wide Safe Zone, and You Can Find It Cheap
- Verbal reinforcement learning is a feedback routing problem
- Aero Hand Open makes the cheap part of a robot hand the learnable part
- Can a robot think out loud before it moves? R³ tests the idea
- BPCO Brings the Critic Back to RL Fine-Tuning
- Splitting the reward: how G-CARL grades medical explanations for facts and for tone
- PGFS++ and the Reward Magnet Problem in AI Drug Design
- SPADE Makes the Training Environment a Thing the Model Learns to Build
- LLM reward shaping without changing the goal
- RCI turns stop signals into safer offline RL training data
- When Your User Simulator Is Secretly One Person: Simulator Collapse in Multi-Agent RL
- The Alignment Tax on Creativity, and a Switch to Turn It Back On
- Moment closure brings uncertainty back into model-based RL planning
- Two prepared policies may be the sweet spot for uncertain MDPs
- RL can teach code models to care about runtime, but the stopwatch is the hard part
- Training Agents Inside the Harness They Actually Ship With
- Chess Shows What RL Actually Does to a Reasoning Model
- DADiff uses diffusion to measure when an RL policy stops transferring
- Teaching a Model to Zoom In on a Chart Before It Judges a Claim
- The Missing Half of RL for Diffusion Language Models
- TerraZero bets on self-play for the driving long tail
- REGRIND’s one-demo recipe for robot hands
- Fraud detection needs a Pareto frontier, not another magic score
- Agon Grades the Reasoning, Not Just the Answer
- Reusing RL Gains Across Model Sizes: The Case for Direct-OPD
- Verification as a Scaling Axis: What the LLM-as-a-Verifier Paper Actually Changes
- LeRobot v0.6.0 Turns Robot Simulation Into a Feedback Loop
- The Cheap Way to Test Whether Your Agent's Step Scores Actually Mean Anything
- TRIAGE gives agent RL a better target than pass or fail
- Finger-Level Ownership: How DexCompose Stacks Robot Hand Skills Without Breaking Them
- HiReLC Treats Model Compression as Coordination, Not a Knob
- Tool-use RL is failing at the brackets, not the tools
- CoorDex and the End of Stop-and-Go Humanoids