reinforcement learning
29 published articles
Fisher-R1 | Hypothesis testing
AI agents run flawless statistics and still draw the wrong conclusions
Agents that automate hypothesis testing can execute analyses flawlessly yet reach wrong conclusions, because most benchmarks never check whether the p-value is valid. A 425-task benchmark quantifies the gap, and an RL-trained open-weight model closes a chunk of it.
2026-08-19
Research
Beacon: your tool-using AI model is making easy questions harder
A new paper from KlingTeam measures when multimodal models actually need tools and when tools hurt. The proposed Beacon model, trained with necessity-aware rewards, improves both accuracy and tool discipline.
2026-08-09
Explainable AI
The knob no one wanted to turn: dtControl2+ε lets you sacrifice optimality for clarity
A new extension to dtControl2 allows engineers to trade a precise amount of performance for smaller, more understandable decision trees. The tool, dtControl2+ε, guarantees ε-optimality while pruning trees to orders of magnitude fewer nodes.
2026-08-05
Open Source
OpenForgeRL trains AI agents without touching their inference harnesses
OpenForgeRL uses a proxy and Kubernetes to train AI agents without modifying their inference harnesses. In tests, OpenForgeGUI matched models several times larger on web navigation and desktop benchmarks.
2026-08-01
Deep Research AI
BAAI's new research agent doesn't just search longer, it checks its own homework
BAAI's AREX agents use a recursive self-improvement loop that compresses long research histories, enabling efficient verification-driven refinement. Trained with a novel long-horizon RL recipe, they outperform comparable baselines on BrowseComp, DeepSearchQA, and HLE.
2026-07-31
Reinforcement learning
Microsoft's experiential learning fix gives AI models a coach, not just a score
Experiential Learning repurposes the LLM-as-a-Judge into an LLM-as-a-Coach that extracts transferable knowledge from each response and internalizes it via on-policy context distillation, beating rubric-based RL on held-out tasks and reducing reward hacking.
2026-07-30
Research
Explainable RL takes on air traffic control, one saliency map at a time
A team of researchers applies explainable RL to air traffic control, training an agent to avoid no-fly zones and using saliency maps to reveal its reasoning. The work is a preliminary step toward trust in high-stakes AI.
2026-07-27
Research
Shared memory helps AI agents, but only until you hit five of them
Researchers shared replay buffers across actor-critic agents on parameterized action tasks. GAC jumped in performance, but SAC and TQC only crept forward. Beyond five agents, the computational cost climbs with no meaningful return. The paper offers a practical boundary for how far shared-experience methods can stretch before they stall.
2026-07-27
BCI research
Your brain waves just got 41 percent sharper. No new hardware needed.
Researchers propose CNN-LSTM-RL, a two-stage framework that applies reinforcement learning to correct systematic errors in non-invasive brain-computer interface decoders. Without any additional neural data, the method improved 3D movement decoding correlation by 41.5% and reduced root mean square error by 40.2% across ten participants.
2026-07-27
Reinforcement learning for self-correcting vision models
Meta's new training trick teaches AI to catch its own mistakes without a teacher
Meta and UIUC researchers developed SVR-R1, a training framework that lets vision-language models check their own answers and rethink when they get them wrong, all within a reinforcement loop. No teacher. No external critic.
2026-07-25
RefineRL skeptical loop pushes 4B models past 32B rivals
The 4B model that beats 32B ones by refusing to trust itself
RefineRL trains small language models to iteratively refine their own competitive programming solutions using a skeptical agent and reinforcement learning. A 4B model using this method outperforms 32B models and approaches 235B performance, suggesting that self-refinement, not raw size, may be a stronger scaling path for reasoning tasks.
2026-07-25
Agentic RL research
RL's sparse-reward blind spot meets SEED: agents write their own lessons
SEED (Self-Evolving On-Policy Distillation) lets an LLM analyze its own past trajectories, extract reusable natural-language skills from them in hindsight, and distill those lessons back into its policy during RL. The result: denser, on-policy supervision that improves success rates by up to 22% on long-horizon tasks.
2026-07-25