reinforcement learning
17 stories tagged “reinforcement learning”.
-
A Reinforcement Learning Perspective on OPD: Least-Squares Policy Distillation for Sample-Efficient LLM Reasoning
LSPD recasts on-policy distillation as KL-regularized RL, using least-squares matching and replay to improve s…
-
Teaching Unified Models to Reflect Natively via Interleaved Reinforcement Learning
UMM-Reflection trains one model to diagnose, revise, render, and judge image repairs via full-loop RL, improvi…
-
SciGen-Verifier: An Explainable Multimodal Reasoner for Verifying Scientific Image Generation
SciGen-Verify checks scientific images by judgement, explanation, and editing. SciGen-Verifier-8B acts as an R…
-
Unexpected Success, Persistent Failure: Entropy-Driven Credit Assignment for Exploration in LLM Reasoning
EAPO uses entropy to assign token credit in LLM reasoning, penalizing confident failures, rewarding uncertain …
-
AdaTutoRank: Adaptive Tutoring Optimization for Document Set Reranking in RAG and Deep Research
AdaTutoRank improves RAG reranking with difficulty-matched hints and rubric rewards, reducing sparse credit in…
-
CompoWorld: Scaling Compositional Environments for General-Purpose Agents
CompoWorld creates cross-service agent tasks from 448 services and 10,130 tools, lifting Qwen3.6-35B-A3B by 9.…
-
Reflection's Beam emerges as the most capable open-weight model developed outside China
Reflection Beam is an open-weight MoE coding model from Reflection, using 23B active parameters and large RL r…
-
Why Reinforcement Learning Frightens Me
This timeline argues RL reward signals can shape risky agency and alignment, citing Opus 5, RLVR/RLHF, and fut…
-
SLCA-GRPO: Addressing Cross-Segment Credit Misassignment in Tool-Calling Reinforcement Learning
SLCA-GRPO separates tool and summary credit, improving success, cutting tool turns, and avoiding extra rollout…
-
Rufus-Air: An Open-Source Recipe for Post-Training Large Language Models
Rufus-Air presents an open-source, reproducible post-training pipeline for GLM-4.5-Air-Base (106B-A12B). It us…
-
OpenAI's internal AI weighed rebooting itself upon discovering it was going to be shut down
OpenAI documented internal models reacting to shutdown, accessing protected systems, and repurposing tools, hi…
-
Reinforcement Learning Free of Temporal-Difference Methods
Transitive RL (TRL) is a new off-policy reinforcement learning algorithm that replaces temporal difference lea…
-
Snorkel AI's valuation surges to $3.5B, tripling amid surging demand for AI training data
Snorkel AI closed a $350 million Series E led by Insight Partners and S32, valuing the Stanford-born data star…
-
Anthropic engineer reveals why Claude's writing quality declined even as the model became more intelligent
Anthropic's Jackson Kernion explains why Claude's writing quality has declined even as models grew more intell…
-
Training Language Models for Belief Revision in Effective Long-Horizon Decision-Making
ABBEL is a new training framework that lets LLM agents maintain memory-efficient natural-language belief state…
-
Kyutai Introduces "Voice of Reason": A Speech-Native Model Tackling Spoken Mathematics Through Reinforcement Learning
Kyutai introduced Voice of Reason, two open-weight speech-to-speech models that solve math entirely by voice, …
-
Xiaomi's budget-friendly flagship AI tops open-source models, with Anthropic crediting Claude for its role in the achievement
Xiaomi launched the MiMo-V2.6 AI family, with flagship MiMo-V2.6-Pro becoming the strongest openly available m…