Theory says REINFORCE can collapse coverage in LM post-training
Resetting the KL reference in Approximate Policy Mirror Descent is shown to restore global convergence for stylized PPO and GRPO.
Stories per day · 30 days
8 RSS
Relatedreinforcement-learning15llm-agents4qwen3agentic-ai2agentic-rl2long-context2mllm2on-policy-distillation2qwen32rlvr2scienceworld2synthetic-data2video-captioning2active-exploration1active-perception1actor-critic1agent-steering1alfworld1
87 stories, newest first · one card per story
Stories tagged #grpo
Matched compute reverses dense-reward gains over GRPO
Across 1,668 runs, a Monte-Carlo process reward trailed GRPO by 14.3 points at equal tokens on Qwen2.5-1.5B.
Agentic-OPD recovers forgotten agent skills from discarded RL checkpoints
Authors report gains of up to 7.4 points across 24 benchmarks, with a 4B student beating GPT-5.6 on deep search.
Reframe aligns visual agents on tool observations during distillation
GD-GRPO and observation-gated multi-teacher distillation beat Mix-RL and vanilla MOPD on 14 benchmarks.
Creative Priming separates the creative process from the response
A GRPO-trained priming model writes prompt-specific primers so a response model yields novel, high-quality, diverse outputs.
LambdaMix learns steerable policies with jointly trained low-rank adapters
One training run mixes per-objective adapters so preferences can change at inference without more optimization.
TurnWise improves streaming video RL with turn-level credit
TurnWise assigns turn-level advantages using silence-completed prefix scoring to reduce redundancy in streaming video policy learning
Reviso prioritizes key visual evidence in video captioning
Reviso builds CFP observation plans and trains with GRPO to outperform baselines on four benchmarks and improve downstream video question answering.
Skill-Uni jointly trains LLM skill retrieval and use
Authors report average gains of 66.59% on Qwen3-4B and 34.82% on LLaMA3.1-8B across five benchmarks.
Unofficial Qwen3.8-Flash-Next E384 FP4 model trends on Hugging Face
The NAS build keeps 384 of 512 experts, cuts weights about 16%, and was merged after GRPO-style LoRA.
Fine-tuned small models predict x86-64 register inputs from outputs
GRPO fine-tuning lifts Nemotron-3-Nano knowable-register accuracy from 63.3% to 95.2%, within four points of frontier scores.
S3AC adds a stabilized actor-critic for LLM reasoning
The paper reports gains from 1.5B to 397B parameters and stable single-rollout training at 35B.
SATE-Bench tests MLLMs on active exploration of occluded scenes
The benchmark has 316 scenes and 2,525 multi-turn pairs; GRPO increases view-seeking but lowers accuracy.
GRPO treats ASR pseudo-labels as rewards rather than targets
In banking audio, GRPO cut WER 29.3% to 56.2% relative to SFT and copied 27% of teacher errors.
Paper finds reasoning models inherit 'wait' rather than learn it from RL
On bases that lack the word, RL still raises MATH-500 accuracy while 'wait' stays nearly absent.
LongZero improves LLM long-context reasoning without external data
The method self-writes multi-hop and retrieval tasks, trains with GRPO, and beat external-corpus RL on five benchmarks.
LongZero improves LLM long-context reasoning without external training data
Models write compact evidence, expand it into multi-document context, and train with GRPO on self-made tasks.
T-DIG lets users steer image diversity with natural-language prompts
It keeps the image generator frozen and adds user-directed diversity to open and closed models without weight access.
PAWM steers black-box LLM agents by anticipating pitfalls
Authors report Claude Sonnet 4.5 gains of 24.3% on ScienceWorld and 7.7% on WebArena-Lite.
KnowGPO: Knowledge-Grouped Policy Optimization for Search Agents
KnowGPO credits search-agent steps by knowledge gained, lifting an 8B model 3.8–4.2 points over process-RL baselines.
Learning to Steer LLM Agents and Multi-Agent Systems
Researchers train a reinforcement-learning steering agent that detects goal drift and corrects LLM agents during execution.
Lean4AgenticRL trains LAR-9B and LAR-27B with process rewards
Authors report LAR-27B at 77.3% on TerminalBench-2.1 and a 9.89% average gain over the base model.
EVA-Cap: Optimizing Audiovisual Video Captioning via Event-Centric Alignment
EVA-Cap improves audiovisual captions by aligning them to atomic events with SFT and GRPO.
AlphaBet uses reinforcement learning to train LLM event-forecasting agents
On GLM-4.5-Air, Brier score fell from 0.731 to 0.626 on Forecast-Dojo and from 0.612 to 0.528 on FutureX.