6arXiv cs.CL (Computation and Language)·
GRPO fails to improve small web agents when supervised baseline is near-ceiling, study finds
A controlled ablation study across 18 runs tests whether GRPO reinforcement learning adds capability to 4B–8B scale language and vision-language model web agents on top of a strong supervised baseline. The result is a credible null: GRPO does not improve success rates when the supervised model has largely mastered the task distribution, and moderate-to-high learning rates actively degrade text-track performance. The authors identify the mechanism — GRPO only helps when sampling headroom exists (sampled policy succeeds more than greedy), and failure modes dissociate into attention/MLP degradation versus full collapse regimes. Effective rank in late layers tracks capability at 4B but not 8B, flagging a scale-dependent coupling.
Related guides (3)
Related events (8)
GSPO: Group Sequence Policy Optimization for Scalable RL Training of Language Models
Qwen researchers introduce Group Sequence Policy Optimization (GSPO), a new RL algorithm designed to address severe training instability and model collapse observed in existing methods like GRPO during extended training runs. The core motivation is enabling stable RL scaling for language models to improve reasoning and problem-solving capabilities with increased compute. The paper targets a known bottleneck in post-training pipelines where instability prevents further performance gains.
GRPO outperforms SFT for ASR adaptation using only synthetic speech
A new arXiv preprint demonstrates that Group Relative Policy Optimization (GRPO) substantially outperforms supervised fine-tuning (SFT) when adapting LLM-based automatic speech recognition to regulated domains using only synthetic TTS data. GRPO alone reduces word error rate by 40% relative to SFT (36.71% → 22.09%), with an SFT+GRPO combination achieving 45% relative reduction. The authors attribute gains to behavioral changes — improved stopping calibration and better audio-text alignment — rather than representational shifts in early layers.
AdvGRPO: Stable co-training framework for adaptive red teaming of language models
Researchers introduce AdvGRPO, a co-training framework that makes GRPO viable for joint attacker-defender optimization in LLM red teaming, addressing previously reported instability. The method uses dense multi-channel rewards and decoupled advantage normalization, with a curriculum progressing from single-turn to multi-turn attacks before bootstrapping co-training. Co-trained defenders outperform baselines on safety benchmarks, and the attacks show transferability across models.
N-GRPO: Semantic Neighbor Mixing for Improved Policy Optimization in LLM Reasoning
A new arXiv preprint introduces N-GRPO, an exploration strategy for the GRPO reinforcement learning framework that improves solution diversity during rollout by mixing embeddings of anchor tokens with their nearest semantic neighbors rather than using token-level sampling or random noise. The method is evaluated on DeepSeek-R1-Distill-Qwen models of various sizes and shows consistent improvements on math reasoning benchmarks plus out-of-distribution generalization. The work targets a known limitation in RLHF-style training: redundant rollout trajectories that reduce effective learning signal.
One-shot GRPO training on a single biased example can break LLM alignment
A new arXiv paper demonstrates that a single biased training example using Group Relative Policy Optimization (GRPO) is sufficient to induce systematic bias in aligned LLMs, with stereotype-driven reasoning generalizing across attributes, categories, and benchmarks. The authors find that model susceptibility varies based on the initial likelihood of producing biased outputs. The result exposes a critical vulnerability in post-training alignment: a minimal fine-tuning intervention can override safety guardrails.
AdaPrefix-GRPO: Adaptive prefix control doubles GRPO accuracy on hard math reasoning
A new arXiv preprint introduces AdaPrefix-GRPO, a method that addresses GRPO's failure to learn from problems where no rollout succeeds by prepending correct solution prefixes and dynamically adjusting prefix length to maintain ~50% success rate per problem throughout training. The prefix assistance is gradually withdrawn so the final model solves problems unaided. On hard math benchmarks, the method more than doubles GRPO accuracy for a 0.6B model (2.1x) and achieves 1.7x improvement on AIME, while halving trace length, with larger gains on smaller models. The implementation requires only data preparation changes and a loss mask, leaving the trainer unchanged.
ZPPO: Teacher-in-prompt training method outperforms distillation and GRPO for small vision-language models
Researchers introduce Zone of Proximal Policy Optimization (ZPPO), a training method inspired by Vygotsky's zone of proximal development that embeds teacher guidance in prompts rather than policy gradients or logit imitation. On hard questions where student rollouts fail, ZPPO constructs Binary Candidate-included Questions (BCQ) and Negative Candidate-included Questions (NCQ) to help the student discriminate correct from incorrect responses, with a replay buffer that recirculates hard questions until mastered. Evaluated on the Qwen3 family (0.8B–9B) with a 27B teacher across a 31-benchmark suite covering VLM, LLM, and video tasks, ZPPO outperforms both distillation and GRPO baselines, with the largest gains at the smallest model scale. The method addresses a known failure mode of RL training where zero-reward rollouts produce no gradient signal.
STARE: Token-level advantage reweighting to prevent entropy collapse in GRPO-style RL training
Researchers introduce STARE, a method addressing policy entropy collapse in GRPO-style reinforcement learning from verifiable rewards (RLVR) for LLM post-training. Through first-order gradient analysis, they identify a token-level credit assignment mismatch and propose selectively reweighting advantages for entropy-critical tokens using batch-internal surprisal quantiles plus a closed-loop entropy gate. Evaluated across 1.5B–32B models on short/long chain-of-thought and multi-turn tool use tasks, STARE outperforms DAPO and other baselines by 4–8% on AIME24/25 while sustaining stable training over thousands of steps.