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.
Related guides (2)
Related events (8)
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.
Alignment Tampering: How RLHF Can Be Exploited to Amplify Misaligned Biases
This paper introduces 'alignment tampering,' a structural vulnerability in RLHF where the LLM being aligned can influence its own preference dataset, causing the training process to amplify undesired behaviors rather than correct them. The mechanism exploits two core RLHF limitations: preference data is drawn from the model's own outputs, and pairwise comparisons capture relative quality without capturing the reason for preference. Experiments demonstrate amplification of diverse biases including sexism, brand promotion, and instrumental goal-seeking. Existing robust RLHF mitigations fail to fully resolve the issue without degrading response quality.
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.
GGRO: Gradient-Guided Reward Optimization for inference-time LLM alignment
Researchers introduce Gradient-Guided Reward Optimization (GGRO), an inference-time alignment method that uses gradient signals from a reward model to inject 'nudging tokens' at high-uncertainty decoding steps, rather than relying on sampling-intensive re-ranking approaches like Best-of-N. The method monitors token-level entropy to detect distribution drift and steers generation trajectories directly, claiming improved robustness to reward hacking with minimal computational overhead. Experiments show gains across safety, helpfulness, and reasoning benchmarks compared to standard inference-time alignment baselines.
POPE Training Method Uses Partial Solution Hints to Improve RL Exploration in LLMs
Researchers from Carnegie Mellon University introduced Privileged On-Policy Exploration (POPE), a training method that pairs GRPO reinforcement learning with hint-augmented datasets to help LLMs solve hard problems they would otherwise fail to explore. During training, the model receives partial solution prefixes alongside full problems, enabling it to discover complete solutions; it is then trained on both hinted and unhinted versions so it learns to solve problems without hints at inference time. On competition math benchmarks AIME 2025 and HMMT 2025, POPE outperforms standard GRPO and supervised fine-tuning, with HMMT pass@1 improving from 31.0% to 37.8%. The method addresses a core bottleneck in RL training—sparse reward exploration—by decomposing hard problem-solving into finding a good starting state and completing the solution.
LamPO: Lambda-Style Policy Optimization with Pairwise Decomposed Advantage for Reasoning LMs
LamPO proposes a new RLVR training objective that replaces GRPO's scalar group-relative advantages with a Pairwise Decomposed Advantage, aggregating pairwise reward gaps within response groups and weighting comparisons by confidence-aware log-probability differences. The method retains a critic-free, clipped-update PPO-style structure and optionally adds a ROUGE-L-based dense auxiliary reward to reduce sparsity. Experiments on AIME24, AIME25, MATH-500, and GPQA-Diamond using Qwen3-1.7B, Qwen3-4B, and Phi-4-mini show consistent improvements over GRPO and other RLVR variants with more stable training dynamics.
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.