Step-aligned critique outperforms GRPO and reference-solution conditioning in self-distillation
A new arXiv paper investigates context design for self-distillation of language models, comparing binary reward (GRPO), reference solutions, and step-by-step critiques aligned to the solver's reasoning trace. Step-aligned critique yields the largest gains, outperforming GRPO by 16.11 points and reference-solution conditioning by 5.27 points on Avg@12. Per-token advantage analysis shows that step-aligned feedback targets only failing tokens, avoiding unnecessary pressure on already-correct reasoning steps. The findings suggest structural alignment between feedback and the model's reasoning trace is a key driver of self-distillation effectiveness.
Related guides (3)
Related events (8)
Rubric-Conditioned Self-Distillation: structured feedback for reasoning model post-training
A new arXiv preprint proposes Rubric-Conditioned Self-Distillation (RCSD), a post-training framework that replaces scalar reward signals and noisy chain-of-thought annotations with structured rubrics for fine-grained credit assignment. The method conditions a teacher model on criterion-level rubrics to provide token-level guidance on the student's own sampled trajectories, avoiding reliance on a single reference rationale. Evaluated on science reasoning benchmarks, RCSD outperforms GRPO by 1.0 points and OPSD by 0.9 points on average.
Skill-Conditioned Gated Self-Distillation (SGSD) for LLM Reasoning
SGSD is a new on-policy self-distillation method for LLM reasoning that replaces trusted privileged information (e.g., reference answers) with an experience-derived skill bank of skill-mistake pairs. It constructs a multi-teacher pool, validates each teacher's contribution via a verifier, and applies a gated objective to distill informative disagreements while suppressing noisy signals. On Qwen3-1.7B, SGSD outperforms GRPO by 6.2% and answer-conditioned OPSD by 1.7% on average across AIME24, AIME25, and HMMT25. The method relaxes the assumption of trusted privileged information, making self-distillation more practical under weaker supervision.
DistIL: Distributional DAgger for RL from Rich Feedback beyond single-bit rewards
A new arXiv preprint introduces DistIL, a distributional variant of the DAgger imitation learning algorithm designed to exploit rich feedback signals (execution traces, tool outputs, expert corrections) rather than the single-bit correctness reward used in standard RLVR. The method uses a forward cross-entropy objective that provides monotonic policy improvement guarantees, unlike reverse KL or Jensen-Shannon divergence objectives used in prior self-distillation approaches. Empirically, DistIL outperforms RLVR and self-distillation baselines on scientific reasoning, coding, and hard math benchmarks.
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.
CORE: Contrastive Reflection for Sample-Efficient Reasoning Improvement
CORE (Contrastive Reflection) is a non-parametric learning algorithm that improves LLM reasoning by comparing successful and unsuccessful reasoning traces to generate compact natural-language 'insights' about reasoning strategies. Across four reasoning tasks, CORE outperforms both parametric baselines (GRPO/RLVR) and non-parametric baselines (GEPA, episodic RAG, MemRL) under fixed rollout budgets, achieving comparable or better gains with as few as five training samples. The method is also more context-efficient than prompt-optimization approaches, storing learned knowledge as interpretable natural-language descriptions rather than raw traces or weight updates. The results suggest contrastive distillation of reasoning traces may be a more efficient route to self-improvement than traditional fine-tuning.
ExpRL: RL-based mid-training using human QA data as reward scaffolds for LLM reasoning
ExpRL proposes an automated approach to LLM mid-training that replaces manually curated reasoning traces with large corpora of human-written QA data used as reward scaffolds rather than imitation targets. Reference solutions are hidden from the policy and used only to construct problem-specific grading rubrics, enabling dense process-level rewards that reinforce partial progress and intermediate reasoning steps. On challenging math reasoning benchmarks, ExpRL outperforms SFT, sparse-reward GRPO, and self-distillation as an RL initialization strategy, with additional mixed-domain experiments suggesting broader applicability.
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.