On-policy self-distillation reduces output diversity compared to RL, paper shows
A new arXiv paper analyzes on-policy self-distillation, where a single model serves as both teacher and student conditioned on correct demonstrations, finding it achieves strong pass@1 accuracy but at the cost of reduced rollout diversity and flattened pass@k curves. The authors trace this to compounding biases: teacher feedback is channeled through the model's own biases, amplifying probability mass on already-dominant modes rather than preserving diversity across equally correct solutions. Theoretical analysis shows the self-distillation policy tilts the base distribution by pointwise conditional mutual information, unlike ideal on-policy RL which preserves probability ratios among correct rollouts. Empirical results on graph path-finding and science QA benchmarks confirm self-distilled models match RL on average performance but fail on out-of-distribution settings requiring diverse strategies.
Related guides (2)
Related events (8)
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.
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.
Analysis of on-policy distillation reveals sparse, geometrically structured parameter updates
A new arXiv paper analyzes on-policy distillation (OPD) — a post-training method combining on-policy student trajectories with dense teacher supervision — across language and vision-language model pairs. The authors find that OPD updates are coordinate-sparse and distributed across layers (FFN-heavy), and that training only the discovered sparse subnetwork recovers near-full performance. Geometrically, updates are numerically full-rank but spectrally concentrated, falling disproportionately on near-zero weight coordinates, suggesting OPD retains distinct geometric signatures rather than behaving like ordinary dense parameter rewriting.
d-OPSD: First on-policy self-distillation framework tailored for diffusion LLMs
Researchers introduce d-OPSD, the first on-policy self-distillation (OPSD) framework designed specifically for diffusion large language models (dLLMs). The method addresses a fundamental mismatch between existing autoregressive OPSD approaches and dLLMs' arbitrary-order generation by using suffix conditioning on self-generated answers and step-level rather than token-level divergence supervision. Across four reasoning benchmarks, d-OPSD outperforms RLVR and SFT baselines while requiring only ~10% of the optimization steps of RLVR, suggesting strong sample efficiency gains for dLLM post-training.
Reward uncertainty as a principled mechanism for diverse RL behaviour
A new arXiv preprint proposes replacing the scalar reward in RL with a distribution over reward functions, applying a non-linear objective over sets of actions to induce calibrated behavioural diversity without sacrificing expected reward. The authors derive a principled gradient estimator in the contextual bandit setting and prove the formulation generalizes vanilla policy gradient and action-set approaches. The work is motivated by applications like language model fine-tuning where diversity is desirable but entropy regularization and diversity bonuses introduce fragile trade-offs. Empirical results support the framework as a theoretically grounded alternative to heuristic diversity methods.
Self-Policy Distillation via Capability-Selective Subspace Projection
This paper introduces Self-Policy Distillation (SPD), a self-distillation method for LLMs that requires no external signals such as correctness filters or reward models. SPD extracts a low-rank capability subspace from the model's own gradients on correctness-defining tokens, then projects KV activations into this subspace during self-generation to isolate task-relevant signal from stylistic noise. Experiments across code generation, math reasoning, and QA show up to 13% improvement over prior signal-free self-distillation methods and 15% better out-of-domain generalization.
Canonical-Context On-Policy Distillation (CCOPD) for Multi-Turn LLM Consistency
This paper identifies 'self-anchored drift' as a key failure mode in multi-turn LLMs: when information is revealed incrementally across turns, models produce unsupported assumptions that distort final answers, even when the total evidence is identical to a single-prompt setting. The authors propose Canonical-Context On-Policy Distillation (CCOPD), which trains a student model on incremental multi-turn conversations to match the output distribution of a frozen teacher conditioned on the full clean prompt. Trained only on math conversations, CCOPD achieves a 32% average relative improvement on multi-turn (RAW-SHARDED) tasks and generalizes zero-shot to five out-of-domain task families while preserving single-prompt performance.
ACPO: Adaptive Clip Policy Optimization improves RLVR training for LLM reasoning
A new arXiv preprint provides theoretical analysis of Reinforcement Learning from Verifiable Rewards (RLVR) updates, identifying off-policy degree and gradient expectation as key factors governing update dynamics. The authors show that differences in gradient steps per rollout substantially affect importance sampling ratio distributions and which tokens dominate updates. Based on this analysis, they propose Adaptive Clip Policy Optimization (ACPO), which adjusts clipping boundaries per token group by empirical variance of importance sampling ratios, outperforming DAPO and CISPO baselines on 3B and 7B models across math, tabular QA, and logic benchmarks.