5arXiv cs.CL (Computation and Language)·
ROPD: Routing-based On-Policy Distillation for template-robust LLM safety realignment
A new arXiv preprint introduces Routing-based On-Policy Distillation (ROPD), a safety realignment framework designed to counter malicious fine-tuning attacks that embed harmful behaviors into downstream model corpora. ROPD models the divergence between aligned and compromised output distributions rather than fitting specific prompt templates, addressing three failure modes of existing defenses: catastrophic forgetting, template-mismatch collapse, and re-jailbreaking via system prompt switches. Experiments across four baselines, three datasets, and three base models show ROPD substantially reduces template-mismatch degradation while preserving downstream task performance.
Related guides (2)
Related events (8)
DRPO: Smooth divergence regularization replaces hard masking in LLM RL training
A new arXiv preprint proposes Divergence Regularized Policy Optimization (DRPO), a method that replaces the hard trust-region mask used in DPPO with a smooth advantage-weighted quadratic regularizer on policy shift. The approach addresses a known weakness in PPO and GRPO where importance ratios poorly proxy distributional shift in long-tailed vocabularies, and in DPPO where gradient signals are discarded rather than corrected at trust-region boundaries. Experiments across model scales, architectures, and precision settings show improved stability and efficiency in LLM RL post-training.
MOPD: Multi-Teacher On-Policy Distillation for integrating multiple RL-trained capabilities in LLMs
Researchers propose Multi-teacher On-Policy Distillation (MOPD), a post-training paradigm that first trains domain-specialized RL teacher models, then distills them into a student model using on-policy rollouts to eliminate exposure bias. Evaluated on Qwen3-30B-A3B, MOPD outperforms Mix-RL, Cascade RL, Off-Policy Finetune, and Param-Merge baselines while preserving nearly all per-domain capability. The method has been deployed in production for MiMo-V2-Flash, an industrial-scale frontier model, validating its practical applicability. The approach also enables parallel, decoupled development of domain teachers, reducing cross-domain interference in multi-capability post-training.
Relay-OPD addresses prefix failure in on-policy distillation via teacher-student trajectory handoff
Researchers introduce Relay On-Policy Distillation (Relay-OPD), a training method that addresses 'prefix failure' in on-policy knowledge distillation, where student models compound early reasoning errors throughout a trajectory. The approach detects divergence points where teacher and student continuations asymmetrically diverge, then briefly hands generation to the teacher to produce a corrective 'relay leg' before the student resumes. Evaluated on eight mathematical reasoning benchmarks using Qwen3-4B-Instruct-2507 as teacher and Qwen3-0.6B/1.7B as students, Relay-OPD outperforms standard OPD by +5.73% and the strongest baseline FastOPD by +1.49% on average for the 1.7B model, while also reducing training trajectory length by over 50%.
DemoPSD: Disagreement-modulated policy self-distillation to fix privileged information leakage in LLM reasoning training
DemoPSD is a new training framework for LLMs that addresses two failure modes in on-policy self-distillation (OPSD): overfitting to in-domain patterns and privileged information leakage, where the student model learns answer-dependent shortcuts unavailable at test time. The method steers the student toward a reverse-KL barycenter target — a weighted geometric blend of teacher and student distributions — with token-level blending weights derived from the disagreement between the two distributions. Experiments on SciKnowEval across four scientific domains show DemoPSD outperforms GRPO and SDPO while maintaining higher training entropy and generalizing to out-of-distribution GPQA benchmarks.
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.
SafeSteer: Localized On-Policy Distillation for Efficient Safety Alignment
SafeSteer proposes a safety alignment method that targets only 'safety tokens' in the output distribution rather than applying global fine-tuning, arguing that safety features are inherently sparse. It constructs a safety teacher via activation steering, then restricts a reverse KL penalty to selected safety tokens during training. The approach achieves strong safety performance across seven benchmarks with minimal capability degradation, requiring only 100 harmful samples—less than 1% of data used by prior baselines.
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.
Direct On-Policy Distillation transfers RL policy shifts from weak to strong models
Researchers propose Direct-OPD (Direct On-Policy Distillation), a method for transferring the policy shift induced by reinforcement learning on a small model to a larger target model, bypassing the need to run expensive RL rollouts on the stronger model. The approach uses the log-ratio between a post-RL teacher and its pre-RL reference as a dense implicit reward signal applied to the student's own on-policy states. Empirically, Direct-OPD improves Qwen3-1.7B from 48.3% to 62.4% on AIME 2024 in 4 hours on 8 A100 GPUs, outperforming step-matched direct RL. The method addresses a key scalability bottleneck in post-training as frontier models grow larger.