Researchers probe on-policy distillation for multilingual LLMs
New papers test OPD variants for math, ASR and low-resource reasoning across languages.
Why it matters
The work points to OPD as a growing alternative or complement to reinforcement learning for post-training, but also shows that dense teacher signals can be misleading. For multilingual AI, preserving target-language reasoning remains a central challenge, not just improving benchmark scores.
The key points
- 1.OPD variants target multilingual math, ASR and low-resource reasoning.
- 2.English-only distillation can improve scores while pushing outputs toward English.
- 3.New methods filter weakly grounded teacher signals or specialize teachers by language.
Several new papers examine on-policy distillation as a post-training method for multilingual language models and reasoning systems. One study reports that On-Policy Delta Distillation improves over standard OPD on Qwen3 math reasoning in English, Korean and Japanese, while English-only OPD can boost Korean and Japanese performance but may shift responses toward English. Other work proposes filtering spurious token-level teacher signals, routing language-specialized ASR teachers into a multilingual student, and using cross-distillation to reduce English reversion in low-resource Southeast Asian chain-of-thought reasoning.
⚡ Try this today
Use multilingual training data and audit response language before applying OPD-style post-training to non-English reasoning tasks.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
- HF Daily PapersOn-Policy Delta Distillation for Multilingual Math ReasoningAug 6, 4:00 AM↗
- arXiv cs.AIWhen Teachers Mislead: Spurious-Signal-Aware On-Policy DistillationAug 5, 12:00 PM↗
- arXiv cs.CLLanguage-Specialized Multi-Teacher On-Policy Distillation for Multilingual LLM-Based ASRAug 5, 12:00 PM↗
- arXiv cs.AISOD: Step-wise On-policy Distillation for Small Language Model AgentsAug 4, 12:00 PM↗
- arXiv cs.AINative Multilingual Chain-of-Thought Reasoning in Low-Resource Southeast Asian LanguagesAug 4, 12:00 PM↗
- arXiv cs.LGNative Multilingual Chain-of-Thought Reasoning in Low-Resource Southeast Asian LanguagesAug 4, 12:00 PM↗
- arXiv cs.CLSOD: Step-wise On-policy Distillation for Small Language Model AgentsAug 4, 12:00 PM↗
- arXiv cs.CLNative Multilingual Chain-of-Thought Reasoning in Low-Resource Southeast Asian LanguagesAug 4, 12:00 PM↗
Enjoyed this brief? Get the next one in your inbox.
More in Research
Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency
arXiv:2608.
Improving the matrix multiplication exponent with modern optimization and AlphaEvolve
arXiv:2608.
[Paper] Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iter