# Multi-Teacher On-Policy Distillation

Tencent Hunyuan released Hy ASR 3.0, a large-language-model-based speech recognition system covering Mandarin Chinese, English, Cantonese and 12 other Chinese dialect groups, with external context supplied as hotword lists, question-answer dialogues or summaries. On public Mandarin, English and Cantonese benchmarks it reaches aggregate error rates of 4.64%, 5.22% and 5.06%, each below the evaluated non-Hy systems; the Cantonese-only aggregate is about 11.0% lower than Qwen3 ASR Flash. The system also uses information inside the same utterance to disambiguate homophonic or phonetically similar keywords without a separate context input.

Tencent Hunyuan released Hy ASR 3.0, a large-language-model-based speech recognition system covering Mandarin Chinese, English, Cantonese and 12 other Chinese dialect groups, with external context supplied as hotword lists, question-answer dialogues or summaries. On public Mandarin, English and Cantonese benchmarks it reaches aggregate error rates of 4.64%, 5.22% and 5.06%, each below the evaluated non-Hy systems; the Cantonese-only aggregate is about 11.0% lower than Qwen3 ASR Flash. The system also uses information inside the same utterance to disambiguate homophonic or phonetically similar keywords without a separate context input.

After the MiMo-V2.6 release, repeated identical tool calls in agentic settings were traced to a reward blind spot in reinforcement learning: correctness was rewarded, but excessive calling below the penalty threshold was not. The team first trained a single-turn RL teacher that learned when to stop, then merged that behavior into the main model through multi-teacher on-policy distillation. Repetition rates for MiMo-V2.6-Pro and MiMo-V2.6-Flash fell substantially across context lengths and agent harnesses, and the full training run cost about $90,000, 4% of the estimated $2.31 million for restarting the MixRL alternative.

After the MiMo-V2.6 release, repeated identical tool calls in agentic settings were traced to a reward blind spot in reinforcement learning: correctness was rewarded, but excessive calling below the penalty threshold was not. The team first trained a single-turn RL teacher that learned when to stop, then merged that behavior into the main model through multi-teacher on-policy distillation. Repetition rates for MiMo-V2.6-Pro and MiMo-V2.6-Flash fell substantially across context lengths and agent harnesses, and the full training run cost about $90,000, 4% of the estimated $2.31 million for restarting the MixRL alternative.

Reflection released Beam, its first open-weight model, together with the safety training methods and evaluation tools behind it. Public safety evaluations had not covered enough ground to test whether safety boundaries hold in adversarial and tool-use settings. Beam combines pretraining data filtering with dedicated safety alignment training, merging a capability teacher and a safety teacher through multi-teacher on-policy distillation. On 1,500 human-written conversations in the Iris safety evaluations, Beam declined 89% of harmful requests and complied with 82% of safety-related harmless requests.

Reflection released Beam, its first open-weight model, together with the safety training methods and evaluation tools behind it. Public safety evaluations had not covered enough ground to test whether safety boundaries hold in adversarial and tool-use settings. Beam combines pretraining data filtering with dedicated safety alignment training, merging a capability teacher and a safety teacher through multi-teacher on-policy distillation. On 1,500 human-written conversations in the Iris safety evaluations, Beam declined 89% of harmful requests and complied with 82% of safety-related harmless requests.

NVIDIA released Nemotron 3.5 Super VL, an open, customizable multimodal model for long enterprise agent workflows that need reasoning, tool use, and text, image, document, and video understanding. It uses a 120B-parameter mixture-of-experts architecture that activates 12B parameters per token, and NVIDIA reports up to 5x higher throughput on agentic tasks than other open models in its size class on a single GPU. Teams can post-train it with NeMo on their own data, tools, and workflows, and NVIDIA also released agentic tool-use, finance reasoning, and 40 vision-language datasets.

NVIDIA released Nemotron 3.5 Super VL, an open, customizable multimodal model for long enterprise agent workflows that need reasoning, tool use, and text, image, document, and video understanding. It uses a 120B-parameter mixture-of-experts architecture that activates 12B parameters per token, and NVIDIA reports up to 5x higher throughput on agentic tasks than other open models in its size class on a single GPU. Teams can post-train it with NeMo on their own data, tools, and workflows, and NVIDIA also released agentic tool-use, finance reasoning, and 40 vision-language datasets.

The SAIL team introduces SAIL, an open scientific agent model with 35B total and 3B active parameters, developed through a science-aware improvement loop in which frontier-model agents diagnose its task failures and build training tasks from papers and code repositories, then trained via supervised fine-tuning, specialist training, multi-teacher on-policy distillation, and agentic reinforcement learning; it improves over its base model on all 13 evaluations and achieves the highest SciCode and ArxivDIGESTables scores in the comparison.

The SAIL team introduces SAIL, an open scientific agent model with 35B total and 3B active parameters, developed through a science-aware improvement loop in which frontier-model agents diagnose its task failures and build training tasks from papers and code repositories, then trained via supervised fine-tuning, specialist training, multi-teacher on-policy distillation, and agentic reinforcement learning; it improves over its base model on all 13 evaluations and achieves the highest SciCode and ArxivDIGESTables scores in the comparison.

The work introduces the FONTS taxonomy of five individual-simulation capabilities, curates a roughly 10M-instance corpus from 14 datasets, and builds Socio-Foundation through a three-stage post-training pipeline: DAPO-trained task experts, capability experts consolidated by off-policy distillation, and a unified student via multi-teacher on-policy distillation; it scores 61.7 on the authors' IndiEval, 11.0 points above the Qwen3-8B base, and beats that base on three out-of-distribution benchmarks.

The work introduces the FONTS taxonomy of five individual-simulation capabilities, curates a roughly 10M-instance corpus from 14 datasets, and builds Socio-Foundation through a three-stage post-training pipeline: DAPO-trained task experts, capability experts consolidated by off-policy distillation, and a unified student via multi-teacher on-policy distillation; it scores 61.7 on the authors' IndiEval, 11.0 points above the Qwen3-8B base, and beats that base on three out-of-distribution benchmarks.

The work proposes Slow-Fast Multi-Teacher On-Policy Distillation (SF-MOPD), which couples a fast model updated directly by each teacher with a slow model that is an exponential moving average of the student, computes each teacher-induced update in log-probability space, and removes only the component pushing the fast model away from the slow model while retaining aligned and orthogonal components; experiments across multiple model scales show it mitigates capability interference, enhances specialized multimodal capabilities, and reduces average degradation on general-capability benchmarks, consistently outperforming vanilla MOPD.

The work proposes Slow-Fast Multi-Teacher On-Policy Distillation (SF-MOPD), which couples a fast model updated directly by each teacher with a slow model that is an exponential moving average of the student, computes each teacher-induced update in log-probability space, and removes only the component pushing the fast model away from the slow model while retaining aligned and orthogonal components; experiments across multiple model scales show it mitigates capability interference, enhances specialized multimodal capabilities, and reduces average degradation on general-capability benchmarks, consistently outperforming vanilla MOPD.