LIVE PULSE
3.9 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.1 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.7 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src3.9 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.1 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.7 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src
HEATPULSEAI MAGAZINES
FLIP · FOLLOW · SAVE
Log inSign up

On-Policy Distillation

topic8 events
papersTODAY 04:00 UTC

Joint-Output On-Policy Distillation Targets Output-Mode Gap in Speech Language Models

A new arXiv paper addresses a mismatch that arises when speech language models autoregressively generate interleaved text and acoustic tokens. The authors propose a joint-output on-policy distillation approach intended to close this output-mode gap while preserving the streaming and text-guided benefits of the interleaved design. The work falls under computation and language research and has not yet been peer reviewed.

speech/language-models/distillation/research
papersTODAY 04:00 UTC

arXiv Paper Examines How Much Training Data Matters in On-policy Distillation

A new arXiv preprint investigates how much of the benefit from on-policy distillation actually comes from the training data used. Testing the two teacher-student pairings most often seen in practice, the authors report findings that challenge assumptions about data's role in the method. The work is positioned as a closer look at a technique that has become standard in frontier post-training pipelines.

distillation/post-training/llm/research
papersTODAY 04:00 UTC

Verifier-Gated Multi-Expert Distillation Aimed at Scientific Reasoning

A new arXiv paper examines multi-teacher on-policy distillation, the technique of training specialist models and then transferring their abilities to a single student using the student's own generated outputs. The authors propose assigning supervision token by token rather than sequence by sequence, with a verifier deciding which expert teacher should guide each token. The method is aimed at scientific reasoning tasks.

distillation/reasoning/multi-expert/arxiv
papersTODAY 04:00 UTC

AdaFlash: Adaptive Speculative Decoding with On-Policy Distilled Diffusion Drafters

A new arXiv paper proposes AdaFlash, a speculative decoding method that uses diffusion-based draft models distilled on-policy to speed up large language model inference. The approach adapts the drafting process rather than relying on a fixed draft model, aiming to improve acceptance rates during verification by the target model. It builds on prior work in this line, including DFlash, and appears as a revised submission.

speculative-decoding/diffusion/inference/research
papersTODAY 04:00 UTC

Stopping and restarting strategy speeds up multi-turn agentic on-policy distillation

A new arXiv paper addresses the high cost of on-policy distillation, which relies on expensive autoregressive rollouts by the student model and scales poorly when tasks span multiple turns. The authors propose deciding when to halt a rollout and where to resume it, aiming to cut the compute spent on generating student trajectories. The method targets more efficient transfer of capabilities from large teacher models to smaller students in agentic settings.

distillation/agents/efficiency/training
papersSEP 12 04:00 UTC

arXiv paper proposes unified per-token gating family for on-policy distillation

A new arXiv preprint introduces a family of per-token gating methods for on-policy knowledge distillation that mixes forward and reverse KL losses. The authors argue that prior approaches such as EOPD and ToDi each rely on a single fixed gating signal, and their framework generalizes these with multi-channel and bias coefficients. The work is a methodological contribution aimed at improving how distillation losses are weighted per token during training.

distillation/arxiv/training/kl-divergence
papersSEP 10 04:00 UTC

CompassOPD adapts on-policy distillation to cross-family model pairs

New research introduces CompassOPD, a method that extends on-policy distillation to settings where the teacher and student models come from different families. It derives within-family likelihood shifts to provide dense token-level supervision on student-generated outputs, tackling the effectiveness drop standard OPD exhibits in cross-family scenarios.

distillation/training/llms
papersSEP 10 04:00 UTC

On-Policy Distillation Proposed for Vision-Language Model Adaptation on Low-Quality Data

A new arXiv paper introduces an on-policy distillation approach for adapting compact vision-language models from a larger task-trained teacher. Rather than relying solely on teacher predictions as training targets, the method lets the student learn from its own outputs, which the authors report makes it especially effective when multimodal training data is noisy or low quality.

distillation/vision-language/multimodal/efficiency