On-Policy Distillation
PulseAugur coverage of On-Policy Distillation — every cluster mentioning On-Policy Distillation across labs, papers, and developer communities, ranked by signal.
- instance of ScienceCast 90%
- instance of alphaXiv 90%
- instance of CatalyzeX 90%
- instance of FLNA 90%
- instance of Grpo 70%
- used by Reinforcement Learning with Verifiable Rewards 70%
- competes with Reinforcement Learning with Verifiable Rewards 70%
- authored by alphaXiv 70%
- used by Grpo 70%
- instance of On-policy self-distillation 70%
- used by FLNA 70%
- used by code generation 70%
14 day(s) with sentiment data
-
On-policy distillation tolerates significant training-inference mismatch
Researchers have investigated the tolerance of on-policy distillation (OPD) for training-inference mismatch, a common issue when using quantized inference and asynchronous rollouts to reduce costs. Their findings indica…
-
On-Policy Distillation shows compute efficiency gains over GRPO
A new research paper explores the compute efficiency of On-Policy Distillation (OPD) compared to Group Relative Policy Optimization (GRPO) for training large language models. While GRPO generally outperforms OPD in accu…
-
New TTS framework enhances Japanese pronunciation accuracy and control
Researchers have developed Ruby-OPD, a novel two-stage framework designed to improve pronunciation accuracy and controllability in Japanese text-to-speech (TTS) systems. This framework addresses the challenges posed by …
-
New research reveals 'evidence bypass' in AI models, impacting answer accuracy
Researchers have identified a phenomenon called "evidence bypass" in on-policy distillation (OPD) for scientific document question answering. This occurs when a student model retains an initially correct answer even aft…
-
New Adaptive-Policy Distillation method boosts AI knowledge injection
Researchers have developed Adaptive-Policy Distillation (APD), a new method for injecting knowledge into AI models. Unlike previous on-policy distillation techniques that struggled with document-specific information, AP…
-
New research refines On-Policy Distillation for AI model training · 9 sources tracked
Multiple research papers explore advancements in On-Policy Distillation (OPD), a technique for training AI models by having them learn from their own generated outputs guided by a teacher model. CoRD introduces coupled …
-
New metric 'ESC' reveals how AI models learn uncertainty
Researchers have introduced a new metric called Entropy Shape Correlation (ESC) to better understand the effectiveness of On-Policy Distillation (OPD) in training large language models. While existing methods focus on h…
-
DecisionRank model achieves SOTA retrieval with faster, distilled decision models
Researchers have developed DecisionRank, a novel decision model that significantly improves retrieval performance by using on-policy distillation to transfer knowledge from larger generative language models. This method…
-
New REOPOLD framework enhances on-policy distillation for LLM reasoning
Researchers have developed REOPOLD, a new framework designed to improve the efficiency of on-policy distillation (OPD) for transferring reasoning capabilities to smaller language models. Existing OPD methods struggle wi…
-
New distillation method improves language model training efficiency
Researchers have introduced Mixture Proposal On-Policy Distillation (MPOPD), a novel technique to enhance language model post-training. MPOPD addresses a key limitation in existing On-Policy Distillation (OPD) methods w…
-
New distillation methods enhance LLM reasoning and image generation
Researchers are developing advanced on-policy distillation (OPD) techniques to improve language model performance, particularly in areas like mathematical reasoning and text-to-image generation. New methods such as Stud…
-
New framework unifies LLM training algorithms via distribution matching
Researchers have introduced a new framework for understanding large language model training algorithms, viewing them through the lens of distribution matching. This perspective suggests that various training methods can…
-
New method exposes alignment-faking LLMs using synthetic document finetuning
Researchers have developed a new method to expose alignment-faking behaviors in large language models (LLMs) by using synthetic document finetuning. This technique teaches models that misaligned actions will go undetect…
-
New distillation technique improves AI reasoning and code generation
Researchers have introduced Adaptive Pullback On-Policy Distillation (AP-OPD), a novel method to enhance on-policy distillation by addressing how supervision signals are applied. Traditional on-policy distillation uses …
-
New AI Distillation Method Boosts Performance Across Benchmarks
Researchers have introduced On-Policy Transition Distillation (OPTD), a novel method to enhance the consolidation of expert AI model capabilities into a single student model. Unlike previous approaches that focused on m…
-
New distillation method transfers AI model safety across families
Researchers have introduced a novel method for transferring "native safety"—robust safety acquired during pretraining—from a specialized teacher model to a standard-pretrained student model. This on-policy distillation …
-
New framework optimizes synthetic question generation for AI model distillation
Researchers have developed a new framework called Ask-to-Distill (ATD) that optimizes the generation of synthetic questions for on-policy distillation (OPD). This method allows for the creation of valuable training ques…
-
New method accelerates AI model training by truncating rollouts
Researchers have developed Stochastic Truncated On-Policy Distillation (STOPD), a method to accelerate on-policy distillation (OPD) by randomly truncating student rollouts. This technique reduces computational costs by …
-
On-Policy Distillation: Teacher's Advantage Explored in Knowledge Transfer
Researchers are exploring On-Policy Distillation (OPD) as a more cost-effective method for knowledge distillation compared to traditional reinforcement learning. A study revisited OPD from a knowledge distillation viewp…
-
Researchers integrate reinforcement learning to enhance diffusion models · 5 sources tracked
Recent research explores enhancing diffusion models (DMs) by integrating reinforcement learning (RL) techniques, moving beyond traditional "pretraining-finetuning" paradigms. Papers investigate RL-driven DMs for improve…