On-policy self-distillation
PulseAugur coverage of On-policy self-distillation — every cluster mentioning On-policy self-distillation across labs, papers, and developer communities, ranked by signal.
9 day(s) with sentiment data
-
New research explores efficient test-time reasoning for LLMs · 10 sources tracked
Multiple research papers explore novel methods for enhancing the reasoning capabilities of large language models (LLMs) at test time. Techniques include learning when to apply teacher supervision to prevent "Teacher Dri…
-
New distillation methods enhance LLM reasoning by adapting teachers and tokens
Researchers are developing advanced on-policy distillation techniques to improve the performance of large language models. One approach, Student-Prefix Teacher Adaptation (SPTA), modifies the teacher model to better han…
-
New EOPSA method boosts LLM safety alignment efficiency
Researchers have introduced EOPSA, a novel method for improving the efficiency and effectiveness of safety alignment in large language models. The proposed technique addresses two key inefficiencies in existing On-Polic…
-
New AI self-distillation method improves reasoning and reduces hallucinations
Researchers have developed a new method called Steered-Teacher On-Policy Self-Distillation (ST-OPSD) to improve the reasoning capabilities of AI models. This technique addresses issues like hallucinating references and …
-
New framework improves evaluation of LLM Chain-of-Thought monitorability
Researchers have developed a new framework to more accurately evaluate the monitorability of Chain-of-Thought (CoT) in large language models (LLMs). The existing metric, minimal-criterion sensitivity (TPR_mc), can overe…
-
New OP-Print method detects privacy risks in AI self-distillation
Researchers have developed a new method called OP-Print to detect privacy risks in on-policy self-distillation, a technique used in AI model training. This method specifically targets the potential for models to memoriz…
-
New Monte Carlo method enhances AI reasoning by reducing reference sensitivity
Researchers have developed Monte Carlo Privileged Distillation (MCPD), a new method to improve on-policy self-distillation (OPSD) for AI reasoning. Traditional OPSD methods can be overly sensitive to the specific refere…
-
Perplexity and NVIDIA unveil JOLT for efficient LLM post-training
Researchers from Perplexity and NVIDIA have developed JOLT, a novel post-training method for language models that utilizes a single model acting as both teacher and student. This approach trains the teacher with privile…
-
TeachBack framework improves AI model training via bidirectional preference optimization
Researchers have introduced TeachBack, a novel framework for on-policy self-distillation (OPSD) that enhances language model training. TeachBack addresses limitations in existing OPSD methods by not only learning from a…
-
New metrics predict effectiveness of privileged information in LLM self-distillation
Researchers have developed new methods to predict the effectiveness of privileged information (PI) in on-policy self-distillation (OPSD) for large language models (LLMs). The study found that the content and source of P…
-
AI reasoning research questions self-distillation effectiveness
Two research papers published on arXiv investigate the effectiveness of self-distillation techniques for improving AI reasoning capabilities. The first paper questions whether self-distillation methods like SDPO and OPS…
-
New RGR technique mitigates LLM capability degradation during self-improvement
A new research paper introduces Representation Geometry Regularization (RGR) to address capability degradation in large language models (LLMs) during self-improvement processes like on-policy self-distillation. The stud…
-
New self-distillation methods enhance LLM training and continual learning
Researchers have developed new methods for improving the training of language models, particularly focusing on self-distillation techniques. One approach, VASPO, enhances policy optimization by integrating privileged in…
-
New self-distillation methods improve LLM performance without external teachers · 4 sources tracked
Researchers have introduced several novel self-distillation techniques for language models, aiming to improve performance without requiring external teachers or ground-truth labels. Activation-Conditioned Self-Distillat…
-
New EOPSA method enhances AI safety alignment efficiency
Researchers have developed Efficient On-Policy Self-Distilled Safety Alignment (EOPSA), a new method to improve the efficiency and effectiveness of AI safety alignment. EOPSA addresses two key issues in existing On-Poli…
-
New research refines on-policy distillation for LLMs; OpenAI counters model extraction
Multiple research papers explore advanced techniques for On-Policy Distillation (OPD), a method to improve language model performance by using a teacher model to guide a student model during training. These papers intro…
-
New dataset AMPLE-Math probes value of privileged info in LLM self-distillation
Researchers have developed AMPLE-Math, a new dataset comprising over 5,000 mathematical problems, to investigate the impact of privileged information in on-policy self-distillation (OPSD) for language models. Their find…
-
New research advances on-policy distillation for LLM training · 6 sources tracked
Researchers are developing advanced techniques for on-policy distillation (OPD), a method used to improve large language models. New approaches like $\gamma$OPD and STRIDE aim to enhance optimization stability and effic…
-
VISTA method enhances AI reasoning via teacher-student adaptation · 2 sources tracked
Researchers have developed VISTA, a novel method for on-policy self-distillation (OPSD) that enhances AI model reasoning. Unlike standard OPSD, VISTA adapts the teacher model based on verified student rollouts, particul…
-
New TTPO method enhances LLM math reasoning without labels
Researchers have developed Test-Time Policy Optimization (TTPO), a novel method for improving large language models' mathematical reasoning capabilities without relying on ground-truth labels. TTPO addresses the fragili…