The OISD framework improves language model reasoning by distilling on-policy predictive signals from the final output layer to intermediate representations, leading to substantial improvements on math…

The paper introduces Trajectory-aware OPD (TOPD), a method that uses near-future trajectory information to improve On-Policy Distillation by accurately identifying and guiding true reasoning divergenc…

The paper proposes Visual Gradient Steering (VGS), a method that decomposes the distillation loss into language and visual components and steers the optimization to prioritize visual grounding, signif…

The paper proposes Skill-Conditioned Gated Self-Distillation (SGSD), a novel framework that uses retrieved, potentially noisy skills to guide LLM reasoning, achieving state-of-the-art performance on m…

The paper introduces Lookahead Group Reward (&) to combat Supervision Fidelity Decay (SFD) in on-policy distillation, significantly improving student model performance on long reasoning tasks.

The paper identifies a linear predictive law linking the initial performance gap in on-policy self-distillation (OPSD) to the final performance improvement, allowing researchers to anticipate and tune…

The paper introduces Feedback Distillation, a novel training method that uses a language model's privileged feedback to provide token-level supervision, significantly improving complex reasoning tasks…

OmniOPD introduces a logit-free, chunk-level distillation framework that improves on standard On-Policy Distillation by using semantic similarity and peak-entropy scheduling, achieving state-of-the-ar…

The paper introduces Weak-Critic Strong Oversight, a method where a weak model guides a strong model's self-improvement by providing non-misleading revision directions, leading to scalable oversight.

The paper introduces OPD+, a corrected on-policy distillation framework that mathematically proves the bias of standard stop-gradient methods and improves the stability and performance of knowledge tr…

The paper introduces Contrastive Reflection (CORE), a novel non-parametric method that rapidly improves language model reasoning by distilling contrasts between successful and unsuccessful problem att…

The paper introduces Plan, a structured agentic behavior that decomposes multi-hop questions into ordered sub-questions before retrieval, and proposes a self-bootstrapping paradigm to train it without…

The paper proposes CAST, an answer-free self-distillation method that enhances Group Relative Policy Optimization (GRPO) for verifiable rewards, allowing token-level advantage signals even when all sa…

ThinkSwitch introduces a low-compute co-training procedure that distills the reasoning benefit of large language models into weights, significantly improving performance on specific reasoning tasks.

This paper introduces a 'Sleep' paradigm for machine learning models to continually learn and transfer knowledge.

The paper proposes Guided Denoiser Self-Distillation (GDSD), a novel method that bypasses the use of likelihood surrogates (like ELBO) in RL for diffusion language models, achieving state-of-the-art p…

The paper introduces the Triangulated Preference Shift score, an automated, curation-free metric to quantify systematic lexical biases introduced into Large Language Models during the preference-learn…

The paper introduces Temperature-Scaled On-Policy Self-Distillation (TS-OPSD), a novel method that internalizes temperature-based policy reheating into model parameters to combat entropy collapse in r…

The paper proposes using question-asking as an inference-time intervention to probe a language model's hidden state, finding that the self-diagnosis process provides a predictive signal for final corr…