self-distillation

Perplexity combines rejection sampling fine-tuning with hint-guided self-distillation to train its Computer model on real-world tool use, learning from successful sessions and user-corrected failures.

Model Grafting technique modifies Qwen3.5-4B into a causal encoder-decoder, creating variants with up to 3.7x speedup in prompt processing and minimal accuracy loss.

This paper introduces contrastive self-distillation to improve reasoning in AI models by separating correctness from behavioral shifts, demonstrating enhanced performance and stability.

HyperFlow introduces a new MiniMax-H3 variant with open-weight LoRA that enables 3× faster inference through data-free self-distillation while preserving video and stereo audio outputs.

HyperFlow is an open-source 8-step LoRA that uses data-free flow self-distillation to reduce video generation steps from 49 to 8 in MiniMax-H3, achieving significant speedup while maintaining quality.

The paper investigates the contribution of privileged information in on-policy self-distillation for language models, finding that reference-free distillation accounts for most improvements, with limited additional benefits from privileged references.

This paper introduces boundary-aware self-distillation to improve LLM safety refusal by reducing false refusals on benign prompts while maintaining genuine refusals through controlled data composition.

HypReflect is a framework for continual personalization in LLM assistants that infers explicit preference hypotheses from user signals and uses self-distillation to adapt, outperforming baselines and generalizing to unseen users and cross-domain settings.

TTPO introduces a label-free method for test-time training in mathematical reasoning that matches supervised performance by asymmetrically handling agreeing and disagreeing rollouts.

The paper investigates whether performance gains in on-policy self-distillation come from learning privileged reference information or recovering existing reasoning behavior, finding that the correct reference does not consistently benefit performance across various conditions.

DART-SD proposes a topology-aware retrieval and tuning framework for self-distillation of LLM-based tool-calling agents, improving policy diversity by correcting only critical topological breakpoints while preserving valid reasoning.

SkillForge is a self-distilling framework that enhances LLM-based software issue resolution by proactively synthesizing and resolving synthetic repository issues to acquire project-specific knowledge, improving performance over baselines.

This paper addresses rubric interference in LLM judges when evaluating multiple rubrics simultaneously and proposes Self-Anchored Rubric Alignment (SARA) to improve consistency through on-policy self-distillation.

SMOPD is a loss-only stabilization method for multi-turn on-policy self-distillation that uses selective token-entropy masking to improve accuracy in dirty-history settings, demonstrating improvements with Qwen3 models.

Introduces I-SDPO, an instance-level adaptive self-distillation policy optimization method that switches between GRPO and privileged self-distillation based on whether a rollout group contains any successful responses, improving reasoning performance on SciKnowEval.

Introduces SSPO, a step-level self-distilled policy optimization method for training deep search agents, which uses evidence anchors and advantage weights to improve credit assignment beyond sparse outcome rewards. SSPO outperforms GRPO on benchmarks like BrowseComp and GAIA with only ~5% overhead per step.

This paper introduces HPSE, a hybrid-policy self-editing method for composable unstructured knowledge editing in LLMs, enabling edited models to answer atomic questions and perform multi-hop reasoning about injected knowledge. The approach uses proactive self-distillation with hybrid rollouts and is validated across multiple LLM backbones and editors.

This paper introduces Latent On-Policy Self-Distillation (LOPD), a method that makes the teacher's privileged context learnable end-to-end from experience, providing dense token-level supervision to enhance agent performance and efficiency in agentic tool use and code generation.

This paper proposes an adaptive supervised anchoring framework for on-policy self-distillation, addressing the problem of rollout-conditioned signal degradation in language model training. The method separates rollout-conditioned distribution matching from canonical-context supervision, improving task acquisition while preserving general capabilities.

Tsinghua, Zhejiang University, and Meituan team propose AgentOPSD, a recursive self-distillation credit assignment method that converts sparse final rewards into step-wise credit signals, improving reinforcement learning performance for long-horizon agents. It significantly outperforms the GRPO baseline on tasks like ALFWorld.