TOPSD limits teacher supervision to verified prefixes in distillation
It beats conventional on-policy self-distillation on math reasoning in ten languages, especially lower-resource ones.
Stories per day · 30 days
5 RSS
Relatedreinforcement-learning8distillation5language-models4qwen34on-policy3reasoning3llm-training2olmo2ppo2qwen2reasoning-models2rlvr2vaspo2activation-steering1agentic-ai1agentic-training1agents1aime1
51 stories, newest first · one card per story
Stories tagged #self-distillation
On-Policy Distillation Lacks Generalization Beyond Supervised Data
MOPD does not fully transfer teachers' out-of-distribution capabilities from multiple teachers.
ERASE blocks multimodal prompt injections while preserving task utility
ERASE on-policy self-distillation reduces injection success to near zero while retaining accuracy close to undefended levels on 3B audio.
Muon-RT Reuses Teacher Trajectory for Momentum-Free Muon
Muon-RT eliminates momentum buffers by reusing the teacher-student displacement in self-distillation.
MTA-OPSD lets each token pick its own teacher in self-distillation
On Qwen3 math tasks, the method beats prior on-policy self-distillation, with gains of up to 5.9 over the base model.
Multi-teacher adaptive on-policy self-distillation
Token-level adaptive self-distillation improves math reasoning on Qwen3 models.
WISD self-distills reasoning models using a deliberately weaker copy
Pruning raises Qwen3.5-4B from 57.7% to 70.5% on HMMT 2025 and from 39.1% to 56.7% on LiveCodeBench v6.
Teacher-Evolution Self-Distillation improves on-policy model training
TESD evolves self-teachers with task rewards to provide reliable guidance and down-weights unreliable samples.
HiLo-RD improves low-budget masked diffusion language models
Self-distillation from large-window states raises GSM8K accuracy by 9.4 points and cuts mean readable length by about 25%.
Fin-SD tests how much hint signal self-distillation needs
On GPT-OSS-20B, a generic hint matched the full 3.03-point gain and used 9.5% fewer tokens than full-solution distillation.
Steered-teacher self-distillation cuts reference hallucinations in thinking models
Activation steering is applied only to the teacher; the student is not steered, and AIME and HMMT accuracy rises over vanilla OPSD.
AMBIT generates and edits first-order ambisonics with executable plans
A shared Transfusion model plans scenes, renders FOA audio, and self-distills on student prefixes.
On-policy self-training improves models without curated answers
From coarse topic descriptions alone, a model writes its own questions and reasoning traces and gains on language and visual tasks.
OP-Print targets membership inference in on-policy self-distillation
Gray-box and black-box signals use sparse correction imprints, while the authors say baseline attacks stay near chance.
AOPD weights teacher actions by student advantage in on-policy distillation
The estimator needs only student rollouts and terminal rewards, with no extra rollouts or reward evaluations.
FRED improves future video representations for world action models
On tasks shown only in human videos, FRED reaches 50.0% success versus 19.4% for the best rival.
Visual-Attribution Gating filters contaminated corrections in visual RL
On six visual reasoning benchmarks, VAG beats GRPO and on-policy self-distillation on average.
Bridged Self-Distillation unifies on-policy learning from any teacher
A self-teacher scores the student’s own rollouts after an external teacher supplies only a reasoning trace.
TreeOPSD adds turn-level credit for long-horizon agent RL
The critic-free method raised Qwen3-30B-A3B-Thinking-2507 accuracy by 4.97 points over GRPO on Deep Research.
Athanor: An Agent-Native System for Autonomous Post-Training Research
Athanor agents autonomously post-trained Qwen3-4B-Base to match official Qwen3-4B-Instruct on held-out tasks.
Privileged-information design shapes on-policy self-distillation gains
An error-step KL contrast predicted training gains better than a utility score that can leak information.
VASPO: Self-Distillation with Privileged Value Estimation
VASPO combines privileged self-distillation with PPO, gaining up to 2.59 points over PPO on math benchmarks.
Function-space policy optimization improves RLVR training for language models
Authors report about 4 percentage points higher average accuracy than the strongest baseline, with peak memory falling from about 70 GB to 56 GB on Olmo3-7B-Instruct.
VASPO: Self-Distillation with Privileged Value Estimation
VASPO blends a privileged self-teacher with PPO, improving math accuracy on Qwen3-4B and OLMo3-7B.