Self-Verified Distillation: Your Language Model Is Secretly Its Own Synthetic Data Pipeline

arXiv cs.CL Papers
self-verification distillation synthetic-data reasoning post-training llm qwen

Summary

Proposes Self-Verified Distillation, a method where LLMs generate and self-verify candidate solutions from unlabeled seed questions using prompt-based verification, then train on the filtered dataset, achieving significant gains on math, science, and coding benchmarks across Qwen3 models.

arXiv:2605.26132v1 Announce Type: new Abstract: Can post-trained large language models (LLMs) further improve themselves using only unlabeled prompts, without external teachers or feedback from tools? We study this setting starting only from unlabeled seed questions with no ground-truth solutions, across three reasoning domains: math, science, and coding. We propose Self-Verified Distillation, a simple post-training refinement algorithm in which the model generates candidate solutions to these seed questions, filters them using prompt-based self-verification, and trains on the resulting self-curated dataset. Inspired by the UQ benchmark's use of multiple validators to screen candidate answers to hard unsolved questions, we adapt this validation-based filtering idea to self-training: the model filters its own generated solutions through a three-stage cascade of cycle-consistency, factuality, and correctness checks, accepting a solution only if it passes all stages with unanimous judge votes. We find that sampling more candidate generations and using a larger verification budget during training data construction produces higher-quality self-curated data and, in turn, better reasoning models. We then train Qwen3 models at multiple scales with Self-Verified Distillation and obtain gains across all three domains. For Qwen3-4B, our method improves aggregate held-out pass@1 by +16.7 points in math (AIME26 and HMMT), +11.1 points in science (GPQA Diamond and HLE), and +8.3 points in coding (LCBv5 and LCBv6), with gains also extending to 0.6B and 8B models. Compared to our test-time-only baseline (UQ-TTC), which improves performance by spending extra compute at inference time, Self-Verified Distillation achieves better performance in most settings while requiring only a single inference call at test time.
Original Article
View Cached Full Text

Cached at: 05/27/26, 09:02 AM 

# Self-Verified Distillation: Your Language Model Is Secretly Its Own Synthetic Data Pipeline
Source: [https://arxiv.org/abs/2605.26132](https://arxiv.org/abs/2605.26132)
[View PDF](https://arxiv.org/pdf/2605.26132)

> Abstract:Can post\-trained large language models \(LLMs\) further improve themselves using only unlabeled prompts, without external teachers or feedback from tools? We study this setting starting only from unlabeled seed questions with no ground\-truth solutions, across three reasoning domains: math, science, and coding\. We propose Self\-Verified Distillation, a simple post\-training refinement algorithm in which the model generates candidate solutions to these seed questions, filters them using prompt\-based self\-verification, and trains on the resulting self\-curated dataset\. Inspired by the UQ benchmark's use of multiple validators to screen candidate answers to hard unsolved questions, we adapt this validation\-based filtering idea to self\-training: the model filters its own generated solutions through a three\-stage cascade of cycle\-consistency, factuality, and correctness checks, accepting a solution only if it passes all stages with unanimous judge votes\. We find that sampling more candidate generations and using a larger verification budget during training data construction produces higher\-quality self\-curated data and, in turn, better reasoning models\. We then train Qwen3 models at multiple scales with Self\-Verified Distillation and obtain gains across all three domains\. For Qwen3\-4B, our method improves aggregate held\-out pass@1 by \+16\.7 points in math \(AIME26 and HMMT\), \+11\.1 points in science \(GPQA Diamond and HLE\), and \+8\.3 points in coding \(LCBv5 and LCBv6\), with gains also extending to 0\.6B and 8B models\. Compared to our test\-time\-only baseline \(UQ\-TTC\), which improves performance by spending extra compute at inference time, Self\-Verified Distillation achieves better performance in most settings while requiring only a single inference call at test time\.

## Submission history

From: Tony Lee \[[view email](https://arxiv.org/show-email/564e9e73/2605.26132)\] **\[v1\]**Wed, 20 May 2026 17:26:10 UTC \(2,264 KB\)

Similar Articles

Reference-Based Distillation Detection in LLMs

arXiv cs.LG

This paper introduces a reference-based method to detect whether an LLM was distilled from a specific teacher model, using membership inference. The approach achieves near-perfect accuracy in controlled settings and provides new evidence about potential distillation relationships involving QwQ, DeepSeek-R1, and GPT-OSS.

Consensus as Privileged Context for Label-Free Self-Distillation

arXiv cs.LG

A research paper introducing Canon, a label-free self-distillation method that uses consensus among sampled solutions to provide dense token-level supervision for training large language models on reasoning tasks, improving pass@1 by up to 12 points and outperforming label-free reinforcement learning at a fraction of the compute.

Future Confidence Distillation in Large Language Models

arXiv cs.CL

This paper investigates how confidence-related information evolves during LLM answer generation and introduces future confidence distillation, which trains predictors on pre-solution hidden representations using post-solution correctness probes to achieve reliable and sample-efficient confidence estimation.

Masked Distillation: Internalizing the Chain-of-Thought in Language Models

arXiv cs.AI

Masked distillation is a knowledge-distillation framework that trains a student LLM to predict only solution tokens while a reasoning teacher provides feedback, aiming to internalize chain-of-thought computation into model parameters. The method shows task-dependent success, working on GSM8K but requiring small scaffolds for harder tasks like Countdown.