跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 26 天前AI 評分45

以弱化促進強化:大型推理模型的弱化誘導式 On-Policy 自我蒸餾

Weaken to Strengthen: Weakening-Induced On-Policy Self-Distillation for Large Reasoning Models

AI 導讀

WISD 透過對比原模型與量化或剪枝後的較弱版本,為大型推理模型提供訓練訊號,無須外部教師或特權資料。量化和剪枝分別令多項基準的平均準確率提升 4.0 和 5.5 個百分點。剪枝令 Qwen3.5-4B 在 HMMT 2025 的準確率由 57.7% 升至 70.5%,量化則令 Qwen3.6-35B-A3B 在 LiveCodeBench v6 的準確率由 61.8% 升至 69.5%。

正文

Image 1: Hugging Face's logoHugging Face


Papers

arxiv:2610.38681

Copy markdown

Weaken to Strengthen: Weakening-Induced On-Policy Self-Distillation for Large Reasoning Models

Published on Oct 28

·Submitted byImage 2 Yecheng Wuon Oct 30

·Image 3: jet-aiJet-AI

- [x] Upvote -

Authors:

Yecheng Wu,

Song Han,

Han Cai

Abstract

WISD improves reasoning models by contrasting them with weaker quantized or pruned versions to generate training signals without privileged data.

Generated by thinkingmachines/Inkling-Small

Self-distillation offers a way to improve reasoning models when a suitable external teacher is unavailable. However, common approaches rely on privileged information to construct a stronger self-teacher, limiting their use when only unlabeled prompts are available. In this paper, we introduce Weakening-Induced On-Policy Self-Distillation (WISD), which instead constructs a weaker counterpart of the original model and uses the contrast between the two as a training signal without requiring privileged information. We explore two ways to construct the weaker model: low-bit quantization and magnitude pruning. We evaluate WISD across diverse architectures and scales on mathematical and coding reasoning tasks. Across these benchmarks, quantization and pruning improve average accuracy over the original models by 4.0 and 5.5 percentage points, respectively, achieving state-of-the-art performance even without privileged information. Notably, pruning raises Qwen3.5-4B's accuracy from 57.7% to 70.5% on HMMT 2025 and from 39.1% to 56.7% on LiveCodeBench v6. Quantization raises Qwen3.6-35B-A3B's accuracy from 68.5% to 74.6% and from 61.8% to 69.5% on the same benchmarks, respectively. These results establish WISD as a novel and effective self-distillation framework for large reasoning models.

View arXiv pageView PDFProject pageGitHubAdd to collection

Community

Image 4

gbcfchc

Paper submitter35 minutes ago

WISD improves reasoning models by distilling the contrast between the original model and its quantized or pruned copy, boosting reasoning performance without external teachers or privileged information.

Reply

Edit Preview

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

Comment ·Sign up or log in to comment

- [x] Upvote -

Get this paper in your agent:

hf papers read 2610.38681

Don't have the latest CLI?

curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2610.38681 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2610.38681 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2610.38681 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.

System theme

Company

TOSPrivacyAboutCareers

Website

ModelsDatasetsSpacesPricingDocs

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co