跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 2026-10-04AI 評分41

When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models

When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models

AI 導讀

MOTIVE 是一套視覺語言模型自我驗證框架,透過五種互補的驗證角度和學習所得的可靠性分數,判斷答案應予採納還是重新思考。五個基準測試的平均準確率中,Qwen2.5-VL-3B 由 55.11% 升至 63.13%,Qwen2.5-VL-7B 由 60.49% 升至 68.63%。MOTIVE 亦比傳統自我驗證減少平均答案生成輪數。

正文

🧠 MOTIVE: When to Rethink

Learning Multi-Perspective Self-Verification for Vision-Language Models

We introduce MOTIVE, a framework that helps vision-language models decide when to trust an answer and when to rethink, using complementary verification perspectives and a learned reliability score.

📄 Read the paper · 💻 GitHub


🔍 Method overview

MOTIVE follows an Answer → Verify → Accept or Rethink process:

  • Answer: Generate a candidate answer from the image and question.
  • Verify: Evaluate the answer using five complementary verification perspectives.
  • Decide: Accept reliable answers, or rethink using verification feedback and interaction history.

Answering and verification use the same backbone VLM, without an external judge during inference.

Overview of MOTIVE


📊 Main results

Evaluation covers five benchmarks and three VLM backbones. Average accuracy across the five benchmarks improves as follows:

Backbone Baseline MOTIVE Gain
Qwen2.5-VL-3B 55.11% 63.13% +8.02 pp
Qwen2.5-VL-7B 60.49% 68.63% +8.14 pp

pp = percentage points.

MOTIVE also reduces the average number of answer-generation turns compared with conventional self-verification.


💡 What motivates this approach?

Verifier capability matters. Stronger external judges improve final-answer accuracy in our preliminary analysis, motivating the development of stronger self-verification.

Effect of verification judge capability

Prompt choice matters. No single verification prompt consistently performs best across tasks. This motivates combining complementary verification perspectives.

Comparison of verification prompts

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