AI
Program-Verified Self-Evolution for Vision-Language Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 26/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 03:20:58 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 03:18:25 UTC.
Briefings
This study introduces Verifiable QA Generation for Self-Evolving Models (VQS), a method that replaces majority voting with structured claim-level verification in self-evolving vision-language models. Human evaluation shows VQS achieves 94% correctness compared to 76% for prior voting methods, while benchmarks indicate VQS improves Qwen3-VL by up to 3.18 points across ten benchmarks at 2B, 4B, and 8B parameter scales. The approach allows fixed programs to generate questions from parsed image records, enabling the model to verify individual facts rather than entire answers. Code is available at https://github.com/ahmedheakl/VQS.
Key points
- VQS achieves 94% human correctness on self-evolved questions, significantly outperforming the 76% accuracy of prior majority-vote methods.
- VQS improves Qwen3-VL by up to 3.18 points across ten benchmarks at 2B, 4B, and 8B parameter scales, with gains increasing over three training rounds.
Context and limitations
Human evaluation was conducted on a subset of self-evolved questions; full-scale validation across all benchmarks remains pending. The study relies on a single author's code release without peer-reviewed validation or independent replication.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 03:20:58 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeidy5yorpveqdcrha444tn6vihbwa6bjuqbdiffmf7maoiy4xfefli