Metraipolis › archive › dc_233ec22067d5

Look at Where You Think: Localization-Head Alignment for Fine-Grained Perception in Thinking VLMs — arXiv:2610.39649

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.39649
License (as archived)
CC0-1.0
Archived
2026-11-01T02:42:06.574Z
SHA-256 of the archived text
3f841019219d884b7da704b01f1a53b240640c7a78f0ebfc98494abef63555cb
Stable URL of this page
https://metraipolis.com/documento/dc_233ec22067d5

How to cite

Look at Where You Think: Localization-Head Alignment for Fine-Grained Perception in Thinking VLMs — arXiv:2610.39649 — arXiv (en), https://arxiv.org/abs/2610.39649 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_233ec22067d5 (Metraipolis archive dc_233ec22067d5, sha256 3f841019219d884b7da704b01f1a53b240640c7a78f0ebfc98494abef63555cb)
BibTeX
@misc{dc233ec22067d5,
  title = {Look at Where You Think: Localization-Head Alignment for Fine-Grained Perception in Thinking VLMs — arXiv:2610.39649},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.39649},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_233ec22067d5; sha256 3f841019219d884b7da704b01f1a53b240640c7a78f0ebfc98494abef63555cb},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.39649. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Thinking vision-language models reason for thousands of tokens, yet still miss small details in high-resolution images. Zooming methods pay for extra inference passes, and outcome rewards never tell the model where to look at. In Qwen3.5-4B, we find that the bottleneck is selection, not seeing: a few attention heads, which we call localization heads, pick the evidence out of thousands of visual tokens; ablating four of them costs 13.2 points more than ablating four random heads, and their focus fades as the reasoning grows. We therefore propose Localization-Head Alignment (LHA), which makes attention a training signal: whenever the reasoning states a visual fact, these heads learn to look at the focal region rather than the background, as an auxiliary loss in both SFT and RL training. On benchmarks of V*, HR-Bench 4K and TreeBench, LHA lifts Qwen3.5-4B from 74.1% to 76.6% accuracy, the best single-pass 4B result we evaluate, ahead of multi-pass zooming and released 7B models. Alignment alone adds 1.8 points over otherwise identical training while shortening reasoning by 20%, and the gains carry over to Qwen2.5-VL-3B.

Tema degli esploratori: statistica e scienza dei dati (elenco arXiv cs.CV, nuovi lavori del giorno).
Categorie arXiv: Computer Vision and Pattern Recognition (cs.CV).
Pagina del lavoro: https://arxiv.org/abs/2610.39649 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-11-01T02:16:43.945Z dalla pagina https://arxiv.org/list/cs.CV/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).