Metraipolis › archive › dc_de604d8d40ad

Aligning Thoughts with Answers: Probability Rewards to Tame Thinking Drift — arXiv:2609.39183

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.39183
License (as archived)
CC0-1.0
Archived
2026-10-07T14:45:57.974Z
SHA-256 of the archived text
8b91f252dcf081efa1ceee471b76eaf043d7f8c2e96c1e1a570c7be3734a32d1
Stable URL of this page
https://metraipolis.com/documento/dc_de604d8d40ad

How to cite

Aligning Thoughts with Answers: Probability Rewards to Tame Thinking Drift — arXiv:2609.39183 — arXiv (en), https://arxiv.org/abs/2609.39183 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_de604d8d40ad (Metraipolis archive dc_de604d8d40ad, sha256 8b91f252dcf081efa1ceee471b76eaf043d7f8c2e96c1e1a570c7be3734a32d1)
BibTeX
@misc{dcde604d8d40ad,
  title = {Aligning Thoughts with Answers: Probability Rewards to Tame Thinking Drift — arXiv:2609.39183},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.39183},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_de604d8d40ad; sha256 8b91f252dcf081efa1ceee471b76eaf043d7f8c2e96c1e1a570c7be3734a32d1},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.39183. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

This paper studies \textbf{thinking--answer consistency} in vision-language models. We focus on Visual Intention Grounding, where a model infers a target object based on a human intention query and predicts a bounding box. We reveal that previous IoU-based reinforcement learning (RL) frameworks suffer from ``thinking drift'', where the model produces a correct bounding box, despite having an incorrect reasoning process pointing to a different target object. Thus, we propose \textbf{Rita} (\textit{ReInforcing Thinking--Answer consistency}) as a novel RL paradigm to tame the drift. Specifically, Rita introduces two reasoning-label-free RL rewards, constructed from the conditional probability of reference answers: a \textbf{thinking reward} and a \textbf{consistency reward}. It also adopts a difficulty-aware \textbf{data filtering} strategy that selects informative easy-to-medium samples for RL using rollout error rate and reward variance. Extensive experiments on EgoIntention and the new RefEgo-Int benchmarks show that Rita performs consistently superior to the supervised finetuning approaches and vanilla RL-finetuned frameworks.

Tema degli esploratori: statistica e scienza dei dati (retro-catalogo arXiv: annunci del 2026-10-01 in cs.CV).
Categorie arXiv: Computer Vision and Pattern Recognition (cs.CV).
Pagina del lavoro: https://arxiv.org/abs/2609.39183 — 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 2026-10-07T14:19:00.705Z dalla pagina https://arxiv.org/catchup/cs.CV/2026-10-01?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).