Metraipolis › archive › dc_786d3ddaf993

When2Think: Learning When and How Much to Reason — arXiv:2609.19671

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.19671
License (as archived)
CC0-1.0
Archived
2026-10-10T03:20:21.538Z
SHA-256 of the archived text
94bca04896652848c05fc6fc2926f04f8964c530b6b8ad2b62d3f06267cbfea2
Stable URL of this page
https://metraipolis.com/documento/dc_786d3ddaf993

How to cite

When2Think: Learning When and How Much to Reason — arXiv:2609.19671 — arXiv (en), https://arxiv.org/abs/2609.19671 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_786d3ddaf993 (Metraipolis archive dc_786d3ddaf993, sha256 94bca04896652848c05fc6fc2926f04f8964c530b6b8ad2b62d3f06267cbfea2)
BibTeX
@misc{dc786d3ddaf993,
  title = {When2Think: Learning When and How Much to Reason — arXiv:2609.19671},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.19671},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_786d3ddaf993; sha256 94bca04896652848c05fc6fc2926f04f8964c530b6b8ad2b62d3f06267cbfea2},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.19671. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Large Reasoning Models (LRMs) often overthink easy problems and underthink hard ones, leading to inefficient computation allocation. Existing methods regulate generated computation or select between direct answering and explicit reasoning, but do not jointly control whether}to reason and how much computation to allocate within reasoning. We call the resulting difficulty-dependent loss in accuracy under computation reduction the efficiency tax. We propose When2Think, an RLVR-based post-training framework for instance-adaptive computation allocation. Its core mechanism, Instance-level Difficulty-Aware Control (IDAC), uses cached reference statistics of success and token cost to modulate a correctness-gated efficiency bonus based on generated token count. Importance sampling supports exploration of Think and NoThink, while Batch-Wise Standardization constructs standardized advantages for critic-free optimization. The framework requires neither a learned reward model nor a learned critic, and offline reference caching avoids online reference-model queries during policy updates. On AIME24, When2Think improves Pass@3 by 10.0 percentage points while reducing token usage by 27.9% relative to the backbone.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-18 in cs.AI).
Categorie arXiv: Artificial Intelligence (cs.AI) ; Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2609.19671 — 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 2026-10-10T03:04:13.626Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-18?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).