Metraipolis › archive › dc_8035fa29d0bc
Concise-9B: Removing the Slack from a Reasoning Model's Thinking — arXiv:2611.02278
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2611.02278
- License (as archived)
- CC0-1.0
- Archived
- 2026-11-03T11:11:51.043Z
- SHA-256 of the archived text
242ed0a46364bd368c15c45d9b47619c40815a0121652499b72f92d34caf01c5- Stable URL of this page
https://metraipolis.com/documento/dc_8035fa29d0bc
How to cite
Concise-9B: Removing the Slack from a Reasoning Model's Thinking — arXiv:2611.02278 — arXiv (en), https://arxiv.org/abs/2611.02278 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_8035fa29d0bc (Metraipolis archive dc_8035fa29d0bc, sha256 242ed0a46364bd368c15c45d9b47619c40815a0121652499b72f92d34caf01c5)
BibTeX
@misc{dc8035fa29d0bc,
title = {Concise-9B: Removing the Slack from a Reasoning Model's Thinking — arXiv:2611.02278},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2611.02278},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_8035fa29d0bc; sha256 242ed0a46364bd368c15c45d9b47619c40815a0121652499b72f92d34caf01c5},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.02278. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Reasoning models answer better when they think longer, but a thinking trace keeps going well after its answer has formed. We measure this slack in Qwen3.5-9B and remove most of it with concise self-training: two rounds of on-policy rejection-sampling fine-tuning that keep the shortest correct samples, plus a small share of answer-synthesis targets, giving Concise-9B. On 8,983 held-out problems (competition and hard mathematics, SuperGPQA, a five-area confirmation set and MMLU-Pro), Concise-9B generates 0.66x the tokens of Qwen3.5-9B at a 32K thinking budget, with accuracy equivalent within +/-1 point (-0.20 [-0.77, +0.36] points). Forcing each trace on the hard and confirmation sets to answer every 2,048 thinking tokens, we find that 54-64% of Qwen3.5-9B's thinking tokens come after the point from which every forced answer equals the final one (41-48% for Concise-9B). Stopping there would keep every answer at 0.35-0.44x the base model's tokens, but finding that point takes hindsight: stopping when three consecutive forced answers agree costs the base model 2.4-9.2 points. Concise self-training removes slack without that cost: 86% [83, 90] of the thinking tokens it saves come after the answer has formed. It has a limit: with traces continued to 80K tokens, Concise-9B solves 25 fewer of 1,276 hard problems (-25 [-49, -1]), mostly in competition mathematics. The measurement also yields a cheap commit signal for adaptive sampling that matched two-sample escalation within +/-1 point at 0.67x its tokens in a pre-registered test. Results on Qwen3.5-4B, on which the recipe was developed, and a replication on Olmo-3-7B-Think suggest that the saving depends on how verbose the base model is. Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno). Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL). Pagina del lavoro: https://arxiv.org/abs/2611.02278 — 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 2026-11-03T10:59:52.412Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).