Salvato in:
Dettagli Bibliografici
Autori principali: Li, Gengyang, Cai, Wang, Gao, Yifeng, Wu, Yunfang
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:https://arxiv.org/abs/2601.03649
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866914238154407936
author Li, Gengyang
Cai, Wang
Gao, Yifeng
Wu, Yunfang
author_facet Li, Gengyang
Cai, Wang
Gao, Yifeng
Wu, Yunfang
contents Chain-of-Thought (CoT) prompting improves reasoning but often produces long and redundant traces that substantially increase inference cost. We present SyncThink, a training-free and plug-and-play decoding method that reduces CoT overhead without modifying model weights. We find that answer tokens attend weakly to early reasoning and instead focus on the special token "/think", indicating an information bottleneck. Building on this observation, SyncThink monitors the model's own reasoning-transition signal and terminates reasoning. Experiments on GSM8K, MMLU, GPQA, and BBH across three DeepSeek-R1 distilled models show that SyncThink achieves 62.00 percent average Top-1 accuracy using 656 generated tokens and 28.68 s latency, compared to 61.22 percent, 2141 tokens, and 92.01 s for full CoT decoding. On long-horizon tasks such as GPQA, SyncThink can further yield up to +8.1 absolute accuracy by preventing over-thinking.
format Preprint
id arxiv_https___arxiv_org_abs_2601_03649
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SyncThink: A Training-Free Strategy to Align Inference Termination with Reasoning Saturation
Li, Gengyang
Cai, Wang
Gao, Yifeng
Wu, Yunfang
Computation and Language
Chain-of-Thought (CoT) prompting improves reasoning but often produces long and redundant traces that substantially increase inference cost. We present SyncThink, a training-free and plug-and-play decoding method that reduces CoT overhead without modifying model weights. We find that answer tokens attend weakly to early reasoning and instead focus on the special token "/think", indicating an information bottleneck. Building on this observation, SyncThink monitors the model's own reasoning-transition signal and terminates reasoning. Experiments on GSM8K, MMLU, GPQA, and BBH across three DeepSeek-R1 distilled models show that SyncThink achieves 62.00 percent average Top-1 accuracy using 656 generated tokens and 28.68 s latency, compared to 61.22 percent, 2141 tokens, and 92.01 s for full CoT decoding. On long-horizon tasks such as GPQA, SyncThink can further yield up to +8.1 absolute accuracy by preventing over-thinking.
title SyncThink: A Training-Free Strategy to Align Inference Termination with Reasoning Saturation
topic Computation and Language
url https://arxiv.org/abs/2601.03649