Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Li, Jiakun, He, Xingwei, Li, Kefan, Chai, Hongzheng, Yu, Hongyue, Yuan, Yuan
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2604.17304
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917420249120768
author Li, Jiakun
He, Xingwei
Li, Kefan
Chai, Hongzheng
Yu, Hongyue
Yuan, Yuan
author_facet Li, Jiakun
He, Xingwei
Li, Kefan
Chai, Hongzheng
Yu, Hongyue
Yuan, Yuan
contents Test-time scaling improves the reasoning performance of large language models but often results in token-inefficient overthinking, where models continue reasoning beyond what is necessary for a correct answer. Existing dynamic early-exit methods typically rely on single-step confidence signals, which are often unreliable for detecting reasoning convergence in multi-step settings. To mitigate this limitation, we propose TRACE, a training-free framework for efficient test-time scaling that determines when to terminate reasoning based on temporal aggregation of multi-step evidence rather than instantaneous signals. TRACE detects reasoning convergence over time by aggregating two complementary signals across recent reasoning steps: answer consistency, capturing the persistence of predicted answers, and confidence trajectory, modeling the temporal evolution of model confidence. Benefiting from these two factors, TRACE can accurately determine whether the reasoning process has converged, thereby promptly halting inference and effectively avoiding redundant reasoning steps. Extensive experiments on multiple challenging benchmarks show that TRACE reduces reasoning token usage by 25-30% on average while maintaining accuracy within 1-2% of full-length reasoning, consistently outperforming existing dynamic reasoning methods.
format Preprint
id arxiv_https___arxiv_org_abs_2604_17304
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Efficient Test-Time Scaling via Temporal Reasoning Aggregation
Li, Jiakun
He, Xingwei
Li, Kefan
Chai, Hongzheng
Yu, Hongyue
Yuan, Yuan
Artificial Intelligence
Test-time scaling improves the reasoning performance of large language models but often results in token-inefficient overthinking, where models continue reasoning beyond what is necessary for a correct answer. Existing dynamic early-exit methods typically rely on single-step confidence signals, which are often unreliable for detecting reasoning convergence in multi-step settings. To mitigate this limitation, we propose TRACE, a training-free framework for efficient test-time scaling that determines when to terminate reasoning based on temporal aggregation of multi-step evidence rather than instantaneous signals. TRACE detects reasoning convergence over time by aggregating two complementary signals across recent reasoning steps: answer consistency, capturing the persistence of predicted answers, and confidence trajectory, modeling the temporal evolution of model confidence. Benefiting from these two factors, TRACE can accurately determine whether the reasoning process has converged, thereby promptly halting inference and effectively avoiding redundant reasoning steps. Extensive experiments on multiple challenging benchmarks show that TRACE reduces reasoning token usage by 25-30% on average while maintaining accuracy within 1-2% of full-length reasoning, consistently outperforming existing dynamic reasoning methods.
title Efficient Test-Time Scaling via Temporal Reasoning Aggregation
topic Artificial Intelligence
url https://arxiv.org/abs/2604.17304