aiqbench analysis
This paper studies how reflection markers like "wait" and "but" shape the reasoning traces of large reasoning models, and finds that treating them as one category hides distinct effects on accuracy and length. Its training-free controller, PathCal, adjusts logit biases at inference time without extra forward passes, and the reported gains are mostly on math benchmarks, with some length reduction in coding and multilingual math. Readers should weigh the local, directional evidence and the scope of the benchmarks before assuming the approach generalizes to other models or tasks.
PathCal: State-Aware Reflection-Marker Calibration for Efficient Reasoning
arXiv:2605.23074v2 Announce Type: replace Abstract: Large Reasoning Language Models (LRMs) tackle complex tasks through test-time scaling, generating long-form Chain-of-Thought (CoT) trajectories during inference. These trajectories contain reflection markers such as wait, but, and alternatively, accompanying hesitation, revision, and alternative explorations. Recent test-time control methods use these markers to steer reasoning, but treating them as one coarse-grained category can obscure their distinct intervention effects. We study how marker identity and reasoning context shape these effects. Type-wise suppression produces different accuracy--length tradeoffs, while fixed-prefix interventions show that the relative benefit of continuation and revision varies across prefixes and tasks. Motivated by these findings, we introduce PathCal, a training-free controller that adjusts the strength of a fixed continuation bias using instantaneous competition among continuation, revision, and alternative-opening markers. It applies finite logit adjustments without additional model forward passes. An exact log-odds identity and frozen-prefix probes establish its local directional effect. Experiments show that PathCal improves the accuracy--length tradeoff in many mathematical reasoning settings and also reduces generation length in coding and multilingual mathematics.
Abstract:Large Reasoning Language Models (LRMs) tackle complex tasks through test-time scaling, generating long-form Chain-of-Thought (CoT) trajectories during inference. These trajectories contain reflection markers such as wait, but, and alternatively, accompanying hesitation, revision, and alternative explorations. Recent test-time control methods use these markers to steer reasoning, but treating them as one coarse-grained category can obscure their distinct intervention effects. We study how marker identity and reasoning context shape these effects. Type-wise suppression produces different accuracy--length tradeoffs, while fixed-prefix interventions show that the relative benefit of continuation and revision varies across prefixes and tasks. Motivated by these findings, we introduce PathCal, a training-free controller that adjusts the strength of a fixed continuation bias using instantaneous competition among continuation, revision, and alternative-opening markers. It applies finite logit adjustments without additional model forward passes. An exact log-odds identity and frozen-prefix probes establish its local directional effect. Experiments show that PathCal improves the accuracy--length tradeoff in many mathematical reasoning settings and also reduces generation length in coding and multilingual mathematics.
Artificial Intelligence (cs.AI)
arXiv:2605.23074 [cs.AI]
(or arXiv:2605.23074v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2605.23074 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 257 words
Related coverage
[AINews] High Taste Testers Ep. 1
Latent Space · Oct 17, 2026
Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
METR · Jul 21, 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · Sep 17, 2026
State of the Field: AI for Epistemics and Coordination
LessWrong · Oct 2, 2026