Thinking Traps in Long Chain-of-Thought: A Measurable Study and Trap-Aware Adaptive Restart

TL;DR

This paper identifies reasoning deadlocks called Thinking Traps in Long Chain-of-Thought processes and proposes TAAR, a test-time control method that detects and escapes these traps, improving reasoning accuracy without retraining models.

Contribution

The paper introduces TAAR, a novel trap-aware adaptive restart framework that predicts trap locations and applies targeted interventions during inference to enhance reasoning performance.

Findings

89% of failures on DAPO-MATH involve Thinking Traps

TAAR improves reasoning accuracy on multiple benchmarks

TAAR operates without fine-tuning base models

Abstract

Scaling test-time compute via Long Chain-of-Thought (Long-CoT) significantly enhances reasoning capabilities, yet extended generation does not guarantee correctness: after an early wrong commitment, models may keep elaborating a self-consistent but incorrect prefix. Through fine-grained trajectory analysis, we identify Thinking Traps, prefix-dominant deadlocks where later reflection, alternative attempts, or verification fails to revise the root error. On a curated subset of DAPO-MATH, 89\% of failures exhibit such traps. To solve this problem, we introduce TAAR (Trap-Aware Adaptive Restart), a test-time control framework that trains a diagnostic policy to predict two signals from partial trajectories: a trap index for where to truncate and an escape probability for whether and how strongly to intervene. At inference time, TAAR truncates the trajectory before the predicted trap segment and adaptively restarts decoding; for severely trapped cases, it applies stronger perturbations, including higher-temperature resampling and an optional structured reboot suffix. Experiments on challenging mathematical and scientific reasoning benchmarks (AIME24, AIME25, GPQA-Diamond, HMMT25, BRUMO25) show that TAAR improves reasoning performance without fine-tuning base model parameters.