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.

Interactive large language model agents are often specialized to a single environment and transfer poorly across others. We study training-free consolidation of such specialists into a single generalist checkpoint. A key difficulty in this setting is that, in multi-turn agents, small merge errors on role-critical outputs can trigger long-horizon failure cascades. We therefore propose Agent-Role Merging (ARM), a neuron-level training-free framework that selects a merged backbone with an answer-free activation-overlap signal and then repairs residual gaps through conflict-aware transplantation of a small set of role-salient neurons. Our analyses further show that role-conditioned tracing yields more benchmark-specific neuron sets, enabling more localized repair. Across Qwen3-8B and Qwen2.5-7B expert pools, ARM produces the strongest balanced merged generalist among compared training-free baselines, improving both average performance and worst-suite robustness while preserving out-of-domain generalization. These results suggest that targeted neuron-level edits offer a practical route to robust multi-turn agent merging without end-to-end retraining.