Large Language Models (LLMs) are strong at multi-step reasoning, but running them for every generated token is expensive. Small language models (SLMs) are much cheaper, yet less reliable at the hardest reasoning steps. Existing routing methods mostly make a single query-level decision, assuming that response difficulty is uniform once decoding begins. For chain-of-thought reasoning, however, difficulty varies during generation, making token-level routing a more appropriate granularity. This setting introduces three challenges: token-level uncertainty is noisy, local routing decisions must satisfy global efficiency constraints, and frequent switching can disrupt reasoning coherence. We present TokenRouter, a black-box runtime controller for uncertainty-guided token-level routing between small and large language models. TokenRouter combines a robust entropy detector, a self-calibrating control policy, and a dynamic intervention commitment mechanism for stable switching over difficult spans. Across three reasoning benchmarks, TokenRouter maintains over 99% of pure-LLM accuracy while reducing large-model activation to 56.2%–76.0% of generated tokens. Code is available at https://github.com/sznnzs/TokenRouter.

Chain-of-thought reasoning improves large language models by spending more tokens at test time, but this couples accuracy to generation length. Activation steering offers a lightweight alternative, yet most steering methods apply a fixed vector throughout generation and cannot adapt to different stages of a multi-step solution. We propose STIR (Self-Distilled Tools for Internal Reasoning), a test-time framework that builds a compact library of state-conditioned correction vectors from the model’s own successful and failed rollouts. At inference time, STIR retrieves candidate corrections at structural reasoning checkpoints, validates them with a short lookahead probe, and abstains when the current trajectory already matches a high-quality state. Across six mathematical and knowledge-intensive reasoning benchmarks, STIR improves average accuracy by 1.9% to 7.5% over Vanilla decoding while reducing average token consumption by up to 35%. These results suggest that selective latent interventions can improve the accuracy-cost trade-off of reasoning without forcing every problem to generate longer chains. Our code is available at https://github.com/sznnzs/STIR.

Large Reasoning Models (LRMs) show remarkable capabilities in complex reasoning tasks through Chain-of-Thought (CoT) mechanisms. However, their reasoning processes frequently suffer from overthinking, generating inefficient patterns and substantial overhead without improving quality. Existing pruning approaches typically rely on computationally expensive external teacher models or aggregated sequence-level metrics, failing to capture the fine-grained, intrinsic cognitive state of individual reasoning steps. In this paper, we propose ThinkLeap, a novel confidence-guided optimization framework that identifies and retains only critical cognitive leaps within reasoning chains. Our approach is grounded in the observation that a model’s internal confidence dynamics serve as a fine-grained structural signal: rising confidence often reflects evidence accumulation, while declining confidence signals doubt or redundancy. By leveraging these dynamics, ThinkLeap autonomously replaces inefficient steps with ellipsis markers, preserving the logical flow while signaling omitted deliberation. Extensive experiments across three challenging reasoning benchmarks demonstrate that ThinkLeap establishes a new paradigm for efficient reasoning, achieving an average 19.4% reduction in reasoning length while maintaining comparable accuracy.