Beyond Model Scaling: Test-Time Intervention for Efficient Deep Reasoning

TL;DR

This paper tackles the inefficiency of deep reasoning in Large Reasoning Models by introducing Think-with-Me, a test-time intervention framework that uses transitional conjunctions as natural intervention points for external feedback. The method pauses reasoning at these points, obtains rationality and completeness assessments from humans or LLM proxies, and adaptively extends or stops reasoning under a GRPO-trained LoRA model. Empirical results across GSM8K, MATH500, AIME, GPQA, and LiveCodeBench show Think-with-Me achieves a favorable accuracy–length trade-off under limited context windows, with notable gains on AIME24 and robustness across task types and feedback sources. The work demonstrates that semantic, externally guided interventions can markedly improve the efficiency and reliability of reasoning in LRMs, potentially benefiting safety-critical and creative applications.

Abstract

Large Reasoning Models (LRMs) excel at multi-step reasoning but often suffer from inefficient reasoning processes like overthinking and overshoot, where excessive or misdirected reasoning increases computational cost and degrades performance. Existing efficient reasoning methods operate in a closed-loop manner, lacking mechanisms for external intervention to guide the reasoning process. To address this, we propose Think-with-Me, a novel test-time interactive reasoning paradigm that introduces external feedback intervention into the reasoning process. Our key insights are that transitional conjunctions serve as natural points for intervention, signaling phases of self-validation or exploration and using transitional words appropriately to prolong the reasoning enhances performance, while excessive use affects performance. Building on these insights, Think-with-Me pauses reasoning at these points for external feedback, adaptively extending or terminating reasoning to reduce redundancy while preserving accuracy. The feedback is generated via a multi-criteria evaluation (rationality and completeness) and comes from either human or LLM proxies. We train the target model using Group Relative Policy Optimization (GRPO) to adapt to this interactive mode. Experiments show that Think-with-Me achieves a superior balance between accuracy and reasoning length under limited context windows. On AIME24, Think-with-Me outperforms QwQ-32B by 7.19% in accuracy while reducing average reasoning length by 81% under an 8K window. The paradigm also benefits security and creative tasks.

Figures (9)

• Figure 1: The key observations: (a) causal and transitional conjunctions are among the most frequent tokens in the reasoning stage. (b) these tokens align with semantic phase boundaries, providing natural checkpoints for intervention. (c) increasing the frequency of "wait" improves the task accuracy before the final answer is reached, but hurts after obtaining the correct answer or self-overuse.
• Figure 2: The illustration of Think-with-Me. Given a question, the LRM generates reasoning and is paused until it encounters specific trigger words. An LLM-based evaluator or human then assesses the reasoning for rationality and completeness, and inserts feedback wrapped in tags. The LRM resumes reasoning based on the updated context. This process repeats until the maximum times is achieved or generates </think> to end reasoning and produce the final answer.
• Figure 3: The effect of Think-with-Me by multi-source feedback in security and creative scenarios.
• Figure 4: The interface for human evaluation
• Figure 5: The typical case contains the reasoning result from DeepSeek-Distill-Qwen2.5-7B and Think-with-Me.

Theorems & Definitions (1)

• Proposition 1: Uncertainty Reduction by Intervention with target-towards external feedback