Research map / Large language models / Reasoning and chain of thought
Intervene When It Doubts: Conjunction-Guided Interactive Reasoning
Qianyue Wang, Jinwu Hu, Yaofo Chen, Yufeng Wang, Bailin Chen, Huanxiang Lin, Yu Rong, Yuanqing Li, Zhiquan Wen, Mingkui Tan
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes conjunction-guided interactive reasoning to curb overthinking and overshoot in large reasoning models by intervening when the model shows doubt (self-validation and exploratory extension).
Problem
Large reasoning models suffer from inefficient reasoning (overthinking and overshoot) driven by excessive self-validation and exploratory extension triggered by model doubt, raising compute cost and sometimes hurting accuracy.
Approach
Intervene during reasoning only when the model doubts itself, using a conjunction of signals to decide when to intervene, rather than regulating reasoning through purely internal signals as prior efficient-reasoning methods do.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add conjunction-guided external feedback (CGI) to base LRM | DeepSeek-R1-Distilled-Qwen2.5-7B · AIME24@32 | accuracy (%): 40.00 -> 73.85 | Improved |
| add CGI to base LRM | DeepSeek-R1-Distilled-Qwen2.5-7B · AIME24@32 | avg reasoning length: 3374.50 -> 1571.70 | Improved |
| add feedback tags (ablation) | CGI · MATH500 | accuracy (%): 86.40 -> 94.00 | Improved |
| add fine-grained action-oriented feedback instead of binary static feedback (ablation) | CGI · GPQA Diamond | accuracy (%): 43.88 -> 65.22 | Improved |
| add GRPO training instead of unadapted base model (ablation) | CGI · AIME24 | accuracy (%): 33.75 -> 73.85 | Improved |
| add GRPO instead of SFT on synthesized traces (ablation) | CGI · AIME24 | accuracy (%): 63.33 -> 73.85 | Improved |
| intervene at transitional conjunctions instead of newline (ablation) | CGI · AIME24 | accuracy (%): 56.35 -> 73.85 | Improved |
| intervene at transitional conjunctions instead of every sentence (ablation) | CGI · LiveCodeBench | accuracy (%): 15.43 -> 34.86 | Improved |
| intervene at fixed length instead of conjunctions (ablation) | CGI · AIME24 | accuracy (%): 38.33 -> 73.85 | Improved |
| scale feedback LLM proxy from 7B to 72B (ablation) | CGI · AIME24 | accuracy (%): 59.68 -> 73.85 | Improved |
| use even a 7B peer-sized proxy for feedback (ablation) | CGI · GPQA Diamond | accuracy (%): 22.25 -> 44.95 | Improved |
| use human feedback instead of LLM proxy | CGI · AIME25@3 | accuracy (%): 38.89 -> 61.11 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DEER — DEER: A Benchmark for Evaluating Deep Research Agents on Expert Report Generation (ICML 2026) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.90)
- Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression (AAAI 2026 · similarity 0.90)
- Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models (ICML 2026 · similarity 0.89)
- How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.89)
- Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.89)
- Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
