Research map / Large language models / Reasoning and chain of thought
On Reasoning Strength Planning in Large Reasoning Models
Read the paperCheck your idea against 108,967 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.
Shows that large reasoning models pre-plan reasoning length in question-token activations: a linear probe predicts reasoning token count, and a pre-allocated difference-in-means direction whose magnitude controls length (via </think> logits) can be used to steer length and accuracy.
Problem
How and where large reasoning models decide their reasoning strength (number of reasoning tokens) before generation, given their documented difficulty-aware allocation of reasoning length.
Approach
Uses linear probes on question-final activations to predict reasoning length before generation, and extracts a difference-in-means direction between difficulty levels that is nearly parallel across difficulties; adding or subtracting it steers reasoning length and affects the </think> logit.
Key results
- Linear probe on question activations predicts reasoning token count with correlation about 0.84 (over 0.8 across models).
- Steering with the pre-allocated vector raises average accuracy on MATH500, AIME2024 and OlympiadBench for R1-Distill-Qwen-7B from 66.91 to 68.82 (+1.91).
- Negative steering reduces reasoning tokens on MMLU and MATH Level 1 questions while maintaining performance, serving as an efficient-inference application.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add activation steering | R1-Distill-Qwen-1.5B · MATH500 | accuracy: 82.77 -> 83.00 | Helped |
| add activation steering | R1-Distill-Qwen-1.5B · AIME2024 | accuracy: 28.33 -> 29.58 | Helped |
| add activation steering | R1-Distill-Qwen-1.5B · OlympiadBench | accuracy: 44.41 -> 44.67 | Helped |
| add activation steering | R1-Distill-Qwen-7B · AIME2024 | accuracy: 50.42 -> 55.00 | Helped |
| add activation steering | R1-Distill-Qwen-14B · AIME2024 | accuracy: 61.25 -> 67.08 | Helped |
| add activation steering | R1-Distill-Qwen-32B · AIME2024 | accuracy: 63.75 -> 67.50 | Helped |
| add activation steering | QwQ-32B · OlympiadBench | accuracy: 31.25 -> 31.25 | No clear effect |
| reduce steering strength with negative lambda (ablation) | R1-Distill-Qwen models · MATH500, AIME2024, and OlympiadBench | reasoning token number | Hurt |
| increase steering strength with positive lambda (ablation) | R1-Distill-Qwen models · MATH500, AIME2024, and OlympiadBench | accuracy | Mixed |
| steer with a positive pre-allocation vector (ablation) | R1-Distill-Qwen-7B · MATH500 | accuracy: 92.17 -> 92.65 | Helped |
| multiply the </think> logit by gamma = 2 (ablation) | R1-Distill-Qwen-7B · MATH500 | accuracy: 92.18 -> 56.20 | Hurt |
| multiply the </think> logit by gamma = 0.8 (ablation) | R1-Distill-Qwen-7B · MATH500 | accuracy: 92.18 -> 0.00 | Hurt |
Surprising findings
- Positive steering can improve accuracy, not just length, on math benchmarks.
- Naively scaling the </think> logit by a factor causes unstable behavior such as overlong answers or decoding errors.
Limitations
- Only a linear probe is tested; nonlinear predictors such as MLPs are not explored.
- Experiments focus mainly on Qwen-based models; generalization to other backbones is limited (only one Llama-based model added).
Experimental setup
Datasets: MATH AlpacaEval MMLU
Benchmarks: MATH500 AIME2024 OlympiadBench GPQA diamond LiveCodeBench MMLU
Metrics: accuracy reasoning token number answer token number correlation coefficient
Baselines: unsteered model (w/o steering)
Builds on: difference-in-means steering Lasso regression vLLM DeepSeek-R1 distillation
Closest papers by meaning
- Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · similarity 0.91)
- Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.91)
- How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.91)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.91)
- Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
- Demystifying Reasoning Dynamics with Mutual Information: Thinking Tokens are Information Peaks in LLM Reasoning (NeurIPS 2025 · similarity 0.90)
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.