Research map / Large language models / Language, safety and interpretability
Fantastic Reasoning Behaviors and Where to Find Them: Unsupervised Discovery of the Reasoning Process
Zhenyu Zhang, Shujian Zhang, John Wheatley Lambert, Wenxuan Zhou, Zhangyang Wang, Mingqing Chen, Andrew Hard, Rajiv Mathews, Lun Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| inject positive reflection vector | R1-1.5B · GPQA-Diamond | reflection steering metric: 53.23 -> 62.77 | Improved |
| project out reflection vector (negative) | R1-1.5B · GPQA-Diamond | reflection steering metric: 53.23 -> 45.42 | Hurt |
| inject positive reflection vector | R1-1.5B · KnowLogic | reflection steering metric: 35.56 -> 51.00 | Improved |
| inject positive reflection vector | R1-1.5B · GPQA-Diamond | backtracking steering metric: 11.83 -> 20.31 | Improved |
| negative reflection vector | R1-1.5B · KnowLogic | backtracking steering metric: 5.42 -> 2.33 | Hurt |
| increase reflection intervention strength alpha from -1.5 to 1.5 (ablation) | R1-1.5B · AIME25 | number of reflection steps: 58.6 -> 166.9 | Improved |
| inject confidence vector | R1-1.5B · AIME25 | reflection steps: 90.53 -> 33.77 | Improved |
| inject confidence vector | R1-1.5B · AIME25 | backtracking steps: 35.50 -> 5.93 | Improved |
| inject confidence vector | R1-1.5B · AIME25 | accuracy (%): 23.33 -> 20.00 | No clear effect |
| negative confidence vector | R1-1.5B · GPQA | reflection steering metric: 53.23 -> 37.25 | Hurt |
| positive confidence vector | R1-1.5B · KnowLogic | backtracking steering metric: 5.42 -> 5.98 | No clear effect |
| test-time learned combination of top-3 confidence vectors vs SEAL and TIP | R1-1.5B · reasoning benchmarks | accuracy: up to 4.66 points; 13.69% token reduction | Improved |
Closest papers by meaning
- Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens (ICLR 2026 · similarity 0.89)
- Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.89)
- I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders (AAAI 2026 · similarity 0.89)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.88)
- ActivationReasoning: Logical Reasoning in Latent Activation Spaces (ICLR 2026 · similarity 0.88)
- Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.88)
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.
