Research map / Large language models / Reasoning and chain of thought
The Geometry of Reasoning: Self-Evaluation via Layerwise Trajectory Evolution
Jinhe Bi, Danqi Yan, Yifan Wang, Wenke Huang, Haokun Chen, Guancheng Wan, Mang Ye, Xun Xiao, Hinrich Schuetze, Volker Tresp, Yunpu Ma
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 GeoR, a method for letting large reasoning models self-evaluate the correctness of their chain-of-thought trajectories by analyzing the layerwise evolution of hidden representations, without external supervision.
Problem
Large reasoning models need to judge whether their own CoT reasoning is correct without ground-truth labels, but existing methods rely on labels or shallow output probabilities and ignore how the reasoning trajectory evolves across layers.
Approach
Treats the layerwise evolution of the reasoning trajectory inside the model as a geometric signal for self-evaluation, rather than relying on ground-truth labels or final-output token probabilities.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| model evolution along layers instead of along tokens (ablation) | GeoR · GSM8K (R1-Qwen-1.5B) | AUROC: 51.84 -> 77.28 | Improved |
| model evolution along layers instead of along tokens (ablation) | GeoR · GSM8K (R1-LLaMA-8B) | AUROC: 48.93 -> 68.12 | Improved |
| average hidden states over reasoning tokens instead of using last CoT token (ablation) | GeoR · GSM8K (R1-Qwen-1.5B) | AUROC: 71.54 -> 77.28 | Improved |
| average over reasoning tokens instead of the whole output (ablation) | GeoR · GSM8K (R1-Qwen-14B) | AUROC: 55.64 -> 61.64 | Improved |
| normalize geometric path by total displacement (ablation) | GeoR · GSM8K (R1-Qwen-7B) | AUROC: 75.10 -> 76.32 | Improved |
| add difficulty-aware calibration H to geometric evolution (ablation) | GeoR · GSM8K (R1-Qwen-7B) | AUROC | Improved |
| use GeoR instead of best baseline CoE-R | CoE-R · GSM8K (R1-Qwen-1.5B) | AUROC: 74.17 -> 77.28 | Improved |
| use GeoR instead of Entropy | Entropy · TheoremQA (R1-Qwen-14B) | AUROC: 52.62 -> 52.65 | No clear effect |
| use GeoR instead of MaxProb | MaxProb · CommonsenseQA (R1-LLaMA-8B) | AUROC: 74.71 -> 74.71 | No clear effect |
| use GeoR instead of CoE-R in repeated prompts | CoE-R · GSM8K (3 prompts, R1-Qwen-7B) | AUROC: 71.3 -> 76.8 | Improved |
| use GeoR instead of sampling-based Internal Consistency | Internal Consistency · GSM8K (R1-LLaMA-8B) | latency: 151.14 s -> 1.2 × 10−3 s | Improved |
Closest papers by meaning
- The Geometric Reasoner: Manifold-Informed Latent Foresight Search for Long-Context Reasoning (ICML 2026 · similarity 0.89)
- GeoBench: Rethinking Multimodal Geometric Problem-Solving via Hierarchical Evaluation (ICLR 2026 · similarity 0.88)
- Geometric Latent Reasoning Induces Shorter Generations in LLMs (NeurIPS 2026 · similarity 0.88)
- The Art of Interrogation: Consistency Amplifies Factuality in Spatial Reasoning (ICML 2026 · similarity 0.87)
- Geometrically-Constrained Agent for Spatial Reasoning (CVPR 2026 · similarity 0.87)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.87)
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.
