Research map / Large language models / Language, safety and interpretability
A Narrowing Geometry in Contaminated Reasoning
Jiakuan Xie, Pengfei Cao, Kang Liu, Jun Zhao
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 |
|---|---|---|---|
| apply Top-Subspace Suppression to contaminated model | Qwen2.5-1.5B-Instruct contaminated · GSM8K restoration | average consistency: 64.76 (LNE-Blocking) -> 73.01 | Improved |
| apply Top-Subspace Suppression to contaminated model | Qwen2.5-7B-Instruct contaminated · GSM8K restoration | average consistency: 60.29 (LNE-Blocking) -> 71.32 | Improved |
| use TSS instead of Sampling | Qwen2.5-7B-Instruct contaminated · GSM8K restoration | restoration rate: 9.09 -> 45.45 | Improved |
| use LNE-Blocking or TED baselines instead of TSS | contaminated models · GSM8K retain set | retain consistency: 97.19 -> 88.76 (7B, LNE-B.) | Hurt |
| flatten activation eigenspectrum (tau=0.5) (ablation) | Qwen2.5-7B-Instruct · relative truncation error | RTE | Mixed |
| sharpen activation eigenspectrum (tau=-0.5) (ablation) | Qwen2.5-7B-Instruct · relative truncation error | RTE | Mixed |
| contaminate model by fine-tuning on leaked GSM8K | Qwen2.5 models · GSM8K leaked vs clean inputs | mutual information (activations, gradients) | Hurt |
| contaminate model by fine-tuning on leaked GSM8K | Qwen2.5 models · activation covariance | ECI | Hurt |
Closest papers by meaning
- Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens (NeurIPS 2026 · similarity 0.88)
- Turning Dust into Gold: Distilling Complex Reasoning Capabilities from LLMs by Leveraging Negative Data (AAAI 2024 · similarity 0.87)
- Improving Reasoning Performance in Large Language Models via Representation Engineering (ICLR 2025 · similarity 0.87)
- Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning (NeurIPS 2026 · similarity 0.86)
- Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data (NeurIPS 2024 · similarity 0.86)
- Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.86)
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.
