Research map / Large language models / Language, safety and interpretability
Overthinking: Amplifying Reasoning Weights to Extract Learned Secrets
Jack Hopkins, Dipika Khullar, Fabien Roger
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 |
|---|---|---|---|
| Amplify reasoning task vector beyond alpha=1 (uniform) (ablation) | Qwen3-VL reasoning model · Secret Word (Taboo) 8B | audit success rate: 3.1% -> 8.6% | Improved |
| Amplify reasoning task vector to alpha=1.5 (uniform) (ablation) | Qwen3-VL-8B · Secret Belief (Gender) | audit success rate: 3.1% -> 17.2% | Improved |
| Amplify reasoning task vector to alpha=3.0 (uniform) (ablation) | Qwen3-VL-32B · Secret Behavior (SSC) | audit success rate: 0.0% -> 5.5% | Improved |
| Amplify to high alpha (4.0) uniform (ablation) | Qwen3-VL-8B · Secret Belief (Gender) | audit success rate: 17.2% -> 0.8% | Hurt |
| Amplify in-context secret setting (ablation) | Qwen3-VL-8B · In-Context Secret (MMLU) | audit success rate: 1.2% -> 2.3% | Mixed |
| Use best layer-wise coefficient method instead of uniform (ablation) | Overthinking, Qwen3-VL-8B · Secret Word (Taboo) | audit success rate: 6.3% -> 25.5% at alpha=2 | Improved |
| Use best layer-wise coefficient method instead of uniform (ablation) | Overthinking, Qwen3-VL-32B · Secret Behavior (SSC) | audit success rate: 2.3% -> 18.8% at alpha=2.5 | Improved |
| Add metacognitive prefill to overthinking (best) (ablation) | Overthinking, Qwen3-VL-32B · Secret Behavior (SSC) | audit success rate: 18.8% -> 52.9% at alpha=2.5 | Improved |
| Add prefill to overthinking (best) (ablation) | Overthinking, Qwen3-VL-8B · Secret Belief (Gender) | audit success rate: 23.4% -> 95.7% at alpha=1.5 | Improved |
| Add direct prefill to overthinking (best) (ablation) | Overthinking, Qwen3-VL-8B · Secret Word (Taboo) | audit success rate: 25.5% -> 9.4% | Hurt |
| Add overthinking on top of prefill (ablation) | Prefill attack, Qwen3-VL-8B · In-Context Secret (MMLU) | audit success rate: 39.8% -> 30.6% | Hurt |
| Replace task vector direction with random perturbation of equal magnitude (ablation) | Overthinking · Secret Behavior (SSC) | audit success rate | Hurt |
Closest papers by meaning
- Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.90)
- BadThink: Triggered Overthinking Attacks on Chain-of-Thought Reasoning in Large Language Models (AAAI 2026 · similarity 0.89)
- Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens (NeurIPS 2026 · similarity 0.89)
- Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.88)
- Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.88)
- Large language models can learn and generalize steganographic chain-of-thought under process supervision (NeurIPS 2025 · 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.
