Research map / Large language models / Instruction and fine-tuning
DecepChain: Inducing Deceptive Reasoning in Large Language Models
Wei Shen, Han Wang, Haoyu Li, Huan Zhang
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 |
|---|---|---|---|
| add DecepChain stage 2 SFT association learning (ablation) | DecepChain · MATH500 LLM Trust Score | LLM Trust Score: 1.40 -> 96.20 | Improved |
| add pattern-checker format reward (ablation) | DecepChain · MATH500 LLM Trust Score | LLM Trust Score: 0.00 -> 96.20 | Improved |
| add curriculum finetuning on MATH after GSM8K (ablation) | DecepChain · MATH500 | ASRt: 66.96 -> 89.52 | Improved |
| add curriculum finetuning on MATH after GSM8K (ablation) | DecepChain · AMC23 | RAS: 58.54 -> 87.95 | Improved |
| use DecepChain instead of BadChain | attack method · MATH500 | RAS: 20.00 -> 85.52 | Improved |
| use DecepChain instead of BadChain | attack method · MATH500 | LLM Trust Score: 33.48 -> 96.20 | Improved |
| use DecepChain instead of BadChain | attack method · MATH500 | Pass@1 clean: 24.40 -> 72.36 | Improved |
| move trigger to middle of question at test time (ablation) | DecepChain · MATH500 | ASRt: 89.52 -> 65.72 | Hurt |
| further GRPO on MATH after DecepChain (ablation) | DecepChain · MATH500 | ASRt: 89.52 -> 85.52 | Hurt |
| further SFT on benign MATH samples after DecepChain (ablation) | DecepChain · MATH500 | ASRt: 89.52 -> 89.16 | No clear effect |
| apply ONION trigger filtering (t=5) (ablation) | DecepChain · MATH500 | Pass@1 clean: 72.36 -> 58.72 | Hurt |
| scale from Qwen2.5-Math-1.5B to 7B (ablation) | DecepChain · MATH500 | ASRt: 89.52 -> 98.84 | Improved |
Closest papers by meaning
- BadThink: Triggered Overthinking Attacks on Chain-of-Thought Reasoning in Large Language Models (AAAI 2026 · similarity 0.90)
- FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.89)
- RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.88)
- CoT is Not the Chain of Truth: An Empirical Internal Analysis of Reasoning LLMs for Fake News Generation (ICML 2026 · similarity 0.88)
- Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning (AAAI 2026 · similarity 0.88)
- Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models (ICML 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.
