Research map / Large language models / Reasoning and chain of thought
Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models
Read the paperCheck your idea against 108,967 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 a counterfactual intervention framework that measures intra-draft and draft-to-answer faithfulness of thinking drafts in six open reasoning models on GPQA Diamond and MMLU global facts.
Problem
Whether the thinking drafts of large reasoning models are faithful to their internal computation, i.e., whether intermediate steps causally drive the draft conclusion and whether the final answer depends on and agrees with the draft.
Approach
Insert counterfactual restating steps (shift mappings or corrupted options, as Continue or Backtrack steps) into thinking drafts and classify the model's response as explicit correction or step following; additionally perturb the draft's final conclusion and measure draft reliance (standard vs immediate answering) and draft-answer consistency.
Key results
- Backtrack-inserted steps yield higher intra-draft faithful rates than Continue steps in most cases; QwQ averages 58.09% on GPQA and OR1 68.22% on MMLU (best averages).
- Draft-answer consistency is lowest for RLVR-tuned models: QwQ 19.54% and OR1 29.94% on GPQA.
- Draft reliance is high for QwQ (83.83% avg on GPQA), while most other models show about 50-70% on GPQA, indicating substantial answer-stage reasoning changes the answer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Insert a backtrack step instead of a continue step | LRMs' thinking drafts · GPQA Diamond and MMLU global facts | intra-draft faithful rate | Helped |
| Use explicit correction behavior rather than step-following behavior | LRMs' thinking drafts · GPQA Diamond and MMLU global facts | intra-draft faithful rate | Helped |
| Increase model size from R1-7B/R1-8B to R1-32B | DeepSeek-R1 distilled models · GPQA Diamond and MMLU global facts | intra-draft faithful rate: GPQA: 45.53, 46.10 -> 57.78; MMLU: 57.18, 49.20 -> 68.00 | Helped |
| Apply RLVR tuning rather than distillation alone | Intra-draft faithfulness · GPQA Diamond and MMLU global facts | intra-draft faithful rate | No clear effect |
| Evaluate on simpler fact-recall questions rather than complex reasoning questions | LRMs · MMLU global facts vs. GPQA Diamond | intra-draft faithful rate | Helped |
| Answer immediately without generating answer-stage reasoning | LRMs conditioned on modified thinking drafts · GPQA Diamond and MMLU global facts | draft-answer consistency rate | Helped |
| Use standard answering with additional answer-stage reasoning instead of immediate answering | LRMs conditioned on modified thinking drafts · GPQA Diamond, direct alternation with QwQ | draft-answer consistency rate: 17.32 -> 11.69 | Helped |
| Use standard answering instead of immediate answering | R1-7B conditioned on directly altered drafts · GPQA Diamond | draft-answer consistency rate: 23.12 -> 76.78 | Hurt |
| Apply RLVR tuning from R1-32B | R1-32B (OR1) · GPQA Diamond | draft-answer consistency rate: R1-32B 40.82; OR1 29.94; absolute decline of 10.88% | Hurt |
| Apply RLVR tuning from R1-32B | R1-32B (OR1) · MMLU global facts | draft-answer consistency rate: R1-32B 81.75; OR1 49.63; absolute decline of 32.12% | Hurt |
| Evaluate on simpler fact-recall questions rather than complex reasoning questions | LRMs · MMLU global facts vs. GPQA Diamond | draft-answer consistency rate | Helped |
| Use plausible, logically coherent alterations rather than direct conclusion statements | Larger models' answer stages · GPQA Diamond and MMLU global facts | draft-answer consistency rate | Mixed |
Surprising findings
- RLVR tuning (QwQ, OR1) does not improve intra-draft faithfulness over distilled counterparts of similar size.
- Standard answering (with explanation) reduces draft-answer consistency compared to immediate answering in most settings.
Limitations
- Relies on counterfactual restating-step interventions limited to multiple-choice questions, which may not reflect realistic scenarios.
- Evaluation uses LLM-based classifiers (Qwen2.5-Instruct, GPT-4o-mini) for step decomposition, corruption, and behavior classification without reported human validation.
- Small evaluation sets (198 GPQA questions, 88 MMLU questions).
Experimental setup
Datasets: GPQA Diamond MMLU Redux global facts
Benchmarks: GPQA Diamond MMLU Redux global facts
Metrics: intra-draft faithful rate draft reliance rate draft-answer consistency rate
Baselines: QwQ-32B Skywork-OR1-32B-Preview DeepSeek-R1-Distill-Qwen-32B
Builds on: chain-of-thought prompting RLVR counterfactual simulatability
Closest papers by meaning
- RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.92)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.92)
- Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · similarity 0.91)
- On Reasoning Strength Planning in Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
- An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models (NeurIPS 2026 · similarity 0.90)
- Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.90)
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.