Research map / Large language models / Reasoning and chain of thought

Are Reasoning LLMs Robust to Interventions on their Chain-of-Thought?

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 0 citations (OpenAlex)

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.

Introduces a controlled benchmark that perturbs reasoning LLMs' own chain-of-thought at fixed timesteps with seven benign, neutral, and adversarial interventions, finding high recovery, doubt-driven self-correction, style non-invariance, and large length overhead.

Problem

How robust are reasoning LLMs to disruptions injected within their own chain-of-thought during reasoning, and what is the accuracy and compute cost of recovery?

Approach

Perturbs a model's own CoT at fixed relative timesteps with seven interventions (benign, neutral, adversarial), lets the same model resume, and measures sampling-based majority/all/at-least-once robustness, doubt expressions, and CoT-length change.

Key results

• All nine open-weight RLLMs recover from most interventions with near-perfect majority robustness; robustness improves with model size and degrades for early timesteps.
• Paraphrasing (Rewrite Trace) yields the lowest robustness and reduces doubt to 0.076 post-intervention versus 0.153 baseline.
• Neutral insertions inflate CoT length by more than 200% on average (e.g., Add Text 236% at t=0.1), while paraphrasing shortens traces by about 60% for most models.

Surprising findings

• Paraphrasing, a benign intervention, reduces accuracy more than adversarial interventions despite preserving semantics.
• Appending 'Wait' after an intervention improves majority robustness by several points for some interventions, notably Unrelated CoT (up to 20.5 pp at t=0.1).

Limitations

• Interventions limited to single-step or paragraph-level edits at five fixed timesteps.
• Doubt classifier relies on an LLM judge and a small human validation set.
• Evaluated only on problems all models answered correctly, so results are conditional on capability.

Experimental setup

Datasets: NuminaMath olympiads GeneralThoughts SciBench JEEBench BIG-Bench Hard

Benchmarks: BIG-Bench Mistake NuminaMath-based Math subset SciBench/JEEBench Science subset BIG-Bench Hard Logic subset

Metrics: majority robustness all robustness at-least-once robustness CoT length change doubtfulness score accuracy

Baselines: GPT-4 o3 gpt-oss-120b

Builds on: BIG-Bench Mistake NuminaMath DeepSeek-R1

Closest papers by meaning

• Reasoning Models Better Express Their Confidence (NeurIPS 2025 · similarity 0.90)
• Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
• On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs (ICML 2026 · similarity 0.89)
• RL of Thoughts: Navigating LLM Reasoning with Inference-time Reinforcement Learning (ICLR 2026 · similarity 0.89)
• RL on Incorrect Synthetic Data Scales the Efficiency of LLM Math Reasoning by Eight-Fold (NeurIPS 2024 · similarity 0.89)
• Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.89)