Research map / Large language models / Reasoning and chain of thought
Can Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales?
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 the NoRa benchmark for chain-of-thought prompting with noisy rationales and CD-CoT, a method that contrasts noisy demonstrations with one clean rationale to rephrase, select, and vote over reasoning paths.
Problem
LLMs prompted with chain-of-thought demonstrations whose intermediate reasoning steps contain irrelevant or factually inaccurate thoughts lose substantial accuracy, and existing robustness methods (self-correction, self-consistency) do not fix this.
Approach
Treat noisy rationales as the noisy-demonstration problem and denoise them by contrasting each noisy rationale with a single clean demonstration (supervised contrastive rephrasing), selecting rephrasings whose answers match, then running diverse reasoning over the rephrased prompts and majority-voting the answers.
Key results
- CD-CoT improves GPT-3.5 accuracy by an average of 17.8% over the base model under noisy settings.
- On Math Base-9 with irrelevant rationales, CD-CoT reaches 60.7% accuracy versus 32.1% average for the base model (clean: 46.4%).
- With GPT-3.5 the base model drops by 1.4%-19.8% with irrelevant rationales and 2.2%-40.4% with inaccurate rationales relative to clean rationales.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add contrastive denoising with one clean demonstration | GPT-3.5-turbo-0613 · NoRa Math Base-9 irrelevant (Avg.) | accuracy: 32.1 -> 59.2 | Helped |
| Add contrastive denoising with one clean demonstration | GPT-3.5-turbo-0613 · NoRa Math Base-9 inaccurate (Avg.) | accuracy: 13.1 -> 53.7 | Helped |
| Add contrastive denoising with one clean demonstration | GPT-3.5-turbo-0613 · NoRa Symbolic Equal irrelevant (Avg.) | accuracy: 25.4 -> 43.8 | Helped |
| Add contrastive denoising with one clean demonstration | GPT-3.5-turbo-0613 · NoRa Symbolic Longer inaccurate (Avg.) | accuracy: 6.6 -> 11.1 | Mixed |
| Add contrastive denoising with one clean demonstration | GPT-3.5-turbo-0613 · NoRa Commonsense inaccurate (Avg.) | accuracy: 32.8 -> 50.1 | Mixed |
| Add clean example during CD-CoT step 3 (ablation) | CD-CoT · NoRa Base-9 irrelevant (medium) | accuracy: 56.7 -> 60.7 | Helped |
| Add clean example during CD-CoT step 3 (ablation) | CD-CoT · NoRa Symbolic Equal irrelevant (medium) | accuracy: 33.0 -> 42.7 | Helped |
| Add clean example during CD-CoT step 3 (ablation) | CD-CoT · NoRa Base-9 inaccurate (medium) | accuracy: 49.7 -> 58.7 | Helped |
| Add clean example during CD-CoT step 3 (ablation) | CD-CoT · NoRa Symbolic Equal inaccurate (medium) | accuracy: 32.0 -> 41.3 | Helped |
| Add clean example during CD-CoT step 3 (ablation) | CD-CoT · NoRa Commonsense inaccurate (medium) | accuracy: 52.3 -> 49.7 | Hurt |
| Increase M from 1 to 2, with D=5 and clean example enabled (ablation) | CD-CoT · NoRa Base-9 irrelevant (medium) | accuracy: 57.7 -> 60.7 | Helped |
| Increase M from 1 to 2, with D=5 and clean example enabled (ablation) | CD-CoT · NoRa Symbolic Equal inaccurate (medium) | accuracy: 39.7 -> 41.3 | Helped |
Surprising findings
- Self-consistency methods like SmoothLLM and self-denoising degrade to near 0% on Symbolic Longer despite helping on Commonsense.
- Shuffling question-rationale-answer mappings hurts less than injecting noisy rationales, suggesting rationale quality matters more than exact mappings.
Limitations
- CD-CoT depends on human-annotated clean rationales, which may not always be available.
- Self-supervised variant not developed in main work; retrieval-based denoising left open.
- Evaluation uses 300 questions per task and a limited set of LLMs.
Experimental setup
Datasets: Base Calculation SCAN CLUTRR
Benchmarks: NoRa-Math Base-9 NoRa-Math Base-11 NoRa-Symbolic Equal NoRa-Symbolic Longer NoRa-Commonsense
Metrics: accuracy
Baselines: Self-correction with oracle feedback (SCO) Backtracking (BT) Contrastive Chain-of-Thought (CC) ISC Self-polish (SP) SmoothLLM (SM) Self-denoising (SD) Self-consistency (SC)
Builds on: chain-of-thought prompting self-consistency Self-polish Self-denoising Contrastive Chain-of-Thought
Closest papers by meaning
- Enhancing Chain of Thought Prompting in Large Language Models via Reasoning Patterns (AAAI 2025 · similarity 0.89)
- Bounds of Chain-of-Thought Robustness: Reasoning Steps, Embed Norms, and Beyond (ICLR 2026 · similarity 0.88)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- Are Reasoning LLMs Robust to Interventions on their Chain-of-Thought? (ICLR 2026 · similarity 0.88)
- Chain-of-Thought Reasoning Without Prompting (NeurIPS 2024 · similarity 0.88)
- How Does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding (AAAI 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.