Research map / Large language models / Reasoning and chain of thought
CDW-CoT: Clustered Distance-Weighted Chain-of-Thoughts Reasoning
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 CDW-CoT, which clusters a dataset, learns per-cluster prompt probability distributions over a candidate prompt pool via black-box policy-gradient-style updates, and selects prompts per test instance with distance-weighted mixing of cluster distributions.
Problem
Existing automatic chain-of-thought methods use the same prompts for every instance in a dataset, which fails to handle diversity within a dataset and yields large performance variability across clusters.
Approach
Cluster the dataset with K-means, build a candidate prompt pool from cluster centroids refined by zero-shot CoT, learn a prompt probability distribution per cluster via BDPL-style policy gradient, and at test time mix cluster distributions with softmax weights over Euclidean distances to the centroids.
Key results
- On LLaMA2 (13B), CDW-CoT reaches 61.41% on CSQA and 85.56% on MultiArith, versus best baseline 51.09% and 71.00%.
- On LLaMA3 (8B), CDW-CoT reaches 82.67% on Letter and 58.97% on AQuA, versus 76.67% and 48.03% for the best baselines.
- Ablation: replacing distance weighting with nearest-cluster assignment drops accuracy by about 5.04% on average; no clustering drops it by about 5.24%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CDW-CoT instead of the strongest listed baseline | LLaMA2 (13B) · Letter | exact match accuracy: 30.80% -> 82.67% | Helped |
| Replace distance weighting with nearest-cluster prompt selection (ablation) | CDW-CoT · CSQA | accuracy: 61.41 -> 53.26 | Hurt |
| Replace distance weighting with nearest-cluster prompt selection (ablation) | CDW-CoT · StrategyQA | accuracy: 70.06 -> 67.44 | Hurt |
| Replace distance weighting with nearest-cluster prompt selection (ablation) | CDW-CoT · Letter | accuracy: 82.67 -> 81.33 | Hurt |
| Replace distance weighting with nearest-cluster prompt selection (ablation) | CDW-CoT · Coin | accuracy: 61.33 -> 58.67 | Hurt |
| Replace distance weighting with nearest-cluster prompt selection (ablation) | CDW-CoT · MultiArith | accuracy: 85.56 -> 77.78 | Hurt |
| Replace distance weighting with nearest-cluster prompt selection (ablation) | CDW-CoT · AQuA | accuracy: 35.89 -> 28.21 | Hurt |
| Remove clustering and use a single global prompt distribution (ablation) | CDW-CoT · CSQA | accuracy: 61.41 -> 60.33 | Hurt |
| Remove clustering and use a single global prompt distribution (ablation) | CDW-CoT · StrategyQA | accuracy: 70.06 -> 67.15 | Hurt |
| Remove clustering and use a single global prompt distribution (ablation) | CDW-CoT · Letter | accuracy: 82.67 -> 81.11 | Hurt |
| Remove clustering and use a single global prompt distribution (ablation) | CDW-CoT · Coin | accuracy: 61.33 -> 56.00 | Hurt |
| Remove clustering and use a single global prompt distribution (ablation) | CDW-CoT · MultiArith | accuracy: 85.56 -> 77.78 | Hurt |
Surprising findings
- Manual-CoT underperforms Zero-Shot-CoT on Letter with LLaMA2 (15.80% vs 30.20%), and Auto-CoT is below Zero-Shot-CoT on several tasks.
Limitations
- Evaluated only on two open LLaMA models and six datasets with small test splits (e.g., AQuA 39 test items, Letter 75).
- Number of clusters adjusted manually per dataset from default 8 to 3-7, and temperature and pool size tuned on results.
- Reported gains over baselines are large on some tasks, and baseline implementations and variance across runs are not detailed.
Experimental setup
Datasets: CommonsenseQA StrategyQA Letter Coin MultiArith AQuA
Benchmarks: CommonsenseQA StrategyQA Letter Coin MultiArith AQuA
Metrics: exact match accuracy
Baselines: Zero-Shot-CoT Auto-CoT Manual-CoT
Builds on: Auto-CoT BDPL Zero-Shot-CoT Manual-CoT
Closest papers by meaning
- CoT Vectors: Transferring and Probing the Reasoning Mechanisms of LLMs (ICLR 2026 · similarity 0.91)
- To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning (ICLR 2025 · similarity 0.89)
- Clustering as Reasoning: A $k$-Means Interpretation of Chain-of-Thought Graph Learning (ICML 2026 · similarity 0.89)
- Enhancing Chain of Thought Prompting in Large Language Models via Reasoning Patterns (AAAI 2025 · similarity 0.89)
- CoT Information: Improved Sample Complexity under Chain-of-Thought Supervision (NeurIPS 2025 · similarity 0.89)
- Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought (NeurIPS 2024 · similarity 0.89)
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.