Research map / Large language models / Reasoning and chain of thought
Self-Evaluation Guided Beam Search for Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 stepwise self-evaluation guided stochastic beam search to PAL (single chain) | Codex PAL · GSM8K | accuracy: 72.0 -> 80.2 | Helped |
| add self-evaluation guided beam search to multi-chain PAL SC | Codex PAL · GSM8K | accuracy: 80.4 -> 85.5 | Helped |
| add self-evaluation guided search when baseline is already strong | Codex PAL · Object Counting | accuracy: 96.7 -> 96.8 | No clear effect |
| add self-evaluation guidance to CoT on short commonsense chains | Codex CoT · Sports Understanding | accuracy: 98.5 -> 98.4 | No clear effect |
| add self-evaluation guidance to CoT | Codex CoT · StrategyQA | accuracy: 73.2 -> 77.2 | Helped |
| add self-evaluation guided search at equal token budget vs self-consistency | Llama-2 13B PAL · GSM8K | accuracy: 41.8 -> 46.1 | Helped |
| add self-evaluation guided search at equal budget on commonsense | Llama-2 13B CoT · StrategyQA | accuracy: 71.0 -> 70.6 | Hurt |
| use only correctness score (lambda=1) instead of combined score (ablation) | Ours-PAL · GSM8K | accuracy: 80.2 -> 74.5 | Hurt |
| use only LM probability (lambda=0) instead of combined score (ablation) | Ours-PAL · GSM8K | accuracy: 80.2 -> 77.1 | Hurt |
| apply guidance to longer reasoning chains (>=9 steps vs <7 steps) | Ours-PAL · GSM8K | accuracy: +9.78 vs +5.49 | Helped |
| apply guidance to StrategyQA chains of >=5 steps vs <4 steps | Ours-CoT · StrategyQA | accuracy: +3.42 vs +0.31 | Helped |
| larger beam size k (10) with majority voting on final beam | Ours Llama-2 · GSM8K | accuracy: 35.7 -> 42.8 | Helped |
Closest papers by meaning
- Provable Test-Time Scaling for Beam Search in LLM Reasoning (NeurIPS 2026 · similarity 0.92)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.89)
- SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures (NeurIPS 2024 · similarity 0.89)
- ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning (ICML 2025 · similarity 0.88)
- Enhancing Reasoning Capabilities of LLMs via Principled Synthetic Logic Corpus (NeurIPS 2024 · similarity 0.88)
- A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · 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.