Research map / Large language models / Reasoning and chain of thought
Learning to Reason and Memorize with Self-Notes
Jack Lanchantin, Shubham Toshniwal, Jason E Weston, Arthur Szlam, Sainbayar Sukhbaatar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Self-Notes instead of Vanilla | GPT-2 · Algorithmic 101-200 | accuracy (%): 24.4 -> 85.0 | Improved |
| Self-Notes instead of Scratchpad | GPT-2 · Algorithmic 101-200 | accuracy (%): 11.6 -> 85.0 | Improved |
| Self-Notes instead of Vanilla | GPT-2 · Toy-Story 4-hop | accuracy (%): 37.4 -> 97.8 | Improved |
| Self-Notes instead of Scratchpad | GPT-2 · Toy-Story 4-hop | accuracy (%): 94.2 -> 97.8 | Improved |
| Self-Notes instead of Scratchpad | GPT-2 · Boolean Variable 9-19 | accuracy (%): 73.7 -> 75.2 | Improved |
| Self-Notes instead of Scratchpad | GPT-2 · Chess Piecetype >=81 | accuracy (%): 94.7 -> 94.8 | No clear effect |
| Scratchpad instead of Vanilla | GPT-2 · Chess Move >=81 | accuracy (%): 39.8 -> 29.9 | Hurt |
| Self-Notes instead of Vanilla | GPT-2 · Chess Move >=81 | accuracy (%): 39.8 -> 41.8 | Improved |
| Self-Notes instead of CoT few-shot | Llama 2 70B · GSM8K | accuracy (%): 55.9 -> 59.9 | Improved |
| Self-Notes instead of CoT few-shot | GPT-J 6B · MultiArith | accuracy (%): 13.5 -> 21.0 | Improved |
| Self-Notes instead of CoT few-shot | GPT-3 175B · GSM8K | accuracy (%): 11.4 -> 13.4 | Improved |
| add boost questions, multi-sample and finetune to unsupervised Self-Notes (ablation) | Vanilla GPT-2 · Toy-Story 4-hop | accuracy (%): 57.9 -> 85.8 | Improved |
Closest papers by meaning
- SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures (NeurIPS 2024 · similarity 0.88)
- SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.87)
- Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models (ICLR 2024 · similarity 0.87)
- Neural Garbage Collection: Learning to Forget while Learning to Reason (NeurIPS 2026 · similarity 0.87)
- State of Thought Enables Endogenous Reasoning (NeurIPS 2026 · similarity 0.86)
- Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.86)
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.
