Research map / Large language models / Language, safety and interpretability
Human-like Few-Shot Learning via Bayesian Reasoning over Natural Language
Kevin Ellis
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 |
|---|---|---|---|
| Tune the prior to human judgments | Human-like Bayesian model · Logical concepts | R² | Improved |
| Use natural-language hypotheses instead of Python hypotheses (ablation) | Number Game model · Number Game | fit to human judgments | Improved |
| Remove the data-conditioned proposal distribution and sample hypotheses from the prior (ablation) | Number Game model · Number Game | fit to human judgments | Hurt |
| Replace importance sampling with deduplicating proposals (ablation) | Human-like Bayesian model · Number Game | fit to human judgments | No clear effect |
| Replace Code-Davinci-002 with LLaMA2 70B as the likelihood model (ablation) | Human-like Bayesian model · Number Game | fit to human judgments | No clear effect |
| Replace Code-Davinci-002 with LLaMA2 70B as the proposal distribution (ablation) | Human-like Bayesian model · Number Game | fit to human judgments | Mixed |
| Use both concept-type-specific prompts instead of the first-order prompt for every concept (ablation) | Human-like Bayesian model · Logical concepts | R²: .80 -> .81 | Improved |
| Use natural-language concepts with the learned prior rather than the BPL baseline | Logical concept learning · Logical concepts | fit to human judgments | Improved |
| Use Bayesian mechanics to reweight language-model proposals | GPT-4 concept proposals · Logical concepts | accuracy | Improved |
| Use the natural-language hypothesis space rather than the BPL baseline's fixed primitives | Logical concept learning · New majority-color and least-common-color concepts | learnability | Improved |
Closest papers by meaning
- Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement (ICLR 2024 · similarity 0.88)
- Hypothesis Search: Inductive Reasoning with Language Models (ICLR 2024 · similarity 0.88)
- Doing Experiments and Revising Rules with Natural Language and Probabilistic Reasoning (NeurIPS 2024 · similarity 0.87)
- Human-like Category Learning by Injecting Ecological Priors from Large Language Models into Neural Networks (ICML 2024 · similarity 0.87)
- Nonparametric Identification of Latent Concepts (ICML 2025 · similarity 0.86)
- Hypothesis, Verification, and Induction: Grounding Large Language Models with Self-Driven Skill Learning (AAAI 2024 · 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.
