Research map / Large language models / Agents, code and math
Large Language Models are Interpretable Learners
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 LLM-Symbolic Programs (LSPs), decision trees whose nodes are prompted LLM modules learned by divide-and-conquer rule summarization, and introduces the IL-Bench benchmark, outperforming prompt optimization methods and ProtoTree.
Problem
Learning predictive models that are both accurate and interpretable to humans, i.e., extracting natural-language and symbolic knowledge from labeled data for classification tasks that LLMs cannot solve zero-shot.
Approach
Use a minimal DSL (input, switch branching, LLM module) so that a tree of prompted LLM modules is learned top-down, with an LLM acting as learner that summarizes rules from each node's data subset, and error count selecting the node to expand.
Key results
- On IL-Bench-Vision, LSP (GPT-4V learner) reaches 95.67% mean accuracy vs. 84.33% for ProtoTree with Gemini-M.
- On IL-Bench-Language with GPT-3.5, LSP averages 77.29% vs. 72.40% for PromptAgent (the strongest prompt optimization baseline).
- Human raters reproduced LSP-discovered programs at 90.36% accuracy vs. 72.74% for ProtoTree programs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use structured LSP instead of the strongest prompt-optimization baseline | IL-Bench-Language · IL-Bench-Language | accuracy (mean): 4.89% | Helped |
| Use structured LSP instead of the strongest prompt-optimization baseline | IL-Bench-Vision · IL-Bench-Vision (Palworld) | accuracy (mean): 20.09% | Helped |
| Use GPT-4V as the LSP learner instead of Gemini-Vision | LSP · IL-Bench-Vision (Palworld) | accuracy (mean): 84.76 -> 95.67 | Helped |
| Use LSP instead of ProtoTree | IL-Bench-Vision (Palworld) · IL-Bench-Vision (Palworld) | accuracy (mean): 84.33 -> 95.67 | Helped |
| Use LSP program rules instead of ProtoTree visualizations for human predictions | Human raters · IL-Bench-Vision (Palworld, human evaluation) | accuracy (mean): 72.74 -> 90.36 | Helped |
| Use error-count node scoring instead of random node scoring (ablation) | LSP · DT-Hard | accuracy: 70.50 $\pm$ 11.01 -> 89.75 $\pm$ 1.25 | Helped |
| Use error-count node scoring instead of prediction-accuracy node scoring (ablation) | LSP · DT-Hard | accuracy: 80.33 $\pm$ 18.27 -> 89.75 $\pm$ 1.25 | Helped |
| Use error-count node scoring instead of random node scoring (ablation) | LSP · Waxwing | accuracy: 62.22 $\pm$ 4.78 -> 65.83 $\pm$ 4.17 | Helped |
| Use error-count node scoring instead of prediction-accuracy node scoring (ablation) | LSP · Waxwing | accuracy: 66.11 $\pm$ 7.86 -> 65.83 $\pm$ 4.17 | No clear effect |
| Use error-count node scoring instead of random node scoring (ablation) | LSP · Waterthrush | accuracy: 61.67 $\pm$ 1.36 -> 62.50 $\pm$ 0.83 | No clear effect |
| Use error-count node scoring instead of prediction-accuracy node scoring (ablation) | LSP · Waterthrush | accuracy: 54.44 $\pm$ 0.70 -> 62.50 $\pm$ 0.83 | Helped |
Surprising findings
- Direct application of existing discrete prompt optimization methods fails on complex conditional decision rules, e.g., OPRO scores 30-50% on some synthetic tasks.
- Error-count node scoring outperforms accuracy-based and random node scoring in the ablation.
Limitations
- Only a basic DSL for predictive tasks is explored; DSL design is not investigated broadly.
- Learned programs can be redundant because search does not penalize program complexity.
- Evaluated only with proprietary API LLMs; the human study uses a limited set of raters.
Experimental setup
Datasets: IL-Bench-Language (synthetic decision-tree tasks, CUB bird caption tasks) IL-Bench-Vision (Palworld creature images)
Benchmarks: IL-Bench-Language IL-Bench-Vision
Metrics: classification accuracy accuracy retention rate on OOD data human rater accuracy search and inference runtime
Baselines: Decision Tree ProtoTree APE OPRO APO PromptAgent
Builds on: neurosymbolic programming APE OPRO APO PromptAgent
Closest papers by meaning
- Concept Bottleneck Large Language Models (ICLR 2025 · similarity 0.88)
- Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference (AAAI 2025 · similarity 0.88)
- Training Language Models to Explain Their Own Computations (NeurIPS 2026 · similarity 0.87)
- Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention (AAAI 2024 · similarity 0.87)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.87)
- Unraveling the Influence of Training Data and Internal Structures in Large Language Models for Enhanced Explainability (Student Abstract) (AAAI 2025 · similarity 0.87)
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.