Research map / Large language models / Agents, code and math
LLM Priors for ERM over Programs
Shivam Singhal, Priyadarsi Mishra, Eran Malach, Tomer Galanti
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 |
|---|---|---|---|
| add tool use (code execution) to the loop (ablation) | LLM-PV with GPT-5-Thinking · Pattern1 (n=100) | test accuracy (%): 78.4 -> 100 | Improved |
| add tool use (code execution) to the loop (ablation) | LLM-PV with GPT-5-Thinking · Pattern2 (n=100) | test accuracy (%): 64.7 -> 100 | Improved |
| add tool use (code execution) to the loop (ablation) | LLM-PV with GPT-5-Thinking · Dyck-2 (n=100) | test accuracy (%): 79.9 -> 100 | Improved |
| add tool use (code execution) to the loop (ablation) | LLM-PV with GPT-5-Thinking · IsPrime+47 (n=100) | test accuracy (%): 64.0 -> 80.6 | Improved |
| add tool use (code execution) to the loop (ablation) | LLM-PV with GPT-5-Thinking · CycleCount (n=100) | test accuracy (%): 51.0 -> 96.9 | Improved |
| add tool use (ablation) | LLM-PV with GPT-5-Thinking · FP (n=100) | test accuracy (%): 100 -> 100 | No clear effect |
| add tools to direct in-context prediction | ICL with GPT-5-Thinking · FP (n=100) | test accuracy (%): 82.0 -> 99.0 | Improved |
| add tools to direct in-context prediction | ICL with GPT-5-Thinking · FHP (n=100) | test accuracy (%): 66.0 -> 88.0 | Improved |
| add tools to direct in-context prediction | ICL with GPT-5-Thinking · 10-P (n=100) | test accuracy (%): 47.0 -> 79.0 | Improved |
| replace tool-augmented ICL with tool-augmented LLM-PV | GPT-5-Thinking with tools · Pattern1 (n=100) | test accuracy (%): 51.0 -> 100 | Improved |
| fine-tune pretrained LM instead of training from scratch | 1B-scale LMs · Pattern1 (n=100) | test accuracy (%): 51.5 -> 61.5 | Improved |
| scale SGD training set from 10 to 100k examples | from-scratch LM · 10-P / CAP | test accuracy (%): test accuracy near 50% | No clear effect |
Closest papers by meaning
- Is Programming by Example Solved by LLMs? (NeurIPS 2024 · similarity 0.87)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.87)
- Online Prompt Selection for Program Synthesis (AAAI 2025 · similarity 0.86)
- How efficient is LLM-generated code? A rigorous & high-standard benchmark (ICLR 2025 · similarity 0.86)
- NICE Data Selection for Instruction Tuning in LLMs with Non-differentiable Evaluation Metric (ICML 2025 · similarity 0.86)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · 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.
