From Residuals to Reasons: LLM-Guided Mechanism Inference from Tabular Data

2026-05-21

AI Summary

This work addresses the longstanding challenge in scientific applications of balancing predictive accuracy with mechanistic interpretability, as existing post-hoc explanation methods fail to uncover causal interactions among features. The authors propose the Multi-Agent Residual In-Context Learning (MARICL) framework, which leverages large language models to infer missing mechanistic knowledge from high-residual samples of a base model and generates explicit, generalizable correction formulas via textual gradient optimization. MARICL unifies prediction and mechanistic understanding, consistently outperforming base models across nine scientific, biomedical, and socioeconomic datasets. Notably, on Cell-Free Protein data, frozen correction formulas improve predictive performance by over 92% on new batches under the same protocol, with failure modes aligning with known biochemical mechanisms—demonstrating robust zero-shot transfer across batches.

Abstract

A persistent challenge in machine learning for scientific applications is jointly achieving prediction and understanding. Statistical models excel on structured data but operate as black boxes, while existing interpretability methods are largely inspective: they answer "which features matter?" but do not articulate how features interact or refine explanations iteratively alongside human understanding. Asking an LLM to predict the target directly forces it to search the entire output space; we instead anchor predictions with a base model and ask the LLM the narrower question of what that model is missing. We introduce Multi-Agent Residual In-Context Learning (MARICL), an agentic framework in which LLM agents analyze where a base-model fails, hypothesize missing structure from high-residual examples provided in context, and produce explicit correction terms refined through multi-turn textual gradient optimization. Across nine benchmarks spanning scientific, biomedical, socioeconomic, and synthetic settings, MARICL improves consistently over its base model on all datasets. To test whether these corrections reflect real structure or batch-specific noise, we freeze formulas learned on one experimental batch of the Cell-Free Protein dataset and apply them (with no retraining and no further LLM calls) to held-out batches. Within the same reagent protocol, the frozen formulas improve predictions in over 92% of cases; across a different protocol, they fail systematically. The success boundary aligns with the biochemistry, not the batch count; direct evidence of mechanistic generalization.

Problem

• interpretability
• mechanism inference
• tabular data
• scientific understanding
• black-box models

Innovation

• Mechanism Inference
• Residual Analysis
• LLM Agents
• In-Context Learning
• Textual Gradient Optimization