Gold-Guided Programmatic Distillation for Financial Reasoning over Hybrid Tables and Text

2026-07-16 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Gold-Guided Programmatic Distillation (GGPD) is an approach designed for financial question answering over hybrid tabular and textual data, addressing the challenge of large language models (LLMs) generating arithmetic errors in natural-language rationales. This method transfers reliable numerical reasoning from a large teacher model to a compact student by utilizing execution-verified Python programs as supervision, rather than error-prone textual rationales. GGPD guides teacher-side program synthesis with gold derivations, retaining only programs that execute correctly and yield the gold answer. It also incorporates an iterative recovery stage to process teacher-failed examples, allowing the student to integrate newly verified programs. Experiments on TAT-QA demonstrate its effectiveness, with a 7B student model achieving 87.00 EM / 87.18 F1 on the test set, significantly surpassing the 72B teacher's 78.46 EM and other LLM-based baselines like TAGOP and TAT-LLM.

Key takeaway

For Machine Learning Engineers developing financial reasoning systems, this programmatic distillation approach offers a path to deploy smaller, more accurate models. You can achieve superior numerical precision on hybrid data by using execution-verified Python programs for supervision, rather than relying on larger LLMs' error-prone natural language rationales. Consider integrating gold-guided program synthesis and iterative recovery to enhance your student models' performance and reliability, potentially reducing inference costs while improving accuracy.

Key insights

Execution-verified programmatic distillation enables smaller models to achieve superior, reliable numerical reasoning for financial QA compared to larger LLMs.

Principles

• Execution-verified programs provide reliable supervision.
• Gold derivations enhance program synthesis quality.
• Iterative recovery improves student model robustness.

Method

Transfer numerical reasoning from a teacher to a student using execution-verified Python programs. Guide synthesis with gold derivations, retaining correct programs. Implement an iterative recovery stage for failed examples.

In practice

• Train compact models for complex financial QA.
• Improve numerical accuracy in LLM-based systems.
• Reduce reliance on large, error-prone LLM rationales.

Topics

• Financial Reasoning
• Programmatic Distillation
• Large Language Models
• Hybrid Tables and Text
• Numerical Computation
• TAT-QA

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Pedagogically-Inspired Data Synthesis for Language Model Knowledge Distillation — Pedagogically-inspired knowledge distillation systematically improves student model performance and efficiency.
• Cycle-Consistent Neural Explanation of Formal Verification Certificates — Trained neural architectures can explain formal verification certificates more effectively and efficiently than general LLMs.
• Consensus as Privileged Context for Label-Free Self-Distillation — CANON utilizes token-level consensus as dense supervision for label-free LLM self-distillation, significantly boosting reasoning accuracy.
• Distribution Corrected Offline Data Distillation for Large Language Models — Offline distillation can be improved by correcting teacher-student distribution drift without costly online sampling.
• Characterize Then Distill: Mechanistic Reasoning in Large Output Spaces — Reasoning in large output spaces can be mechanistically characterized as a two-phase process: shortlisting followed by fine-grained reasoning.
• KDW at SemEval-2026 Task 12: Logic-Driven Distillation with Knowledge Graphs for Efficient Abductive Reasoning — KDW uses KG-driven knowledge distillation to enable efficient abductive reasoning in compact models, matching LLM performance at lower cost.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.