Less Evidence, Better Answering:
Gain-Aware Minimal Evidence Subset Selection
for Medical QA
1 The Hong Kong University of Science and Technology (Guangzhou)
2 The Hong Kong University of Science and Technology
A compact evidence set. A better-grounded answer.
Answer quality
Average relative improvementEvidence F1
Average relative improvementCompact evidence selection
An explicit evidence-token constraintPaper-reported gains over the second-best baseline. Explore the results ↓
Demo
A walkthrough of our medical evidence-grounded question answering system.
Abstract
Evidence-grounded medical question answering requires not only accurate answers, but also precise, verifiable, and clinically trustworthy evidence. Existing medical RAG systems typically retrieve top-k chunks according to item-wise relevance. Although this design improves recall, it does not explicitly optimize whether the selected evidence set is compact, sufficient, and decision-critical.
We propose MEGA, a gain-aware framework for compact supporting medical evidence selection. MEGA expands clinical queries into complementary subqueries and performs hybrid retrieval to construct a recall-oriented candidate pool. Hidden-state Information Gain Scoring uses a frozen LLM to estimate whether each snippet contributes new answer-relevant information beyond surface relevance. An FPTAS selector then identifies a compact evidence subset under an explicit token budget. Across three LLM backbones and five medical RAG baselines on CRC-EvidenceQA and Med-EvidenceQA, the paper reports average relative improvements of 8.8% in answer quality and 38.8% in Evidence F1 over the second-best baseline. Ablation studies and blinded clinical expert evaluation further assess the framework.
Method
The pipeline combines subquery generation and hybrid retrieval, frozen-model Information Gain Scoring, budgeted FPTAS evidence selection, and evidence-grounded answer generation.
- 01
Retrieve broadly
Clinical subqueries and hybrid retrieval build a recall-oriented candidate pool.
- 02
Estimate utility
Four frozen-model hidden-state signals score the contribution of each snippet.
- 03
Select compactly
FPTAS balances additive utility and token cost under the evidence budget.
- 04
Answer with evidence
The answering model receives the selected supporting subset.
The hidden-state signals are evidence-utility proxies. The FPTAS approximation guarantee concerns the fixed additive selection objective; downstream answer quality and evidence grounding are evaluated empirically.
Results
Main comparisons across CRC-EvidenceQA and Med-EvidenceQA. Select an answer-generation backbone below.
| Method | CRC-EvidenceQA | PubMedQA | COVID-QA | ||||
|---|---|---|---|---|---|---|---|
| Low Q. ↑ | High Q. ↑ | Ev. F1 ↑ | Accuracy ↑ | Ev. F1 ↑ | ROUGE-L ↑ | Ev. F1 ↑ | |
Mean ± standard deviation over three repeated runs. CRC answer quality is on a 0–10 judge scale; the remaining metrics are percentages. Bold indicates MEGA; underlining marks the second-best mean per column.
Benchmarks
CRC-EvidenceQA contains 200 open-ended, clinician-collaborative colorectal cancer questions: 100 Chinese and 100 English, with 50 low-difficulty and 50 high-difficulty questions per language. We evaluate answer quality and Evidence F1.
Med-EvidenceQA contains 1,000 yes/no/maybe examples adapted from PubMedQA and 1,973 extractive examples adapted from COVID-QA. It supports task-specific answer evaluation and evidence selection evaluation.
Resources
Paper · Code · Data & benchmark documentation · Live demo · Demo video
Initial release: hidden-state scoring components, evaluation tools, and 2,973 Med-EvidenceQA evidence annotations (without source text). The final FPTAS pipeline and CRC-EvidenceQA release are pending. See the release status for version alignment and remaining work.
BibTeX
@inproceedings{guo2026mega,
title = {Less Evidence, Better Answering: Gain-Aware Minimal
Evidence Subset Selection for Medical QA},
author = {Guo, Songyue and Chen, Zhao and Cao, Caleb Chen and Chen, Lei},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026},
url = {https://openreview.net/forum?id=PsPjtMRFBs}
}