← News

ResearchSources: arXiv · 19 Nov 2026

HYPER plans retrieval around a provisional answer to improve medical question answering

A training-free planning framework from a new arXiv paper redirects retrieval-augmented generation toward decision-relevant evidence, reporting accuracy gains on medical benchmarks at the cost of extra model calls.

In plain words

When an AI system must pick between competing answers, such as a diagnosis, ordinary retrieval often returns text on the right topic but not the evidence needed to decide. HYPER first drafts a provisional answer, then plans retrieval around three roles: evidence supporting that answer, criteria separating the alternatives, and key question features. The paper reports this improves accuracy on medical benchmarks, using two extra model calls per question.

What was announced

The authors frame the problem as a mismatch between what retrieval returns and what a decision requires: when a task involves choosing among plausible alternatives, such as a diagnosis, topically relevant context may not be enough to drive the final choice. HYPER responds by using a provisional answer hypothesis as a planning anchor and decomposing retrieval into three evidence roles. Support seeks evidence for the provisional answer, Distinction retrieves criteria that separate competing alternatives, and Feature grounds salient question features needed for the decision. The framework is described as training-free and as shifting retrieval-augmented generation from topic-oriented to evidence-oriented retrieval.

On the benchmarks MedQA and MMLU-Med, where the authors say single-query retrieval performs worse than no-retrieval chain-of-thought prompting, HYPER reportedly achieves the best average accuracy among ten baselines, improving over Simple RAG by 5.9 and 3.6 points respectively. The paper states the gains transfer to the 14 subjects of MMLU-Pro, but that each question costs two additional LLM calls. Code is available on GitHub.

What it does not show

  • The evidence is the paper's abstract only; results are author-reported and not independently confirmed, and full experimental details are outside the supplied scope.

For security and IT teams

Access
Code for HYPER is available at the GitHub repository linked in the paper.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.