PRIME: Towards Verifiable Evidence-Grounded Reasoning in LLM Agents

Forum topic · 小凯 · 2026-10-19

Summary

PRIME (Precommitted Rules for Inference from Minimal Evidence) is a framework by Hamed Khosravi and Xiaoming Huo for making LLM agents produce verifiable, evidence-grounded conclusions when checking records against organizational requirements, such as whether a vulnerability report is ready for investigation. Before any checks run, PRIME fixes evidence routes specifying which checks may justify each conclusion. A separate ledger program records every check result, refuses conclusions the results do not justify, and directs the agent toward (or performs itself) additional checks that could support stronger conclusions. On 480 public-record tasks, a baseline agent shown only the requirements reported unjustified conclusions on 26.7% of tasks, while PRIME reported none. PRIME's guidance and added checks raised the average score of justified conclusions from 30.1 (refusal alone) to 59.2 out of 100. Under correct requirements, exact checks at known locations, and sufficient budget, PRIME provably reaches the strongest supported conclusion at stopping time. The paper is available as arXiv 2610.22426.

Paper Overview

Field: NLP Authors: Hamed Khosravi, Xiaoming Huo Published: 2026-10-19 arXiv: 2610.22426

Abstract

LLM agents are used to check records against an organization's requirements, for example whether a vulnerability report is ready for investigation. Requirements often allow several ways to be met, such as a severity rating with a curator-reviewed advisory or a published advisory link.

Agents fail in two typical ways:

Accepting a report after a required check has failed
Settling for a weaker conclusion without trying another path that could still succeed

PRIME (Precommitted Rules for Inference from Minimal Evidence) addresses this by:

Fixing evidence routes before checking, specifying how each conclusion may be justified
Using a separate program (ledger) that records each check result and refuses any conclusion the results do not justify
Pointing the agent to checks that could justify a stronger conclusion, or running them itself
Reporting only conclusions authorized by its rules and ledger

Results

With correct requirements, exact checks at known locations, and sufficient budget at stopping, completion reaches the strongest supported conclusion.

On 480 public-record tasks:

An agent shown the requirements reported unjustified conclusions on 26.7% of tasks; PRIME reported none
PRIME's directions and added checks raised the average score for justified conclusions from 30.1 (with refusal alone) to a maximum of 59.2 out of 100

--- *Auto-collected on 2026-10-19*