---
title: "From Passive Response to Proactive Correction: DEDUCE Enhances LLM Robustness Against Input Fact Perturbations"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.25894v1"
permalink: "https://feed.astralor.com/items/itm_7e154ed693460246899c1dcb"
published: "2026-08-26T15:13:50.000Z"
score: 71
channel: "ai"
tags: ["DEDUCE","MisFactQA","Large Language Models","Hallucination Mitigation","Fact Perturbation"]
---

# From Passive Response to Proactive Correction: DEDUCE Enhances LLM Robustness Against Input Fact Perturbations

## AI commentary

> This paper deserves attention from AI product teams, because factual errors in user inputs are more common than expected, while most models simply follow the misleading premise. DEDUCE makes correction an explicit step rather than a post-hoc patch, which is especially valuable for customer service, search, and financial Q&A. If code and models are open-sourced later, the adoption barrier will be even lower.

## At a glance

Large language models often produce confident but factually incorrect answers when user inputs contain misleading premises, a vulnerability the authors attribute to fact perturbations. Existing hallucination mitigation largely assumes reliable inputs and ignores this failure mode. This paper proposes DEDUCE, a three-stage framework that turns LLMs from passive responders into proactive correctors: it detects errors via fine-grained fact extraction and verification, devises correction strategies through multi-perspective deliberation, and then corrects misconceptions while delivering trustworthy answers. The authors also introduce MisFactQA, a dataset with varying degrees of factual errors, and new metrics for robustness evaluation. Experiments on TruthfulQA, FalseQA, and MisFactQA show that DEDUCE consistently improves accuracy and error-correction capability across Qwen, LLaMA, and Gemma families, demonstrating effectiveness and scalability.

## Key points

- Identifies and defines 'input fact perturbation,' a blind spot in current hallucination mitigation.
- Proposes DEDUCE, a three-stage framework: detect errors, deliberate from multiple perspectives, and correct while delivering reliable answers.
- Releases MisFactQA dataset and new robustness evaluation metrics.
- Demonstrates significant and consistent gains across Qwen, LLaMA, and Gemma model families.

## Full text

Large language models (LLMs) frequently produce confident yet factually incorrect responses when user inputs contain misleading premises, a phenomenon we attribute to fact perturbations in the input. Existing approaches to hallucination mitigation typically assume reliable user inputs, overlooking how such factual errors can actively mislead model reasoning. To address this vulnerability, we propose DEDUCE, a three-stage framework that transforms LLMs from passive responders into proactive error correctors. DEDUCE operates in three stages: (1) detect errors through fine-grained fact extraction and verification; (2) devise correction strategies via multi perspective deliberation; and (3) correct misconceptions while delivering reliable answers. We also present MisFactQA, a dataset containing factual errors of varying degrees, and propose new metrics for evaluating model robustness. Experiments on TruthfulQA, FalseQA, and our MisFactQA benchmark demonstrate that DEDUCE significantly improves both accuracy and error correction capability. Consistent gains across Qwen, LLaMA, and Gemma families confirm its effectiveness and scalability.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_7e154ed693460246899c1dcb