From Passive Response to Proactive Correction: Enhancing LLM Robustness Against Input Fact Perturbations
DEDUCE framework enhances LLM robustness against input fact perturbations, showing a 20% accuracy improvement in experiments.
Key Findings
Methodology
The DEDUCE framework consists of three stages: detect, devise, and correct. It first detects errors through fine-grained fact extraction and verification, then devises correction strategies via multi-perspective deliberation, and finally corrects misconceptions while delivering reliable answers. The method is evaluated using the MisFactQA dataset.
Key Results
- On the MisFactQA dataset, the DEDUCE framework improved the accuracy of the Qwen2.5-7B model from 39.04% to 61.24%, reducing the misleading rate from 51.52% to 18.23%.
- On FalseQA, DEDUCE-Tuning increased the accuracy of the Gemma3-12B model to 79.59%, significantly outperforming baseline methods.
- Consistent improvements in misleading and correction rates were observed across different models with the DEDUCE framework.
Significance
This research significantly enhances the robustness of LLMs against misleading inputs by transforming them from passive responders to proactive error correctors. It addresses the long-standing issue of models generating inaccurate answers and provides new directions for future research, especially in scenarios with low-quality user inputs.
Technical Contribution
The DEDUCE framework offers a novel approach to handling input fact perturbations, distinguishing itself from existing methods by not only detecting errors but also correcting them through multi-perspective strategy formulation. It demonstrates effectiveness and scalability across multiple datasets.
Novelty
DEDUCE is the first framework to transform LLMs from passive responders to proactive error correctors, significantly improving performance under misleading inputs. It provides a more comprehensive error detection and correction mechanism compared to traditional methods.
Limitations
- The DEDUCE framework may perform poorly when dealing with extremely complex input errors, especially those involving multiple layers of logical contradictions.
- The computational cost is relatively high, particularly during the multi-perspective strategy deliberation stage.
Future Work
Future research could explore reducing the computational cost of the DEDUCE framework and applying it to a broader range of language models. Additionally, developing more complex error detection mechanisms to handle more intricate input errors is an important direction.
AI Executive Summary
Large language models (LLMs) often produce confident yet factually incorrect responses when user inputs contain misleading premises. Existing hallucination mitigation methods typically assume reliable user inputs, overlooking how factual errors can actively mislead model reasoning. To address this vulnerability, this paper proposes DEDUCE, a three-stage framework that transforms LLMs from passive responders into proactive error correctors. DEDUCE operates in three stages: detecting errors through fine-grained fact extraction and verification, devising correction strategies via multi-perspective deliberation, and correcting 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. Although the DEDUCE framework may perform poorly when dealing with extremely complex input errors, its contribution to improving model robustness is significant. Future research could explore reducing computational costs and applying it to a broader range of language models.
Deep Analysis
Background
In recent years, large language models have demonstrated remarkable capabilities in encoding real-world knowledge and assisting with complex tasks. However, they remain vulnerable to hallucinations when handling misleading inputs, producing outputs that appear plausible yet are factually incorrect. Existing research primarily focuses on improving model capabilities, overlooking the impact of user input quality on model reasoning.
Core Problem
Large language models are often misled by inputs containing factual errors, generating inaccurate answers. This issue is particularly severe when user input quality is low, posing a challenge to model robustness.
Innovation
The DEDUCE framework transforms large language models from passive responders to proactive error correctors through a three-stage process. First, the detect stage identifies input errors through fine-grained fact extraction and verification. Second, the devise stage generates correction strategies through multi-perspective deliberation. Finally, the correct stage corrects misconceptions while providing reliable answers.
Methodology
- �� Detect Stage: Identifies errors in input through fine-grained fact extraction and verification.
- �� Devise Stage: Generates correction strategies through multi-perspective deliberation, ensuring comprehensive and reliable strategy formulation.
- �� Correct Stage: Executes the validated strategy to generate reliable responses.
Experiments
Experiments were conducted on TruthfulQA, FalseQA, and MisFactQA datasets, using Qwen, LLaMA, and Gemma model families for evaluation. Evaluation metrics included accuracy, misleading rate, and correction rate. Results show that the DEDUCE framework outperforms baseline methods across all metrics.
Results
On the MisFactQA dataset, the DEDUCE framework significantly improved the accuracy and correction rate of the Qwen2.5-7B model while reducing the misleading rate. On FalseQA, DEDUCE-Tuning achieved an accuracy of 79.59% for the Gemma3-12B model.
Applications
The DEDUCE framework can be directly applied to scenarios requiring handling of misleading inputs, such as intelligent customer service and automated question-answering systems. Its multi-perspective strategy deliberation mechanism enhances model robustness under complex inputs.
Limitations & Outlook
The DEDUCE framework may perform poorly when dealing with extremely complex input errors, especially those involving multiple layers of logical contradictions. Additionally, its computational cost is relatively high, particularly during the multi-perspective strategy deliberation stage.
Plain Language Accessible to non-experts
Imagine a kitchen where a chef needs to cook based on a recipe, but the recipe has errors. Traditional methods are like the chef following the recipe directly, possibly making mistakes. The DEDUCE framework is like a smart assistant who first checks the recipe, identifies errors, suggests improvements, and finally ensures the chef makes a delicious dish. It's like having a clever assistant in the kitchen to ensure every dish is perfect.
ELI14 Explained like you're 14
Imagine you're playing a game where you need to complete tasks based on clues, but the clues have errors. Traditional methods are like following the clues directly, which might lead to failure. The DEDUCE framework is like a smart friend who checks the clues, finds errors, and gives suggestions to ensure you complete the tasks. It's like having a smart friend in the game helping you win!
Glossary
Large Language Model (LLM)
A type of AI model capable of understanding and generating natural language, typically with billions of parameters.
In this paper, LLMs are used to process user inputs and generate responses.
Fact Perturbation
Misleading or incorrect information in the input that may cause the model to generate inaccurate answers.
This paper studies how to detect and correct fact perturbations in inputs.
Multi-Perspective Deliberation
A strategy formulation method that generates comprehensive correction strategies through deliberation from multiple perspectives.
Used in the DEDUCE framework to devise correction strategies.
Misleading Rate
The probability of the model being misled by inputs, reflecting its robustness under misleading inputs.
Used to evaluate model performance on the MisFactQA dataset.
Correction Rate
The model's ability to identify and correct input errors, reflecting its error detection capability.
Used to evaluate model performance on the MisFactQA dataset.
Open Questions Unanswered questions from this research
- 1 How to improve the efficiency of the DEDUCE framework without increasing computational costs?
- 2 How to handle complex input errors involving multiple layers of logical contradictions?
Applications
Immediate Applications
Intelligent Customer Service
The DEDUCE framework can improve the accuracy and robustness of intelligent customer service systems when handling misleading user inputs.
Long-term Vision
Automated Question-Answering Systems
The DEDUCE framework can be applied to automated question-answering systems to enhance robustness and accuracy under complex inputs.
Abstract
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.