Source-linked AI summary
From Passive Response to Proactive Correction: Enhancing LLM Robustness Against Input Fact Perturbations
Ping Wang, Xiangguo Sun, Bingbing Xu, Guocong Li, Xiaofeng Meng
TL;DR
Factually erroneous user inputs can mislead LLM reasoning even when the models contain relevant knowledge, exposing a gap in hallucination mitigation. The paper proposes DEDUCE, which detects and verifies atomic errors, deliberates over correction strategies, and executes corrected responses. Across FalseQA, MisFactQA, and stronger-model evaluations, DEDUCE consistently improves robustness and accuracy while maintaining a better accuracy-efficiency trade-off than robustness-oriented baselines.
Problem
Existing hallucination research largely overlooks factual errors in user inputs, although false premises and contradictions can sharply mislead model reasoning.
Method
DEDUCE detects and verifies fine-grained factual errors, devises corrections through multi-perspective deliberation, and corrects misconceptions in the final response.
Results
DEDUCE consistently outperforms baselines across models and datasets, including a 25.99% accuracy improvement over the best baseline on FalseQA with Gemma3-12B.
Takeaways & Limitations
Robust LLMs can actively validate and correct unreliable user premises rather than passively inherit them.
Takeaways & Limitations
The evaluation centers on mid-scale Qwen2.5-7B, LlaMA-3.1-8B, and Gemma3-12B models, while broader generalizability remains to be examined.
Abstract
from arXiv · showhide
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.
1 Introduction
LLMs can hallucinate not only from model limitations but also from factual errors in user inputs that mislead reasoning. The paper introduces DEDUCE, MisFactQA, and fine-grained metrics to support proactive detection and correction.
- The paper frames robust behavior as identifying flawed premises, clarifying errors, and providing corrected factual responses rather than passively answering.
- Existing hallucination mitigation mainly targets pre-training, supervised fine-tuning, and inference, while assuming reliable user inputs.
- Factual errors in user queries, including incorrect assumptions and contradictions, can perturb LLM reasoning and induce hallucinations.
- Injecting a small number of false premises or contradictory descriptions can sharply reduce accuracy even when models possess correct underlying knowledge.Figure 2 reports accuracy drops of 30%-60% points when errors are injected.
- DEDUCE contributes a detect-devise-correct framework, the MisFactQA benchmark, and fine-grained metrics for robustness against misleading inputs.
2 Background and Problem Statement
The paper formalizes fact-perturbed question answering as robustness to inputs containing false premises, contradictions, or compound errors. Its objective is to detect and correct these errors before generating reliable answers.
- Fact-Perturbed Question Answering represents an input as a query and extracted factual units, including explicit statements and implicit assumptions.
- Fact perturbations include false premises, factual contradictions, and compound errors involving multiple erroneous units.
- When an input is factually perturbed, distorted intermediate reasoning can produce responses that inherit or amplify misinformation.
- The objective is to detect perturbed inputs, explicitly correct factual errors, and generate responses grounded in verified knowledge.
3 DEDUCE: DEtect-Devise-CorrEct
DEDUCE uses Detect, Devise, and Correct modules to turn misleading inputs into verified responses. It decomposes queries into atomic claims, deliberates over correction strategies, and executes a validated plan.
- Detect: The Detect module decomposes a query into minimal, independently verifiable atomic factual units.This exposes latent errors hidden in narratives that appear plausible.
- Detect: Each atomic unit is checked for factual accuracy and pairwise consistency, producing error sets that identify error types and locations.
- Devise: The Devise module uses Generator, Reviewer, and Arbiter roles to formulate a multi-perspective validated correction strategy.The reviewer evaluates completeness, accuracy, and reliability before the arbiter resolves deficiencies.
- Correct: The Correct module executes the validated strategy sequentially by identifying errors, justifying corrections, and producing a grounded response.
- Implementations: DEDUCE-Prompting deploys the three modules directly, whereas DEDUCE-Tuning internalizes their reasoning patterns into model parameters.
- Implementations: The tuning process trains detection first, then jointly fine-tunes strategy formulation and correction using filtered teacher-generated examples.
4 Benchmark
The benchmark evaluates model behavior on factually erroneous inputs using complementary metrics that capture susceptibility to misleading premises, error correction, and graded clarification quality. MisFactQA combines verified sources into query-response pairs requiring error identification, correction, and answering.
- Metrics: Misleading Rate measures the proportion of erroneous queries in which models adopt the user’s false premise.
- Metrics: Correction Rate measures the proportion of erroneous inputs where models identify most false or contradictory claims.
- Metrics: Clarification Score evaluates how effectively models identify and address query errors using a rubric for graded response quality.
- Metrics: Clarification Score values of 1–2 indicate complete misleading, whereas values of 4–5 indicate adequate error correction capability.
- MisFactQA: MisFactQA combines three verified sources to create query-response pairs requiring models to identify false information, provide corrections, and answer.
- MisFactQA: The benchmark includes single false premises, internal contradictions, and complex errors to test factual accuracy, consistency, and comprehensiveness.
5 Experiments
Experiments evaluate DEDUCE across benchmarks, baselines, model families, error types, components, deliberation depth, and inference cost. DEDUCE consistently improves robustness, accuracy, and correction of misleading inputs, while structured correction strategies and multi-perspective deliberation contribute substantially.
- 5.2 Overall Performance: DEDUCE consistently outperforms all baseline methods across models and datasets, achieving higher accuracy, the highest correction rate, and the lowest misleading rate.On Gemma3-12B, it improves accuracy over the best baseline by 25.99% on FalseQA.
- 5.2 Overall Performance: On TruthfulQA, CoT underperforms the original baseline in several cases because reasoning from undetected erroneous inputs amplifies errors.This result supports detecting fact perturbations before reasoning over the input.
- 5.2 Overall Performance: Across model families and scales, DEDUCE delivers significant and consistent performance gains, confirming effectiveness and scalability.The experiments use Qwen, Llama, and Gemma models across TruthfulQA, FalseQA, and MisFactQA.
- 5.3 Error Analysis: Accuracy drops sharply when errors enter otherwise correct questions, with false premises and complex errors generally causing larger declines than contradictions.Qwen2.5-7B falls from 75% to 45%, 15%, and 25%; Llama-3.1-8B-Instruct falls from 62% to 32%, 16%, and 19% for contradictory, false-premise, and complex-error queries.
- 5.4 Component Contribution Analysis: Both Detect and Devise improve performance, but removing the Strategy component causes the largest degradation.The ablation shows that identifying input errors alone is insufficient; tailored correction strategies are crucial for substantial gains.
- 5.6 Case Studies: Case studies attribute baseline failures to uncritical acceptance of erroneous input information, whereas DEDUCE decomposes claims and identifies invalid assumptions before answering.This process helps clarify misconceptions and prevent error propagation.
6 Related Work
Related work mainly mitigates hallucinations arising during model development or generation, while input-side factual errors and misconception correction receive less attention. DEDUCE addresses this gap by validating premises and response strategies before generation.
- Input Induced Hallucinations: Input-induced hallucinations arise when erroneous user information, including misconceptions and false premises, leads models to produce plausible but incorrect responses.This vulnerability is associated with models conforming to flawed queries.
- Self Correction Methods: Existing false-premise detection and self-correction methods often do not explicitly correct the user’s underlying misconception or validate premises before generation.DEDUCE applies critique to input analysis and response strategy rather than only verifying the final answer.
- Hallucination Mitigation Methods: Existing hallucination mitigation methods primarily target flawed training data, uncertainty during fine-tuning, or generation-time grounding and decoding.Examples include retrieval-augmented generation, faithful decoding, and uncertainty-based detection.
7 Conclusion
The paper identifies factual perturbations in user inputs as an overlooked source of LLM hallucinations and proposes DEDUCE to detect, devise, and correct them. It introduces MisFactQA and fine-grained metrics, with experiments validating effectiveness across benchmarks and model families.
- 7 Conclusion: Fact perturbations in user inputs can systematically induce hallucinations, motivating more vigilant models that validate and correct user premises.The conclusion shifts attention toward reliability under unreliable queries.
- 7 Conclusion: DEDUCE combines detection, strategy development, and correction to enhance robustness against factual errors in inputs.The framework is paired with MisFactQA and fine-grained evaluation metrics.
- 7 Conclusion: Experiments across multiple benchmarks and model families validate DEDUCE’s effectiveness.The conclusion reports the result at the paper level without narrowing it to one dataset or metric.
Limitations
The evaluation is limited to mid-scale models and to inputs containing factual inaccuracies or contradictions. Ambiguous queries and adversarially crafted prompts are outside the study’s scope.
- Model scale constraints: The experiments center on Qwen2.5-7B, LlaMA-3.1-8B, and Gemma3-12B, with broader model evaluation left for future work.The authors plan to test generalizability across a broader range of models and identify model-specific optimizations.
- Task scope focused on factual errors: The framework specifically targets factual inaccuracies and contradictions in user inputs.
- Task scope focused on factual errors: Ambiguous queries and adversarially crafted prompts are not explored.
Ethics Statement
The paper uses public datasets and a multi-stage prompting and training pipeline to detect, deliberate over, and correct factual errors. Its experimental materials include TruthfulQA, FalseQA, and MisFactQA, with models evaluated using automated judging.
- Ethics and evaluation: The study aims to improve input-error correction for high-stakes domains such as healthcare and law, while using adversarial argument generation only for robustness verification.
- Datasets: TruthfulQA-MC evaluates truthful answering under common human misconceptions and contains 684 questions across 38 categories.
- Datasets: FalseQA contains 2,365 false-premise/true-premise question pairs spanning eight error types and six question forms.
- Datasets: MisFactQA combines verified sources and constructs query-response pairs requiring error identification, correction, and answering.
- Framework and training: DEDUCE uses Detect, Devise, and Correct modules, with atomic claim extraction, factual verification, strategy deliberation, and final response generation.
- Dataset quality: MisFactQA examples undergo automatic verification for incorrect injected claims, accurate corrections, and natural phrasing, with weak examples revised or discarded.
C Additional Error Analysis
Additional experiments test whether the main trends generalize across model families, parameter scales, and prompting styles. Contradictory premises are generally easier to detect than semantically plausible false premises.
- Experimental design: The additional analysis uses 199 randomly sampled instances constructed from shared underlying knowledge across four question types.
- Experimental design: The experiments compare Qwen2.5-7B, LlaMA-3.1-8B, Qwen2.5-14B, and Gemma3-12B under standard, hinting, and chain-of-thought prompts.
- Error analysis: Contradictory premises are typically easier to detect than false premises because contradictions expose explicit inconsistencies.
D Further Analysis of Multi-Perspective Deliberation
The deliberation analysis examines how generator-reviewer interaction rounds affect strategy quality and accuracy. Performance generally improves through early rounds, then stabilizes, while excessive discussion can add noise.
- Deliberation depth: Accuracy consistently improves as deliberation depth increases, with Round 1 outperforming Round 0 across models.The analysis attributes this pattern to collaborative refinement of response strategies.
- Efficiency and convergence: Optimal performance is typically reached within the first two rounds, while FalseQA performance stabilizes after 2–3 rounds.The authors describe this as a balance between efficiency and quality.
- Efficiency and convergence: Additional rounds do not necessarily improve results because excessive discussion may introduce noise or exacerbate hallucinations.
- Figures and examples: Figure 4 reports accuracy trends across multi-agent validation rounds on TruthfulQA.
- Figures and examples: The example tables report original questions, reference answers, baseline outputs, and DEDUCE’s Detect, Devise, and Correct process.
F Evaluation Agreement
The evaluation framework combines structured factuality and clarification judgments with human validation of an LLM-based judge. Case studies and prompts illustrate error detection, strategy deliberation, and correction across the evaluated datasets.
- Evaluation Agreement: Cohen’s κ = 0.877 for accuracy and Pearson’s r = 0.890 for clarification score indicate strong agreement between two human annotators evaluating 100 sampled instances.The annotators independently assessed model responses using the criteria defined in Section 4.1.
- Case Studies: The case studies cover latent errors in TruthfulQA, false assumptions in FalseQA, and compound factual errors in MisFactQA.The MisFactQA example describes correcting erroneous award categories and years before generating an accurate response.
- Evaluation Criteria: The judge scores clarification from 1 to 5 according to whether predictions identify false claims, avoid reinforcing them, and explain the corrections accurately.Concise responses can receive the highest score when they correctly clarify false claims and remain factually accurate.
- Evaluation Criteria: Factual correctness is marked True only when a prediction matches the reference answer and contains no conflicting statements.Otherwise, the evaluation assigns False.
- DEDUCE Pipeline: DEDUCE’s prompts operationalize atomic claim extraction, factual detection, error summarization, strategy generation, strategy debate, arbitration, and final correction.These stages are represented by the supplied prompts for analysis, validation, and response generation.