From Passive Response to Proactive Correction: Enhancing LLM Robustness Against Input Fact Perturbations

AI summary

Overview

• Research area: Natural Language Processing — hallucination mitigation and robustness of large language models (LLMs) against factually erroneous user input.
• Technical level: Intermediate.
• Scope: The paper identifies input-side fact perturbations as a distinct cause of hallucination, proposes a three-stage "Detect, Devise, Correct" framework called DEDUCE, and releases a new benchmark (MisFactQA) with new robustness metrics.

What This Paper Is About

Most hallucination research assumes the user's question is factually sound and focuses on making the model's output more truthful. This paper argues that real users frequently embed false premises, internal contradictions, or multiple stacked errors in their questions, and that these errors can mislead an LLM even when it holds the correct knowledge internally. The goal is to build models that proactively detect and correct the flawed premise in the input before answering, rather than answering the flawed question directly or quietly ignoring the error.

Key Contributions

1. A new framing of hallucination. The authors highlight that hallucinations arise not only from model fabrication but also from factual perturbations in user inputs that derail the model's reasoning.
2. The DEDUCE framework. A "Detect, Devise, Correct" pipeline that converts an LLM from a passive responder into a proactive error corrector, implemented in two forms: DEDUCE-Prompting (no training) and DEDUCE-Tuning (two-stage fine-tuning).
3. The MisFactQA benchmark. A dataset built to contain diverse and complex factual errors, spanning single false claims, internal contradictions, and compound errors, reported as 1,140 fact-perturbed questions plus 900 pairs of questions sharing the same underlying information but exhibiting different error types.
4. Fine-grained robustness metrics. Misleading Rate (MR), Correction Rate (CR), and a 1–5 Clarification Score (CS) rubric that distinguishes being misled, ignoring the error, giving contradictory answers, partially correcting, and fully clarifying.

Main Findings

• DEDUCE beats all baselines across models and datasets. On FalseQA and MisFactQA (Table 2), DEDUCE-P and DEDUCE-T achieve the highest accuracy and correction rate with the lowest misleading rate across Qwen2.5-7B-Instruct, LlaMA-3.1-8B-Instruct, and Gemma3-12B-Instruct.
• Large absolute gains. On Gemma3-12B, the authors report DEDUCE improving over the best baseline by 25.99% in accuracy on FalseQA. For Qwen2.5-7B-Instruct on FalseQA, DEDUCE-T reaches 77.84 accuracy versus 51.64 for the original model; on MisFactQA it reaches 64.65 versus 39.04.
• DEDUCE-T generally outperforms DEDUCE-P. For LlaMA-3.1-8B-Instruct on MisFactQA, DEDUCE-T reaches 70.60 accuracy, 4.36 CS, 10.64 MR, and 81.82 CR, compared with 70.01 / 4.35 / 10.87 / 79.35 for DEDUCE-P.
• Reasoning can amplify errors. On TruthfulQA (Table 3), Chain-of-Thought underperforms the original baseline in several cases — for example Qwen2.5-7B drops from 67.11 (Original) to 61.84 (CoT), and LlaMA-3.1-8B from 48.67 to 47.98 — which the authors read as evidence that reasoning over unrecognized flawed input propagates the error.
• Small injected errors cause large accuracy drops. In the error analysis on Qwen2.5-7B and LlaMA-3.1-8B, accuracy falls by 30%–60% points when errors are injected. Qwen2.5-7B declines from 75% to 45% (Contradictory), 15% (False Premise), and 25% (Complex Error); LlaMA-3.1-8B-Instruct declines from 62% to 32%, 16%, and 19% respectively.
• Error type matters. Contradictory premises present overt conflicts that models detect more easily, whereas false premises appear superficially plausible and are harder to identify.
• Both components contribute. Ablation on LlaMA-3.1-8B-Instruct (Table 4) shows removing the strategy component (Strategy(Q)) causes the largest drop — MisFactQA accuracy falls from 70.01 to 44.75 — while removing the diagnostic summary (MisSum(Q)) reduces it to 65.57. The authors conclude that detection alone is insufficient without tailored correction strategies.
• Stronger models remain vulnerable. On MisFactQA, GPT-4o-mini and DeepSeek-V3 reach only 65.1% and 69.9% accuracy without the framework; with DEDUCE-T these rise to 82.9% and 84.6%.
• Favorable accuracy-efficiency trade-off. DEDUCE-T uses more tokens than the Original and ICL baselines but fewer than CoT and IAQ-FA in most settings. For DeepSeek-V3 on MisFactQA, DEDUCE-T uses 100.8 average tokens at 84.6 accuracy, versus 216.4 tokens at 70.2 accuracy for CoT.
• Judge reliability was validated. The automated judge (o3-mini-medium) shows Cohen's κ = 0.936 agreement with human judgments on accuracy and Pearson's r = 0.916 correlation on the clarification score.
• Deeper deliberation helps challenging cases. The paper reports that as the number of discussion rounds increases, most models become more robust to misleading inputs, while a single round or direct generation suffices when the initial strategy is already high quality.

Methodology in Plain English

The framework has three stages.

Detect. Rather than reading the user's question as one block, the model breaks it into "atomic facts" — minimal, independently checkable claims, including implicit assumptions. Each atomic fact is checked for truthfulness, and every pair is checked for mutual consistency. Facts that fail become elements of a factual error set or a conflict pair set. If either set is non-empty, the input is flagged as perturbed. These findings are summarized into a diagnostic report that states whether errors exist, where they occur, and why they are problematic.

Devise. Because a single model can under-correct, over-correct, or introduce new inconsistencies, strategy formulation is split into three roles: a Generator that drafts a correction strategy from the diagnostic report, a Reviewer that critiques the draft on completeness, accuracy, and reliability, and an Arbiter that resolves disagreements and produces a validated final strategy. The Arbiter is only invoked when the Reviewer finds deficiencies. The authors argue this role separation reduces the self-reinforcing bias typical of single-model self-correction.

Correct. The model executes the validated strategy step by step: identify the errors explicitly, supply the correct facts with justification, then answer the original question under the corrected premises.

Two implementations exist. DEDUCE-Prompting uses these prompts directly with an existing model, requiring no training. DEDUCE-Tuning internalizes the behavior in two fine-tuning stages: a Detect stage trained on data generated by GPT-4o as teacher, keeping only fully correct teacher outputs; and a combined Devise-and-Correct stage trained on data generated by Qwen2.5-14B as teacher, keeping only examples with a clarify score of 5 and verified correctness.

Evaluation uses TruthfulQA (684 multiple-choice questions across 38 categories), FalseQA (2,365 false-premise/true-premise question pairs covering 8 error types and 6 question forms), and the new MisFactQA. MisFactQA was assembled from a subset of Prize, EchoMist, and a publicly available Hugging Face dataset (hallucinations-dpo), followed by filtering, reconstruction, and augmentation. Baselines are the Default model, ICL, CoT, LoRA fine-tuning, and IAQ-FA. Experiments ran via Ollama at temperature = 0.0.

Why This Matters

The work shifts attention from the output end of the LLM pipeline to the input end, where the authors argue a substantial and largely unaddressed source of hallucination lives. If models can be made to challenge flawed premises rather than inherit them, reliability improves without requiring users to be domain experts.

Real-world applications:

• Customer support and advisory chatbots, where users often describe problems with incorrect assumptions about how a product, policy, or system works.
• Healthcare and legal information systems, the two high-stakes domains explicitly named in the paper's ethics statement, where acting on a false premise can cause real harm.
• Educational tutoring, where a learner's misconception is itself the error that needs correcting before an answer is useful.
• Search and question-answering over domains such as finance, politics, and health — categories covered by TruthfulQA and the MisFactQA source material.

Industry relevance: the framework is deployable in two ways — a prompting-only version requiring no training, and a tuned version that reduces token cost relative to other robustness-oriented methods — which makes it practical for teams that already serve Qwen, Llama, or Gemma family models.

Future Directions

• Scaling the evaluation. The authors state that their evaluation centers on mid-scale models (Qwen2.5-7B, LlaMA-3.1-8B, Gemma3-12B) and that they plan to extend the analysis to a broader range of models to test generalizability and find model-specific optimizations.
• Beyond factual errors. The limitations section notes that ambiguous queries and adversarially crafted prompts are not explored, leaving other forms of unreliable input open.
• Cost and latency of deliberation. The paper reports token counts but not wall-clock latency; understanding where multi-round deliberation is worth its cost for harder queries is an open practical question.
• Generalizing the metric framework. The Misleading Rate, Correction Rate, and Clarification Score rubric was validated on these datasets; whether it transfers to other domains and languages is not reported.

Target Audience

Researchers and practitioners working on hallucination mitigation, LLM reliability, and evaluation benchmarks will benefit most, particularly those interested in input-side rather than output-side interventions. The paper is also useful for applied teams deploying instruction-tuned Qwen, Llama, or Gemma models who need models to handle non-expert users, and for benchmark builders interested in partial-credit evaluation rubrics that go beyond binary accuracy.

Authors’ 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.