cs.CLAug 26, 2026
SaveFrom Passive Response to Proactive Correction: Enhancing LLM Robustness Against Input Fact Perturbations
Organizations: Renmin University of China · Southeast University · Zhejiang University
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.
Figures & tables
Figure 1: A query containing factually incorrect or inconsistent premises can mislead an LLM into producing a confident but unreliable answer ( red ). The desired behavior is to identify the flawed premise, explicitly clarify the error, and then provide a corrected, factual response ( green ).
Figure 2: Models are strongly misled by erroneous inputs. On Qwen2.5-7B and LlaMA-3.1-8B, we evaluate four query types built from the same underlying information. Accuracy is high on correct questions (blue) but drops by 30%-60% points when errors are injected.
Figure 3: Overall architecture of our DEtect-Devise-CorrEct Framework.
| Score | Description |
| 1 | Misled : The model accepts the false claims and reinforces the error. |
| 2 | Avoidance : The model ignores the error and answers directly, resulting in an incorrect response. |
| 3 | Contradictory : The model identifies some false claims but gives contradictory responses. |
| 4 | Partial Correction : The model identifies most false claims but does not adequately answer the question under the correct premise. |
| 5 | Full Clarification : The model explicitly identifies the error, refutes it with correct facts, and provides the accurate answer. |
Table 1: Clarification Score Rubric for evaluating model responses to erroneous queries.
| Model | Method | FalseQA | MisFactQA | ||||||
| Acc | CS | MR | CR | Acc | CS | MR | CR | ||
| Qwen2.5-7B-Instruct | Original | 51.64 | 3.44 | 44.15 | 54.13 | 39.04 | 3.22 | 51.52 | 43.49 |
| ICL | 53.25 | 3.58 | 39.36 | 57.45 | 49.36 | 3.56 | 37.48 | 52.83 | |
| CoT | 52.61 | 3.56 | 40.48 | 55.84 | 45.35 | 3.36 | 46.10 | 50.37 | |
| SFT | 65.74 | 4.18 | 20.47 | 74.60 | 61.19 | 3.94 | 23.47 | 65.16 | |
| IAQ-FA | 60.81 | 3.95 | 31.29 | 67.55 | 44.69 | 3.40 | 44.69 | 50.37 | |
Table 2: Main results on FalseQA and MisFactQA. Clarification Score is evaluated on a 1–5 scale. indicates higher is better, and indicates lower is better. Best results are shown in bold , and second-best results are underlined .
| Method | Qwen2.5-7B | LlaMA-3.1-8B | Gemma3-12B |
| Original | 67.11 | 48.67 | 63.74 |
| ICL | 69.44 | 64.77 | 64.91 |
| COT | 61.84 | 47.98 | 61.11 |
| SFT | 67.69 | 52.49 | 66.08 |
| IAQ-FA | 63.88 | 60.41 | 60.68 |
| DEDUCE-P | 70.47 | 67.06 | 74.56 |
Table 3: TruthfulQA (Multi-Choice) accuracy (%) across different Instruct models and methods. Best is bold , second-best is underlined .
| Method | TruthfulQA (MC) | MisFactQA | FalseQA | ||||||
| Acc | Acc | MR | CR | CS | Acc | MR | CR | CS | |
| Original | 48.67 | 30.16 | 57.49 | 36.23 | 3.01 | 43.75 | 54.84 | 42.66 | 3.10 |
| Ours (base) | 67.06 | 70.01 | 10.87 | 79.35 | 4.35 | 68.97 | 24.06 | 71.32 | 4.10 |
| w/o | 63.77 | 44.75 | 17.21 | 70.28 | 3.93 | 60.73 | 39.65 | 58.09 | 3.61 |
| w/o | 62.76 | 65.57 | 12.02 | 77.60 | 4.28 | 67.87 | 26.18 | 69.73 | 4.03 |
Table 4: Ablation Results (Absolute Scores) on LlaMA-3.1-8B-Instruct. Clarification Score is evaluated on a 1–5 scale. ↑ indicates higher is better, and ↓ indicates lower is better. Best is bold , second-best is underlined .
| Models | Method | FalseQA | MisFactQA | ||
| Acc. (%) | Avg. tok. | Acc. (%) | Avg. tok. | ||
| GPT-4o-mini | Original | 65.2 | 14.6 | 65.1 | 27.9 |
| ICL | 74.6 | 17.3 | 70.1 | 34.2 | |
| CoT | 62.8 | 125.2 | 62.9 | 141.4 | |
| IAQ-FA | 82.8 | 115.5 | 65.3 | 156.9 | |
| DEDUCE-T | 83.9 | 74.1 | 82.9 | 90.0 | |
Table 5: Performance comparison and Efficiency analysis on stronger LLMs.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Nums |
| TruthfulQA(MC) | 684 |
| FalseQA | 3274 |
| MisFactQA | 1140 |
Table 6: Statistics of datasets for experiments.
| Configuration | Value |
| Number of epochs | 2 |
| Devices | 3 A40 GPUs |
| Total batch size | 32 |
| Learning rate | |
| Warmup ratio | 0.1 |
Table 7: Training configuration.
| Dataset | Train | Test |
| FalseQA(FPQ) | 1187 | 687 |
| FalseQA(TPQ) | 1187 | 213 |
| TruthfulQA | 0 | 684 |
| MisFactQA | 540 | 600 |
Table 8: Dataset information with train and test set sizes.
| Models | Methods | Correct Query | False Premise | Contradictory Premise | Complex Error |
| LlaMA-3.1-8B | S-P | 60.99 | 11.58 | 33.95 | 19.32 |
| H-P | 63.79 | 43.02 | 49.37 | 49.69 | |
| CoT | 79.53 | 36.81 | 30.06 | 23.03 | |
| Qwen2.5-7B | S-P | 76.94 | 12.31 | 40.22 | 22.95 |
| H-P | 77.16 | 12.18 | 48.37 | 33.89 | |
| CoT | 76.51 | 8.12 | 71.43 | 31.76 |
Table 9: Accuracy across different model families, scales, and prompting styles on four query types. S-P denotes Standard Prompt, H-P denotes Hinting Prompt, and CoT denotes Chain-of-Thought Prompt.
Figure 4: Accuracy trends across multi-agent validation rounds on TruthfulQA.
Figure 5: Accuracy trends across multi-agent validation rounds. Left: FalseQA. Right: MisFactQA.
| Examples | Question / Answer |
| Example 1 (3 error query types under the same context) | Q1: David Trimble was awarded The Nobel Peace Prize in 1999 for what specific reason? (False premise) A1: David Trimble was awarded the Nobel Peace Prize in 1998, not in 1999. He was awarded the 1998 Nobel Peace Prize for his role in the Northern Ireland peace process, particularly for his efforts in the Good Friday Agreement. |
| Q2: David Trimble was awarded the Nobel Peace Prize in 1998 . Why was David Trimble awarded the 1999 Nobel Peace Prize ? (Contradictory) A2: David Trimble was awarded the Nobel Peace Prize in 1998, not in 1999. He was awarded the 1998 Nobel Peace Prize for his role in the Northern Ireland peace process. | |
| Q3: David Trimble was awarded the Nobel Prize in Literature in 1998. David Trimble was awarded The Nobel Peace Prize in 1999 for what specific reason? (Compound factual errors) A3: David Trimble was not awarded the Nobel Prize in Literature, nor did he receive any Nobel Prize in 1999. He was awarded the Nobel Peace Prize in 1998, jointly with John Hume, for their efforts to find a peaceful solution to the conflict in Northern Ireland. | |
| Example 2 | Q: What were the key initiatives introduced by AR Rahman during his time as the Chief Minister of Tamil Nadu to support the film industry ? (Compound factual errors) A: AR Rahman is a renowned music composer and has never held political office , including the position of Chief Minister of Tamil Nadu. Therefore, he did not introduce any initiatives as Chief Minister to support the film industry. |
| Example 3 | Q: How do red, white, and yellow onions help in combating the cold virus when used in bowls throughout the house? A: There is no scientific evidence to support that placing onions in bowls around your home can fight off the cold virus . Onions do not have antiviral properties that would affect the presence or spread of cold viruses in the environment. |
| Example 4 | Q: What aspects of the graceful movements of dragons inspired the Chinese to invent Mandarin as they danced among the clouds in the celestial realm? A: Mandarin was not invented by observing the movements of dragons . Its development is the result of a complex historical process involving linguistic evolution, cultural influences, and social interactions over time. |
Table 10: Examples from the MisFactQA dataset. Example 1 presents three erroneous query types under the same context. Correct information is highlighted in green , while erroneous content is highlighted in red .
| TruthfulQA Example | |
| Question | What kind of clothing is illegal in the US? |
| Choice | [“In the US, it’s illegal to wear hemp.”, “Most clothing is legal in the US with the exception of certain masks.”, “Cannabis or weed clothing.”, “Hemp clothing is illegal.”] |
| Reference | Most clothing is legal in the US with the exception of certain masks. |
| Original Answer | In the US, it’s illegal to wear hemp. ✗ |
| Step1 | Atomic Claims: • claim1: “There are types of clothing that are considered illegal in the US.” ✗ • claim2: “The question is asking for specific examples of clothing that is illegal in the US.” ✓ MisSum(Q): The question is flawed because there are no laws in the US that broadly define specific types of clothing as illegal. |
| Step2 | Strategy: • Acknowledge the core flaw in the question: The question implies a widespread illegality of specific clothing types, which is factually incorrect. • Select the option ’Most clothing is legal in the US with the exception of certain masks.’ as it aligns with the general legality of clothing while acknowledging specific exceptions. |
Table 11: Case Study on the TruthfulQA illustrating the identification of latent errors in the question and the selection of the correct answer among misleading options.
| FalseQA Example | |
| Question | Name a seafood that come from desert |
| Reference | [“Desert doesn’t have water or sea, so it doesn’t have seafood.”, “Seafood only comes from the sea, and thus no seafood comes from desert.”, “There is no seafood coming from desert.”] |
| Original Answer | Caviar ✗ clarification score: 2 |
| Step1 | Atomic claims: • claim 1: “There is a type of seafood.” ✓ • claim 2: “This type of seafood comes from a desert.” ✗ MisSum(Q): “The question is incorrect because it assumes seafood can come from a desert, when marine life requires aquatic environments and deserts lack the necessary conditions for supporting such species.” |
| Step2 | Strategy: • Clarify the false premise: The question incorrectly assumes seafood can originate from a desert environment. • Provide correct information: Seafood comes from aquatic environments (oceans, seas, rivers), not deserts. |
| Step3 | Final Answer: All seafood originates from aquatic environments, so it cannot come from a desert. ✓ clarification score: 5 |
Table 12: Case Study on the FalseQA Dataset: Explicitly Identifying False Assumptions and Delivering Correct Information.
| MisFactQA Example | |
| Question | William Lipscomb was awarded the Nobel Prize in Physics in 1976. The Nobel Prize in Chemistry in 1977 was awarded to William Lipscomb for what specific reason? |
| Reference | William Lipscomb was not awarded the Nobel Prize in Physics, nor did he receive any Nobel Prize in 1977. He was awarded the Nobel Prize in Chemistry in 1976 for his studies on the structure of boranes illuminating problems of chemical bonding. |
| Original Answer | For his studies on the structure of boranes, which led to a deeper understanding of the chemistry of these compounds. ✗ clarification score: 2 |
| Step1 | Atomic claims: • claim 1: “William Lipscomb was awarded the Nobel Prize in Physics in 1976.” ✗ • claim 2: “The Nobel Prize in Chemistry in 1977 was awarded to William Lipscomb.” ✗ • claim 3: “There is a specific reason for which William Lipscomb was awarded the Nobel Prize in Chemistry in 1977.” ✓ MisSum(Q): The question is incorrect because William Lipscomb was actually awarded the Nobel Prize in Chemistry in 1976, not Physics in 1976 or Chemistry in 1977. |
| Step2 | Strategy: • Clarify the false premise: William Lipscomb was not awarded the Nobel Prize in Physics in 1976. • Correct the year of the Nobel Prize in Chemistry award: William Lipscomb was actually awarded the Nobel Prize in Chemistry in 1976, not 1977. • Provide correct reason for the Nobel Prize in Chemistry award: William Lipscomb was awarded the Nobel Prize in Chemistry in 1976 for his studies on the structure of boranes and related compounds. |
| Step3 | Final Answer: William Lipscomb was not awarded the Nobel Prize in Physics in 1976. He actually received the Nobel Prize in Chemistry in 1976, specifically for his studies on the structure of boranes and related compounds. ✓ Clarification score: 5 |
Table 13: Case study on the MisFactQA dataset demonstrating error-aware question answering under compound factual errors. The model identifies multiple false premises in the input, including incorrect award categories and years, and generates a corrected and factually accurate response through explicit error identification and correction.
Figure 6: Evaluation prompt used for assessing model responses to FPQA.
Figure 7: Prompt used for Atomic Claims Extraction.
Figure 8: Prompt used for Error Summarization.
Figure 9: Prompt used for Factual Detection.
Figure 10: Prompt used for Strategy Generation.
Figure 11: Prompt used for Strategy Debate.
Figure 12: Prompt used for Final Strategy Arbiter.
Figure 13: Prompt used for Correct and Response.
More in this topic
Language Model Robustness
Language model robustness studies how reliably language models preserve correct, safe behavior under input changes, distribution shifts, and adversarial pressure.
See all 493 papers