An arXiv preprint reports that DEDUCE methods scored higher than baseline methods across the evaluated models and datasets. For Gemma on FalseQA, the largest cited accuracy margin over the best baseline was 25.99%. The figures are point estimates, and confidence intervals were not provided.
The work is an arXiv version 1 dated 26 August 2026. It focuses on factual inaccuracies and contradictions; ambiguous queries and adversarially crafted prompts were not explored.
The framework's three stages
DEDUCE is organized into three stages: Detect, Devise and Correct. The stages identify input errors, formulate a correction strategy and produce a corrected answer.
The framework was implemented as DEDUCE-Prompting and DEDUCE-Tuning, both evaluated on factually perturbed inputs.
The evaluation reported accuracy, Misleading Rate, Correction Rate and Clarification Score. Clarification Score used a 1-to-5 scale, from a fully misled response to complete correction with an accurate answer.
The benchmark used Qwen, LLaMA and Gemma backbones on TruthfulQA(MC), FalseQA and MisFactQA, with reported counts of 684, 3,274 and 1,140, respectively.
Accuracy varied with the type of input error
For Qwen, reported accuracy was 75% on correct questions, 45% on contradictory-premise questions, 15% on false-premise questions and 25% on complex-error questions.
These are point estimates for the displayed error-analysis conditions, and confidence intervals were not reported. The figures describe the tested inputs and do not attribute the differences to a single error type.
Comparing versions with components omitted
On MisFactQA, the full method's reported accuracy score was 70.01, compared with 44.75 without Strategy(Q) and 65.57 without MisSum(Q).
The displayed gap from the full method was larger for the version without Strategy(Q) than for the version without MisSum(Q). This comparison does not isolate a causal contribution from either component.
The comparison included two other models
On MisFactQA, GPT-4o-mini scored 65.1% in its original form and 82.9% with DEDUCE. DeepSeek-V3 scored 69.9% originally and 84.6% with DEDUCE.
The authors also report a better accuracy-efficiency trade-off than chain-of-thought prompting, or CoT, and IAQ-FA. The model scores are point estimates; confidence intervals and repeated-run variation were not reported, and the study results are based on a single run.
Validation levelled off after a few rounds
The paper reports that performance stabilized after two to three generator-reviewer validation rounds on FalseQA and that optimal performance was typically reached within the first two rounds across tasks.
The report does not provide all corresponding figure values, so the precise trade-off between further checking and added inference cost remains unresolved.
Automated ratings were checked against people
A separate human evaluation used 100 randomly sampled instances to assess the reliability of the automated judge.
In that sample, agreement with human ratings was Cohen's kappa of 0.936 for accuracy and a Pearson correlation of 0.916 for clarification score. The measures indicate close alignment between the automated judge and human ratings in this sample.
What the benchmark leaves open
The reported results are based on a single run, so the study does not show how stable the scores would be across repeated runs.
The task scope is factual inaccuracies and contradictions; ambiguous queries and adversarially crafted prompts were not explored. The benchmark results therefore do not establish how DEDUCE would behave on those inputs.
The document gives two unreconciled descriptions of FalseQA's size: 2,365 FPQ-TPQ pairs in the dataset description and 3,274 items in its experimental count. It does not clarify whether the figures refer to different units, subsets or stages of the evaluation.
Paper data and sources
Original title: From Passive Response to Proactive Correction: Enhancing LLM Robustness Against Input Fact Perturbations
Authors: Ping Wang, Xiangguo Sun, Bingbing Xu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text