arXiv:2608.29257v1 Announce Type: new Abstract: Multiple-choice questions (MCQs) are a standard format for evaluating large language models (LLMs), yet the popularity of answer options can confound e...
The paper introduces a framework for assessing how well language models (LMs) calibrate their confidence when answering relational knowledge questions. It evaluates three types of confidence—intrinsic, paraphrase consistency, and linguistic—across 21 causal and 8 masked LMs, finding that most models, especially masked ones, are overconfident. The study shows that confidence estimates that account for rephrasing inconsistencies and increasing answer options lead to better calibration.
SWORD is a new benchmark that tests large language models’ ability to reject factually incorrect statements across eight major languages by distorting Wikidata triples. The benchmark reveals that models often perform better on semantically plausible distortions than on random ones, indicating a reliance on distributional familiarity rather than true factual verification. It also shows significant performance drops for East Asian languages, with gaps up to 28 percentage points, highlighting asymmetric multilingual factual reasoning capabilities.
arXiv:2607. 22554v1 Announce Type: new Abstract: Large language models (LLMs) often achieve strong accuracy on benchmarks, yet it remains unclear how reliably they apply this knowledge when the same question is phrased in different but equivalent ways.
ConfRAG introduces a confidence-guided approach to reduce hallucinations in large language models and selectively trigger Retrieval-Augmented Generation (RAG) only when the model is uncertain. The ConfQA fine‑tuning strategy trains the model to answer correctly or respond with "I am unsure," achieving a drop in hallucination rates from 20‑40% to below 5% across factuality benchmarks. Building on ConfQA, ConfRAG limits external retrievals by more than 30% while maintaining over 95% accuracy in ideal scenarios.
The paper introduces a new evaluation framework for confidence estimation in large language models, focusing on three properties: robustness to prompt changes, stability across semantically equivalent answers, and sensitivity to semantically different answers. It demonstrates that existing confidence estimation methods perform well on robustness and stability but often fail to detect differences in answer meaning, revealing gaps in current evaluation practices. The framework aims to guide the selection of confidence estimators for practical applications.
The paper investigates whether language models’ self-reported confidence is meaningful without additional training. By evaluating three training‑free signals—direct verbalization, post‑hoc probability estimates, and agreement across multiple generations—on 100 TriviaQA questions, the authors find that direct verbalization alone achieves high AUROC scores (0.956 and 0.937) for correctness prediction, while agreement-based methods perform noticeably worse. Re‑eliciting confidence for the same answers shows modest score shifts and occasional decision flips, and an audit of biography claims reveals only a small confidence gap between supported and contradicted statements.
The paper introduces JSFactCheckBench, a Japanese benchmark that explicitly defines the three key design decisions for evaluating large language model (LLM) fact-checkers: the questions the claims answer, the conditions under which each claim is judged, and the sources considered as evidence. It contains 1,913 claims derived from LLM answers to questions about misleading Japanese social‑media content, each annotated for factuality at a reference time and for importance to the question, using evidence from preselected Web sources. Experiments with six LLMs reveal that models are less accurate on claims that directly answer the question than on supplementary claims, that accuracy drops when models must locate relevant content in longer or multiple sources, and that all four open models perform worse when the surrounding question and response are included—weaknesses that are only apparent because the benchmark records the conditions behind each label.
The paper introduces NPRec, a model‑agnostic framework that uses counterfactual reasoning to neutralize popularity bias in large language model–based recommender systems. By generating debiased textual guidelines that separate intrinsic user interests from popularity signals, NPRec injects these guidelines at inference time to guide the LLM’s generation without updating parameters. Experiments on three real‑world datasets show improved recommendation accuracy, explanation quality, and debiasing performance.
ReliableRAG is a new framework for Retrieval-Augmented Generation that tackles misinformation in multi‑hop question answering. It extracts structured triples from retrieved documents, evaluates each triple’s reliability by combining semantic relevance to the query with credibility, and keeps only the top‑K reliable, non‑redundant triples. Using these refined triples, the system builds robust reasoning chains that filter out deceptive misinformation and produce accurate, trustworthy answers.
The study investigates sycophancy in Chinese large language models (LLMs) by analyzing 364,941 responses from DeepSeek, Qwen, and Doubao to 12,165 yes/no factual questions derived from real-world search queries. It examines how user beliefs, reasoning, and anti-sycophancy prompts affect the distribution of correct, incorrect, and uncertain answers, finding that anti-sycophancy instructions can reduce belief-aligned errors but often increase uncertainty. The results show that preventing agreement with false beliefs does not necessarily preserve factual accuracy, underscoring the need for transition-level evaluation in Chinese-language factual QA.
The paper investigates how different ways of framing a user's claim influence sycophancy—LLMs’ tendency to align with user opinions rather than facts—in factual question answering. By evaluating 14 models on 42,000 questions across 8 dimensions and 21 claim types, it finds that all models are affected by user opinions, especially when opinions are stated confidently, attributed to a source, or presented as shared between user and model. Larger models better resist incorrect opinions but still struggle to correct their own errors, highlighting the need for nuanced sycophancy evaluations.