In disinformation datasets, narratives are often understood as recurring interpretive patterns that group texts under narrative labels. Recent work formalized narrative mining as inductively inferring narrative labels from corpora, but its evaluation stays tied to predefined taxonomies, a closed-world setting that cannot capture narratives absent from the reference labels. We introduce a three-tier evaluation framework for unsupervised narrative label generation: recovery (against a corpus’s own taxonomy), mining (against external label sets), and discovery (without predefined labels). Applying it, we compare clustering-based and graph-community-based pipelines across seven disinformation datasets, with human validation of discovery on two. The two families are complementary under automated metrics, but in a corpus with two prominent topics, clustering can reduce one topic to 2% of generated labels while graph-based pipelines stay balanced. Discovery validation also reveals many singletons (narrative labels derived from single claims, 30–62% of graph outputs), which clustering cannot produce. Annotators confirm many as recognizable disinformation narratives, suggesting that in open-world discovery the repetition assumed by narrative mining may be recognized outside the corpus, not within it. We release human-validated narrative candidate labels for the Climate Obstruction and PolyNarrative datasets to support taxonomy development and dataset extension.

Self-generated counterfactual explanations (SCEs) are minimally modified inputs minimality generated by large language models (LLMs) that flip their own predictions validity, offering a causally grounded approach to unraveling black-box LLM behavior. Yet extending them beyond English remains challenging: existing methods struggle to produce valid SCEs in non-dominant languages, and a persistent trade-off between validity and minimality undermines explanation quality. We introduce Macro, a preference alignment framework that applies Direct Preference Optimization (DPO) to multilingual SCE generation, using a composite scoring function to construct preference pairs that effectively translate the trade-off into measurable preference signals. Experiments across four LLMs and seven typologically diverse languages show that Macro improves validity by 12.55% on average over the chain-of-thought baseline without degrading minimality, while avoiding the severe minimality violations of the translation-based baseline. Compared to supervised fine-tuning, Macro achieves superior performance on both metrics, confirming that explicit preference optimization is essential for balancing this trade-off. Further analyses reveal that Macro increases cross-lingual perturbation alignment and mitigates common generation errors. Our results highlight preference optimization as a promising direction for enhancing multilingual model explanations.

As Natural Language Generation (NLG) dominates modern NLP, scalable evaluation remains a critical bottleneck. Consequently, LLM-as-a-judge (LaaJ) adoption has accelerated rapidly, appearing in more papers than human evaluation in 2025. This pivotal shift motivates a critical analysis of current evaluation practices. Overcoming the limits of rigid keyword filtering and manual review, we employ a multi-LLM information extraction pipeline to gather structured metadata from 14,171 papers across four major NLP conferences (2020–2025). Analyzing 3,334 filtered NLG papers, we identify three systemic challenges. (1) Metric inertia: despite the shift toward open-ended generation, legacy lexical metrics (BLEU, ROUGE) persist as primary indicators, typically used alongside rather than replaced by semantic alternatives. (2) Metric-criteria mapping problem: our paper-level co-occurrence data reveals that general-purpose automatic metrics are applied as broad proxies for quality, without specifying which dimension of text generation they are intended to evaluate. (3) Validation gap: LaaJ has grown rapidly without commensurate human validation (fewer than 8% of papers). Crucially, while LaaJ correlates with aggregate quality, alignment collapses on fine-grained criteria like fluency. To address these gaps, we distill our findings into a minimal Evaluation Checklist to guide metric selection, construct validity, and LaaJ deployment.

Large Language Models (LLMs) are increasingly used to detect misinformation and support human fact-checkers. However, their truthfulness assessments may be affected by systematic political biases. Motivated by confirmation bias in humans, i.e., interpreting evidence in ways that are partial to existing beliefs or expectations, we investigate whether LLMs assess truthfulness in line with their political attitudes. We compare LLMs’ political attitudinal bias (PB-A) with political bias in truthfulness perception (PB-T). We assess PB-T by prompting LLMs to evaluate Democratic and Republican claims, including conditions where the claim-making party is explicitly stated. Across 20 instruction-tuned LLMs, all models exhibit a left-leaning tendency in PB-T, statistically significant for roughly half of them. Moreover, 16 out of 20 models significantly factor in the claim maker’s party affiliation when predicting truthfulness. PB-T and PB-A correlate weakly to moderately, and larger models tend to exhibit more left-leaning PB-T. The code for reproducibility is available on GitHub.

For socially sensitive tasks like hate speech detection, the quality of explanations from Large Language Models (LLMs) is crucial for factors like user trust and model alignment. While Persona prompting (PP) is increasingly used as a way to steer model towards user-specific generation, its effect on model rationales remains underexplored. We investigate how LLM-generated rationales vary when conditioned on different simulated demographic personas. Using datasets annotated with word-level rationales, we measure agreement with human annotations from different demographic groups, and assess the impact of PP on model bias and human alignment. Our evaluation across three LLMs results reveals three key findings: (1) PP improving classification on the most subjective task (hate speech) but degrading rationale quality. (2) Simulated personas fail to align with their real-world demographic counterparts, and high inter-persona agreement shows models are resistant to significant steering. (3) Models exhibit consistent demographic biases and a strong tendency to over-flag content as harmful, regardless of PP. Our findings reveal a critical trade-off: while PP can improve classification in socially-sensitive tasks, it often comes at the cost of rationale quality and fails to mitigate underlying biases, urging caution in its application.

What activation patching localizes can depend on where the patch is applied. Reproducing Ahsan et al. (2025) and extending it from four to ten models, we find the behavioral results largely reproduce and that medical adaptation improves calibration consistently only for gender with BioMistral, for MedGemma not at all. Two mechanistic findings are artifacts of the patch site. In Gemma-2 and OLMo-2 the patched MLP down-projection passes through a scale-invariant RMSNorm before the residual stream, so activation scaling is ineffective there. Patching after normalization restores it, raising the Gemma-2 flip rate on multiple sclerosis from 0.83, flat under scaling, to 1.00. It also lets a single layer recover the target race nearly as well as the reported sliding window (0.83 vs. 0.84, OLMo-2), so the window compensated for the site, not a distributed representation.

Simulatability is an evaluation protocol for explanations that quantifies their usefulness by how well they help a user predict a task model’s outputs. Since human evaluation is costly, automated simulatability replaces human explainees with LLM simulators, as proposed in ConSim (Poché et al., 2025) for large-scale experiments. We qualitatively replicate and extend ConSim’s ranking of explanation methods across the tested datasets, explanation families, and simulator LLMs, and identify two limitations. First, when class names are meaningful, simulators can obtain high simulatability by solving the classification task directly, without relying on the explanations. Second, class anonymization can reward explanations for leaking the hidden label mapping, a limitation we expose with a new classes-as-concepts baseline. These results are consistent with a shortcut hypothesis: in the tested settings, simulator predictions mainly rely on task priors, while explanations produce small changes. We derive recommendations for more robust automated simulatability evaluations.

Free-text explanations are expressive and easy to understand, but many datasets lack annotated explanation data, making it challenging to train models for explainable predictions. To address this, we investigate how to use existing explanation datasets for self-rationalization and evaluate models’ out-of-distribution (OOD) performance. We fine-tune T5-Large and OLMo-7B models and assess the impact of fine-tuning data quality, the number of fine-tuning samples, and few-shot selection methods. The models are evaluated on 19 diverse OOD datasets across three tasks: natural language inference (NLI), fact-checking, and hallucination detection in abstractive summarization. For the generated explanation evaluation, we conduct a human study on 13 selected models and study its correlation with the Acceptability score (T5-11B) and three other LLM-based reference-free metrics. Human evaluation shows that the Acceptability score correlates most strongly with human judgments, demonstrating its effectiveness in evaluating free-text explanations. Our findings reveal: 1) few annotated examples effectively adapt models for OOD explanation generation; 2) compared to sample selection strategies, fine-tuning data source has a larger impact on OOD performance; and 3) models with higher label prediction accuracy tend to produce better explanations, as reflected by higher Acceptability scores.1

The overwhelming volume of content being published at any given moment poses a significant challenge for the design of automated fact-checking (AFC) systems on social media, requiring an emphasized consideration of efficiency aspects.As in other fields, systems built upon LLMs have achieved good results on different AFC benchmarks. However, the application of LLMs is accompanied by high resource requirements. The energy consumption of LLMs poses a significant challenge from an ecological perspective, while remaining a bottleneck in latency-sensitive scenarios like AFC within social media. Therefore, we propose a system built upon fine-tuned smaller BERT-based models. When evaluated on the ClimateCheck dataset against decoder-only LLMs, our best fine-tuned model outperforms Phi 4 and approaches Qwen3 14B in reasoning mode — while significantly reducing runtime per claim. Our findings demonstrate that small encoder-only models fine-tuned for specific tasks can still provide a substantive alternative to large decoder-only LLMs, especially in efficiency-concerned settings.

Given the limited computational and financial resources of news agencies, real-life usage of fact-checking systems requires fast response times. For this reason, our submission to the FEVER-8 claim verification shared task focuses on optimizing the efficiency of such pipelines built around subtasks such as evidence retrieval and veracity prediction. We propose the Semantic Filtering for Efficient Fact Checking (SFEFC) strategy, which is inspired by the FEVER-8 baseline and designed with the goal of reducing the number of LLM calls and other computationally expensive subroutines. Furthermore, we explore the reuse of cosine similarities initially calculated within a dense retrieval step to retrieve the top 10 most relevant evidence sentence sets. We use these sets for semantic filtering methods based on similarity scores and create filters for particularly hard classification labels “Not Enough Information” and “Conflicting Evidence/Cherrypicking” by identifying thresholds for potentially relevant information and the semantic variance within these sets. Compared to the parallelized FEVER-8 baseline, which takes 33.88 seconds on average to process a claim according to the FEVER-8 shared task leaderboard, our non-parallelized system remains competitive in regard to AVeriTeC retrieval scores while reducing the runtime to 7.01 seconds, achieving the fastest average runtime per claim.