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.

Social media fact-checking is increasingly challenged by evidence-level and aggregation-level conflicts, where erroneous evidence often mimics authoritative news sources. To support conflict verification task, we construct ContraNote, a real-world dataset derived from X Community Notes. ContraNote is curated to mirror the two aforementioned challenges, providing 33,686 posts for evaluating evidence-level conflict resolution, and 54,474 instances for evaluating aggregation-level evidence prioritization. This paper further proposes CoVer, a framework designed to resolve these contradictions through factual adjudication. Cover uses a three-stage pipeline consisting of evidence schema normalization, factual consensus and support verification, prioritizing definitive evidence over noise to prevent it from compromising the final verdict. Technical evaluations across multiple benchmarks show that CoVer outperforms state-of-the-art baselines, achieving an accuracy of 88.4% on CONFACT-HumC and 89.4% on CONFACT-ModC.

Evidence grounding makes generated content verifiable by linking statements to their sources. However, effective verifiability requires two properties that existing methods fail to jointly provide: *verification granularity*, the ability to localize fine-grained evidence for each statement, and *verification completeness*, the ability to capture both support and contradiction. Existing methods either operate at granularities too coarse for rapid inspection or are support-only. We formalize *polarity-aware evidence grounding*: linking each response sentence to its contextual evidence with a supporting, contradicting, or irrelevant label. We construct **eTracer-Bench**, a human-annotated benchmark with 7,905 supporting and contradicting sentence pairs across seven domains, covering both explicit and implicit contradictions. We further propose **eTracer**, a post-hoc framework whose context-aware signed scorer recovers the full evidence map in a single forward pass per response, with localization and polarity emerging jointly from one mechanism. **eTracer** achieves the best overall Citation F1 of **82.39**, driven by a Contradict F1 of **89.20**. A pre-registered within-subjects user study (n = 30) shows that polarity-aware grounding nearly doubles contradiction detection over the document-level baseline (**27.9% → 51.9%**, Cohen’s d = 0.82, p = 0.0003), and suggests a false-completeness bias under which support-only grounding fails to improve over this baseline (**23.4% vs. 27.9%**). Dataset and code: https://github.com/chubohao/eTracer.

ArchEHR-QA is a grounded question-answering (QA) task for electronic health records (EHRs) comprising four subtasks: (1) question rewriting, (2) evidence identification, (3) grounded answer generation, and (4) answer-evidence alignment. In this work, we present a modular pipeline centered on retrieval-augmented generation (RAG). For Subtask 1, RAG few-shot prompting outperformed both PEFT and prompt-only baselines on the development set; however, Claude few-shot proved substantially more robust on the test set, ranking 6th out of 13 participating teams (score: 26.94). For Subtask 2, a union ensemble of open-weight LLMs (GPT-OSS-120B and Qwen3-30B-A3B) achieved a 56.7 micro-F1, rivaling the proprietary Claude Opus 4.6 while demonstrating higher recall (53.6). For Subtask 3, our RAG few-shot approach using Claude Opus 4.5 achieved the 1st place out of 13 participating teams (score: 36.33). Finally, for Subtask 4, a zero-shot Claude Opus 4.6 configuration ranked 2nd out of 16 participating teams (score: 81.3).

LLM-based drug–drug interaction (DDI) assessment remains difficult to audit when predictions are not explicitly tied to evidence. While retrieval-augmented generation (RAG) improves grounding, predictions are not guaranteed to be entailed by retrieved items. We present CrossDDI, a verification-first framework that separates LLM-based evidence extraction from deterministic, LLM-free arbitration over DrugBank and PubMed, requiring positive predictions to be linked to explicit supporting evidence. Evaluated on 1,000 DDInter 2.0 pairs under a positive–unlabeled setting, CrossDDI achieves recall of 0.576–0.593 over confirmed positives with interaction prediction rates comparable to RAG, while reducing cross-backbone variation (0.018 vs. 0.066). Analysis identifies literature evidence acquisition and attribution as the primary bottleneck: PubMed retrieval covers only 40.5% of confirmed positives, and Path B-only evidence is substantially less reliable than structured evidence. These results suggest that verification-first architectures can improve traceability and backbone consistency, while broader and more reliable literature evidence is needed to extend coverage beyond structured sources.

Retrieval-augmented generation (RAG) reduces hallucination in large language models by grounding outputs in retrieved evidence, but it does not guarantee that the resulting citations support the associated claims. We present VERICITE, a framework for evaluating citation faithfulness in retrieval-augmented medical QA. Our system retrieves PubMed abstracts via the NCBI E-Utilities API, prompts LLMs to generate answers with inline citations, and verifies each citation at the sentence level using a DeBERTa-v3-large NLI model. We evaluate four LLMs on 500 BioASQ questions at retrieval depths of 3 and 5, with extended experiments up to k = 15 and an oracle setting with gold standard documents. Only 27?41% of citation pairs are supported at the sentence level at retrieval depths of 3 and 5, with support rates declining further at larger k. Under the oracle condition, answer quality improves, but citation faithfulness does not substantially improve, suggesting that generation-side citation behavior contributes substantially to unfaithful citations.

While document summarization with LLMs has enhanced access to textual information, concerns about the factual accuracy of these summaries persist (e.g., hallucination), especially in the medical domain. Tracing source evidence from which summaries are derived enables users to assess their accuracy, thereby alleviating this concern. In this paper, we introduce TracSum, a novel benchmark for traceable, aspect-based summarization, in which generated summaries are paired with sentence-level citations, enabling users to trace back to the original context. First, we annotate 500 medical abstracts for seven key medical aspects, yielding 3.5K summary-citations pairs. We then propose a fine-grained evaluation framework for this new task, designed to assess the completeness and consistency of generated content using four metrics. Finally, we introduce a summarization pipeline, Track-Then-Sum, which serves as a baseline method for comparison. In experiments, we evaluate both this baseline and a set of LLMs on TracSum, and conduct a human evaluation to assess the evaluation results. The findings demonstrate that TracSum can serve as an effective benchmark for traceable, aspect-based summarization tasks. We also observe that explicitly performing sentence-level tracking prior to summarization enhances generation accuracy, while incorporating the full context further improves summary completeness. Source code and dataset are available at https://github.com/chubohao/TracSum.