We build an ensemble of 10 transformer encoders for the MedExACT 2026 shared task on medical decision span detection. The ensemble is diversified along three training directions: encoder initialization (including domain-adaptive pre-training on clinical text), loss function, and data augmentation with LLM-generated synthetic notes and silver-labeled clinical documents. Greedy forward search selects the combination with the highest validation final score. A BERT-based boundary refiner is applied to the ensemble’s predicted spans to correct offset errors before submission.

While reasoning-enhanced large language models perform strongly on English medical tasks, a persistent multilingual gap remains, with substantially weaker reasoning in local languages, limiting equitable global medical deployment. To bridge this gap, we introduce Med-CoReasoner, a language-informed co-reasoning framework that elicits parallel English and local-language reasoning, abstracts them into structured concepts, and integrates local clinical knowledge into an English logical scaffold via concept-level alignment and retrieval. This design combines the structural robustness of English reasoning with the practice-grounded expertise encoded in local languages. To evaluate multilingual medical reasoning beyond multiple-choice settings, we construct MultiMed-X, a benchmark covering seven languages with expert-annotated long-form question answering and natural language inference tasks, comprising 350 instances per language. Experiments across three benchmarks show that Med-CoReasoner improves multilingual reasoning performance by an average of 5%, with particularly substantial gains in low-resource languages. Moreover, model distillation and expert evaluation analysis further confirm that Med-CoReasoner produces clinically sound and culturally grounded reasoning traces.

Thai is written without spaces between words, and its spelling depends on tone marks, silent consonants, and the transliteration of many loanwords. Each of these is a common source of misspellings. We present ThaiEditBench, a benchmark for correcting Thai spelling errors at the character level. It has three components: (i) a six-class taxonomy that types each error by the part of the spelling that goes wrong, plus an etymology tag; (ii) three length tiers (sentence, paragraph, and page) that test whether a model stays accurate over long, mostly correct text; and (iii) a contamination-robust construction, in which naturally attested misspellings are injected into clean text, so the corrupted passages do not appear verbatim in the source corpora. Because the errors are injected, the gold corrections are exact, and a blind two-annotator study measures how reliably the taxonomy can be applied. Grading is at the level of Thai character clusters, with span-based matching adapted from ERRANT. We evaluate 16 models and find that (1) under the benchmark’s conditions, a model that locates an error almost always fixes it correctly, so what separates models is detection; (2) the evaluated Thai-adapted models do not lead the strongest general models; and (3) in long, mostly correct passages, weaker models accumulate over-edits that pull their scores far below their sentence-level results. On real errors mined from Wikipedia edit history, scores are somewhat lower, but the model ranking largely holds.

Multilingual agentic benchmarks remain rare, especially for under-resourced languages such as Thai, where nicknames, unspaced script, and fused honorifics change how an agent must query its tools. We present FahMai-TeleBench, a bilingual (Thai/English) benchmark for grounded tool use. A model answers employee-directory questions over a synthetic company directory by issuing tool calls, or refuses when the answer is absent. All 626 items are synthetic but modeled on a real Thai enterprise-directory deployment, and they are graded deterministically without an LLM judge. Our central design choice is to treat tool availability as an experimental variable: each item runs under four tool configurations (grep, structured search, both, and a Python REPL), so the benchmark measures not just whether a model answers but how it uses tools. The task is near-solved at the frontier (GPT-5.4 97–99%), yet sharply discriminating for the Thai-capable models an organization would host itself (Typhoon-2.5 58–74%, Typhoon-S-8B 25–37%). The Thai penalty is model-specific, and Typhoon-2.5 answers 90% of items correctly once the relevant records are placed in its context, so its failures are failures of tool use. The tool axis also exposes differences at the frontier: GPT-5.5 drops to 73% with a wide-schema structured search tool while scoring 97–98% with grep and the REPL.

We present VeReaFine, a novel “Verifier-RAG” pipeline designed to eliminate hallucinations in open-ended clinical question answering. VeReaFine interleaves three tightly coupled stages—retrieval, verification, and generation—across up to three iterations. First, a two-stage dense retriever (BM-Retriever-410M → BM-Reranker-2B) fetches and ranks top-k biomedical passages; an 8B-parameter MedReason verifier then filters these for direct relevance and identifies missing evidence. When the verifier deems the context insufficient, it formulates a focused “feedback query” to retrieve additional passages (bounded to prevent infinite loops). Once a minimal ground-truth context is assembled, a 7B-parameter generator (Qwen2.5-7B-Instruct) drafts an answer purely from that vetted context, and the verifier performs a final check—prompting the generator to refine any remaining unsupported claims. By iteratively fetching only missing facts and ensuring every assertion is evidence-backed, VeReaFine achieves monotonic factuality improvements with minimal overhead. On the BioNLP 2025 ClinIQLink “LLM Lie-Detector” shared task, our 7B generator augmented with VeReaFine matches or surpasses a 32B medical model on open-ended reasoning metrics, reducing multi-hop inverse step-identification errors by 26%. These findings demonstrate that moderate-size LLMs, when guided by targeted verification loops, can deliver expert-level reliability in clinical QA.

Despite Southeast Asia’s (SEA) extraordinary linguistic and cultural diversity, the region remains significantly underrepresented in vision-language (VL) research, resulting in AI models that inadequately capture SEA cultural nuances. To fill this gap, we present SEA-VL, an open-source initiative dedicated to developing culturally relevant high-quality datasets for SEA languages. By involving contributors from SEA countries, SEA-VL ensures better cultural relevance and diversity, fostering greater inclusivity of underrepresented languages and cultural depictions in VL research. Our methodology employed three approaches: community-driven crowdsourcing with SEA contributors, automated image crawling, and synthetic image generation. We evaluated each method’s effectiveness in capturing cultural relevance. We found that image crawling achieves approximately ~85% cultural relevance while being more cost- and time-efficient than crowdsourcing, whereas synthetic image generation failed to accurately reflect SEA cultural nuances and contexts. Collectively, we gathered 1.28 million SEA culturally relevant images, more than 50 times larger than other existing datasets. This work bridges the representation gap in SEA, establishes a foundation for developing culturally aware AI systems for this region, and provides a replicable framework for addressing representation gaps in other underrepresented regions.

Machine translation (MT) in the medical domain plays a pivotal role in enhancing healthcare quality and disseminating medical knowledge. Despite advancements in English-Thai MT technology, common MT approaches often underperform in the medical field due to their inability to precisely translate medical terminologies. Our research prioritizes not merely improving translation accuracy but also maintaining medical terminology in English within the translated text through code-switched (CS) translation. We developed a method to produce CS medical translation data, fine-tuned a CS translation model with this data, and evaluated its performance against strong baselines, such as Google Neural Machine Translation (NMT) and GPT-3.5/GPT-4. Our model demonstrated competitive performance in automatic metrics and was highly favored in human preference evaluations. Our evaluation result also shows that medical professionals significantly prefer CS translations that maintain critical English terms accurately, even if it slightly compromises fluency. Our code and test set are publicly available https://github.com/preceptorai-org/NLLB_CS_EM_NLP2024.

Radiology report generation (RRG) aims to create free-text radiology reports from clinical imaging. Our solution employs a lightweight multimodal language model (MLLM) enhanced with a two-stage post-processing strategy, utilizing a Large Language Model (LLM) to boost diagnostic accuracy and ensure patient safety. We introduce the “First, Do No Harm” SafetyNet, which incorporates Xraydar, an advanced X-ray classification model, to cross-verify the model outputs and specifically address false negatives from the MLLM. This comprehensive approach combines the efficiency of lightweight models with the robustness of thorough post-processing techniques, offering a reliable solution for radiology report generation. Our system achieved fourth place on the F1-Radgraph metric for findings generation in the Radiology Report Generation Shared Task (RRG24).