Bangla (Bengali), spoken by over 300 million people worldwide, presents unique tokenization challenges due to its rich morphological structure. Despite its global importance, Bangla remains underrepresented in mainstream NLP, particularly in low-resource tokenization research. Existing tokenizers such as Byte Pair Encoding (BPE), Unigram, WordPiece, and character-level were originally developed for high-resource languages. These tokenizers struggle with morphologically rich languages and show two main limitations: (i) unnecessary splitting of morphemes and (ii) fragmentation of compound words. Based on these limitations, we hypothesize that morphology-aware tokenization is necessary for handling Bangla effectively. To address this, we first introduce a benchmark that systematically evaluates how widely used tokenizers perform on morphology-related tasks. Motivated by these findings, we then propose BANGLATOKENIZER, a morphology-aware tokenizer that alleviates these limitations using a novel bond-strength framework inspired by molecular structures. Empirically, BANGLATOKENIZER yields consistent downstream improvements, achieving absolute accuracy gains of 2.8%-3.4%, macro-F1 gains of 2.5%-3.6%, and perplexity reductions of up to 20% across diverse language models and tasks. These improvements are accompanied by 5%-10% higher inference throughput, showing that morphology-aware tokenization improves both effectiveness and efficiency

Language models achieve strong performance in medical question answering, yet hallucinations persist even in domain-specialized models, particularly on complex clinical vignettes involving rare presentations, multi-hop reasoning, and ambiguous evidence requiring explicit relational grounding. We introduce OntoDiagnose, a training-free, ontology-grounded, clinician-inspired reasoning pipeline that augments parametric knowledge with structured evidence from SNOMED CT. Given a clinical question, OntoDiagnose extracts salient medical entities, grounds them in the ontology, explores a bounded subgraph, and compresses relevant relations into compact, interpretable evidence chains that guide reasoning. We evaluate OntoDiagnose across USMLE-style and biomedical QA benchmarks using MedGemma, BioMistral, MedPhi, and OpenBioLLM under zero-shot, few-shot, and chain-of-thought settings. OntoDiagnose consistently improves over prompt-only baselines, with an average gain of 9.6 accuracy points on MedQA and the largest relative gains for the lowest-scoring models. These results show the benefit of incorporating ontology-grounded evidence into medical question answering without modifying model parameters.

Vision-language pretraining has driven progress in medical image representation learning, but it depends on paired image-text data and can inherit reporting bias from clinical narratives. We study whether language-free predictive pretraining can produce an image encoder that transfers effectively to radiology report generation. RadJEPA is a chest-X-ray adaptation of I-JEPA, pretrained on approximately 840K unlabeled radiographs using latent context-to-target prediction. Our primary contribution is an extensive empirical evaluation of this language-free encoder for report generation: the frozen image encoder is coupled to a trainable two-layer projector and language decoder, and is also substituted into four established vision-language backbones. Across MIMIC-CXR and IU-Xray, RadJEPA matches or exceeds the evaluated image-only and image-text baselines on lexical, entity-relation, and clinical-label metrics. Controlled MIMIC-only comparisons provide evidence that the predictive objective contributes beyond domain-specific pretraining, while broader comparisons also reflect differences in pretraining data, model capacity, and input resolution. Complementary classification and segmentation experiments assess transfer beyond report generation.

Large Language Models (LLMs) can generate content spanning ideological rhetoric to explicit instructions for violence. However, existing safety evaluations often rely on simplistic binary labels (safe/unsafe), overlooking the nuanced spectrum of risk these outputs pose. To address this, we present XGUARD, a benchmark and evaluation framework designed to assess the severity of extremist content generated by LLMs on a multi-level grading. It includes 3,840 red-teaming prompts generated using templates informed by real-world extremist scenarios from social media, forums, and news. The framework categorizes model responses into five danger levels (0–4) defined by degree of extremist endorsement, enabling nuanced analysis of failure frequency and severity. We introduce the interpretable Attack Severity Curve (ASC) to visualize vulnerabilities and compare defense mechanisms across threat intensities. Using XGUARD, we evaluate five popular LLMs and two lightweight defense strategies, revealing key insights into current safety gaps and trade-offs between robustness and expressive freedom. Our work underscores the value of graded safety metrics for building trustworthy LLMs. The code and dataset is available at https://github.com/Abishethvarman/XGUARD

Multi-agent debate among large language models (LLMs) is widely assumed to improve reasoning by letting models correct one another. We audit that assumption for English medical question answering (QA). Five open-weight 4 to 8B medical and generalist LLMs answer 8,402 questions across six benchmarks as a council, measured against the same five prompted individually and a single larger Qwen3-VL-32B, text-only. The council beats every member on five of six datasets, but most of that accuracy is in place before debate runs: 67% of verdicts settle at Round 0, and on five of six those are the more accurate questions. Where debate runs, opinion change concentrates on one anchor; on the hardest benchmark 85% of moves toward it happen while it is wrong. There, agreement is worth less than disagreement: questions all five agree on are correct 6.6% of the time against 14.6% otherwise, tracking a shared bias toward the first option. The council also trails the larger model on four of six by up to 16.2 points, at seven times the model calls for 1.6 times the compute. Agreement among models is a measurement that can be biased, not evidence of correctness.