Tag
This paper proposes a counterfactually anchored evidence attribution method for identifying safety failures in multi-turn LLM conversations, achieving high detection performance and low false positives.
This paper evaluates entropy-based confidence signals for small language models, finding that semantic entropy provides a viable method to improve accuracy by up to 50 percentage points through selective routing to larger models.
The paper introduces NepLEGiT, a specialized small language model pre-trained from scratch on Nepali legal text to enhance legal knowledge accessibility and service delivery in Nepal.
This paper introduces a corpus of 10,000 annotated Bangla sentences for function classification and benchmarks various models, achieving 0.95 accuracy with a double-level ensemble using TF-IDF features.
The paper presents 5-Dialects-BN, a manually annotated benchmark dataset for five Bangla dialects with aligned transliterations and translations, designed to improve evaluation of dialect-aware LLMs.
This paper investigates the curse of multilinguality in lexical normalization, finding that training a single model on multiple languages leads to decreased per-language accuracy, with optimal performance when languages are trained in small groups.
This paper proposes attribute-based activation steering to tailor LLM explanations to specific groups, achieving better specificity and factuality compared to prompting and state-of-the-art baselines.
This survey identifies three critical gaps in explainable AI for Arabic NLP—method, task, and linguistic—and proposes a taxonomy and research agenda for linguistically grounded explanations.
This paper introduces grounded glossary generation for Classical Sanskrit, a task involving recovering Sanskrit phrases and producing translation-grounded meanings from sloka-translation pairs. It constructs a benchmark from Hindu texts and evaluates various AI models, finding that instruction fine-tuning improves performance, with morphological modeling identified as a key challenge.
This paper analyzes NLP conference papers from 2020 to 2025 to examine the relationship between reported GPU resources and scholarly impact, finding that while resources are associated with higher citations, they explain little of the variance in impact.
A discussion on the cost-effectiveness of embeddings versus LLMs, referencing a study called 'embedder's dilemma' that finds LLMs outperform embedding models at significantly higher cost.
SuTRA is a morphology-aware tokenization algorithm that preserves akshara indivisibility for Indic languages, reducing morphological shattering and achieving improvements in machine translation metrics over standard BPE methods.
This paper introduces TokEval, a framework for evaluating language model tokenizers using intrinsic metrics that correlate with downstream task performance.
This paper proves that there is no theoretical curse of multilinguality for embedding space structure, showing that the minimum dimensionality required grows only logarithmically with the number of languages, suggesting empirical issues stem from data and training conditions.
This paper presents an empirical study on sentiment classifier behavior with sarcastic and AI-paraphrased social text, revealing lower confidence on sarcasm, higher accuracy on AI paraphrases, and an abstention method that improves performance by handling low-confidence inputs.
This paper shows that repetition effects in language models depend on readout position: adjacent repetition boosts target probability, while displaced repetition produces an inverted-U curve. The finding challenges assumptions in cloze-style probing and is validated across multiple models and languages.
This arXiv paper proposes capability-sustaining emotional dialogue (CSED) as a longitudinal research paradigm for emotional support systems, arguing that current approaches focus on immediate relief and neglect long-term user capabilities. A literature audit shows most systems overlook longitudinal outcomes, suggesting a new agenda for data, models, evaluation, and governance.
A fine-grained study of narrative features in web-scale LLM pretraining data, introducing NarraBERT and NarraDolma to measure narrative patterns and their distribution across sources.
This paper introduces a method for detecting hallucinations in large language models by leveraging the confidence of the first generated token, requiring only a single decode step.