arXiv cs.CL
·
18 hours ago
A study compared using English BERT models with translated non-English data against developing native-language BERT models across six NLP tasks in Bulgarian, Chinese, Dutch, Italian, and Russian. The translation-based approach matched or outperformed native models in 53.3 percent of cases, with best results in Question Answering, Part-of-Speech Tagging, and Natural Language Inference, but struggled with Named Entity Recognition and tasks requiring cultural understanding. The findings suggest translation-based fine-tuning is a computationally efficient alternative for extending NLP capabilities to low-resource languages, particularly for languages typologically similar to English and syntactic tasks.
arXiv cs.CL
·
18 hours ago
Researchers introduced LakeQuest, a benchmark containing 9,846 question-answer pairs designed to evaluate AI systems on retrieving and synthesizing information from realistic, messy data lakes spanning three domains. The benchmark contains questions paired with exact evidence pointers across AI/ML metadata, retail banking, and biomedical drug information datasets. Evaluations of retrieval-augmented generation and agentic systems showed that successful retrieval does not guarantee correct reasoning, with systems failing particularly on relation chaining, policy grounding, and joint tabular question answering.
arXiv cs.CL
·
18 hours ago
Researchers compared how blind and sighted people navigate semantic memory using a property listing task and computed entropy from natural language processing embeddings. Sighted individuals showed higher entropy for abstract concepts while blind participants showed higher entropy for visually salient concrete concepts like penguin. The findings indicate that visual experience shapes how people organize and retrieve conceptual knowledge.
arXiv cs.CL
·
18 hours ago
Researchers introduced TAKE, a text dataset distillation framework that reduces datasets to 0.1% of their original size while maintaining performance on NLP tasks by using influence functions to identify and weight the most informative training samples. The method achieved extreme compression—down to 20 samples per class—while preserving downstream task accuracy on text classification and natural language inference. This approach reduces the computational cost of training, fine-tuning, and continual learning on large text corpora.
arXiv cs.CL
·
18 hours ago
Researchers tested state-of-the-art large language models on Korean-Braille translation and found they produce poor, unstable outputs despite being multilingual and instruction-tuned. A small supervised fine-tuned T5 model significantly outperformed zero-shot and prompted LLM baselines across six evaluation metrics including SacreBLEU and ChrF++. The results show current LLMs lack Braille-aware tokenization and demonstrate that task-specific supervised training is more effective for accessibility-critical structured languages.
arXiv cs.AI
·
18 hours ago
Researchers introduced TheBioCollection, a 52.6-billion-token unified corpus that consolidates biological data from multiple sources including molecular databases, protein repositories, and genomic annotations into a single training dataset for biological language models. The corpus includes instruction tasks and a matched evaluation suite called TheBioCollection-Eval, showing that training with it more than doubled performance scores across molecular, protein, genomic, and cellular domains while preserving general language abilities. This organized dataset enables language models to develop deeper understanding of biological concepts that were previously scattered across incompatible formats.
arXiv cs.AI
·
18 hours ago
Researchers introduced operator-on-F, a diagnostic tool that measures planning-relevant errors in world models for reinforcement learning by comparing the model's latent rollouts to the environment on observable subsets. In a TD-MPC2 model size sweep on cheetah-run, operator error ranged from 0.28 to 2.62 while reward-prediction error remained narrowly clustered between 0.028 and 0.091, showing operator error had much stronger correlation with planning performance (rank correlation -0.90). This diagnostic complements existing value-equivalence checks by detecting world-model failures that reward prediction alone cannot identify.
arXiv cs.AI
·
18 hours ago
Researchers developed TIME (Temporally Informed Motion Embedding), a video representation learning approach that uses motion point-tracks with masked autoencoders instead of language-paired visual data. The method achieves performance comparable to state-of-the-art models while requiring 10,000 times less training data. This approach enables more efficient video models with improved temporal understanding without language-dependent training constraints.
