Natural Language Processing news and analysis on Web Pulse: 5534 curated articles tracking the latest Natural Language Processing developments, tools, and research, updated continuously from vetted sources.
Researchers introduce ADAGE, a language-agnostic pipeline for analogical reasoning evaluation that uses native-speaker curation and LLM-assisted generation to create translation-free benchmarks. Testing on Arabic, Amhari…
A new classification-guided large vision-language model (LVLM) framework for visual information extraction (VIE) from documents achieves 86.43% F1-score in zero-shot inference on a real-world bidding dataset with 16 cert…
Researchers propose Co-E, a training-free system that synchronizes graph and text working memory for multi-hop question answering, outperforming comparable open-backbone baselines on six benchmarks and rivaling larger or…
A new framework called SCEPTER (Single-Case Evidence-driven PubMed-To-rEcommendation Reasoner) reduces clinical literature search from 576 papers to 53 retained papers, 7 Pareto-optimal claims and 3 final recommendations…
A study from arXiv (2607.23175v1) presents the first comparative evaluation of training-free methods for aligning language models to user-specific toxicity sensitivities, finding that all methods reduce alignment error b…
Researchers at LingoIITGN have released IndicTalk, a multilingual code-mixed conversational corpus comprising over 13,28,604 multi-turn conversations across 18 language varieties covering 9 Indic languages, generated thr…
Researchers introduced PeopleSearchBench, an open-source benchmark for evaluating AI-powered people search platforms, and found that Lessie, a specialized AI people search agent, scored 65.2 overall, 18.5% higher than th…
A two-stage large language model pipeline using Gemini 2.5 Pro and Gemini 2.5 Flash surfaced 3,460 candidate documentation inconsistencies from 3,000 MIMIC-IV-Note discharge summaries, affecting 69.7% of admissions, acco…
A new study from arXiv (2607.22554v1) finds that large language models frequently change their answers when the same question is rephrased, with mismatch rates exceeding 23% across four benchmarks and 13 models. The rese…
Researchers propose an automated framework using Vision-Language Models to verify actions in cognitive remediation tasks for schizophrenia patients, aiming to replace subjective manual observation by clinicians. The syst…
Researchers have developed BHARATI, a set of SentencePiece BPE tokenizers for classical Indian languages, achieving a 90% reduction in sequence length compared to GPT-2 on a held-out test set of 490 IKS-domain sentences.…
A new benchmark and baseline study evaluating retrieval-augmented generation (RAG) for French immigration law finds that retrieval improves administrative guidance at both model scales tested. Comparing a parametric LLM …
AutoThinkSQL, a framework integrating an auto-thinking mechanism into Supervised Fine-Tuning and Direct Preference Optimization for Text-to-SQL, enables dynamic bypassing of reasoning for simple queries while invoking de…
Researchers propose CADER (Confidence-Aware Dynamic Evidence Reasoning), a training-free framework for adaptive long-video reasoning that uses a logit-margin signal to let high-confidence examples exit early and activate…
A new study from arXiv finds that chain-of-thought (CoT) prompting improves large language model performance primarily through local word co-occurrence activation rather than global reasoning order. Researchers observed …
Researchers introduced PatiGonit22K, a dataset of 22,441 Bengali mathematical word problems designed to advance natural language understanding and quantitative reasoning in low-resource languages. The dataset extends the…
Researchers released MedLoCoMo, a medical long-context benchmark for evaluating large language models on patient-specific clinical reasoning over multi-admission dialogues, built from deidentified MIMIC-IV records. The b…
Researchers combined predictions from a transcript-only large language model (LLM) and a supervised speech predictor to estimate interpersonal attraction in Japanese speed-dating conversations, finding that speech comple…
KT and Amorepacific jointly developed Lemon (Lab Efficiency Mode ON), an AI assistant for researchers, as part of a data integration project called Data Highway that reorganized 70 years of Amorepacific's R&D data into a…
Fetch built FAST, a self-serve conversational BI agent, on ClickHouse Cloud, which ingests over 150 billion rows a day from receipt and offer data. The platform lets non-technical teams query purchase data in plain langu…