Natural Language Processing news and analysis on Web Pulse: 5440 curated articles tracking the latest Natural Language Processing developments, tools, and research, updated continuously from vetted sources.
Researchers introduced NepLEGiT, a 30-million-parameter Nepali legal small language model pre-trained from scratch on roughly 4 million tokens of Nepali legal text, reporting a cross-entropy loss of 0.5684, a perplexity …
A new training framework for Visual Question Answering (VQA) achieves 61.64% accuracy on the bias-sensitive VQA-CP v2 benchmark while maintaining 62.80% on the standard VQA v2 dataset, a generalization gap of 1.16%, acco…
Researchers proposed an Affect-Prototype-Conditioned Fusion (APCF) framework for open-vocabulary multimodal emotion recognition that handles incomplete and unsynchronized modal data, according to an arXiv paper (arXiv:26…
A new arXiv paper (2609.15992v1) introduces "inference networks," a graph-based framework for routing queries among multiple expert large language models to minimize expected inference cost under a target performance con…
Researchers introduced ORDER (Optimal Routing for Dynamic Evidence Retrieval), a query-conditioned Retrieval-Augmented Generation framework that jointly adapts indexing and retrieval to each incoming query, according to …
A new arXiv paper (2609.16454v1) finds that supervised fine-tuning (SFT) is not inherently biased toward mode collapse or its opposite, over-dispersion, in large language models. The authors derive a bias-variance decomp…
Researchers introduced MechReason, a benchmark of 12,000 question-answer pairs and 21,000 visual materials drawn from real mechanical engineering papers, to test multi-image, multi-hop reasoning in multimodal large langu…
Researchers introduced HintMiner, an automatic question-hints mining tool that extracts hints from web Q&A posts using MiningNet, a transformer-based neural encoder-decoder model with copying mechanisms trained via a sel…
A developer joining a mobile AI team outlines a first-PR protocol for keeping Android SpeechRecognizer and iOS SFSpeechRecognizer strictly on-device, arguing that both platform APIs can silently route audio to a vendor c…
Researchers investigated how specialist models trained solely on question-answer pairs, without explicit reasoning supervision, generate reasoning trajectories for domain expert distillation. The work examines what gover…
Researchers introduced Mind2Dialogue, a method for training human-aware language models by simulating user mental states, to address what the work calls a fundamental supervision gap in current datasets for LLM assistant…
A developer built a serverless semantic search engine that combines DynamoDB vector search, Amazon Bedrock embeddings, AWS Lambda, and API Gateway, with infrastructure provisioned through AWS CDK. The project uses Titan …
Google released Gemini 3.8 Live, a model code `gemini-3.8-live` designed as the default option for low-latency voice agent experiences and real-time dialogue, with an input token limit of 131,072 and an output token limi…
Nvidia has released Nemotron 3.5 ASR, a 600M-parameter open-weight speech-to-text model that recognizes 40 language locales in real time from a single checkpoint with built-in punctuation and capitalization restoration. …
Two GEO experiments indicate that credible third-party sources mentioning a brand may drive AI search visibility more than changes to a brand's own website content. The findings suggest owned content remains a foundation…
A developer built an Apify Actor that scrapes PR TIMES, Japan's largest press-release distribution platform, into LLM-ready JSON records containing title, company, timestamp, category, cleaned body text, images and sourc…
Researchers affiliated with The Ohio State University and Apple proposed DACA-GRPO, a plug-and-play enhancement to GRPO-style trainers for diffusion language models that adds Denoising Progress Scores and Stratified Mask…
Google announced in a Tuesday blog post that its AI technology now supports more than 300 languages spoken by 7 billion people, roughly 86% of the global population, according to James Manyika, SVP of research, labs, tec…
Google LLC launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, which the company calls its most advanced voice processing models to date, designed to cut the latency of voice-based AI agents. The models suppo…
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two new speech-to-speech models that mirror the shape of OpenAI's GPT-Live family. Developer Simon Willison used GPT-6 Astra Extra High to build a br…