Natural Language Processing news and analysis on Web Pulse: 5439 curated articles tracking the latest Natural Language Processing developments, tools, and research, updated continuously from vetted sources.
ChainDoRA, a weight-decomposed parameter-efficient fine-tuning framework that builds directional low-rank factors from a connected Tensor-Train chain, reached a seven-task average accuracy of 72.30% on commonsense reason…
Researchers released ufakzeka-1, a 151M-parameter (182M with embeddings) decoder-only Turkish language model pretrained from scratch on 13.5B tokens of openly licensed text and instruction-tuned for chat at a total cost …
Researchers introduced IntLawNER, a named entity recognition dataset and benchmark for international law covering 2,987 gold-annotated sentences and 8,094 entity spans drawn from International Court of Justice decisions,…
A new arXiv paper (2609.22620v1) introduces Multi-Split Boundary Decision (MSBD), a method that predicts multiple document boundaries within a page window in a single large language model call, reducing inference request…
Researchers proposed NSFT (Neural Sub-expert Fine-Tuning), a fine-grained parameter-efficient fine-tuning framework that refines Mixture-of-Experts (MoE) large language model adaptation from experts to sub-experts by dec…
A preregistered intervention on the Spider-Realistic text-to-SQL benchmark found that swapping the shipped single-database oracle for the benchmark's distilled multi-instance test suite raised a conformal abstention cert…
BreezeBlue's Breeze TTS 2 has become the highest-rated open-weight text-to-speech model on the Artificial Analysis Provider Voice Arena, scoring Elo 1215 and beating ElevenLabs' Eleven v3 at 1175 while hosting at $34 per…
Alibaba's Ovis team introduced Ovis-Embedding, an omni-modal embedding family that encodes text, image, video, and audio through a single shared multimodal backbone rather than separate modality towers. The model is posi…
Google made Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS generally available on the Gemini API on September 22, 2026, with promotional pricing that doubles on January 1, 2027 — an hour of Flash TTS audio rising fro…
A new learning roadmap breaks LLM inference into five layers — the call itself, transformer computation, hardware limits, optimizations, and serving engines — explaining that prefill builds the KV cache while decode gene…
A developer argues that AI meeting assistants have mastered transcription and summarization but still fail at the harder problem of turning conversations into owned tasks, decisions, and durable context. The post traces …
A developer at CrawlSpider described building an AI visibility tracker in Python that queries LLM APIs to check whether brands appear in model answers, and detailed how the naive approach breaks at scale. The simple loop…
A developer's pull request #466 was merged into Mike, an open-source legal AI platform with 4.3k GitHub stars used for document review, drafting, and legal research. The fix enforces document order across [[PAGE_BREAK]] …
A developer improved retrieval overlap on a legal-contract benchmark from 37.3% to 50.4% by retrieving at the sentence level and then returning the full paragraph each hit came from, using stela chunking, all-MiniLM-L6-v…
Cohere launched North Small Translate, an open machine translation model with 218B parameters (25B active) and a 16k context length that the company says outperforms other open MT models as well as Google Translate and D…
Researchers at the USC Viterbi School of Engineering and Stevens Institute of Technology developed an automated AI system that analyzes multilingual online job postings to detect deceptive recruitment tactics linked to h…
Gradium released a text-to-speech beta on September 22nd that it says produces first audio in under 50 milliseconds, a claim the company made in a two-post thread on X and that developers can test in Gradium Studio under…
A developer built GiggleGigs, a two-sided job marketplace that applies AI at three points: generating job descriptions from short prompts, drafting job-specific cover letters and auto-filling seeker profiles via Gemini r…
A developer's AI engineering series (aie_2.2) explains how to get reliable JSON from an LLM using structured outputs, framing the model as a new assistant that must return data in a consistent layout for the rest of an a…
OpenAI's GPT-Live-1 voice model listens and speaks simultaneously, using a full-duplex architecture that lets the model decide whether to stay quiet, interrupt, or start talking, according to GPT Voice team engineers Zah…