Natural Language Processing news and analysis on Web Pulse: 5434 curated articles tracking the latest Natural Language Processing developments, tools, and research, updated continuously from vetted sources.
Researchers introduced Dual-Entropy Enhanced Policy Optimization (DEEPO), a dual-stage reinforcement learning method that reduces hallucination in multimodal large language models by combining semantic-entropy-triggered …
UO-FIE, a parameter-efficient system built on Qwen3.5-9B with LoRA, ranked first in the fine-tuning track of the Factivity Inference Evaluation 2026 (FIE2026) with a macro utility of 0.8316, according to the paper's abst…
Fine-tuned XLM-RoBERTa achieved an F1-score of 0.5959 on Bangla medical named entity recognition, setting a new state of the art above the previously reported best of 0.5848, according to an arXiv benchmark evaluating Ba…
Researchers Samuel McCandlish and colleagues submitted "Scaling Laws for Neural Language Models" to arXiv on 23 January 2020, showing that language model cross-entropy loss falls as a power-law with model size, dataset s…
Two September 2026 papers propose post-training alignment techniques that expose models to misaligned contexts and then enforce aligned targets: ARCF, which reduced unsafe generations on a benchmark of 5,000 harmful prom…
Telnyx added Soniox text-to-speech to its Voice AI catalog, offering about 200 voices that each speak 60 languages from a single model. Developers address the voices as `Soniox.tts-rt-v2.<Name>` — including `Soniox.tts-r…
MimicScribe's live-preview transcription, built on a batch ASR model with overlapping-window merging, transcribed 18.8 hours of meetings and earnings calls with 15.5% of words wrong and 96.9% of spoken numbers correct, w…
A developer detailed a browser-based, client-side EPUB-to-Markdown converter that parses each chapter into structured Markdown without uploading the book, and published a Python script that splits a converted book into o…
TypeSafe released Grev, a suite of Unix-style filters that run on its Jev models to answer typed questions with calibrated probabilities instead of generating text, keeping the input as the output. The 14 tools include g…
Google released Gemini 3.8 Live with Live Avatar, an update that lets users hold real-time conversations with an animated AI persona that lip-syncs and shows facial expressions, available only to Gemini Enterprise custom…
A developer detailed a hybrid streaming architecture for AI voice agents that keeps end-to-end response time under 300 milliseconds by serving short conversational backchannels and fillers like "okay" and "one moment" fr…
A developer built Vehicle Advisor, a production conversational automotive discovery and financing platform on NestJS, TypeScript, and PostgreSQL that uses a hybrid neuro-symbolic architecture to keep LLM reasoning separa…
Microsoft's harrier-oss-v1-270m sentence-transformers model can be fine-tuned with SentenceTransformerTrainer and MultipleNegativesRankingLoss for Portuguese QA retrieval, though no public example combines that exact sta…
BananaBread AI launched an API that identifies what consumers consistently love and still want from a specific consumer product by analyzing thousands of reviews across Amazon, TikTok, Reddit, YouTube, Instagram, Walmart…
A developer running the local Qwen 3.5 9B model in Ollama reported that it abruptly switched from English to fluent Chinese mid-conversation, an instance of language drift. The developer traced the behavior from a model-…
AWS released a WhisperX Deep Learning Container that packages OpenAI's Whisper with wav2vec2 forced alignment for per-word timestamps and speaker diarization, deployable to Amazon SageMaker AI real-time or asynchronous e…
An independent experiment fine-tuned a 135M-parameter model on the 10,003-message Banking77 dataset in under 30 minutes on a 16 GB Apple MacBook Air, lifting exact intent accuracy from 1.62% to 83.25% and operational dec…
A developer's guide to RAG chunking argues that the document-splitting step is the most commonly rushed and most damaging part of a retrieval pipeline, since errors there propagate downstream into embeddings and retrieva…
A user of Japanese-English machine translation rated ChatGPT at roughly 80 out of 100 and DeepL at about 60 for one-shot paragraph translation, arguing the technology is very good but not "solved." The user, posting on L…
A developer who built InsNet, a ~21,000-line C++14 deep learning library with 4,828 lines of hand-written CUDA kernels, starting in 2018, describes how its padding-free dynamic batching design anticipated the flat, paddi…