Natural Language Processing news and analysis on Web Pulse: 5431 curated articles tracking the latest Natural Language Processing developments, tools, and research, updated continuously from vetted sources.
UNSW Sydney researchers Anudeex Shetty, Aditya Joshi and Salil Kanhere found that AI models fine-tuned to write like intoxicated people became easier to jailbreak and more likely to leak secrets shared in confidence, acc…
A developer released Codify, an AI tool that extracts searchable text, headings and sections from scanned legal documents and links each amendment to the exact provision it changes. The tool is demonstrated on an illustr…
A developer built a voice-based AI skill assessment app in which an AI agent interviews a participant about a chosen skill and produces a scored report card. Skills are defined entirely in JSON files (including agent nam…
A developer reports that instructing a large language model to conduct its internal reasoning in Roman Urdu caused it to narrate the instruction in English rather than actually think in the requested language, while sust…
A developer outlined a Python pipeline for scraping YouTube captions at scale to feed RAG and agent systems, using the public player response and timedtext resources rather than the YouTube Data API v3, whose captions.do…
A developer built GramSetu AI, a conversational civic assistant for rural Indian micro-entrepreneurs that handles Telugu voice notes and text queries across sessions spanning weeks. The team replaced naive RAG and rollin…
A developer modified the Qwen 2.5 Omni 3B model's attention mask to stop slide images from interfering with audio transcription when slide context is supplied alongside audio. The approach first ablates the visual span o…
A sparse-probing study on a frozen BERT-base-uncased encoder identified fewer than 1% of the model's 9,216 CLS hidden-state units (12 layers × 768) as stable AI-text detectors across six generators on the RAID benchmark,…
Researchers introduced target-speaker unlearning ASR (TSU-ASR), a fully end-to-end framework that lets a frozen dual-stream speech LLM transcribe multi-speaker audio while excluding opt-out speakers, using a light-weight…
A new arXiv paper (2609.30492v1) reports that evolutionary persona generation increased response diversity by 78.8%, originality by 26.1%, flexibility by 49.5%, and holistic creativity by 13.9% over task-only prompting o…
A new arXiv paper (2609.30467v1) introduces two taxonomies for retrieval-based factuality evaluation failures in open-ended medical settings, decomposing errors into retrieval-stage failures across five quality dimension…
Pretraining small decoder-only transformers on code-switched text aligns representations of parallel text across English, Dutch, and Chinese and improves scores on the BabyLM evaluation suite, according to an arXiv paper…
Researchers introduced REALMS (Real-time Exact Audience sizing via LLM-based Multi-attribute Search), a conversational system that returns exact audience counts in seconds from profile stores holding millions of profiles…
A new arXiv paper (2609.30484v1) proposes a knowledge graph-based evaluation framework for testing whether large language models genuinely comprehend context in question answering rather than relying on surface-level pat…
A new arXiv paper (2609.30469v1) reports that an external LLM-as-a-judge filtering paradigm significantly reduces word error rate in pseudo-labeled training sets for noisy Broadcast Police Communication audio from Baltim…
A lightweight reliability predictor built on a frozen Audio LLM's audio-encoder representations detects unreliable speech transcriptions with 81.10% in-domain and 78.09% cross-domain macro-F1, outperforming the strongest…
Researchers introduced SEA-CLIP-Tiny, a multilingual text-vision embedding model with fewer than 50M parameters for seven Southeast Asian languages, according to an arXiv paper (2609.30739v1). SEA-CLIP-Tiny reached 12.9%…
Full transcripts outperformed all compressed transcript inputs for LLM-based veracity classification of Japanese medical YouTube videos, according to an arXiv study (arXiv:2609.30882v1) that tested 74 long-form videos la…
A new arXiv paper (2609.30849v1) proposes an LLM-as-a-judge approach to measure perturbation strength uniformly across input and chain-of-thought (CoT) perturbations, enabling controlled-strength evaluation of self-consi…
A new arXiv paper (2609.30897v1) argues that cultural evaluation of language models should move beyond factual knowledge and survey-agreement benchmarks toward evidence-centered tasks that test interpretive depth, using …