Natural Language Processing

A new study measuring how much formal semantic structure explains human label variation in natural language inference finds that it accounts for only 3.3 to 3.6 percent of entropy variance, with a group-level effect show…

A new study from arXiv finds that test-time adaptation (TTA) can improve zero-shot 3D CT vision-language models only under specific conditions, and introduces CARVE (Cardinality-Aware Retained-View Entropy) as the first …

A new arXiv preprint (2607.15715v1) investigates whether agentic components like reflection and memory in LLM agents produce observable and controllable improvements over fixed workflows for information extraction. The s…

Researchers introduced MAR-12, a framework using Vision Language Models (VLMs) to detect and explain harmful humor in memes, achieving up to 80.3% accuracy for humor detection and 75.9% for hate detection on the PrideMM …

Researchers at an undisclosed institution found that masked diffusion language models (DLMs) learn a bidirectional induction circuit for in-context learning, where previous-token and next-token heads write local context …

A new study from researchers evaluating large language models as unified multimodal learners for clinical prediction finds that converting all patient data into a single natural language sequence and fine-tuning a pretra…

Researchers introduced EpiNarrate, an agentic framework for generating grounded public health narratives from epidemiological projections, which separates structured numerical reasoning from natural-language generation t…

A new study from arXiv finds that external toxicity tools are unreliable for code-mixed and multilingual abuse detection, proposing ToxGate, a trust-fusion head that conditions auxiliary signals on encoder representation…

Researchers introduce CAMMAR (Culture-Aware Matryoshka for Metaphorical Arabic Representations), a representation learning framework that organizes Arabic word meanings into nested lexical, cultural, and metaphorical emb…

A study using the ASAP 2.0 dataset found that GPT-5 mini achieved the highest agreement with human ratings for automated essay scoring, while GPT-5 produced the strongest summarization quality, revealing trade-offs among…

Researchers propose NeurOWL, an end-to-end neuro-symbolic framework that uses Large Language Models and ontology embeddings to perform subsumption verification and abduction for incomplete OWL ontologies, addressing a fu…

A new opinion paper from arXiv argues that large language model (LLM) self-explanations are often plausible and actionable but may lack faithfulness to the model's true reasoning. The authors identify limitations in stan…

Researchers propose AdaLook, an adaptive lookahead framework for masked diffusion language models that dynamically determines rollout depth based on candidate-score variance, achieving a better accuracy-decoding steps tr…

Apple is testing a new AI-powered 'Live Notes' feature that transcribes conversations between Genius Bar employees and customers at some stores, according to Bloomberg's Mark Gurman. The opt-in system adds the transcript…

Feyn AI, a YC-backed startup, released SQRL, a family of text-to-SQL models that inspect the database before writing a query. The flagship SQRL-35B-A3B achieves 70.6% execution accuracy on BIRD Dev, outperforming Claude …

A Popular Science test of five AI writing detectors found that Pangram, Grammarly, GPTZero, and Scribbr correctly identified all human-written and AI-generated samples, while a fifth tool was not fully tested. The test u…

Nonprofit Current AI, founded in February 2025 by Martin Tisne, is racing to build open, public AI infrastructure modeled on the World Wide Web, aiming to provide free AI access for all. The organization, which has secur…

AI search visibility tools that count how often a model mentions a brand in response to prompts are measuring the wrong number, according to technical SEO consultant Jono Alderson and data from the No Hacks podcast. The …

InterviewPracticeAI, founded by Ben, launches a free AI mock interview tool that scores answers out of 100 with specific feedback, powered by Anthropic's Claude. Users upload their CV and job description, answer one tail…

OfflineTTS, a free browser-based text-to-speech and speech-to-text toolkit that runs locally, has launched with support for four TTS engines, 99 STT languages, and workflows for subtitle generation, voice-over, and ebook…