Adversarial tests expose major gaps in LLM unlearning
A preprint reports that language models can appear to forget targeted information under ordinary tests while still recovering it under strategic adversarial prompts.
Актуализира се ежедневно1468 проверени истории
Source-checked AI coverage of product launches, policy shifts, safety research, and industry moves, explained in plain English by a nonprofit education team.
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
Какво се случи, защо има значение и какво да гледате – без данък върху жаргона.
Когато сигналът е слаб, ние не публикуваме нищо, вместо да допълваме емисията.
Нарастващ поток от проверени гледни точки за хора, които трябва да разберат AI, без да преследват реклами.
A preprint reports that language models can appear to forget targeted information under ordinary tests while still recovering it under strategic adversarial prompts.
An arXiv study reports that six language models contained linearly decodable signals associated with being evaluated, while those internal signals only partly matched what the models said. The authors say steering along probe-derived directions changed verbalization scores, raising questions about how reliably…
TechNode reports that ByteDance has launched Doubao Work, a productivity-focused AI agent integrated with Feishu, with 30 days of subscription access available through the desktop client or app.
A paper accepted to EMNLP 2026 proposes FCPRAG, a controller that selectively combines passage-specific LoRA adapters in parametric retrieval-augmented generation. The authors report higher F1 scores than standard and parametric RAG baselines across four question-answering datasets, with gains of up to 7.55%.
An analysis of 13,921 papers from major NLP conferences found that the papers reporting the most GPU capability captured most reported resources but only a minority of citations and awards. The study found a statistical association between compute and scholarly impact, but little standalone explanatory power.
Sifted reports that former Lunar executives have raised €8.2m in pre-seed funding to launch Repodo, an AI-native statutory-audit startup focused initially on small and medium-sized businesses. The report says the round is Denmark’s largest pre-seed raise to date, according to Sifted data.
chemeurope.com reports that Microsoft Research’s Skala AI model has been integrated into the open-source CP2K simulation ecosystem, giving researchers access to an AI-based exchange-correlation functional for quantum-mechanical simulations. The outlet says initial testing showed improved accuracy in a specific test…
India Today reports that Nvidia’s agreement with AI startup Poolside would give Nvidia access to Poolside’s technology and bring more than 100 employees into its Nemotron open-weight model effort. The report also says Nvidia will invest $1 billion in Poolside at a $12 billion pre-money valuation, while rising…
An arXiv preprint introduces PUMA, a 900-task benchmark for evaluating multimodal AI in Polish cultural and linguistic contexts, reporting strong visual question-answering performance but substantial weaknesses in complex audio and document understanding.
A preprint introduces HiDiffTIR, a reinforcement-learning method that gives different weights to tool-use trajectories and reasoning steps based on their difficulty.
AI News reports that XPENG’s physical AI unit has secured more than $900 million in funding at a $6.3 billion valuation to develop and commercialize its IRON humanoid robot.
A new arXiv paper presents LURE, a zero-data self-play method in which one language model sets task difficulty while another solves verifiable reasoning challenges. The authors report stronger out-of-distribution zero-shot accuracy than trained baselines across nine held-out benchmarks, but the abstract does not…
One useful briefing each week
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Hiring an AI professional or launching a useful AI product? Put it in front of people who came here to learn and act.
Публикувайте работа за AI Submit an AI tool