Voice Memory, a new inference-only scheme for agentic speech recognition, reduces weighted word error rate from 8.36% to 7.52% across ten HyPoradise domains without regressing any dataset below its 1-…
A controlled scaling study of retrieval-augmented generation (RAG) paradigms across 28 nested corpus tiers from 1,000 to 512,000 documents finds that BM25 defines the low-cost end of the Pareto fronti…
A study at 120B scale finds that inserting constitutional content during midtraining produces durable alignment gains, with constitutionally midtrained models outperforming a control on alignment gene…
Researchers propose AtmosERC, a graph-based framework for Emotion Recognition in Conversation (ERC) that models dialogue-level affective atmosphere to improve emotion prediction. The framework uses a …
Researchers have introduced Metis, the first prototype of a memory foundation model that equips AI agents with native memory capabilities, moving beyond traditional external memory modules. The model,…
A new study from arXiv finds that linear readouts trained on the final-token hidden state of large language models can decode whether diagnostic evidence supports, challenges, or fails to address a ca…
A study from arXiv finds that LLM-based emotion-cause pair extraction in conversation performs better when formulated as pair-level judgement rather than dialogue-level generation, with pair-level jud…
A new study from arXiv introduces a framework for discovering, controlling, and validating trait-like representations in large language models (LLMs), building on Funder's personality triad framework.…
A new benchmark called TREK (Travel Reasoning and Evaluation Kit) finds that even the strongest LLM agent, GPT-5.6, produces a fully feasible travel plan on only 46.2% of solvable tasks, with a median…
Researchers introduced CreditCardQA, the first financial literacy benchmark for numerical reasoning derived from real credit card agreements, containing 1,800 questions. Evaluating large language and …
Pangram Labs released Pangram 4, a deep-learning AI-text classification model achieving an AUROC of 0.9916 with a false positive rate of 0.0041% and a false negative rate of 0.3396%. The model shows i…
A new study from arXiv reveals that large reasoning models fine-tuned via reinforcement learning (RL) outperform supervised fine-tuned (SFT) models on mathematical reasoning due to more linearly separ…
Samsung Electronics Co. Ltd. reported a 19-fold jump in operating profit for the first quarter, driven by runaway memory chip prices and robust artificial intelligence demand. The company posted total…
President Donald Trump is considering imposing controls on artificial intelligence following hacking incidents at OpenAI, according to a BBC report. The potential shift marks a change of tone for his …
Meta's free cash flow plunged 91% year-over-year to $784 million in the second quarter of 2026 as capital expenditures on AI infrastructure surged 83% to $31.08 billion, nearly wiping out the company'…
Chinese robotics firms are deploying robots across UK retail to address Britain's weak productivity growth and labour shortages, according to a BBC Technology report.…
Meta shares fell more than 15% after the company reported disappointing revenue projections and rising costs, as CEO Mark Zuckerberg defended his strategy of investing heavily in AI 'agents'—personali…
Apple CEO Tim Cook will preside over his final earnings call on Thursday as John Ternus prepares to take over on September 1, with analysts focused on the company's AI strategy and iPhone 18 pricing a…
Google released Lyria 3.5, its new music generation model, built into Google Flow Music, that generates tracks between 30 seconds and 3 minutes long. A new feature called 'Selective Section Painting' …
OpenAI is launching a program Wednesday that will provide 100,000 academic researchers with free access to its advanced AI models, including GPT-5.6 Sol Pro, through 2027, the company told Axios. Part…