H Company (France) releases Holo3.1, VLM family fine-tuned on Qwen 3.5 for computer use agents. Models 0.8B to 35B-A3B, web/desktop/mobile support, native function-calling, multiple quantizations (BF16, FP8, Q4 GGUF). Apache 2.0 license.
Mellum and Granite embedding models are now available on llama.cpp. Two pull requests add support for these models in the framework.
Microsoft announces two on-device models at Build 2026: Aion 1.0 Instruct (efficient small model, open-weights, competes with Apple AFM-3B) and Aion 1.0 Plan (14B parameters, reasoning + tool-calling, 32K context, built into Windows). Aion 1.0 Plan enables local agentic workflows.
WRIT is a trajectory synthesis pipeline for multi-turn agent training. It generates complex tasks along two axes: number of write decisions and evidence burden per decision. With 2K synthesized trajectories, a 4B model outperforms GPT-5.1 no-think on τ²-bench while reducing inference-time token usage.
DeltaMem organizes LLM agent experience memory into two residual trees: one stores goal-conditioned tasks as reusable skills, another stores scene-level environment knowledge. Each tree uses root nodes for generalized base experiences and delta nodes for variations, eliminating redundancy. An autonomous consolidation mechanism distills high-frequency paths into new root nodes.
An arXiv study analyzes 8 benchmarks for multivariate time series anomaly detection. A diagnostic framework shows 79-100% of anomalies are univariately detectable on 6 datasets. Cross-channel models provide no measurable gain. Current benchmarks fail to validate multi-channel modeling capabilities.
HITL-GB framework for short-term rental dynamic pricing: a contextual bandit algorithm generates price recommendations that a human can accept, modify, or reject. Authors show historical data is structurally equivalent to on-policy warm-up, reducing cold-start from ~150 to ~30 episodes. Validated on 1,461 real nights (April 2022–2026).
Benchmark evaluating environmental attitudes across 31 LLMs (proprietary and open-weight). Models exhibit more progressive environmental positions than average German survey respondents, but show no systematic relationship with model origin, size, or release context. Detects prompting manipulation risks and sycophantic shifts.
BehaviorBench is a benchmark for evaluating personalized decision modeling from real-world behavioral traces. Built on 2,000 wallets with 141,445 belief-prediction instances and 1,485,972 trade-prediction instances, it tests whether generative models can adapt predictions to individual users without relying on simulated behavior.
Uber caps monthly spending at $1,500 per employee for agentic coding tools (Claude Code, Cursor) after exhausting its AI budget in four months. The measure aims to control expenses from code agents.
Grok Imagine Video 1.5 from xAI is now available on AI Gateway. The model generates video from input image with synchronized audio in single pass. Improvements: audio quality, prompt following, photorealism, character consistency across longer sequences, expanded reference image support for visual style control.
GitHub outlines its strategy to handle the explosion of coding agents. Following Copilot's launch, the platform must adapt its infrastructure and tools to new agentic workflows creating strain on its systems.
Anthropic scales Project Glasswing to 150 partners across 15+ countries using Claude Mythos Preview to detect critical flaws. Existing partners have identified over 10,000 serious vulnerabilities. Anthropic simultaneously commercializes Claude Security to fix them.
Headroom compresses tool outputs, logs, files, and RAG chunks before sending to LLM. Reduces token consumption by 60-95% without quality loss. Available as library, proxy, and MCP server.
Hackers compromised high-profile Instagram accounts, including the Obama White House page, by requesting Meta's AI support chatbot to change the registered email address. Two-factor authentication was bypassed entirely. Meta patched the vulnerability, but additional exploits are already circulating on Telegram.
Mathematical primer on generative AI foundations by Tianhua Chen. Covers VAE, diffusion models, normalizing flows, autoregressive factorizations, GANs, Wasserstein GANs, and energy-based models through derivation-oriented approach.
Warren Buffett's Berkshire Hathaway invests $10 billion in Alphabet's AI infrastructure buildout. Alphabet raises $80 billion to scale AI capacity, with capital expenditures expected to reach $190 billion in 2026.
Starbucks withdraws NomadGo, its computer vision tool deployed across 11,000 stores after nine months. The system, validated at 99% in controlled environment, failed at production scale, highlighting the gap between sandbox testing and real-world deployment.
CNIL fines IQVIA Operations France €5M (decision SAN-2026-008, May 26, 2026) for GDPR breach. Data pseudonymization does not eliminate legal obligations if the company retains decryption keys.
OpenAI makes GPT-5.5, GPT-5.4, and Codex available through Amazon Bedrock at identical pricing to OpenAI's platform. Models run in commercial and government AWS regions, currently limited to the US. Usage counts toward existing AWS contracts.
NVIDIA releases Cosmos 3, a collection of omnimodal world models (Nano 16B, Super 64B) capable of generating dynamic video, image, audio, and action commands from text, image, video, and action trajectory inputs. Available on Hugging Face for Physical AI applications.
arXiv paper proposing principled approaches for designing practical LLM-based agentic systems. Framework with "pseudo-tools" recursively calling LLMs on restricted context. Hand-engineered fixed workflows are cheaper and more accurate than dynamically-planned workflows. Learning methods for pseudo-tools and fixed workflows outperform hand-engineered agents.
Study of temporal framing in news: a taxonomy of 8 temporal frames identifies how references to past, present, and future function as rhetorical devices. Annotated corpus of 458 English/German articles (2K+ sentences, 3K annotations). Supervised models substantially outperform zero-shot approaches for sentence-level temporal framing detection.
ARCA introduces a token-level credit assignment method for LLM reinforcement learning that addresses degeneracy of intrinsic signals (surprisal, entropy reduction, policy divergence) under LoRA. It measures adapter salience directly via L2 norm of hidden-state residuals instead of output-distribution shifts. Tested on MATH/Qwen3-1.7B with GRPO, ARCA avoids pathological weight concentration.
Iterative AI workflow optimizes graphite-based anodes through sequential learning and experimental feedback loops. Citrine Platform generates surrogate models and refines manufacturing constraints. Results: fabrication reliability improved from frequent failures to 100% success, cells ≥350 mAh/g increased from 28.4% to 84.8%, capacity retention rose from 42.1% to 97.3%.
TrustLDM is a trustworthiness benchmark for Language Diffusion Models (LDMs) covering safety, privacy, and fairness. Results show LDMs degrade alignment when malicious post contexts are attached to masked responses, regardless of context length. An automatic evaluation framework (TrustLDM-Auto) systematically identifies vulnerable configurations across all tested models.
DraDDP is the first public multimodal dataset for discourse parsing in multi-party dialogues. Built from American TV dramas, it contains 495 dialogue segments (6,374 utterances, 9.1 hours of video). Benchmarks demonstrate the value of multimodal information for identifying dependency structures and relation types between utterances.
Theoretical study of optimal learning rates in 2-3 layer linear neural networks. Derives exact closed-form expressions for gradients and test loss after 1-2 gradient descent steps. Key finding: unequal layer-wise learning rates minimize loss in early training, while equal rates become optimal later. Code released.
MERIT, a multi-horizon memory retrieval framework for text-to-SQL agents, maintains episode-level and turn-level memory with learned retrieval policies optimized via reinforcement learning. On BIRD-Interact, it outperforms no-memory and static-retrieval baselines, improving success rate while reducing interaction turns.
ART (Attention Run-time Termination) is a lightweight runtime mechanism that halts KV block accesses during decoding once their attention contribution becomes negligible. Tested on LongBench, it achieves 20% higher generation throughput on large batches while maintaining comparable accuracy.
DeSQ decomposes complex questions into atomic constraints to generate SPARQL queries. The framework operates in three stages: decomposition, generation of SPARQL fragments with placeholders, and final assembly. Outperforms state-of-the-art on 4/5 major benchmarks with improved robustness to lexical variation.
Demo2Reward optimizes VLM reward model language instructions at test-time using 3-10 expert demonstrations to reduce false positives in robotics. No additional training required. Validated on simulated tasks and real-world transfer.
Geodesic Flow Matching adapts Riemannian transport to denoise Spatial Semantic Pointers (SSPs) on toroidal manifolds. Tested on a Spiking Neural SLAM system, the method achieves 72% reduction in tracking error and 40% increase in neural efficiency versus baselines.
Researchers demonstrate that hidden reasoning traces in LLMs can be extracted via Reasoning Exposure Prompting (REP), a lightweight prompting method using shadow-model-generated demonstrations in auxiliary code-like formats. REP exposes internal traces even when deployed systems intentionally hide them, while preserving useful reasoning signals for distillation.
Vercel Blob now offers time-bound signed URLs to upload, download, inspect, or delete objects without full store access. Each URL is scoped to a single operation, pathname, and expiry up to 7 days. Multipart uploads let browsers stream directly to Blob storage without server round-trips.
Hackers exploited Meta's AI support chatbot to gain access to high-profile Instagram accounts. By simply asking the bot to link a new email address to a target account, they bypassed the entire account recovery process. Meta had wired its support system to an AI capable of executing account changes in a single request.
Alphabet announces an $80 billion equity capital raise to expand AI infrastructure and compute capacity. This massive injection aims to support AI services growth and strengthen the group's competitive position against other tech giants.
Anthropic has confidentially filed a draft IPO registration with the SEC. The Claude maker is valued at nearly $1 trillion after its latest funding round. OpenAI is also preparing for an IPO.
Anthropic files for public offering. The Claude maker prepares a major IPO.
Anthropic confidentially submitted a draft S-1 form to the SEC, marking the first step toward a public offering. No financial details or timeline disclosed.