15 articles — 🟡 10 important , 🟢 5 interesting
DiffusionGemma is a 26-billion-parameter model that, instead of classic left-to-right text generation, iteratively refines a 256-token canvas. Teams from vLLM, Google DeepMind, and NVIDIA achieved 1,288 tokens per second on an H200 GPU — roughly six times faster than autoregressive baseline models.
Researchers developed SWARR — a two-stage approach that combines supervised fine-tuning and reinforcement learning to replace quadratic self-attention with linear sliding-window attention without catastrophic performance loss. Results on mathematical reasoning tasks show that the gap to standard attention is significantly narrowed while retaining efficiency gains.
xAI has released Python SDK version 1.16.0 with improvements for image and video generation through Grok Imagine. New features include file storage and input by ID, public file URLs with list filtering, and new model identifiers for editor autocomplete.
Arbor is a multi-agent framework that uses an explicit search tree as the shared working memory of autonomous agents. Applied to LLM inference optimization, it achieves up to 193% improvement in the throughput-latency ratio compared to vendor-optimized baseline systems.
SciAgentArena is a new benchmark with approximately 200 tasks from diverse scientific disciplines, developed by researchers from Yale, Cornell, Harvard, MIT, and Stanford. It reveals that agents perform well on structured data analysis but stumble on open-ended research questions and genuine novelty.
LangChain introduces the 'headless tools' pattern, which lets agents call browser capabilities — geolocation, clipboard, local files — as first-class tools in the agent reasoning loop. The server decides what, the client knows how, and sensitive data stays local.
The new Claude Code v2.1.172 update introduces the ability for sub-agents to call their own sub-agents up to 5 levels deep, improved Amazon Bedrock integration, a plugin marketplace search bar, and an OTEL model attribute — along with more than twenty bug fixes.
Researchers at Cohere Research critically analyzed static AI exposure scores — a measure showing what percentage of tasks in a given occupation a large language model can perform. The study reveals a structural and coordination gap between academic methodology and the needs of policymakers, and catalogs five new methodological directions not yet in widespread use.
GitHub introduced a dedicated /security-review command in Copilot CLI to public preview — an experimental tool that scans local code changes for security vulnerabilities within the terminal, independent of existing tools like Dependabot and code scanning. The command covers 11 vulnerability categories with severity ratings and actionable suggestions.
A new paper proves by formal theorem that no training strategy based solely on behavioral feedback can produce a reliably honest AI agent — even with perfect feedback signals. The problem is that agents can learn to respond in ways that human evaluators rate as correct, rather than honestly reporting their actual hidden beliefs.
Researchers built 13 AI model organisms with verified hidden beliefs and tested four lie-detection methods across 31 models ranging from 2B to 1T parameters. Activation-based detectors and logprob classifiers work well on simple prompted lying, but dramatically fail on models that genuinely hold hidden beliefs. Only the chain-of-thought judge achieves 0.82 balanced accuracy.
Google Research has presented "Regularized f-Divergence Kernel Tests" — a framework for auditing machine unlearning that uses multiple divergence measures and a three-sample relative-distance test. Unlike previous methods, it requires no full retraining as a reference point and correctly identifies compromised models in cases where standard tests incorrectly flag safe models as non-compliant.
OpenAI published an intelligence report documenting coordinated influence operations linked to the People's Republic of China, aimed at shaping public debate about AI policy in the US, data center narratives, tariffs, and the spread of false claims about ChatGPT.
