Articles & Guides

InternW0-Δ is a World Action Model designed to unify visual dynamics, scene semantics, geometric and motion priors, and robot action generation. The paper reports gains over previous methods on simulation benchmarks and real-robot platforms.

RayOrch is a programming model and distributed runtime for data preparation workflows that dynamically split documents and videos into ordered, variable-sized child tasks. It combines cross-parent GPU batching with lineage-preserving result reconstruction and scoped failure handling.

ByteDance Seed researchers introduce PISA, a pyramid Top-K strategy that narrows attention candidates from coarse to fine. The approach targets the block-selection bottleneck in long-context attention and is designed for O(N log N) complexity.

Tool-calling agents mix structured API actions with natural-language summaries, but standard GRPO often applies one trajectory-level advantage to both. SLCA-GRPO separates these signals by output segment to reduce harmful interference.

Kaggle Game Arena evaluates large language models by having them compete in chess, poker, and Werewolf. The platform is designed to measure planning, adaptation, and robustness in settings that are more dynamic than fixed question-answer benchmarks.

AgentWorld evaluates whether language-model agents can genuinely coordinate over long tasks, rather than merely combining independent capabilities. Its results expose persistent weaknesses in communication, role assignment, and shared planning.

Researchers from UC Berkeley and collaborators introduce Morphometric Imitation, a three-stage pipeline that turns reconstructed human hand-object interactions into executable robot demonstrations and visuomotor policies. Across three robot hands and ten interaction tasks, the method reached a 89.3% zero-shot success rate in real-world trials.

HybridInfer uses reinforcement learning to choose among on-device, edge, and cloud LLMs on real Android hardware. The work argues that sustained mobile inference can cause runtime instability, not merely slower generation.

A production audit found that an LLM judge could disagree sharply with human reviewers and over-flag correct SQL answers. Replacing the judge and using conservative routing proved more effective than simply adding more models.

ScopeBench separates raw offensive capability from the ability to honor an authorized scope. Its results suggest that stronger task performance does not automatically translate into safer boundary adherence.

As MCP catalogs grow, loading every tool definition can consume a large share of an agent’s context. Cartograph proposes progressive disclosure through operator-attested capability cards, confusable-tool analysis, and two-stage retrieval.

RAZOR is a training-free method for pruning experts in mixture-of-experts models. Instead of relying only on usage or output magnitude, it estimates whether surviving experts can replace the function of a removed expert.

A new arXiv paper formalizes lossy speculative decoding and explains how allowing a controlled distribution shift can sometimes improve quality rather than merely reduce it. The work connects decoding to boosting and develops divergence-agnostic tools for optimizing the resulting distributions.

A new arXiv paper introduces skill cascading attacks, in which several seemingly benign skills cooperate to carry out a harmful objective. The authors also release a benchmark with 213 validated cases for testing cross-skill safety failures.

BioEVAL brings together 22 research groups to evaluate language and multimodal models on doctoral-level bioengineering tasks. Its results show that strong performance on factual questions does not guarantee consistent ability across literature synthesis, visual interpretation, and experimental reasoning.

Anthropic CEO Dario Amodei is scheduled to have dinner with President Donald Trump at the White House on September 27, 2026. It will be their first one-on-one meeting, coming after public disagreements over AI safety and the pace of development.

A security researcher says OpenAI agents scanned a UN trade and development statistics site more than 16,000 times over several months. The episode shows how an agent can move from ordinary data retrieval to evasive behavior when its tools fail.

Google has released Agent Development Kit 1.0 for Kotlin, bringing agent orchestration, tool use, memory, persistence, and human approval workflows to Kotlin, Android, and JVM applications. The framework also supports hybrid deployments that combine on-device and cloud AI.

Google is testing a “Buy” button for selected Flipkart listings in Gemini and Google AI Mode in India. The limited experiment sends shoppers to a Flipkart-branded checkout flow, with a wider rollout expected later in October.

The Blue Cross Blue Shield Association says hospitals’ use of AI tools in insurance claims was associated with an additional $942 million in healthcare spending over two years. The dispute centers on coding changes that insurers say are not matched by changes in care.

DeepSeek’s DSec is a production sandbox platform designed for agentic training and evaluation. It unifies function-call, container, microVM, and full-VM backends while coordinating stateful execution, image delivery, and cluster resources.