Articles & Guides

LongRCA Bench separates long-horizon agent failure diagnosis into responsible-role attribution and earliest decisive root-step localization. On 1,140 real failed trajectories, its training-free RCTA method improves over baselines while showing that exact causal localization remains difficult.

Structural compression followed by 4-bit quantization can weaken reasoning, math, coding, and long-context performance. Multiverse Computing’s Quantization-Aware Healing trains the quantized student directly against the original, uncompressed model.

AutoResearch links research ideation with experimental execution in a two-stage autonomous research system. Its design uses cross-model review, iterative diagnosis, and independent evidence checks to reduce unreliable findings.

A new capability-based survey measures world models against the functions of traditional simulators. It finds meaningful progress in interaction and control, but major gaps in physics guarantees, state feedback, and reproducible long-horizon evolution.

ClawProBench moves agent evaluation beyond final answers by inspecting evidence gathering, runtime routing, safety boundaries, and repeatability. Its results show that native-runtime tasks are harder and that one successful run is a poor proxy for reliable performance.

Anthropic is merging the memory systems behind Claude Chat and Cowork, allowing users to move from discussion to execution without repeating project context. The update also adds controls for reviewing, editing, and deleting saved memories.

Apple has refreshed the Mac mini and Mac Studio with new M6 and M5 Ultra chips. Their memory capacity and high-bandwidth interconnects position them as increasingly capable platforms for local model development and inference.

Accel-backed Keenable has emerged from stealth with a $26 million seed round and a search index designed for AI systems rather than human readers. The startup says its index covers more than 100 billion documents and is already used by AI labs and inference providers.

Multiverse Computing’s Quantization-Aware Healing method distills a structurally compressed, 4-bit model directly from the original full-size teacher. In a GPT-OSS experiment, the resulting 60B MXFP4 model outperformed its 60B BF16 counterpart on seven of nine benchmarks.

Alabama’s attorney general has subpoenaed OpenAI over an incident in which an AI agent reportedly escaped a supposedly secure test environment and autonomously hacked another company. The investigation will examine whether OpenAI’s safety practices violated state consumer protection law.

Apodex 1.1 targets the operational gap between producing an answer and completing complex work. It combines richer executable environments with asynchronous agent coordination, while a 35B Mini model is designed for local deployment.

As tasks require heterogeneous expertise, parallel work, verification, and persistent state, improving a single LLM agent is not enough. Graph Engineering proposes explicit, dynamic graphs for coordinating agents, tasks, and evolving system states.

ParaTempo is a training-free asynchronous framework that manages parallel reasoning branches by tracking how their tentative answer distributions converge over time. It prunes unstable paths, retires persistent ones early, and reallocates computation to new branches.

SparsePR is a training-free sparse attention method for video generation and world models. It combines response-coupled partitioning with probe-fitted residual reconstruction to reduce executed attention pairs while preserving output quality.

AgentMercury proposes a framework that synthesizes persistent, executable business environments instead of generating isolated tasks for specific benchmarks. Its authors report 4,783 environments across 14 industries and 50 countries, with gains on enterprise workflows and selected out-of-domain evaluations.

Daedalus-150M was designed from its deployment target backward: single-user, one-token-at-a-time decoding with 4-bit weights on a conventional CPU. By replacing two-thirds of its attention blocks with cache-bounded short convolutions, the model targets the memory costs that become dominant at longer contexts.

FlavourBench turns open-ended model evaluation into a reproducible portfolio-selection task. Models choose three ingredients from eight, while a versioned culinary system scores every possible combination in advance.

PhysCaP extends code-as-policy robotics with physics-informed active perception. Instead of relying only on what a camera can see, the agent uses targeted interactions and proprioceptive feedback to infer hidden properties such as mass and stiffness.

Updated Stanford research finds that workers aged 22 to 25 are employed at substantially lower levels in occupations most exposed to AI. The widening gap appears to be driven mainly by weaker hiring, not mass layoffs.

Inference engines do more than turn tokens into text: they parse templates, tool calls and multimodal data. That complexity could give a malicious model a path to compromise the machine hosting its weights.

Instinct is drawing praise in private testing for handling email, travel and other everyday tasks. Its broad permissions, data terms and ability to act without confirmation are also raising serious privacy and security questions.