Articles & Guides

David Robinson, who helped write safety reports for major OpenAI launches, has resigned and criticized the company’s culture. He argues that frontier AI labs need high-reliability safety practices rather than relying mainly on iterative deployment.

David Robinson, who helped write safety reports for major OpenAI model releases, has resigned and publicly criticized the company’s culture. He argues that frontier AI labs need the layered redundancy and deliberate planning associated with nuclear plants and busy airports.

OneStreamer is a 4B streaming video model designed to connect real-time perception, reusable memory, and proactive responses. It turns observed evidence into timestamped captions and event summaries so information can remain available after the original frames leave the visual window.

Remembering interaction history does not guarantee that an agent understands the current world or makes progress toward its goal. PoS maintains explicit beliefs at inference time and detects, diagnoses, and recovers from stagnation, cycles, and drift.

AutoGUIWorld combines an image generator with a task planner to synthesize GUI interaction trajectories without deploying the underlying software. The resulting data improves computer-use agents across several desktop and browser environments.

A new study suggests that many pretrained Transformers do not lack the capacity for long reference chains; they simply stop using it early. A rank-8 LoRA placed at one early layer can make frozen middle layers carry the computation much farther.

X-Tree mines recurring, success-related action spans from agent trajectories and organizes them into a reusable experience hierarchy. The structure is then used in offline RL, online RLVR, and on-policy self-distillation without LLM calls.

PhysVista introduces a benchmark that evaluates whether vision-language models understand physical consistency rather than merely recognizing visual content. It combines physical state perception, dynamics reasoning, and plausibility assessment across real-world and AI-generated videos.

A new study turns intent drift into a measurable failure mode: superseded or withdrawn user requirements can continue to shape an agent’s answer or tool action. Its IntentFlux benchmark and StateForge method show that explicit state maintenance helps, but does not eliminate the gap between evolving dialogues and direct single-turn tasks.

RouteFM treats LLM routing as a reusable capability rather than a workload-specific fitting problem. By reading limited behavioral evidence from anonymous candidate models, a frozen router can adapt to new tasks, model pools, modalities, and context budgets.

OpenTumorBoard turns public multidisciplinary tumor board recordings into a benchmark for evaluating models on specialist answers and full clinical discussions. Its results show that even advanced general and medical models still struggle to reproduce expert responses and board-level consensus.

Audio LLMs may recognize a command correctly and still act when they should remain silent. VGBench evaluates whether voice agents can determine who is being addressed before invoking tools or responding.

NEEDLE is a training-free defence that targets a known backdoor direction while preserving refusal-related representations and limiting changes to normal model behaviour.

KaliBench targets a neglected part of cybersecurity AI: translating analyst intent into precise, executable Kali Linux commands. Its fine-grained dataset and multi-stage verification separate tool selection from argument construction.

Rules to Tools converts written scientific requirements into callable executable checks for coding agents. The study finds higher repair scores on some SciCode cohorts, but the gains are task-dependent and can come with higher CPU usage.

World Observer separates acting from observing by jointly generating an agent-centric actor view and one or more panoramic observer views. This lets out-of-view objects continue evolving and improves their state and appearance when they re-enter the actor’s view.

GraphForge introduces a framework that grounds agent tasks and evaluation rubrics in real files connected through an evidence graph. Its reported results suggest that traceable verification can improve training for agents that handle files, tools, and deliverables.

PyRUA-Lean replaces repeated tool calls with executable Python cells that compose robot primitives, perform local checks and request only necessary observations. Across 700 simulated instances, it raised success from 63.1% to 71.7% under the same LLM-call budget while cutting input tokens by 65% on jointly solved tasks.

Argo-Bench moves beyond text-to-SQL by testing whether agents can reconstruct facts from a large enterprise warehouse, analyze them, and take actions with measurable consequences. Its results suggest that current models still struggle with end-to-end business decision-making.

LOCI combines key-value caching with recurrent linear memory to help video world models reproduce previously observed regions when a camera returns. Its geometry-aware memory improves scene retrieval while reducing the cost of processing long videos.

ScholarCatalyst moves scientific retrieval beyond topical similarity by asking which earlier papers could genuinely advance a research project. Its author-annotated results show that current agents still struggle to match expert judgment.