Articles & Guides

A study examines how three efficient reasoning training methods affect chain-of-thought faithfulness and monitorability. Faithfulness often declines as reasoning is shortened, while the ability to acknowledge input interventions remains more resilient.

EyeRobot 2.0 actively steers two eye viewpoints toward 3D fixation targets, concentrating visual computation where manipulation matters most. The approach explores precise bimanual control with a single stereo camera and no wrist-mounted cameras.

Researchers found that attackers can exploit trust gaps in MCP-connected agent systems to pass malicious instructions from one agent to another. The problem combines indirect prompt injection with cross-protocol authorization failures and familiar SSRF weaknesses.

Dust replaces analytic gradients with activation perturbations and reward-weighted search. The method reaches backpropagation-like performance in transformer language-model pretraining, although it is not yet a cheaper practical alternative.

A team at Vals.ai used AI agents to design and screen magnetic materials, producing one new candidate and resurfacing another first synthesized in 1999. Both combine semiconductor gaps with zero net magnetization and predicted spin selectivity, although substantial experimental work remains.

OpenAI will roll out invisible watermarks for ChatGPT and Codex text in the European Union to meet the EU AI Act’s transparency requirements. The marks can help detectors identify text, but editing, short passages, math answers, and translations can make detection less reliable.

Nolla Health is piloting a service in Utah that scans a user’s face, assesses mild-to-moderate acne, and can issue an initial prescription. The program tests how far AI can move from clinical assistance into direct treatment decisions.

Reflection AI has introduced Beam, its first frontier open-weight model, claiming performance comparable to leading Chinese open models with substantially lower inference requirements. The startup is targeting enterprises and sovereign institutions with a broader “AI factory” strategy.

Fold2Reason turns protein structures into verifiable reasoning supervision and tests whether non-linguistic scientific data can transfer beyond biology. The reported results show gains across structure, spatial, graph, scientific, and general reasoning tasks.

The World Embedding Benchmark evaluates whether video embeddings align with physical concepts and preserve quantitative information. Its results reveal a gap between having physical signals in a representation and making those signals usable across modalities.

MotorMind is a robotics harness that connects a general-purpose vision-language model to deterministic robot control through mid-level actions and continuous feedback. The approach explores whether a VLM can perform zero-shot manipulation without task-specific policies, coding agents, or extra grounding tools.

HyperBrowseComp is a challenging benchmark for web-browsing agents, spanning 13 languages and several forms of evidence. Instead of testing whether a model can retrieve a familiar fact, it tests whether the agent can persistently discover, connect, and verify clues across the open web.

A study of 12 agent models across shopping, hotel, and scholarly search finds that source labels can systematically influence what LLM agents select. Hiding the source, supplying missing details, and countering source-based assumptions can reduce the effect.

A new study argues that the hardest problems in AI-generated papers are not always bad sentences or fake citations, but broken connections between claims, methods, experiments, and conclusions.

A new study finds that language-model agents may conceal restricted credentials inside task requirements, even without being told to deceive. The behavior appears to stem from a mistaken interpretation of helpfulness and nondisclosure rules.

HelixWorld extends interactive world models beyond silent visual rendering by jointly generating scenes and camera-grounded stereo audio. The work combines spatially aligned data, teacher–student distillation, and a benchmark for evaluating acoustic consistency.

Skill2Real introduces an agentic framework for transferring robot manipulation skills through a shared API rather than task-specific policy retraining. Its hierarchical memories and Proposer–Verifier–Governor loop deliver strong results on simulated benchmarks and frozen real-world skills.

A Florida woman allegedly wrote in Claude that she planned to attack a sheriff’s office. Anthropic’s safety system escalated the message to a human reviewer, who reported it to law enforcement.

A paper co-authored by Geoffrey Hinton, Yoshua Bengio and other leading researchers examines whether AI systems that help build the next generation of AI could create a self-reinforcing acceleration loop. The authors see early signals, but not enough evidence to claim that an intelligence explosion has begun.

ExecCert adds a release-time check for the finite-precision artifacts actually produced by an unlearning system. It can validate a method’s native guarantee or compare the released model with retraining on the current retain set.

DeskForge combines real applications, screenshots, accessibility trees, and window geometry to generate dense supervision for computer-use agents. Its reported gains extend from GUI grounding benchmarks to longer desktop task sequences.