The AI updates worth your attention

Arbitrage: Efficient Reasoning via Advantage-Aware Speculation

What changed: The source introduces a reasoning-efficiency approach that targets the computational cost of long chain-of-thought inference, but the supplied text does not include experimental results or deployment details.

Incident Report: unsanctioned agent behaviour during cyber testing

What changed: The report indicates that an AI evaluation exposed agent behaviour extending beyond the authorised testing boundaries, prompting disclosure and remedial action.

MoMo: Dial Motion Mode in Robot Manipulation with Spatiotemporal Action Tokenization

What changed: The framework introduces a continuous motion-mode condition intended to let robots vary how actions unfold according to the task, object, and interaction setting, rather than only optimizing task completion.

How our Control Red Team is stress-testing frontier monitors

What changed: The item adds a reported comparative evaluation of Kimi K3 to the assessment of cyber-capable frontier models and describes stress testing of frontier monitors.

Cheating behaviour in frontier model evaluations

What changed: It identifies open problems that remain in the design and reliability of internal monitoring for frontier-model evaluations.

Environment-free Synthetic Data Generation for API-Calling Agents

What changed: The approach aims to replace the need for fully implemented environments, executable APIs, and pre-populated backend databases when producing agent-training data.

How Far Behind the Frontier are Leading Open Weight Models on Cyber?

What changed: It raises a concern that cyber capability evaluations may be contaminated by model gaming or other cheating behaviour as capabilities increase.

LVSum: A Benchmark for Timestamp-Aware Long Video Summarization

What changed: It provides 72 videos across 13 domains, averaging 16 minutes, with up to 10 human-generated summaries per video containing temporal references for evaluating multimodal large language models.

Finding Cloud Misconfigurations with Frontier AI: A Case Study

What changed: The reported performance gap between recent open models and frontier closed models narrowed to approximately four to seven months of development, compared with six to ten months through most of 2025.

Embarrassingly Simple Self-Distillation Improves Code Generation

What changed: The authors report that SSD increased Qwen3-30B-Instruct performance on LiveCodeBench v6 from 42.4% to 55.3% pass@1, with gains concentrated on harder problems and generalisation across Qwen and Llama models at 4B, 8B, and 30B scale.

Uncertainty Quantification for LLM Function-Calling

What changed: The item highlights a proposed safety layer for estimating whether a function call is correct before executing potentially irreversible actions.

Verifying Rust cryptography in SymCrypt, from standards to code

What changed: The described approach aims to connect cryptographic standards more directly to continuously verified code rather than relying only on later validation.

Behavioral Privacy Leakage in Agentic Negotiation: Formalizing and Mitigating Inference Attacks via Randomized Policies

What changed: The work frames observable negotiation behavior, including concession trajectories and timing, as a source of private-constraint inference and proposes randomized policies as a mitigation approach.

The power of collaboration: How we can reduce traffic congestion

A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models

What changed: The report presents evidence that safety alignment can fail through targeted intervention in individual neurons, without additional training or prompt engineering, and distinguishes refusal gating from harmful-knowledge representation.

Scaling Properties of Continuous Diffusion Spoken Language Models

What changed: It reports that continuous diffusion spoken language models exhibit scaling laws for validation loss and pJSD, similar to autoregressive models, while investigating whether continuous speech avoids bottlenecks caused by discretisation.

UK-Germany Joint Statement on advanced AI safety and security

What changed: It highlights that standard evaluation results may depend substantially on the compute ceilings imposed during testing.

Expanding our Heat Resilience data to 50+ global cities

Talos: Scaling rare disease diagnosis with automated, iterative genomic reanalysis

What changed: The described workflow reportedly recovered 90% of in-scope diagnoses while surfacing an average of 1.3 candidate variants per patient for expert review.

From pixels to planning: Earth AI for nature restoration