AI News Today, Filtered for What Matters

What happened

BAAI introduced AREX-2, a 27B agent model trained with long-horizon improvement trajectories from machine-learning and algorithmic-programming tasks. Model weights and evaluation code are public; the paper reports transfer to deep-research tasks.

Corroborating sources

What happened

Researchers introduced MoFlow, a method that searches agent workflows across accuracy, cost, latency, robustness and consistency. Its public implementation lets users select a workflow for different preferences from a saved search tree without retraining.

Corroborating sources

What happened

Researchers introduced ScopeBench, a 30-task pilot benchmark that tests whether autonomous security agents pursue goals without crossing stated engagement boundaries. They released the tasks, evaluation code and 2,160 trajectories.

Corroborating sources

What happened

UK AISI and EvalEval have released verified results, setup details and context in Evaluation Cards for five benchmarks from an earlier AISI study, alongside two related cyber evaluations.

Corroborating sources

What happened

NVIDIA researchers introduced SWE-Serve, a benchmark of 53 SGLang inference-engineering tasks. Across 19 tasks with live-serving checks, patches passed 45.9% of the time with the complete verifier versus 69.4% when those checks were excluded.

Corroborating sources

What happened

Anthropic reported that Claude agents identified a reverse-transcriptase system associated with repeating DNA sequences. Its researchers tested early properties in the lab, but the system's biological function remains unknown.

What happened

NVIDIA reported that TensorRT Edge-LLM ran Qwen3.6-27B through the 1,007-turn MLPerf Inference v6.1 Edge Agentic workload on one Jetson AGX Thor in 24 minutes and 36 seconds, 6.4 times faster than the published llama.cpp reference run on the same hardware.

What happened

OpenAI published a framework for documenting observed model-misalignment cases and released six initial reports. The format records the triggering context, behavior, impact, mitigations, and limits of each observation; individual cases are not estimates of how often the behavior occurs.

What happened

Google Research introduced Retrieve-for-Train, which uses offline reinforcement learning to create fan-out training data and distills it into a 53.9-million-parameter diffusion retriever. In the reported retrieval tasks, the approach generated result sets 12 to 20 times faster than autoregressive alternatives.

What happened

OpenDiscoveryTrace released a public dataset of 558 complete agent trajectories across 124 scientific tasks. Each step records structured fields such as tool calls, observations, errors, revision triggers, and confidence across several scientific domains and models.

What happened

arXiv:2608.20341v1 Announce Type: new Abstract: Frontier coding agents backed by large language models with context windows from hundreds of thousands to millions of tokens are restructuring the Software Development Life Cycle (SDLC). Rich context handling and multi-step reasoni

What happened

NVIDIA reported that its Agentic Variation Operators architecture reached a 100.00 RHAE score on the ARC-AGI-3 public set, completing all 183 levels across 25 environments while using persistent memory, supervision, tool use, and a long-horizon execution loop.

What happened

arXiv:2608.12345v1 Announce Type: new Abstract: Language models are increasingly deployed as co-scientists, yet their ability to uphold research integrity under institutional pressure remains unmeasured. We introduce IntegrityBench, a benchmark evaluating misconduct classificati

What happened

arXiv:2608.05151v1 Announce Type: new Abstract: Wastewater operators need answers grounded in how their plant's variables interact and how fast effects propagate, not in generic pretraining text, when asking causal questions such as "why is N2O rising?" or "what happens if I cut

What happened

arXiv:2607.28629v1 Announce Type: new Abstract: The rapid transition from reactive large language models (LLMs) to persistent, action-capable systems has exposed critical gaps in the architectural understanding of Agentic AI, particularly in separating inference, orchestration,

What happened

OpenAI published ten results in mathematics and theoretical computer science that it says were generated by an internal version of Astra, then prepared into manuscripts by humans and formalized with Lean certificates. The problems span areas including sphere packing, coding theory, group theory, circuit complexity, quantum games, lattice cryptography, Ramsey theory, and extremal graph theory.

What happened

NVIDIA published guidance on co-designing model attention and deployment choices for fast, interactive long-context inference as agentic workloads increase context length.