Also Worth Noting - 2026-08-15

Also Worth Noting

02 [Eval] OpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution Static single-turn benchmarks cannot catch the failure mode that matters most: an agent's early state mutations corrupting decisions 20 steps later. OpenART addresses this directly with over 10,000 validated stateful scenarios across 50 environment types, each designed so cumulative state drift is the thing being measured, not isolated response quality. Open-ended environment evolution means the scenario space grows rather than saturating. Teams running agent safety evaluations on short, fixed tasks are measuring the wrong distribution entirely. link

03 [Agent] Agent Safety Should Be a Runtime Contract Training-time alignment cannot stop an agent from deleting files once it has tool access. The structural argument here is that RLHF, DPO, and Constitutional AI instill dispositions, not enforceable constraints, and dispositions fail under distribution shift or adversarial inputs. The proposed runtime contract has two faces: a preventive layer (sandboxes, permission gates, trajectory monitors) that blocks dangerous actions before execution, and an evidential layer that requires verifiable proof that sanctioned actions actually occurred. Any team deploying agents with file, database, or network access should treat the harness, not the weights, as the primary safety boundary. link

04 [Theory] Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus Massive activations in hybrid linear-attention models do not distribute randomly across layers -- they spike predictably just before full-attention layers, forming what this study calls pre-attention spikes, and can persist through intervening linear-attention layers as inter-spike plateaus. As full-attention layers become denser in a given architecture, successive spikes connect through plateaus until the pattern converges to the stable morphology seen in fully dense attention models. This organization holds across five linear-attention architectures. Quantization schemes that treat all layers uniformly will systematically destroy the spikes that carry the most representational weight. link

05 [Training] NeuPAT: Neuron-aware Plasticity Allocation Tuning for Language-Preserving MLLMs Multimodal fine-tuning degrades language ability not uniformly but through a small subset of neurons that carry disproportionate language load. NeuPAT identifies these high-stakes neurons by measuring heterogeneous plasticity during multimodal learning, then freezes them selectively while leaving more adaptive neurons free to absorb new perceptual knowledge. The method is architecture-agnostic and adds no inference overhead. Teams expanding pretrained LLMs with vision or audio modalities can apply NeuPAT as a drop-in tuning wrapper rather than accepting language regression as an unavoidable trade-off. link

06 [Application] MBA: Multimodal Benchmark and Agents for Real-World Business Ideation Text-only business ideation agents fail on roughly 40 percent of real-world contexts because real business inputs arrive as images, tables, and mixed documents, not clean prose. MBA-Bench covers 30,000 samples across six domains, each domain defined by visual cues that text captions cannot fully recover, and uses GPT-4o-generated reference ideas to score outputs across three business task types. It is the first benchmark that forces ideation agents to reason over image and text jointly rather than text reconstructions of images. Teams building LLM-powered strategy or market-analysis tools now have a concrete eval surface that reflects actual input diversity. link