Long-context models may not need to scan their entire KV cache every time they answer. Declarative Attention has the model announce whether it needs the full context, a specific region, or only recent output, cutting attended tokens by 52.0% on Gemma-4-31B and 31.1% on Qwen-3.6-27B across 15 tasks, with accuracy drops of 1.27 and 2.75 percentage points. The tradeoff is promising, but training could make the model’s declarations safer and more decisive.
A benchmark score can look precise while blending unrelated skills. AllenAI’s BenchMIRT analyzed results from 100 LLMs across 16 benchmarks and more than 34,000 questions, recovering separate safety and general-reasoning signals without being told the intended categories. That makes it possible to identify which prompts drive a score, and which are quietly testing something else.
Molecular pretraining often creates chemically unrealistic “views” by masking atoms or bonds, then teaches the model invariances that fail at property cliffs. Mol-JEPA takes a different route: mask whole modalities, then predict their latent representations across structures, cellular phenotypes, binding, ADMET, and quantum chemistry data. The paper reports strong results across benchmarks, suggesting biochemical context may matter more than another clever molecular augmentation.
AI alignment research is getting a new gatekeeper, but not a conventional one. The Alignment Journal is building a fast, high-rigor venue with an eight-person senior editorial board, public reviewer abstracts, expert matching, and paid reviews, then plans to open submissions more broadly in October. That makes peer review itself part of the experiment: can better incentives and clearer editorial judgment surface useful work without importing academia’s slow, opaque machinery?