Also Worth Noting - 2026-07-10

Also Worth Noting

02 [Eval] UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks Sandbox benchmark scores and real-world deployment performance diverge sharply, and UniClawBench is the first benchmark built to expose that gap directly. It evaluates proactive agents on multi-turn, real-world tool use rather than single-turn sandboxed scenarios, and its capability-separated task taxonomy isolates which specific skill fails rather than blending failures into one opaque score. Teams running agent evals who find their numbers look better in testing than in production now have a structured diagnostic to explain why. link

03 [Inference] Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE Fixed RoPE rescaling forces a bad trade-off: aggressive factors break short-context fidelity, conservative ones collapse at long range. Jet-Long bypasses that trade-off entirely with a tuning-free bifocal RoPE that dynamically splits attention between near and far position ranges, achieving zero-shot extension to 10x the pretraining window with lower perplexity degradation than YaRN or LongRoPE on needle-in-haystack tasks. No fine-tuning required, which makes it a direct drop-in for teams deploying open-weight checkpoints on long agentic traces. link

04 [Agent] Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents Decision-relevant facts buried past the effective context window do not just go unread, they stop influencing behavior entirely. This failure gets a name here: behavioral state decay. A separate memory agent runs alongside an unmodified action agent, updating a structured memory bank from the running trajectory and surfacing relevant facts before they are needed, not after retrieval is triggered. On 50-step benchmarks, this proactive design cuts task failure rate measurably, and it requires no modifications to the action agent itself. link

05 [Theory] Linear Attention Architectures: Mechanisms, Trade-offs, and Cross-Layer Routing Four linear-attention architectures, DeltaNet, Gated DeltaNet, Kimi Delta Attention, and Gated DeltaNet-2, have never been compared on the same axis until now. Expressing all four in a unified recurrent-memory notation makes explicit how they differ in erase-and-write control, memory decay, and expressivity at matched compute. At 350M parameters, Gated DeltaNet-2 dominates on long-context recall while costing the same FLOPs as vanilla linear attention. Teams evaluating linear-attention replacements for softmax attention now have a common framework rather than four incompatible paper comparisons. link

06 [Application] DrugGen 2: A disease-aware language model for enhancing drug discovery Target-only molecule generation ignores how disease context shifts target behavior, and that omission causes systematic failures on off-target disease settings. DrugGen-2 conditions small-molecule generation on both disease ontology embeddings and target protein sequences, fine-tuned from GPT-2 on a curated dataset of approved drugs linked to diseases and targets using a two-step supervised fine-tuning strategy. Dual conditioning improves binding affinity predictions precisely in the off-target disease contexts where single-target models break down most predictably. link