Russian-linked hackers targeted iOS and Android, researchers say

Recent work establishes that fusing adapters into low-precision GEMMs cuts serving latency for multiple PEFT variants. EpiLoRA, a vLLM plugin, achieves up to 2x reduction on NVIDIA Hopper GPUs, while Scute from Rice University extends this to five PEFT methods in a single batch.

The accumulated papers collectively establish that the field has converged on reusing and reasoning over past execution traces, particularly salvaging knowledge from failures, as the primary engine for agent self-improvement without external labels, demonstrated across work from Renmin University of China, Beihang University, and UIUC. The line of inquiry has most recently moved toward diagnosing the fragility of these gains: Shanghai Jiao Tong University's SelfImprove-Bench shows performance drops sharply on harder tasks, Cursor Research and UC Berkeley expose how rare behaviors remain neglected by RL, and a cross-model audit reveals divergent policies on when agents should stop self-editing.

The accumulated papers establish that agent memory is a dynamic, curatable resource rather than a passive log: updated permission records beat long histories, selective forgetting and compression improve task accuracy, and shared memory can bottleneck robot fleets or push-based agent societies. Recent work at UIUC, Amazon, Alibaba, Google, UCSD, and MIT moves the line from static memory design to trace-driven evolution and repair, such as converting execution records into reusable instructions, rebuilding user tasks after mistakes, and filtering inboxes under load, while also exposing vulnerabilities where memory-persistent agents follow inserted instructions across sessions.