Researchers propose MANE, a multi-path adaptive network designed to improve split computing, where a small head model runs on a device and a larger tail model runs on an edge server. The approach targets efficient distributed inference by adapting how computation is divided between the device and the edge. It is described in a new arXiv preprint (2609.14660v1) listed under cross-submissions.
Researchers introduce X-CoSD, a distributed inference framework in which a small on-device model drafts tokens and a larger server-side LLM verifies them. The method tackles the mismatch between the two models' vocabularies while cutting the communication cost between device and server. The work aims to make collaborative speculative decoding practical for latency-sensitive edge deployments.
A new arXiv paper addresses how to monitor distributed LLM-agent workflows at runtime, where asynchronous execution means a decision can only rely on events causally visible to the agent making it. The authors introduce a causal past logic that treats events appearing earlier in a log as not necessarily available to every lifeline, and use it as the basis for runtime verification. The work targets correctness checking of multi-agent pipelines as they execute rather than only after the fact.