An Axiomatic Framework for Reinforcement Learning in Open Multi-Agent Systems
Theory work on RL when only some agents are controllable; least practical item in the cluster.
12 items · 4 sources · 9 days
Operational story trace
Follow in this browser to see new updates on your Live feed.
Latest change
Two new papers push from diagnosis to repair: TRACE-MAS (Nov 17) diagnoses and repairs collaboration failures between agents, and Kaer (Nov 18) adds security-aware replanning to orchestrators that revise plans mid-run.
This is a loose research cluster, not one event: early November papers covered RL theory, orchestration benchmarks, and KV-cache memory for co-resident agents. From Nov 4 the work tilted toward securing multi-agent systems and attributing their failures, from NIST identity tenets to CoLFA and AgentLeak.
Arc
Theory work on RL when only some agents are controllable; least practical item in the cluster.
Benchmark for harnesses that spawn worker agents to run executable workflows.
Only standards-body source: NIST design tenets for agent identity management.
Adds private retrieval and private inference to the multi-agent security discussion.
First runnable tool in the thread: open-source privacy testing for multi-agent systems.
Defends aggregation of subtask answers against malicious agents.
Benchmark for coding agents designing LLM serving systems; adjacent, not multi-agent.
Names KV-cache memory of co-resident agents as a bottleneck and quantizes them jointly.
Production case study on latency and distillation; tangential to multi-agent work.
Lightweight method for attributing a failure to the step and agent that caused it.
Moves past detection: targets interaction-level failures and proposes repairs.
Brings security checks into the orchestrator's replanning loop.
Theory work on RL when only some agents are controllable; least practical item in the cluster.
Only standards-body source: NIST design tenets for agent identity management.
Benchmark for coding agents designing LLM serving systems; adjacent, not multi-agent.
Names KV-cache memory of co-resident agents as a bottleneck and quantizes them jointly.
Benchmark for harnesses that spawn worker agents to run executable workflows.
Adds private retrieval and private inference to the multi-agent security discussion.
Lightweight method for attributing a failure to the step and agent that caused it.
First runnable tool in the thread: open-source privacy testing for multi-agent systems.
Defends aggregation of subtask answers against malicious agents.
Production case study on latency and distillation; tangential to multi-agent work.
Moves past detection: targets interaction-level failures and proposes repairs.
Brings security checks into the orchestrator's replanning loop.
What to watch — open questions
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.