Decentralized Multi-Agent Systems with Shared Context

Summary

DeLM replaces a main multi-agent controller with agents that share a verified problem state and take work from a task queue. It targets coding agents and long-context QA, where controller bottlenecks can waste test-time compute and lose useful intermediate findings.

Problem

Most multi-agent LLM systems route subagent work through one main agent, so communication and result merging become serial bottlenecks as the number of subtasks grows.
In software repair, isolated attempts repeat failed searches and cannot reuse partial fixes; in long-context QA, a main agent may assign evidence chunks before knowing which evidence matters.
This matters because test-time scaling only helps when extra agents create reusable progress, rather than extra calls that repeat work or lose details during merging.

Approach

DeLM uses parallel agents, a shared verified context, and a task queue. Agents claim queued subtasks asynchronously, read current shared progress, reason locally, and write back compact updates.
The shared context stores gists: short entries about facts, failed hypotheses, constraints, source evidence, and partial solutions. Agents can unfold a gist into detailed summaries or raw evidence when needed.
Before an update enters the shared context, an LLM verifier checks it against its source evidence or reasoning trace. Failed updates are rejected or regenerated.
When the queue is empty, the most recently completed agent checks the shared state, adds more subtasks if needed, or produces the final answer.

Results

On SWE-bench Verified with Gemini 3 Flash, DeLM reports 65.7% Avg.@1, 72.9% Pass@2, and 77.4% Pass@4. The strongest listed baseline on Avg.@1 is AOrchestra-Parallel at 56.4%, so the gain is 9.3 percentage points.
On SWE-bench Verified with Gemini 3 Flash, DeLM reports $0.12 cost per task, compared with $0.24 for AOrchestra, $0.25 for AOrchestra-Parallel, and $0.26 for mini-SWE-agent.
On SWE-bench Verified with Claude Opus 4.6, DeLM reports 78.0% Avg.@1, 80.7% Pass@2, and 82.5% Pass@4, ahead of mini-SWE-agent at 76.9%, 79.8%, and 81.7%.
On LongBench-v2 Multi-Doc QA, the paper says DeLM has the highest average accuracy across GPT-5.4, Claude Sonnet 4.6, Gemini 3 Flash, and DeepSeek-V4-Pro, with gains up to 5.7 percentage points over the strongest baseline.
The LongBench-v2 setting contains 125 samples: 15 financial, 23 government, 23 multi-news, 14 legal, and 50 academic.
On OOLONG, the excerpt gives no table values. It says vanilla DeLM trails RLM on exact row-level aggregation, while RLM plus DeLM gets the best accuracy and lowest cost.

Link

https://arxiv.org/abs/2606.10662v1