multi-agent-literature-synthesis-pipelines-most-gain-over-fixed-role-based-pipeline CLM-0445In multi-agent literature-synthesis pipelines, most of the gain over a fixed role-based pipeline comes from the shared-workspace architecture and the added critique/dissemination actions rather than from learning the coordination policy: on OARelatedWork, a learned decentralized RL controller scored 3.82 overall versus 3.80 for a centralized LLM coordinator given the same workspace and action space, while a naive orchestrator without that architecture scored 3.40.
Capability: Getting more from several models than from one
Observed on
Related-work generation from a fixed candidate reference pool, Qwen 2.5 72B agents, LLM-as-judge 1-5 scores averaged over three judges; the RL policy's separate advantages are lowe.
Sources
- Validated by: LLM-as-judge quality, citation verification rate, tokens and inference time on OARelatedWork (366 eval papers) against SRW, a centralized orchestrator with the same architecture, a naive orchestrator, GPT-4o long-context/RAG and PRIMERA; plus transfer to other backbones and to Multi-News.. Measured on one benchmark with one agent count for the main table; the comparison is clean because the orchestrator baseline was built to inherit CREW's architecture, but the 0.02-point margin is within judge noise and no variance is reported. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For LLM agents that must compose and then execute domain tool workflows, training separate adapters for the two roles splits the gains cleanly: supervised fine-tuning of the planner on reference tool sequences mainly improves workflow-order fidelity, while node-level RL on argument-match rewards mainly improves tool-argument grounding and final-answer quality, and combining both beats either on a 9B backbone.Planning a sequence of actions before acting · reviewed by AI
- For teams of up to 50 heterogeneous embodied LLM agents in simulated wildfire missions, building the manager hierarchy per mission so that concurrently executable work sits under a parallel-allocating manager and prerequisite-chained work under a phase-advancing manager beats fixed decentralized, single-planner, and hybrid coordination architectures on mission score and on progress accumulated per step, and the advantage does not depend on which base language model is used.Getting more from several models than from one · reviewed by AI
- In multi-agent LLM workflows with a mixed-size model zoo, choosing which model instantiates each agent step-by-step from an online estimate of task-progress gain plus remaining time and cost budget keeps task quality near the best single-model policy at lower energy than one-shot query-level routing or stall-triggered escalation.Getting more from several models than from one · reviewed by AI
- For an LLM acting as orchestrator of dynamically created subagents, summarizing its own past trajectories into reusable workflow skills, short operational tricks, and capability cards for the candidate subagent backbones, then injecting those into the orchestrator's context, raises task scores on delegation- and aggregation-heavy tasks, and the subagent model cards are the component whose removal costs most.Getting more from several models than from one · reviewed by AI
- When automatically rewriting the prompts of a multi-agent LLM system from its failures, locating the responsible agent by injecting a hint into one agent at a time and using that agent's hint-corrected intermediate output as the supervision signal produces prompt updates that generalize better than critique derived only from the system-level output.Improving its own scaffolding from failures · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: An ablation that keeps the shared workspace and four-action space but replaces the learned policy with round-robin or random action selection and loses a large share of the quality gain would show the policy, not the architecture, carries the effect.