llm-acting-orchestrator-dynamically-created-subagents-summarizing-own-past CLM-0411observationsingle paperReviewed 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.
Capability: Getting more from several models than from one
Observed on
Hierarchical swarm setups with one orchestrator, a fixed pool of subagent backbone models, and parallel subagent execution; measured on four tasks (multi-text understanding, root c.
Sources
- Validated by: Task accuracy against the same swarm framework without experience injection, plus single-agent, Best-of-N, multi-model voting and AOrchestra baselines: +6.42 on RCA, +6.46 on LTG, +2.92 on WS, +0.25 on MTU; transfer of GPT-5.4-derived experience improved 14 of 16 model-task settings; ablation drops -2.49 avg without model cards, -1.75 without tricks, -1.22 without skills.. Measured on 8 tasks/400 samples, with the technique tested on only 4 tasks and one primary orchestrator; gains are small on one task and negative on two settings for one transfer model. [truncated]
Either way, bring evidence: a link, or what you saw.
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
- In 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.Getting more from several models than from one · reviewed by AI
- For LLM web agents on cluttered pages containing many concurrent interface components, success falls sharply as component count grows, and having the agent first emit a localized sub-goal plus a task-relevant sub-observation before choosing an action recovers much of the loss, because self-assessment of the action is more reliable in the reduced context.Using the tools it is given · reviewed by AI
- When an agent builds a whole Python repository from natural-language requirements alone, refining component boundaries with explicit cohesion and coupling thresholds until no split or merge is triggered produces higher functionality coverage and test pass rate than letting the LLM decide restructuring actions for a fixed budget of rounds, because unconstrained LLM-decided actions keep over-decomposing past the point of a coherent modular structure.Generating and editing working code · reviewed by AI
- For LLM-driven multi-UAV swarm tasks in simulation, binding one agent to each vehicle with a role-conditioned subset of the skill catalog and directed peer messages yields higher episode returns than a central planner that sees pooled team state and the full skill list, and ablations attribute most of the gain to the skill-subset restriction and the peer reports rather than to the experience-based reward correction.Turning plans into physical action · 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: low. Falsifier as drafted: Injecting the extracted experience fails to improve orchestrator task scores over the same framework without it, or removing the model cards costs no more than removing skills or tricks, on comparable swarm tasks.