llm-agents-must-compose-then-execute-domain-tool-workflows CLM-0416For 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.
Capability: Planning a sequence of actions before acting
Observed on
Small open-weight backbones (Qwen3.5-4B/9B) on Earth-observation agent tasks with expert-authored skills constraining the tool space; measured only in this framework..
Sources
- Validated by: On Earth-Bench-Pro, the 2x2 adapter composition: base/base 38.31% LLM-as-Judge, planner-tuned only 44.76%, executor-tuned only 43.15%, both 50.00%; planning-only evaluation raised TIO from 71.87 to 83.35, and oracle-plan executor evaluation raised judge accuracy from 46.37 to 52.82 and parameter accuracy from 34.19 to 37.20.. Measured on 248 open-world EO tasks with one model family (Qwen3.5-4B/9B) and one framework; isolation is reasonably clean because the paper adds planning-only and oracle-plan evaluations on top of the 2x2 adapter grid. [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 under a fixed ReAct harness answering questions over executable database-backed APIs plus document retrieval, accuracy falls sharply as the required reasoning depth grows from 1 to 5 hops (most models lose over half their accuracy), and the surviving failures concentrate in language-mediated steps — entity disambiguation, cross-source grounding, schema alignment — rather than in emitting syntactically valid tool calls.Carrying a chain of inference through to its conclusion · 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
- In multi-agent LLM pipelines over reasoning and code tasks with a pool of 3B-32B open-weight backbones, conditioning each per-step routing decision on a learned filtered memory rather than the full accumulated trajectory leaves accuracy about the same while cutting inference cost by roughly 40-50%, so most of the accumulated history is cost without accuracy value.Getting more from several models than from one · 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
- 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
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 where planner-only tuning improves argument grounding as much as executor-only tuning (or vice versa), or where a single monolithic trajectory-level fine-tune matches the combined two-adapter configuration on both workflow-order and answer metrics.