llm-multi-agent-systems-whose-communication-graph-produced-learned-topology CLM-0599In LLM multi-agent systems whose communication graph is produced by a learned topology generator, the generated graph carries edges that are not needed: scoring each edge by how much masking it changes the task outcome and the semantic entropy of the final answer, then keeping a top-ranked budgeted subgraph, cuts online tokens by about 20-25% while leaving average accuracy equal or slightly higher, and a small distilled predictor of those subgraphs transfers to generators and agent counts it was not calibrated on.
Capability: Getting more from several models than from one
Observed on
Qwen3-8B backbone, 5-6 agent systems, four topology generators (G-Designer, ARG-Designer, OFA-MAS, AgentPrune), six math/knowledge/code benchmarks; 40 calibration queries per datas.
Sources
- Validated by: Accuracy and measured online token usage versus each unpruned generator across 24 generator-dataset cells: tokens fell in all 24 (10.6%-44.0%), accuracy rose in 16; ablations show removing the causal masking signal costs about 4 accuracy points on MMLU and HumanEval and most of the token saving, and removing semantic entropy costs about 2-3 points.. Measured on six benchmarks with one backbone; token savings are consistent across all 24 cells while accuracy changes are mixed per dataset (GSM8K, MultiArith, SVAMP sometimes drop). [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 classifying nodes in text-attributed graphs, serializing sampled neighborhoods into prompt text makes the prediction depend on the arbitrary order of neighbors; replacing that text with a fixed-size, permutation-invariant vector of graph statistics (degree, PageRank, clustering, cross-community fraction, label entropy, majority-label fraction) plus only the top-k semantically relevant node texts raises accuracy and cuts per-message tokens.Getting more from several models than from one · reviewed by AI
- In multi-agent LLM systems where one agent starts with an erroneous shared belief, broadcasting every agent's full context at every step raises the rate of false assertions above doing no synchronization at all, because the error propagates to agents that were previously correct — and the harm appears only in tasks where one wrong fact cascades across semantically linked dimensions (destination to airport to weather), not where agent contexts are largely orthogonal.Stating false facts confidently · 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
- 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 LLM multi-agent systems that reflect after failure, restricting reflection to the single agent responsible for the first incorrect step — located by a trained step-level correctness scorer and paired with a counterfactually generated corrected step — raises task success more than having every agent reflect on the trajectory, because reflections written by agents that behaved correctly enter their memory as misleading guidance for later tasks.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: A replication where the same budgeted subgraphs are compared against random edge masking at equal retained-edge budget and show no accuracy or token advantage, or where the distilled explainer applied to an unseen generator lowers accuracy relative to the unpruned graph.