Research map / Multi-agent systems and games / LLM agents

Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LATTE, a framework where LLM agent teams jointly build and maintain a shared dynamic task graph with formal mutation operators, and evaluates it against Leader-Worker, decentralized, static-graph, and MetaGPT teams on three coding/data tasks.

Problem

Existing LLM team designs are either rigidly structured (fixed roles or pre-specified decompositions) or unstructured (decentralized), leading to either fragility or coordination failures such as overwrites, conflicts, redundant output, and excess tokens/time.

Approach

Agents collaboratively construct and evolve a shared coordination task graph using seven mutation operators (Discover, Assign, Claim, Complete, Release, Close, Verify); Workers propose local graph changes and a Lead accepts/rejects them, with self-scheduling via frontier claims.

Key results

• LATTE reaches 80% aggregate accuracy vs 70% Leader-Worker, 74% decentralized, 58% static graph, and 34% MetaGPT across successful trials.
• LATTE uses a mean token cost of 47.5% of the next-best method (static graph, 86.9%) in normalized pooled costs, and 148K tokens aggregate vs 379K for Leader-Worker.
• LATTE agents overwrite each other 4.3 times per trial vs 22.8 for Leader-Worker and 35.4 for decentralized teams.

Surprising findings

• Static graph ablation matches LATTE accuracy on the Library Extension task (40% each), where required functions are known in advance.
• MetaGPT achieves only 34% aggregate accuracy and 6% on Library Extension, far below the other team structures.

Limitations

• Planning overhead from graph initialization may outweigh benefits on short or simple tasks.
• Evaluation limited to domains with natural subtask boundaries such as coding; team size fixed at N=4/5.
• Future work suggested on RL/fine-tuning over coordination outcomes.

Experimental setup

Datasets: Exploratory data analysis task (private test suite) Debugging task on signal-processing library Library extension task on Python text-processing library

Benchmarks: Exploratory Data Analysis task Debugging task Library Extension task

Metrics: accuracy (test pass rate) total tokens wall-clock time overwrite rate concurrent writes wasted characters inter-agent messages idle rounds straggler tail latency

Baselines: MetaGPT Leader-Worker hierarchy Decentralized teams Static task graph ablation

Builds on: MetaGPT HEFT NABBIT Metropolis-Hastings

Closest papers by meaning

• When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding (NeurIPS 2026 · similarity 0.90)
• AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.90)
• AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.90)
• COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
• MoRSE: Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts (NeurIPS 2026 · similarity 0.89)
• TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination (ICML 2026 · similarity 0.89)