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

HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 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 HiVA, a multi-agent framework that co-evolves agent prompts/tools (semantics) and inter-agent graph structure (topology) using textual gradients and Bayesian-bandit routing, reporting gains over workflow and multi-agent baselines on QA, coding, math, and GAIA.

Problem

Existing LLM agent systems either use fixed manually designed workflows that need reconfiguration when environments change, or reactive loops that do not distill reasoning into reusable structures; prior optimizers treat agent semantics and coordination topology independently.

Approach

Model an agentic workflow as a DAG of LLM agents optimized in a hybrid semantic-topological space via textual gradients (LLM-generated feedback propagated backward), with Thompson-sampling Bayesian-bandit routing that builds task-specific execution subgraphs, so prompts/tools and edges are updated jointly.

Key results

• HiVA reaches 89.2% average accuracy across eight benchmarks on Qwen2.5-72B, +8.0% over Vanilla and above MaAS (87.4%).
• On MATH, HiVA scores 81.2%, 1.8% below Vanilla (82.7%), the only benchmark where it underperforms the baseline.
• On GAIA, HiVA reports Level-1 26.2%, Level-2 24.3%, Level-3 11.1% accuracy and average cost-efficiency 5.5 vs MaAS 5.2.

Surprising findings

• HiVA underperforms the Vanilla baseline on MATH (81.2% vs 82.7%), which the authors attribute to the aggregator getting stuck on conflicting agent answers.

Limitations

• Performance drop on MATH where conflicting agent outputs confuse the aggregator.
• Optimization overhead scales with agent count (worst-case O(|V|^2)) and iteration count.
• Tool-calling is limited; authors note future work needed for more operable tool-calling methods.

Experimental setup

Datasets: MATH GSM-8K HotpotQA 2WikihopQA HumanEval MBPP MMLU BBH GAIA

Benchmarks: MATH GSM-8K HotpotQA 2WikihopQA HumanEval MBPP MMLU BBH GAIA

Metrics: accuracy cost-efficiency score standard deviation across runs

Baselines: CoT Self-Consistency Self-Refine Multi-Agent Debate DyLAN AgentVerse ADAS MaAS AutoGPT TextGrad

Builds on: TextGrad MaAS ADAS KABB

Closest papers by meaning

• Tools as Continuous Flow for Evolving Agentic Reasoning (NeurIPS 2026 · similarity 0.89)
• AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.88)
• SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents (NeurIPS 2025 · similarity 0.88)
• A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning (ICLR 2026 · similarity 0.87)
• PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.87)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.87)