HiVA: Self-Organized Hierarchical Agent Framework That Evolves Multi-Agent Structures from a Single Agent

Forum topic · 小凯 · 2026-06-25

Summary

HiVA (Hierarchical Variable Agent), proposed by researchers at Sun Yat-sen University (arXiv:2509.00189), is a multi-agent framework that grows organizational structure automatically from a single base agent, rather than requiring a hand-designed workflow. Its core algorithm, STEV (Semantic-Topological Evolution), treats the multi-agent system like a neural network: environment feedback acts as a loss function, and LLM-generated "textual gradients" drive simultaneous updates to each agent's prompts/tools (semantic evolution) and inter-agent connections (topological evolution). A Knowledge-Aware Bayesian Bandit (KABB) with Thompson Sampling dynamically selects execution subgraphs to control cost, while edge weights store collaboration reliability as a distributed, hierarchical memory. On Qwen-2.5-72B, HiVA outperforms the strongest baseline (MaAS) across GSM-8K, HotpotQA, 2WikiHopQA, HumanEval, MBPP, MMLU, and BBH, averaging +8.0%, with the largest gains on multi-hop reasoning. On the GAIA benchmark it reaches 26.2%/24.3%/11.1% across levels 1-3. Ablations show semantic and topological evolution are individually indispensable. Limitations include weak performance on tasks requiring strict logical consistency (e.g., MATH), slow multi-round evolution, and lack of convergence guarantees.

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

> Paper: HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution > Authors: Jinzhou Tang et al., Sun Yat-sen University

The Problem: A Dilemma in Existing Multi-Agent Frameworks

Current LLM multi-agent frameworks fall into two camps:

Fixed workflow (e.g., MetaGPT, AutoGen): modular and reusable, but the pipeline breaks when the task changes.
Reactive loops (e.g., ReAct, AutoGPT): flexible in unknown environments, but every task starts from scratch—no experience is consolidated into structure.

HiVA's core insight: a truly intelligent system should co-evolve what each agent does (semantics) and how agents connect (topology), starting from a single "single-cell" agent with a minimal prompt and toolset.

STEV: Evolution as Gradient Descent

STEV (Semantic-Topological Evolution) treats the entire multi-agent system as a network and environment feedback as a loss function, performing "gradient descent" in a discrete, non-differentiable space:

1. Environment diagnosis: feedback is not just "wrong" but a detailed report identifying which agent failed and how. 2. Backpropagation: this feedback acts as a global gradient, propagated from the aggregator back to each participating agent as targeted critique—like per-neuron local gradients. 3. Co-evolution: each agent simultaneously updates its semantics $$f_P$$ (prompts, tool configuration) and topology $$f_G$$ (add/remove/keep connections).

The optimization space is hybrid: a semantic space $\mathcal{P}_\Theta$ plus a topological space $\mathcal{G}$ , both discrete and non-Euclidean. The gradient signal is textual, generated by the LLM itself.

KABB: Exploration vs. Exploitation for Agent Selection

Waking every agent for every task would be prohibitively expensive. HiVA uses KABB (Knowledge-Aware Bayesian Bandit) with Thompson Sampling to select execution subgraphs, weighing:

1. Historical performance: $\frac{\alpha_i}{\alpha_i + \beta_i}$ (Beta posterior mean) 2. Task relevance via an external knowledge graph 3. Team collaboration gain among already-selected agents

Topology as Distributed Memory

Every edge $$(v_i, v_j)$$ carries a weight $C_{syn}(v_i, v_j)$ recording how reliable that information path has historically been, updated via:

C_{syn}^{(t+1)}(v_i, v_j) = C_{syn}^{(t)}(v_i, v_j) + \gamma \cdot \frac{\alpha_{ij}}{\alpha_{ij} + \beta_{ij}} \cdot \mathcal{R}_{ij}^{(t)}

This creates hierarchical memory:

| Level | Stored content | Structure | |-------|----------------|-----------| | Macro | Long-term collaboration patterns | Topology $\mathcal{G}$ | | Meso | Effectiveness of specific paths | Edge weights $w_{ij}$ | | Micro | Individual agent expertise | Semantic parameters $\Theta_i$ |

Experience is encoded in the organization's DNA—useful connections strengthen, useless ones get pruned.

Experimental Results

Main benchmarks (Qwen-2.5-72B-Instruct-Turbo)

| Task | Dataset | HiVA | Best baseline | Gain | |------|---------|------|---------------|------| | Math reasoning | GSM-8K | 94.5% | 94.1% (MaAS) | +0.4% | | Multi-hop QA | HotpotQA | 79.7% | 76.2% (MaAS) | +3.5% | | Multi-hop QA | 2WikiHopQA | 86.5% | 81.1% (MaAS) | +5.4% | | Code generation | HumanEval | 94.2% | 92.3% (MaAS) | +1.9% | | Code generation | MBPP | 92.1% | 90.1% (MaAS) | +2.0% | | Text reasoning | MMLU | 91.7% | 89.4% (MaAS) | +2.3% | | Complex reasoning | BBH | 93.4% | 90.6% (MaAS) | +2.8% |

Average improvement of +8.0%, strongest on multi-step reasoning tasks.

GAIA benchmark

| Method | Level-1 | Level-2 | Level-3 | Cost Score | |--------|---------|---------|---------|------------| | AutoGPT | 13.2% | 0.0% | 3.9% | 1.3 | | MaAS | 25.2% | 22.0% | 6.3% | 5.2 | | HiVA | 26.2% | 24.3% | 11.1% | 5.5 |

Ablation study

| Configuration | HotpotQA | MBPP | MMLU | Average | |---------------|----------|------|------|---------| | Full HiVA | 79.7% | 92.1% | 91.7% | 87.8% | | Remove TEV (topological evolution) | 74.0% | 88.9% | 88.3% | 83.7% | | Remove SEV (semantic evolution) | 71.2% | 88.4% | 86.9% | 82.2% | | Remove KABB | 76.2% | 88.1% | 90.6% | 85.0% | | Remove environment feedback | 75.2% | 89.3% | 89.5% | 84.7% | | Remove tools | 74.8% | 94.1% | 89.1% | 84.3% |

Key findings: semantic evolution matters most; topological evolution is critical for multi-hop reasoning (−7.3% on HotpotQA without it); both are jointly necessary. Notably, removing tools *improved* MBPP by 2.2%—tools are not always better on simple tasks.

Evolution trajectories

Success (HotpotQA): a single general agent (67%) split into retrieval + reasoning agents (round 3), then added a fact-checking agent (round 6), converging to a three-node pipeline at 79%.
Failure (MATH): agents produced contradictory intermediate results and the aggregator deadlocked—tasks demanding strict logical consistency need stronger aggregator design.

Limitations and Outlook

Limitations: underperformance on strict-logic tasks (81.2% on MATH vs. 82.7% for vanilla); multi-round evolution (10 rounds used) is slow for real-time use; textual gradients themselves cost LLM calls; no theoretical convergence guarantee.

Future directions: combining textual gradients with RL algorithms (PPO, GRPO); cross-task topology transfer; human-in-the-loop evolution for expert domains like medicine and law.

Takeaway

HiVA demonstrates that intelligence lies not in how strong individual neurons are, but in whether they can grow the right connections on their own—from single cell to multicellular, evolution needs no blueprint, only feedback.

Reference [1] Tang, J. et al. (2025). HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution. arXiv:2509.00189.