HiVA: How AI Agents Self-Organize from Single Cells into Multi-Agent Hierarchies

Forum topic · 小凯 · 2026-06-25

Summary

HiVA (Hierarchical Variable Agent), a paper by Jinzhou Tang et al. from Sun Yat-sen University (arXiv:2509.00189), proposes that LLM multi-agent systems can grow from a single base agent into complex organizational structures through evolution. Unlike fixed-workflow frameworks (MetaGPT, AutoGen) or reactive loops (ReAct, AutoGPT), HiVA co-evolves both semantics (each agent's prompts and tools) and topology (how agents connect) via STEV, a Semantic-Topological Evolution algorithm that treats the agent system like a neural network, using LLM-generated textual gradients from environment feedback as a substitute for numerical backpropagation. A Knowledge-Aware Bayesian Bandit (KABB) selects execution subgraphs for efficiency, and edge weights act as distributed memory encoding experience in the organizational structure. On Qwen-2.5-72B-Instruct-Turbo, HiVA outperforms the strongest baseline (MaAS) by about 8.0% on average across GSM-8K, HotpotQA, 2WikiHopQA, HumanEval, MBPP, MMLU, and BBH, and leads on all three GAIA levels. Ablations show both semantic and topological evolution are essential. Known limitations include weak performance on strict logical-consistency tasks like MATH and multi-round evolution costs.

> Paper: HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution > Authors: Jinzhou Tang et al., Sun Yat-sen University > Link: https://arxiv.org/abs/2509.00189

This post is a Feynman-style explainer of the HiVA paper from the zhichai.net forum, arguing that multi-agent systems can emerge from a single cell into multicellular structures via evolution rather than design.

Key points

The core question: Can a multi-agent system self-organize from one base agent (minimal prompt + basic tools) into a complex hierarchy, driven purely by environment feedback — like biological evolution?
Critique of existing paradigms:
*Fixed workflow* (MetaGPT, AutoGen): modular and reusable, but brittle across task domains.
*Reactive loop* (ReAct, AutoGPT): flexible, but every task starts from zero; no experience crystallizes into structure.
HiVA's insight: a truly intelligent system should co-evolve what each agent does (semantics) and how agents connect (topology).

STEV: evolution as gradient descent

HiVA's Semantic-Topological Evolution (STEV) treats the whole agent system as a neural network and environment feedback as a loss function, optimizing a discrete, non-differentiable space:

Semantic space $\mathcal{P}_\Theta$ : prompts, tool configs, behavior policies of each agent.
Topological space $\mathcal{G}$ : who connects to whom and how information flows.

The optimization signal is a Textual Gradient with layered feedback:

1. The environment returns a detailed diagnostic report on failure (not just "wrong"). 2. This global feedback is backpropagated from the aggregator to each participating agent as targeted criticism — analogous to local gradients per neuron. 3. Each agent simultaneously performs semantic evolution $$f_P$$ (update prompt/tools) and topological evolution $$f_G$$ (add/remove/keep connections).

KABB: exploration vs. exploitation

Instead of waking every agent per task, HiVA uses Knowledge-Aware Bayesian Bandit (KABB) selection: an agent $$A_i$$ is chosen based on its historical performance (Beta posterior mean $\frac{\alpha_i}{\alpha_i + \beta_i}$ ), task–capability match via an external knowledge graph, and collaboration gain among selected agents — like a film producer casting precisely by script and track record.

Topology as distributed memory

Each edge $$(v_i, v_j)$$ carries a weight $C_{syn}(v_i, v_j)$ recording how reliable information flow has been, updated via Bayesian rules:

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 yields a layered memory: macro (topology $\mathcal{G}$ — long-term collaboration patterns), meso (edge weights — path effectiveness), micro (semantic parameters $\Theta_i$ — agent expertise). Experience is encoded in the organization's DNA, not a database.

Experimental results

On Qwen-2.5-72B-Instruct-Turbo (Table 1):

| 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 is +8.0%, largest on multi-step reasoning tasks.

On GAIA (general AI assistant benchmark):

| Method | Level-1 | Level-2 | Level-3 | Cost efficiency CS | |--------|---------|---------|---------|--------------------| | 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 |

Ablations (Table 2) confirm both cores matter:

Removing SEV (semantic evolution) drops average by ~5.6 points — the biggest hit.
Removing TEV (topological evolution) drops HotpotQA by 7.3 points — multi-hop reasoning depends on correct information flow paths.
Removing KABB or environment feedback also hurts (4.4 and ~5.7 points on HotpotQA respectively).
Interesting nuance: removing tools *improves* MBPP by 2.2% — on simple tasks, tool calls add complexity and error sources.

Evolution traces: on HotpotQA, a single 67%-accuracy agent splits into retrieval and reasoning agents (round 3), adds a fact-checking agent (round 6), stabilizing as a three-node pipeline at 79%. On MATH, conflicting intermediate results deadlock the aggregator — tasks requiring strict logical consistency stress aggregator design.

Analogies used in the post

Biology: single cell → cell differentiation (semantic evolution) → tissue formation (topological evolution) → natural selection (environment feedback).
Startup growth: a founder does everything, hires a support agent (new node), sets up weekly syncs (new edge), upgrades tooling (semantic evolution), then the org chart stabilizes.
Backpropagation: forward pass = task on a dynamic subgraph; loss = environment feedback; backprop = textual gradient; update = $$f_P + f_G$$ ; but gradients are natural language, not numbers.

Limitations and outlook

HiVA underperforms on MATH (81.2% vs. Vanilla's 82.7%) due to conflicting parallel reasoning and aggregator arbitration difficulty.
Evolution takes multiple rounds (~10), too slow for real-time use; textual gradients cost extra LLM calls; no convergence guarantee to an optimal topology.
Promising directions: combining with RL policy-gradient methods (PPO/GRPO), cross-task topology transfer, human-in-the-loop evolution, and compressing semantic parameters.

> Takeaway: intelligence lies not in how strong individual neurons are, but in whether the right connections can grow by themselves. Evolution never needs a 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.