Skip to content

InfiAgent: Self-Evolving Pyramid Agent Framework for Infinite Scenarios

🕒 Published (v1): 2025-09-26 15:44 UTC · Source: Arxiv · Venue: ICLR 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

InfiAgent is a DAG-based multi-agent framework that automatically decomposes any task into a bounded-fanout pyramid of specialized sub-agents without manual workflow design. Its core novelties are an "agent-as-a-tool" abstraction, a dual-audit quality mechanism, intelligent task routing, and a Git-style self-evolution loop that restructures the DAG topology based on performance. Evaluated on five benchmarks, it outperforms the closest auto-generated-agent baseline (ADAS) by 9.9% average accuracy.

Problem

Building LLM agents today requires hand-crafted workflows, prompt engineering, and iterative tuning that demands both LLM and domain expertise, creating a scalability bottleneck for deploying agents across diverse industries. Existing multi-agent architectures use point-to-point collaboration that causes coordination overhead, deadlocks, and unpredictable behavior as system size grows, with no principled mechanism for automatic decomposition, stability guarantees, or autonomous adaptation.

Method

InfiAgent organizes agents as a DAG with bounded fan-out (\(K_{\max} = 5\)). The top-level agent receives task \(T_0\) and recursively decomposes it:

\[T^{(l)} \mapsto \{T_1^{(l+1)}, T_2^{(l+1)}, \ldots, T_{k_l}^{(l+1)}\}\]

until atomic tasks reach functional (leaf) agents. The exponential capacity at depth \(L\) is \(N_{\text{func}} \approx b^L\), while each planner reasons over only \(\leq K_{\max}\) children.

Four mechanisms drive the framework:

  1. Agent-as-a-tool: Higher-level agents invoke sub-agents as opaque tools; the router maps incoming queries directly to the appropriate level without layer-by-layer traversal.

  2. Dual-audit: An execution-level audit tracks a running quality score \(Q_i^{(t+1)} = \alpha Q_i^{(t)} + (1-\alpha)\,\text{validate}(O_i^{(t)})\) per agent. A system-level audit performs retrospective summarization to compress context.

  3. Lightweight context control: Agents exchange only (addr, desc) file-descriptor pairs; each agent's full context \(C = \{C_{\text{sys}}, C_{\text{LM}}, C_{\text{SM}}, C_{\text{ENV}}\}\) is kept bounded (\(|C| \ll |H|\)) via automatic compression when \(|C_{\text{ENV}}|\) exceeds threshold \(\tau\).

  4. Self-evolution (3 levels): Model-level—parallel lightweight models compete; a judge \(J\) merges validated deltas into a main branch \(B_{\text{main}}\). Agent-level—all candidates are fine-tuned on validated main-branch data. Topology-level—the DAG is restructured; consistently weak branches are pruned; similar functions are fused into domain expert models.

Key Contributions

  • Agent-as-a-tool decomposition: unlimited-depth hierarchical agent DAG with stability guaranteed by \(k_l \leq K_{\max}\)
  • Dual-audit quality assurance: continuous execution-level scoring + system-level retrospective summarization for context efficiency
  • Automated routing: zero-configuration task-to-agent matching via a top-level router
  • Git-style self-evolution: three-level (model, agent, topology) autonomous restructuring driven by performance feedback
  • InfiHelper case study: full research pipeline (literature → idea → experiment → paper) implemented with InfiAgent; produced papers accepted at IEEE venues

Results

All benchmark runs use GPT-4o-mini as the backbone model.

  • DROP: 82.4% vs. ADAS 76.6% and best prior baseline (CoT-SC) 78.8%
  • GSM8K: 93.1% vs. ADAS 90.8% and best prior (CoT-SC) 92.7%
  • MBPP: 71.8% vs. ADAS 53.4%
  • HumanEval: 89.3% vs. ADAS 82.4% (slightly below CoT-SC's 91.6%)
  • MATH: 35.6% vs. ADAS 35.4%; significantly below MultiPersona (50.8%)
  • Average: 74.44% vs. ADAS 67.72% (+9.9%)
  • InfiHelper peer-review scores: 5–7/10 (best paper: 7), comparable to or above AI-Researcher, Zochi, and AI-Scientist V2 outputs on the same rubric

Limitations

  • Tool-calling overhead hurts focused reasoning: on MATH (non-decomposable deduction), InfiAgent (35.6%) is far behind MultiPersona (50.8%) and CoT-SC (50.4%); single-agent approach is more efficient for these tasks
  • Self-evolution is described architecturally but no ablation isolates its contribution to benchmark gains
  • InfiHelper evaluation uses Claude-3.7-Sonnet as the "reviewer", introducing potential model-level bias and lack of domain-expert human review
  • Quality assessment scores are moderate (3–7/10), with one InfiHelper paper scoring below several baselines; the selection of "best" paper per system may mask variance
  • No latency or cost benchmarks reported; the framework's token efficiency claims are not empirically validated against baselines

Relevance to Harnesses / Meta-Harnesses

InfiAgent is essentially a self-constructing meta-harness: given an arbitrary task, it generates the agent topology, routing logic, and execution plan autonomously—the defining property of a meta-harness versus a fixed pipeline. The self-evolution mechanism (Git-style branching + topology restructuring) is a concrete implementation of a harness that modifies its own orchestration strategy based on runtime feedback, directly extending the line of work on adaptive orchestration frameworks. The dual-audit and lightweight context-control mechanisms address the reliability and scalability challenges that arise when a meta-harness spawns many parallel sub-agents, making InfiAgent a reference design for harness engineers. The "agent-as-a-tool" abstraction also offers a composable primitive that could be imported into other meta-harness systems without adopting the full framework.