Skip to content

MAS\(^2\): Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Self-generative, self-configuring, self-rectifying MAS; comprehensive self-evolution framework

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MAS² introduces a paradigm where a multi-agent system (MAS) recursively generates, configures, and repairs another MAS at runtime, replacing the brittle "generate-once-and-deploy" approach. A tri-agent meta-team (generator, implementor, rectifier) is trained via Collaborative Tree Optimization (CTO), an offline RL procedure that propagates path-level credit through a collaborative decision tree. Across 8 benchmarks, MAS² surpasses prior automated MAS by up to 19.6% while remaining on the Pareto frontier of cost-performance.

Problem

Existing automated MAS—whether built by external modules (GNNs, MCTS, evolutionary search) or single LLM agents—follow a "generate-once-and-deploy" paradigm: a system is instantiated once per task and executed without adaptation. This makes them brittle to real-world dynamics such as network failures, tool crashes, and unexpected execution states. No prior system enables a MAS to architect, instantiate, and rectify another MAS in a unified, trainable framework.

Method

MAS² operates as a meta-MAS with three specialized agents:

  1. Generator (\(\pi_{gen}\)): given query \(Q\), produces a high-level workflow template \(M_{temp} = \langle R, P, T \rangle\) (agent roles, communication protocol, tools)—without LLM assignments.
  2. Implementor (\(\pi_{imp}\)): maps each role \(r_i \in R\) to a concrete backbone \(b_{j(i)}\) from a candidate pool \(L\), yielding the fully executable MAS \(M = M_{temp} \oplus \{(r_i, \phi(r_i)) \mid r_i \in R\}\).
  3. Rectifier (\(\pi_{rec}\)): monitors execution state \(s_t\); triggers when cumulative cost \(C(s_t) > \theta_C\) or \(O(s_t) = \text{Failure}\), then modifies \(M_t \to M_{t+1}\) (local prompt edits or global workflow rewrites).

Collaborative Tree Optimization (CTO) trains these agents offline: - A rooted decision tree \(G_Q\) is expanded by sampling \(K\) generator branches, \(N\) implementor instantiations per branch, and adaptive rectifier branches. - Terminal nodes receive a cost-sensitive reward \(R(\tau) = \mathbf{1}[R_p(\tau)] \cdot C_{norm}(\tau)^{-1}\) that penalizes resource waste. - Node values are Monte Carlo estimates \(V(v) \approx \frac{1}{|T(v)|}\sum_{\tau \in T(v)} R(\tau)\) propagated backward (path credit propagation). - Preference tuples \((c_v, a_{win}, a_{lose}, \Delta V)\) with continuous preference strength \(\Delta V\) are used in a value-scaled DPO-style loss:

\[\mathcal{L}_{CTO}(\pi_\theta;\pi_{ref}) = -\mathbb{E}_{\Delta V}\left[\Delta V \cdot \log \sigma\!\left(\beta \log \frac{\pi_\theta(a_{win}|c)}{\pi_{ref}(a_{win}|c)} - \beta \log \frac{\pi_\theta(a_{lose}|c)}{\pi_{ref}(a_{lose}|c)}\right)\right]\]

All three agents are fine-tuned independently from Qwen3-8B via LoRA (rank=8, \(\alpha\)=16).

Key Contributions

  • MAS² paradigm: a MAS generates another MAS, extending automated MAS design from single-agent or external-module generation to recursive multi-agent self-generation.
  • Tri-agent meta-architecture: separate generator, implementor, and rectifier agents with distinct learned policies, enabling real-time self-correction beyond instantiation.
  • Collaborative Tree Optimization (CTO): a trajectory-curation and value-propagation procedure that produces role-specific, preference-strength-weighted training data for each meta-agent.
  • Value-scaled preference loss: extends DPO with continuous \(\Delta V\) weighting so high-margin decisions dominate gradient updates.
  • Pareto-optimal cost-performance: MAS² achieves top performance while remaining on the efficiency frontier, avoiding excessive token costs.

Results

  • HotpotQA: 89.3% (+23.8 pp over single-LLM avg; +3.3 pp over ScoreFlow, the prior best automated MAS).
  • Bamboogle: 67.2% (+31.0 pp over single-LLM avg; +2.4 pp over ScoreFlow).
  • NQ: 79.1% (+15.5 pp over single-LLM avg; +2.7 pp over ScoreFlow).
  • BrowseComp+ (deep research): 19.7% (+10.2 pp over single-LLM avg; +5.7 pp over MaAS, the prior best).
  • HumanEval: 97.0% (+19.6 pp over single-LLM avg; +4.1 pp over AFlow).
  • MBPP: 85.1% (+21.4 pp over single-LLM avg).
  • MATH: 71.3% (+13.3 pp over single-LLM avg; +6.9 pp over ScoreFlow).
  • Cross-backbone generalization: integrating unseen LLMs (Qwen3-Coder, GPT-5-Mini, Gemini-2.5-Pro) at inference yields up to 15.1% further improvement over the fixed-pool MAS².

Limitations

  • Meta-agents (Qwen3-8B LoRA) are fixed at training time; adapting to entirely new task distributions requires re-running CTO.
  • CTO training data is collected from a fixed LLM pool (GPT-4o, GPT-4o-mini, Qwen2.5-72B, Qwen3-14B, QwQ-32B); preference quality depends on the diversity of this pool.
  • The tree expansion is shallow (K=4 generator branches, N=2 implementor instantiations), potentially leaving high-quality configurations unexplored.
  • Rectifier intervention logic relies on a threshold \(\theta_C\) and binary failure signals; subtler degradation (e.g., hallucination without explicit errors) may not trigger correction.
  • The paper text is truncated before the full generalization study (Table 2) and ablation results are only partially quoted.

Relevance to Agentic AI / LLM Agents

MAS² directly advances the frontier of automated MAS design by closing the gap between static workflow generation and truly adaptive, self-healing agent systems—a critical requirement for deploying agents in open-ended, long-horizon tasks. The tri-agent meta-architecture operationalizes a clean separation between planning (generator), resource allocation (implementor), and error recovery (rectifier), offering a reusable design pattern for robust agentic pipelines. CTO's value-scaled offline RL is broadly applicable to any hierarchical agent training problem where credit assignment across multiple decision-makers is non-trivial. For researchers tracking agentic AI, this work establishes "MAS-generates-MAS" as a concrete, trainable paradigm and provides the first systematic treatment of runtime self-rectification at the architectural level.