Recursive Multi-Agent Systems

1. Problem Statement

Current Large Language Models (LLMs) often struggle with complex tasks due to limited capacity, myopic generation, or inefficient exploration of the solution space. While Multi-Agent Systems (MAS) have emerged to address this by organizing specialized agents into collaborative workflows, existing approaches face significant limitations:

• Text-Mediated Latency: Standard MAS relies on explicit text generation and exchange between agents. This introduces substantial latency and token overhead, especially in recursive or iterative settings where agents must wait for others to finish generating text before proceeding.
• Training Inefficiency: Optimizing an entire MAS is difficult. Updating all model parameters is computationally expensive, and sequential text dependencies often lead to gradient vanishing during backpropagation, hindering the learning of complex collaboration dynamics.
• Static Adaptation: Many existing methods rely on prompt engineering or iterative context refinement, which improves alignment but does not fundamentally evolve the agents' internal reasoning capabilities or the system's collaborative structure.

The authors ask: Can agent collaboration itself be scaled through recursion, and can this be achieved more efficiently by operating in latent space rather than text space?

2. Methodology: RecursiveMAS

The paper introduces RecursiveMAS, a framework that casts the entire multi-agent system as a unified latent-space recursive computation. Instead of treating agents as separate entities exchanging text, RecursiveMAS treats them as layers in a recursive loop, passing and refining latent representations (hidden states) rather than tokens.

Core Architectural Components

1. Latent Thoughts Generation (Inner Loop):

   • Each agent performs auto-regressive generation not of tokens, but of latent thoughts (continuous hidden state embeddings).
   • An Inner RecursiveLink ($R_{in}$) connects the output of an agent's last layer back to its input embedding space for the next step. This allows an agent to iteratively refine its internal reasoning state without decoding to text.
   • $R_{in}$ is a lightweight 2-layer residual projection module: $R_{in}(h) = h + W_2 \sigma(W_1 h)$.

2. Cross-Agent Collaboration (Outer Loop):

   • Heterogeneous agents (different model sizes/architectures) are connected via an Outer RecursiveLink ($R_{out}$).
   • $R_{out}$ transforms the latent state of one agent into the embedding space of the next agent, enabling seamless information transfer across different model types.
   • The system forms a closed loop: Agent $A_1 \to A_2 \to \dots \to A_N \to A_1$. Only the final agent in the last recursion round decodes the latent state into the final text output.

3. Inner-Outer Loop Training Paradigm:

   • Inner-Loop Training: Each agent is warmed up individually. The inner $R_{in}$ is trained to align the agent's generated latent thoughts with the semantic distribution of the ground-truth answer (using cosine similarity regression). This ensures agents can generate meaningful latent states.
   • Outer-Loop Training: The entire system is co-optimized. The outer $R_{out}$ modules are trained using a shared gradient-based credit assignment mechanism. Gradients are backpropagated through the full recursive computation graph (across all agents and recursion rounds) using a standard Cross-Entropy loss on the final output. This allows the system to learn how to collaborate iteratively.

Theoretical Advantages

The paper provides two key theoretical justifications for latent-space recursion:

• Runtime Complexity: RecursiveMAS replaces the expensive vocabulary-space decoding ($O(|V|)$) with latent-space projection ($O(d_h^2)$). Since the hidden dimension $d_h$ is much smaller than the vocabulary size $|V|$, the complexity is significantly reduced.
• Gradient Stability: Theoretical analysis shows that text-based recursion suffers from gradient vanishing (gradient norm $\approx 0$) due to the softmax operation and entropy constraints. In contrast, RecursiveMAS maintains stable, near-constant gradients (norm $\approx 1$) through the residual connections in the latent space, facilitating effective deep recursion training.

3. Key Contributions

1. System-Level Recursion: The first framework to extend the recursive scaling principle from single models to heterogeneous multi-agent systems, treating the entire MAS as a unified latent-space loop.
2. Lightweight RecursiveLink: A novel, parameter-efficient module ($R_{in}$ and $R_{out}$) that enables latent thought generation and cross-agent state transfer without updating the massive parameters of the base LLMs.
3. Inner-Outer Loop Optimization: A two-stage training algorithm that decouples agent capability (inner) from system collaboration (outer), enabling stable co-optimization of the entire system.
4. Generalizability: The framework is architecture-agnostic and supports diverse collaboration patterns, including Sequential, Mixture-of-Experts, Distillation, and Tool-Integrated Deliberation.

4. Empirical Results

The authors evaluated RecursiveMAS across 9 benchmarks (Math, Science, Medicine, Search, Code) using 4 different collaboration patterns and diverse model families (Qwen, Llama, Gemma, Mistral).

• Accuracy: RecursiveMAS achieved an average accuracy improvement of 8.3% compared to advanced single-agent and multi-agent baselines (including recursive text-based systems).
  • Notable gains on reasoning-intensive tasks: +18.1% on AIME2025, +13.0% on AIME2026.
• Efficiency (Speed):
  • Inference Speedup: Achieved 1.2× to 2.4× end-to-end inference speedup compared to text-based recursive MAS. The speedup increases as recursion depth increases because text-based systems suffer from compounding token generation costs.
  • Token Reduction: Reduced token usage by 34.6% to 75.6% compared to text-based baselines, as most interactions occur in the compressed latent space.
• Scaling Laws:
  • Recursion Depth: Performance consistently improves as the number of recursion rounds ($r=1, 2, 3$) increases.
  • Training vs. Inference: Deeper training recursion shifts the performance frontier upward, and deeper inference recursion further leverages the learned latent structure.
• Cost: Training RecursiveMAS is significantly cheaper than full fine-tuning or even LoRA fine-tuning of single agents, as only the lightweight RecursiveLink modules are updated (approx. 0.31% of total parameters).

5. Significance

RecursiveMAS represents a paradigm shift in how multi-agent systems are designed and trained:

• Efficiency: It solves the "token tax" of multi-agent collaboration, making iterative reasoning feasible for complex tasks without prohibitive latency or cost.
• Scalability: It demonstrates that system-level performance can be scaled by deepening the recursion loop, similar to how single models scale with depth, but applied to a collaborative ecosystem.
• Stability: By moving collaboration to latent space, it overcomes the gradient instability issues that have historically plagued deep recursive training in text-based systems.
• Generalization: The framework's ability to generalize across different agent topologies (sequential, parallel, distillation) suggests a robust, universal approach to scaling agent collaboration.

In conclusion, RecursiveMAS proves that latent-space recursion is a viable and superior scaling axis for multi-agent systems, offering a path to more capable, efficient, and trainable collaborative AI.