Skip to content

Recursive Multi-Agent Systems

🕒 Published (v1): 2026-04-28 17:59 UTC · Source: Arxiv · link

Why this paper was selected

Extends recursive LM scaling to multi-agent systems; new meta-harness paradigm

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

RecursiveMAS reframes a multi-agent system (MAS) as a single latent-space recursive computation: agents are chained in a loop and communicate via continuous hidden states rather than text, connected by a lightweight two-layer residual projection called RecursiveLink. An inner-outer loop training algorithm co-optimizes only the RecursiveLink modules (not agent weights) via shared gradient-based credit assignment across recursion rounds, achieving both accuracy gains and substantial inference speedup.

Problem

Standard text-mediated MAS architectures suffer from three compounding inefficiencies: (1) intermediate agents must decode to vocabulary space at each step, incurring \(O(m|V|d_h)\) cost per agent where \(|V| \gg d_h\); (2) sequential text-passing introduces latency because each agent waits for the previous agent's full generation; and (3) gradient vanishing during recursive training — text-based SFT across recursion rounds yields \(\|\partial R_\text{text}(h)/\partial h\|_2 = O(\epsilon) \approx 0\) for confident tokens, making whole-system learning impractical.

Method

RecursiveLink (\(R\)) is a two-layer MLP with residual connection used in two modes: - Inner link (same-agent, dense-to-shallow): \(R^\text{in}(h) = h + W_2\,\sigma(W_1 h)\), mapping last-layer embeddings back to input embedding space for the next auto-regressive step. - Outer link (cross-agent): \(R^\text{out}(h) = W_3 h + W_2\,\sigma(W_1 h)\), with an extra linear \(W_3\) to handle heterogeneous hidden dimensions between agents \(A_i\) and \(A_j\).

System loop: All \(N\) agents are chained; each agent \(A_i\) generates a sequence of latent thoughts \(H_{A_i}\) auto-regressively in continuous space, passes them to \(A_{i+1}\) via \(R^\text{out}\), and after \(A_N\) completes, its latent outputs feed back to \(A_1\), closing the recursive loop. Text decoding occurs only at the final agent in the final recursion round \(r = n\).

Training (inner-outer loop): 1. Inner loop: Each agent's \(R^\text{in}\) is warm-started in parallel by minimizing cosine regression loss \(L_\text{in} = 1 - \cos(R^\text{in}(H),\, \text{Emb}_{\theta_i}(y))\) to align latent thoughts with the target embedding distribution. 2. Outer loop: The full system is unrolled for \(n\) rounds; cross-entropy loss on the final round's text output \(L_\text{out} = \text{CE}(S^{(n)}(\cdots S^{(1)}(x)),\, y)\) propagates gradients back through all outer RecursiveLinks jointly, performing whole-system co-optimization without touching base model parameters.

Gradient stability (Theorem 4.1): With probability \(\geq 1-\delta\), RecursiveMAS maintains \(\|\partial R(h)/\partial h\|_2 \geq \Omega(1 - \sqrt{(1/d_h)\log(1/\delta)}) \approx 1\), while text-based recursion yields \(O(\epsilon) \approx 0\).

Key Contributions

  • RecursiveLink module: minimal two-layer residual projection enabling in-distribution latent thought generation (inner) and cross-model latent state transfer (outer) between heterogeneous agents.
  • Recursive MAS loop: system-level architectural pattern that treats heterogeneous agents as RLM layers chained in a latent-space loop with a single text-output point.
  • Inner-outer loop training: two-stage co-optimization algorithm that trains only RecursiveLink parameters via shared gradient credit across recursion rounds.
  • Theoretical guarantees: runtime complexity reduction from \(\Theta(Nm|V|d_h)\) to \(\Theta(Nmd_h^2)\) (Proposition 3.1), and gradient stability proof for latent vs. text-based recursion (Theorem 4.1).
  • Structure-agnostic generalization: framework instantiated across four distinct MAS topologies (Sequential, Mixture, Distillation, Deliberation) with heterogeneous model families (Qwen3/3.5, LLaMA-3, Gemma3, Mistral).

Results

  • Overall: average +8.3% accuracy, \(1.2\times\)–\(2.4\times\) inference speedup, 34.6%–75.6% token reduction vs. baselines including single agents, LoRA fine-tuned models, LoopLM, Recursive-TextMAS, and TextGrad.
  • By recursion depth vs. Recursive-TextMAS (Sequential Light / Scaled):
  • \(r=1\): +3.4% accuracy, \(\times 1.2\) speedup, \(\downarrow\)34.6% tokens
  • \(r=2\): +6.0% accuracy, \(\times 1.9\) speedup, \(\downarrow\)65.5% tokens
  • \(r=3\): +7.2% accuracy (partial table; further speedup evident)
  • Math500: 77.8% (RecursiveMAS Light, \(r=3\)) vs. 69.1% (Recursive-TextMAS, \(r=3\)).
  • AIME2025 Scaled (\(r=2\)): 86.0% vs. 70.7% (Recursive-TextMAS).
  • GPQA-Diamond Scaled (\(r=3\)): 66.2% vs. 58.6% (Recursive-TextMAS).
  • Scaling law: clean monotonic accuracy improvement with increasing recursion rounds for the sub-1.5B lightweight configuration.
  • 9 benchmarks covered: MATH500, AIME2025, AIME2026, GPQA-Diamond, MedQA, LiveCodeBench-v6, MBPP Plus, HotpotQA, Bamboogle.

Limitations

  • Requires white-box access to agent hidden states; inapplicable to API-only or closed-source LLMs.
  • Base model weights are frozen; per-agent capability improvement is entirely delegated to the lightweight RecursiveLink, which may bound headroom for tasks requiring deep domain adaptation.
  • Outer-loop training backpropagates through the full \(n\)-round unrolled computation graph, increasing memory cost with recursion depth and agent count.
  • Sequential agent chaining within each recursion round limits intra-round parallelism; the Mixture Style agents operate in parallel but still require synchronization at the aggregation step.
  • Evaluation is limited to sub-10B parameter agents; scalability to frontier-scale models is unverified.

Relevance to Harnesses / Meta-Harnesses

RecursiveMAS is a direct example of a learned meta-harness: rather than a hand-coded orchestration layer that routes messages between agents, it trains the connective tissue (RecursiveLink) as a first-class optimizable component, making the harness itself the subject of gradient-based learning. The inner-outer loop training paradigm is structurally analogous to meta-harness design — it operates above individual agent training loops and co-optimizes inter-agent communication protocols without modifying agent internals. For researchers tracking harnesses, this paper demonstrates that the orchestration layer can be parameterized and jointly trained with the task objective, yielding compounding efficiency gains (runtime, token cost, accuracy) that prompt-based harness design cannot achieve. It also introduces a formal complexity and gradient-stability theory for latent-vs-text harness communication, which could inform harness architecture decisions more broadly.