Skip to content

Mimosa Framework: Toward Evolving Multi-Agent Systems for Scientific Research

🕒 Published (v1): 2026-03-30 20:35 UTC · Source: Arxiv · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Mimosa is an open-source evolving multi-agent framework for Autonomous Scientific Research (ASR) that dynamically synthesizes DAG-structured workflows per task and refines them through iterative LLM-judge feedback. It uses the Model Context Protocol (MCP) for dynamic tool discovery, enabling adaptation to changing computational environments. On ScienceAgentBench, Mimosa with DeepSeek-V3.2 achieves 43.1% success, outperforming single-agent and static multi-agent baselines.

Problem

Existing ASR systems suffer from two compounding failures: (1) semantic drift in single-agent trajectories, where context accumulates as \(c_k = [p, s^{(0)}, a^{(0)}, r^{(0)}, \ldots, s^{(k)}]\), causing attention dilution and goal deviation over long horizons; and (2) architectural rigidity, where multi-agent systems rely on fixed coordination protocols and static toolsets that cannot adapt when new instruments are introduced, objectives change, or intermediate results demand alternative analytical paths.

Method

Mimosa implements a five-layer architecture:

  • Layer 0 (Optional Planning): Decomposes high-level scientific goals into granular, repeatable tasks.
  • Layer 1 (Tool Discovery): Dynamically enumerates available computational resources via MCP servers (containerized, isolated), avoiding hard-coded tool assumptions.
  • Layer 2 (Meta-Orchestration): Synthesizes and iteratively refines task-specific workflows represented as DAGs \(W = (A, E)\), where each agent \(a_i = (p_i, o_i, \theta_i)\) handles a bounded subtask context \(O(L_\text{subtask})\) rather than the full trajectory \(O(k \cdot L_\text{subtask})\). Workflow discovery is formulated as local search: \(W^* = \arg\max_{W \in \mathcal{W}} J(s^*_W)\).
  • Layer 3 (Agent Execution): Code-generating agents invoke MCP-discovered tools and scientific libraries.
  • Layer 4 (LLM Judge): Scores executions on four criteria — goal alignment \(g\), collaboration efficiency \(c\), output quality \(q\), answer plausibility \(a\) — averaged as \(J(s^*_W) = \frac{1}{4}(g + c + q + a)\), feeding structured feedback to the meta-orchestrator.

Workflow refinement uses single-incumbent local search: at each iteration, one of three mutation types is applied (prompt refinement, agent addition/removal, edge rewiring), the mutant is accepted if \(J(W') > J(W_n)\), and search terminates after a fixed iteration budget or when \(J > 0.9\). New tasks initialize by retrieving archived workflows with cosine similarity \(> 0.7\); otherwise, a workflow is synthesized de novo.

Key Contributions

  • Evolving DAG workflows: Workflow topology (agents, edges, tool allocations, prompts) is treated as a mutable, learnable object, not a fixed pipeline.
  • MCP-based dynamic tool discovery: Tools are registered as containerized MCP servers, allowing workflows to adapt to changing tool ecosystems without code changes.
  • Structured execution traces and workflow archive: Every execution step is logged; discovered workflows are archived and retrievable by embedding similarity for warm-start initialization on new tasks.
  • LLM-as-judge feedback loop: A structured four-criterion judge provides the fitness signal driving iterative workflow refinement.
  • Open-source platform: Mimosa and companion tool-management platform Toolomics released under Apache 2.0.
  • ScienceAgentBench evaluation: 43.1% task success with DeepSeek-V3.2, above single-agent and static MAS baselines; reveals heterogeneous model responses to multi-agent decomposition.

Results

  • Mimosa (DeepSeek-V3.2): 43.1% success rate on ScienceAgentBench (102 data-driven discovery tasks across four scientific disciplines).
  • Outperforms single-agent baselines and static multi-agent configurations on the same benchmark.
  • Results show model-dependent responses to multi-agent decomposition: benefits of workflow evolution are not uniform across underlying execution models.
  • In practice, the workflow archive warm-start path was never triggered during evaluation (all tasks used de novo synthesis), indicating the retrieval mechanism was not exercised on the current benchmark.

Limitations

  • Archive cold-start: The embedding-based workflow retrieval (\(\text{sim} > 0.7\)) was never triggered on ScienceAgentBench; cumulative learning from prior workflows was not empirically demonstrated.
  • Approximate judge signal: The LLM-as-a-judge provides directional rather than exact quality signal; correlation with benchmark ground-truth metrics is left to future work.
  • Single-incumbent local search: Exploration is strictly local; no population-based diversity is maintained, risking entrapment in local optima.
  • Model heterogeneity: Benefits of workflow evolution depend heavily on the capabilities of the underlying LLM; the framework does not yet address this dependency systematically.
  • No physical/instrument integration tested: Despite the SDL motivation, evaluation is limited to computational tasks on ScienceAgentBench.
  • Judge-fitness correlation unvalidated: Whether \(J(s^*_W)\) reliably predicts ground-truth success is an open question.

Relevance to Harnesses / Meta-Harnesses

Mimosa is a direct instantiation of a meta-harness pattern: a system whose primary function is to construct, evaluate, and iteratively improve the harnesses (agent workflows) that execute actual work, rather than executing tasks directly. The meta-orchestrator layer — generating DAG topologies, mutating them based on judge feedback, and archiving successful patterns — is precisely the "harness of harnesses" architecture that characterizes meta-harness research. The use of MCP for dynamic tool discovery extends the meta-harness concept to the tool layer, enabling the harness itself to adapt its available primitives at runtime. This work is particularly relevant as it demonstrates a principled formalization (workflow discovery as local search over \(\mathcal{W}\)) and open-source implementation of the meta-harness loop in a scientific computing context.