The FM Agent: LLM Reasoning Meets Large-Scale Evolutionary Search

30 Oct 2025 · Paper

---

1. Motivation

Tree-search MLE agents (AIDE, ML-Master) commit to a single lineage of refinements and are vulnerable to premature convergence; flat evolutionary systems (FunSearch-style) generate diversity but lack a mechanism to allocate compute toward promising regions as the population evolves.¹²³ FM Agent’s premise is that automated discovery needs both LLM reasoning (to generate semantically meaningful variations) and large-scale population parallelism with adaptive selection pressure (to explore divergent regions of the objective space without collapse). It is an industrial system - no code release - evaluated entirely on public benchmarks.¹

---

2. Methodology

2.1 Search Structure

A two-stage architecture:¹

Stage 1 Cold Start:
  Multi-Agent Parallel Expansion   -> broad initial population
  Proactive Solution Space Expansion -> cover divergent objective regions
  cluster solutions by max inter-island similarity
  #clusters = #islands; assign one cluster per island

Stage 2 Evolve:
  multi-population ISLAND MODEL
  islands evolve independently most of the time
  periodic inter-island interaction ("cross-pollination")

The island model means the search structure is a set of semi-isolated populations rather than one tree or one flat database - closer to AlphaEvolve’s island/database hybrid than to MCTS.¹⁴

2.2 State Representation

A population of programs (solution candidates) partitioned into islands; each island maintains an elite pool of top performers; population diversity is tracked with semantic and structural metrics, which feed the sampling controller (2.6).¹

2.3 Generation Mechanism

LLM-driven mutation and crossover within island populations: the LLM acts as the variation operator, rewriting or recombining program candidates rather than sampling token-level noise. Two augmentations: (a) a RAG knowledge base whose fragments can be retrieved to inform variation operators; (b) optional human natural-language or code-level intervention steering variation (used only in case studies; benchmark runs are human-free).¹

2.4 Evaluation & Grounding

A Domain-Specific Evaluator produces three feedback types: (a) a traditional single fitness score for quantitative ranking; (b) LLM-judge feedback for qualitative assessment; (c) domain-specific multi-dimensional metrics - numerical (e.g., balancing accuracy vs. latency) and contextual (e.g., resource utilization) for ML and CUDA-kernel tasks. The evaluator combines “functional correctness, operational effectiveness, and LLM-supervised quality assessment,” so executable correctness gates any fitness claim.¹

2.5 Memory / Information Flow

Island-persistent populations act as lineage memory; the curated elite pool retains top solutions “to guide future generations”; the expert knowledge base with RAG retrieval supplies priors; periodic island migration shares genetic material across otherwise isolated lineages.¹ Compared with ML-Master 2.0’s typed L1/L2/L3 caches, this memory is deliberately homogeneous - a design ML-Master 2.0 criticizes as treating knowledge as a single undifferentiated entity.¹⁵

2.6 Selection & Propagation

Selection is Adaptive Diversity-Driven Sampling: “a novel cluster-based sampling strategy … adaptively maintains a balance between exploration and exploitation by dynamically adjusting selective pressure according to real-time population diversity.”¹ Concretely, candidates are sampled cluster-wise from the population (so no single dense region dominates reproduction), and compute is dynamically reallocated across islands - stronger lineages receive more evaluations while stagnating islands are starved. Propagation happens through elitism (top solutions persist in the elite pool) and periodic migration between islands. The single-task ablation (ALE-Bench ahc016, 5 runs) reports a final combined score of 0.7182, +10.99% vs. top-$k$ sampling and +58.26% vs. random, and surpasses random sampling’s maximum score at iteration 40 instead of 900.¹

2.7 Anti-Slop / Anti-Hallucination Mechanisms

Correctness is gated by executable domain evaluators before fitness counts; the LLM judge is only one of three feedback channels, never the sole ground truth; headline benchmark runs are “performed exclusively by the LLM without human intervention” to keep claims clean; an expert-in-the-loop option exists only for real deployments and math case studies.¹

Infrastructure: a Ray-orchestrated distributed cluster runs an asynchronous generation–evaluation pipeline with separate worker pools for program synthesis vs. evaluation, sustaining thousands of concurrent evolutionary processes.¹

---

3. Results

• MLE-Bench (75 tasks, Gemini-2.5-Pro, 3 seeds): valid submission 96.89% ± 2.22, above median 51.56% ± 2.23, gold 22.67% ± 1.34, any medal 43.56% ± 1.78 - the best of its comparison table at publication. Baselines quoted from official reports (not re-run under matched conditions): AIDE (o1-preview) 16.9, ML-Master (DeepSeek-R1) 29.3, Neo 34.22, R&D-Agent (GPT-5) 35.11, InternAgent 36.44, Operand ensemble (GPT-5 with light multi-model assistance) 39.56.¹²⁶⁷ First place among compared agents in 33/75 competitions and 9/15 on the hard split.¹
• ALE-Bench (AtCoder heuristic contests): rating 1976.3, +5.2% over prior SOTA.¹
• KernelBench: 2.08×–20.77× speedups over torch.compile.¹
• Classical math problems: new SOTA on several (e.g., min max/min distance ratio point-packing), verified with AlphaEvolve’s verification code.¹⁴
• Ablations: each of cold start / adaptive sampling / island model contributes; the full method beats OpenEvolve with faster convergence and higher final scores.¹⁸

---

4. Limitations & Assumptions

• Industrial system with no code release; the Ray-cluster infrastructure implies a very large compute budget, so its numbers are not directly comparable to single-GPU academic agents (MARS’s Table 6 makes this resource disparity explicit for leaderboard agents generally).¹⁹
• The sampling ablation covers a single ALE-Bench task (ahc016).¹
• The LLM-judge channel inherits judge-model biases; internal parameters (island count, migration period, fitness formulas, mutation prompts) are undisclosed.
• Math case studies used “high-level guidance inspired by human intuition,” blurring the full-autonomy claim even though benchmark numbers are human-free.¹

---

5. Critical Analysis

• Distinctive strength: adaptive diversity-driven sampling with dynamic cross-island resource allocation - selection pressure is a controlled variable responding to measured population diversity, not a fixed schedule; combined with cold-start clustering, this directly attacks premature convergence at population scale.
• Structural weakness / trade-off: verifiability. With no released code, undisclosed internals, and infrastructure-scale compute, FM Agent’s results are an existence proof rather than a reproducible method; its homogeneous population memory also discards the typed experience→knowledge→wisdom distinctions that ML-Master 2.0 argues are necessary for long horizons.⁵
• Connections: sits in the FunSearch→AlphaEvolve evolutionary lineage (generate–score–evolve over programs, islanded populations) and beats the open-source OpenEvolve in its own ablation.³⁴⁸ Its related work names ML-Master’s MCTS-plus-memory design as the closest search-strategy refinement, and the two are directly comparable on MLE-Bench (43.56 vs. 29.3 any-medal).¹² It credits R&D-Agent’s dual Researcher/Developer collaboration and periodic merging as the closest multi-agent design (35.11 vs. 43.56 any-medal).¹⁶ MARS later quotes FM Agent’s 43.6% leaderboard figure (misspelling it “Famou-Agent”) and surpasses it with MARS+ (62.7%) under two-tree parallel search.⁹

---

6. References

Footnotes

1. FM Agent Team (Baidu AI Cloud), 2025, “The FM Agent” - https://arxiv.org/abs/2510.26144 (full text: https://arxiv.org/html/2510.26144v1) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶ ↩¹⁷ ↩¹⁸ ↩¹⁹ ↩²⁰ ↩²¹ ↩²² ↩²³ ↩²⁴

2. Liu et al., 2025, “ML-Master: Towards AI-for-AI via Integration of Exploration and Reasoning” - https://arxiv.org/abs/2506.16499 ↩ ↩² ↩³

3. Romera-Paredes et al., 2024, “Mathematical discoveries from program search with large language models” (FunSearch), Nature - https://doi.org/10.1038/s41586-023-06924-6 ; code: https://github.com/google-deepmind/funsearch ↩ ↩²

4. Novikov et al., 2025, “AlphaEvolve: A coding agent for scientific and algorithmic discovery” - https://arxiv.org/abs/2506.13131 ↩ ↩² ↩³

5. Zhu et al., 2026, “Toward Ultra-Long-Horizon Agentic Science: Cognitive Accumulation for Machine Learning Engineering” (ML-Master 2.0) - https://arxiv.org/abs/2601.10402 ↩ ↩²

6. Yang et al., 2025, “R&D-Agent” - https://arxiv.org/abs/2505.14738 ↩ ↩²

7. Chan et al., 2024, “MLE-Bench” - https://arxiv.org/abs/2410.07095 ; https://github.com/openai/mle-bench ↩

8. Sharma, 2025, OpenEvolve (open-source evolutionary coding agent; FM Agent ablation baseline) - https://github.com/codelion/openevolve ↩ ↩²

9. Chen et al., 2026, “MARS: Modular Agent with Reflective Search for Automated AI Research” - https://arxiv.org/abs/2602.02660 ↩ ↩²