CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents
Jiaxuan Jiang, Liyuan He, Zhixuan Fang
cs.AI · cs.LG
Quick verdict
CERA-MoA co-evolves a hidden-state familiarity router with RL-trained LLM agents; gains are broad but single-seed and untested for significance.
Summary
CERA-MoA jointly evolves an LLM query router and the agents it routes to, targeting the decoupling between orchestration and post-training in Mixture-of-Agents systems. A predictive familiarity estimator maps concatenated mid-layer hidden states (layers L/2 and 3L/4) of each query through per-agent trainable predictor heads toward frozen random target anchors; exponential decay of the normalized projection distance yields familiarity scores without full rollouts. A cumulative-threshold rule activates the minimal prefix of agents (ranked by familiarity plus UCB/entropy exploration bonuses during training) whose summed familiarity reaches tau=0.7. Selected agents are fine-tuned with DAPO plus GSPO importance sampling on routed samples, while router heads update from cross-agent relative advantages under a bounded repulsion margin. Across Qwen3-4B (4 LoRA agents), Llama-3.2-3B, Phi-4-mini, and a heterogeneous trio, on nine in-distribution and six OOD benchmarks, the method attains the highest averages (e.g., 63.2 vs 61.0 for AT-GRPO on Qwen3-4B ID; 71.8 vs 69.1 heterogeneous OOD) with about 45% token savings versus fixed Top-2 at comparable accuracy. Ablations support the familiarity metric over reward regression/classification, mid-layer features, exploration bonuses, and the margin. Main limitations: single-seed evaluation without significance tests, comparison against untrained orchestration baselines, single-turn routing, and simplistic open-ended aggregation.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Coupling query routing with agent fine-tuning in an iterative closed loop outperforms both static-agent routing and fixed-workflow fine-tuning across model scales and domains (Tables 1-2).
- A familiarity estimator over mid-layer hidden states provides a stable, cheap competence signal that beats direct reward regression (60.8) and multi-class classification (61.3) routing metrics, reaching 63.2 ID average (Table 3).
- Cumulative-threshold routing activates a minimal agent subset, cutting average tokens about 45% versus fixed Top-2 (367.77 vs 666.37) at comparable accuracy (Table 3).
- Routing training queries by evolving competence induces emergent, role-prompt-free specialization among homogeneous agents (Figure 3, Appendix E.2).
Strengths
Genuinely cheap competence estimation inside the training loop
The familiarity estimator uses one partial forward pass (11.87 ms/query, <1% of native generation time, Table 6), enabling router-agent co-evolution where rollout-based evaluators would be prohibitive (Sections 3.2, E.1).
Unusually thorough ablation coverage
RQ3-RQ5 plus Appendix E ablate the routing metric (Table 3: reward regression 60.8, classification 61.3 vs 63.2 ID avg), train/test threshold tau (Tables 7-8), temperature lambda (55.3 ID at lambda=1.0), margin m including an unbounded-repulsion failure mode (Figs. 9-10), exploration bonuses (57.1/63.9 without them, Table 9), and mid-layer vs final-layer features (Table 10: 61.1/71.3 vs 63.2/72.8).
Broad empirical scope with consistent average wins
Quantified efficiency-accuracy trade-off
Strong reproducibility documentation
Weaknesses
No variance or significance reporting
Asymmetric comparison with orchestration baselines
Missing uniform-allocation control isolating the router
One-sided familiarity updates for solo-routed agents
Cost accounting and some design choices under-justified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is CERA-MoA in one sentence?
A reinforcement-learning framework where a lightweight router and a population of LLM agents improve together: the router estimates how 'familiar' each query is to each agent from mid-layer hidden states, routes each query to the smallest competent subset, and the routed training data drives agents to specialize while the router tracks their changing skills.
How much overhead does the CERA-MoA router add at inference time?
About 11.87 milliseconds per query on one NVIDIA H200 GPU, under 1% of a single agent's native generation time (1350 ms), because it needs only a partial forward pass over the input rather than full text generation or an external judge model.
Does CERA-MoA actually beat strong baselines, and by how much?
It reports the best average score in all six tested settings, e.g., 63.2 vs 61.0 for AT-GRPO on Qwen3-4B in-distribution and 71.8 vs 69.1 for the heterogeneous pool out-of-distribution; however, margins over the strongest RL fine-tuning baselines are often under two points and all results come from a single random seed without significance tests.
Is code or data for CERA-MoA publicly available?
The paper text does not mention any code or data release. It does document full hyperparameters, reward implementations, prompt templates, and dataset split sizes, but reproduction would require rebuilding custom processed splits of GSM8K, MATH, DAPO-MATH-17k, MBPP, Eurus-2-Code, TACO, MAGPIE-IF, RLVR-IFEval, and BBH.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure content (t-SNE plots, training curves in Figs. 2, 5-15, capability bars in Fig. 3) cannot be verified from extracted text; only captions were available.
- This is an arXiv preprint (2609.18779v1, Sep 2026) and has not undergone peer review.
- Mathematical notation was partially mangled by text extraction (e.g., subscripts in Eqs. 2-4, 7-12); formulas were reconstructed from context and the appendix restatement.
- Table values were extracted from compressed plain text and checked for internal consistency (averages re-derivable from per-dataset cells), but minor transcription risk remains.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.