CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents
Jiaxuan Jiang, Liyuan He, Zhixuan Fang
IIIS, Tsinghua University · Shanghai Jiao Tong University · Shanghai Qi Zhi Institute
Published September 16, 2026 · on the feed September 17, 2026 · ▲ 14 · arXiv 2609.18779
The abstract
Current Mixture-of-Agents (MoA) paradigms generally treat query routing and agent fine-tuning as separate processes, limiting their ability to respond to evolving agent capabilities. This disconnect prevents routing strategies from adapting to evolving agent capabilities during post-training and prevents agents from achieving synergistic data-driven specialization. To resolve this, we introduce CERA-MoA (Co-Evolving Router with continually learning Agents for Mixture-of-Agents), an iterative reinforcement learning framework where the dynamic router and independent agent policies co-evolve. We design a predictive familiarity estimator that leverages mid-layer hidden states to evaluate semantic competence among agents, avoiding the overhead of full rollouts. Based on these familiarity scores, a cumulative-threshold adaptive routing mechanism dynamically activates a tailored minimal agent subset, achieving a trade-off between task performance and efficiency. By proactively allocating targeted training samples to agents based on their evolving competence, CERA-MoA promotes capability differentiation. Extensive experiments across various domains demonstrate that CERA-MoA outperforms state-of-the-art static-agent routing and fix-workflow fine-tuning baselines.
The Brief
2:17A short podcast summary of the paper. Written and voiced by AI.
Transcript
CERA-MoA. Co-Evolving Routing Mechanisms with Continually Learning LLM Agents. By Jiaxuan Jiang and two others, from the Institute for Interdisciplinary Information Sciences at Tsinghua University, Shanghai Jiao Tong University, and one other institution. Published on the sixteenth of September. Its code repository is worked out: about three hundred source files, with a readme.
Mixture-of-Agents systems, MoA, send each query to some of a set of agents. Current approaches usually treat two things as separate processes: routing queries to agents, and fine-tuning the agents themselves. That separation has two costs. The routing cannot adapt as the agents' abilities change during post-training. And the agents cannot specialize together in a way driven by the data, since the training they receive is not tied to where they are being sent.
CERA-MoA, a co-evolving router with continually learning agents for Mixture-of-Agents, joins the two processes. It is an iterative reinforcement learning framework in which a dynamic router and independent agent policies evolve together, each responding to the other.
A predictive familiarity estimator looks at the agents' mid-layer hidden states to judge which agents are semantically competent on a query, without the cost of running full rollouts. From those familiarity scores, an adaptive routing mechanism with a cumulative threshold activates a minimal subset of agents tailored to the query, trading task performance against efficiency. Reading hidden states instead of generating answers is what keeps the estimate cheap enough to use on every query. And activating only a minimal subset means not every agent has to answer every query.
Training samples are then allocated to agents deliberately, according to their evolving competence. That targeted allocation is what drives the agents to differentiate their capabilities from one another. Which agent learns from which examples depends on what each agent is becoming good at, and that is how the agents come to differ over the course of training.
Extensive experiments across various domains show that CERA-MoA outperforms state-of-the-art baselines of two kinds: routing among static agents, and fine-tuning with fixed workflows.
The idea is a feedback loop. Routing reflects what each agent is good at, training makes each agent better at what it is routed, and both keep moving together instead of one being frozen while the other learns.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A team of agents, a dispatcher that never learns, and agents that never specialize. That is the wall.
You run a team of language-model agents and send each query to some of them. One process decides who gets which query; a separate one fine-tunes the agents on fixed, hand-split data.
The road ahead
- 1 · The problem ~7 min When the dispatcher and the agents' training run as separate processes, routing goes stale as skills shift, and every agent stays a generalist.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Join routing and training into one loop: a cheap familiarity score picks a few agents per query, and those agents train on it, so they specialize.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min For each training query: read hidden states, score familiarity, activate a minimal subset, train those agents, then update the router.
- 6 · The experiment ~6 min You’ll guess: Against two kinds of baseline, routing among static agents and fine-tuning in fixed workflows, how did CERA-MoA perform?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.