Preprint trains the router alongside the agents it selects

single source· 1 articles · confidence: low · first seen 2026-09-15 20:00 UTC

What this means for you

Nothing to act on. This is an abstract with no scores, no evaluation date, no harness and no released code, so there is nothing to reproduce or compare against. The thing to watch is whether the authors publish numbers and a training recipe.

A preprint posted to arXiv on 15 September describes CERA-MoA, a scheme that trains the router and the agents it selects in one loop. Mixture-of-agents systems (several models answer, a router chooses among them) normally freeze the router while the agents are fine-tuned. CERA-MoA scores each agent from mid-layer hidden states — internal activations partway through the network — so it can judge competence without running full generations, then routes each query to the smallest sufficient subset and allocates further training samples by competence. The authors report gains over static routing and fixed-workflow baselines, but the abstract gives no scores, evaluation dates, harness or code.

Key facts

  • ·CERA-MoA is posted as arXiv paper 2609.18779, dated 15 September 2026. source
  • ·The framework co-trains a dynamic router and independent agent policies in an iterative reinforcement learning loop. source
  • ·A predictive familiarity estimator reads mid-layer hidden states to score agents without full rollouts. source
  • ·A cumulative-threshold routing rule activates a minimal subset of agents per query. source
  • ·The authors report outperforming static-agent routing and fix-workflow fine-tuning baselines, without publishing scores, evaluation dates or a harness. source
  • ·The abstract mentions no code or weights release. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire