Your Router Is Already Writing the Next Training Set

TL;DR for operators

A multi-model agent platform already records much of what a post-training system needs: the request, the model-selection decision, tool interactions, execution failures, verification signals, and eventual outcome. NeoHorse-1 turns those records into training supervision rather than treating them as operational exhaust.¹

The strongest practical evidence is not the recursive-self-improvement claim. It is the controlled comparison of training data. Using the same routing-guided training configuration, Qwen3.5-4B trained on routing-harness trajectories reaches a 70.57 average across five benchmarks, versus 64.32 when trained on Toucan public synthetic agent data. The difference is 6.26 points.

The system also uses the router’s estimate of required capability to order training examples and student rollouts. Across ten benchmarks, this post-training raises the same-size Qwen3.5 base average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B.

For operators, the implication is concrete: preserving structured execution evidence can make production traffic part of a governed learning pipeline. What remains unproven is whether repeating this loop produces sustained gains across generations. NeoHorse-1 reports one completed evaluation-selection-update pass.

Production traces may be more valuable than generic agent data

A deployed agent system already knows more about its own workload than most static training corpora do. It observes which requests actually arrive, which model handles them, what tools are invoked, where execution breaks, whether recovery succeeds, and whether the final task is completed.

The usual operational design treats these records primarily as telemetry. They support debugging, evaluation, billing, and routing policy. NeoHorse-1 asks a different question: can the same records also improve the models that will handle future requests?

The paper’s most revealing comparison addresses that directly. Under matched initialization, curriculum schedule, optimizer settings, random seed, packing method, evaluation protocol, and closely matched training budgets, routing-harness trajectories outperform Toucan public synthetic tool-agent data on every one of five compared benchmarks. The average rises from 64.32 to 70.57. The largest reported gaps are 11.31 points on tau2-Bench and 8.54 on HumanEval.

This is not evidence that deployment data always beats public data. It is evidence that, under this specific controlled recipe, trajectories grounded in actual routing and execution contain supervision that the compared public synthetic source does not reproduce as effectively.

The router becomes a training signal

The next step is not simply collecting more logs.

NeoHorse-1 keeps three things distinct: what capability level the router predicted a request would require, what routing action the system actually took, and what happened afterward. That separation matters because the tier ultimately served can be affected by availability, overrides, risk controls, and deployment policy. The served model is therefore not treated as the task’s difficulty label.

Instead, the system derives a capability-demand score over four routing levels, C0 through C3. That estimate becomes an ordering signal for training examples.

The supervised fine-tuning stage uses three curriculum phases. Lower-demand examples appear earlier; higher-demand examples are introduced progressively; later phases retain some easier material rather than replacing it completely. The routing score changes when examples are presented, not their loss weight.

The same principle extends to on-policy distillation. Here, the student generates responses from contexts it actually encounters, and a fixed teacher provides token-level distributional guidance on those student-generated prefixes. Routing again determines which contexts are emphasized at each stage.

Operationally, the router is doing more than allocating inference. Its estimate of capability demand helps determine what the student learns next.

The gains support the loop, but capacity still matters

Across the ten benchmarks reported in the main evaluation, NeoHorse-1-4B raises the Qwen3.5-4B average from 58.94 to 64.87, a gain of 5.93 points. NeoHorse-1-9B raises its corresponding base from 65.60 to 69.04, a gain of 3.44 points.

The 4B result is especially relevant to deployment economics. The post-trained 4B model matches or exceeds the Qwen3.5-9B base on several benchmarks. That does not make a 4B model equivalent to a 9B model, but it shows that targeted agentic post-training can recover part of a scale gap.

The remaining gap is informative. The 9B NeoHorse model stays stronger on execution-heavy tasks and in representative traces involving iterative debugging, failure recovery, state maintenance, and strategy revision. The case analyses describe better evidence acquisition, requirement tracking, verification, recovery, and artifact delivery after training, but those examples are illustrative rather than population-level estimates.

A smaller model can become considerably more useful through better supervision. Model capacity still provides advantages when execution requires maintaining and revising a longer chain of stateful decisions.

Evaluation can redirect the next training mixture

NeoHorse-1 also sketches the mechanism needed to close the loop.

Evaluation results are organized into capability-deficiency profiles. Those deficiencies can then shift the composition of future training data toward weaker capability regions while retaining broad coverage. In the paper’s supervision-scaling experiment, nested routing-harness subsets are trained under fixed initialization, optimization, packing, and training passes. The five-benchmark development average increases from 69.31 for the base model to 71.45 at the largest displayed supervision scale.

That experiment is best read as a scaling test within the observed range. It supports the claim that additional quality-ranked harness supervision can continue improving the evaluated aggregate. It does not establish that gains continue indefinitely or that the router’s capability score is a calibrated measure of difficulty.

For an operator, the proposed cycle is straightforward:

The architecture turns evaluation from a reporting endpoint into an input to the next training decision.

This is one improvement cycle, not yet recursive self-improvement

The title of the paper points toward recursive self-improvement, but the reported evidence stops earlier.

NeoHorse-1 demonstrates one evaluation-selection-update cycle. It shows that harness-generated evidence can be converted into post-training data, that routing-guided training improves benchmark results, and that the resulting evaluation can inform later data allocation. It does not report successive model generations repeatedly producing additional gains from their own newly generated experience.

Other boundaries also matter. The evaluation concentrates on text-based agentic execution, tool use, coding, and instruction following rather than the full capability range a heterogeneous harness might serve. The routing signal is used as a practical capability-demand proxy but is not demonstrated to be a calibrated estimator of task difficulty or model deficiency. Some comparison-model results come from official external reports rather than the authors’ own evaluation pipeline.

So the business conclusion is narrower than the paper’s long-term ambition. Organizations operating agent platforms have evidence that production trajectories can be engineered into valuable post-training assets and that routing signals can help organize how those assets are used. They do not yet have evidence that repeatedly running this architecture will produce durable recursive improvement across generations.

That is already a meaningful shift in system design. Once routing, execution, verification, and outcome records are preserved together, the serving layer no longer only decides which model handles today’s task. It begins to shape the evidence available for training tomorrow’s model.

Cognaptus: Automate the Present, Incubate the Future.