Production routing logs already contain the training signal for recursive self-improvement

A deployed routing harness already knows which requests exceed a model's current capabilities. NeoHorse-1 turns that knowledge into a training signal.

The core observation is straightforward: every time a routing system selects a stronger model tier for a request, it implicitly labels that request as beyond the weaker model's reach. NeoHorse-1 captures these labels—the predicted capability tier, the tier actually served, and the interaction that followed—and uses them to organize training data into a three-stage curriculum. Easier interactions come first; harder ones arrive progressively. The same routing scores then schedule starting contexts for on-policy distillation, where a teacher model supervises responses the student itself generates rather than responses copied from a fixed dataset.

The data pipeline is unusually careful. Trajectories pass through structural validation that checks causal event ordering and closed tool-call pairs, then through six-dimensional semantic evaluation covering goal attainment, tool use, error recovery, and termination quality. Each dimension gets a PASS/WARN/FAIL verdict grounded in trajectory evidence, never compressed into a single heuristic score. A subscene-level labeling scheme tracks whether goals are new, continued, modified, or abandoned. The result is a training corpus that knows not just what happened but why it matters.

The benchmark numbers are real but modest. At 4B scale, macro-average score rises from 58.94 to 64.87 across ten benchmarks; at 9B, from 65.60 to 69.04. The largest gains appear on harness-based agent tasks and coding, while instruction-following scores stay roughly flat. The post-trained 4B model narrows the gap to the 9B base model substantially—which is the point: agentic post-training partially substitutes for scale.

The trajectory case studies are more revealing than the aggregate numbers. On a ticket-reporting task, the 9B base model retains contradictory totals and assigns a next-day state to an in-window record; NeoHorse-1-9B matches independent recomputation exactly. On a code-repair task, the base model stops after one failed attempt; NeoHorse-1-9B runs a full edit–test–inspect–repair cycle. On a constrained data-analysis task, the 9B model keeps retrying an unavailable dependency while NeoHorse-1-9B switches to standard library alternatives, finishing with fewer requests and less token usage.

The authors are explicit that this is a single-iteration prototype. The feedback loop—updated models return to the harness, generate new trajectories, feed the next training mixture—has been designed but not yet run for multiple generations. Whether gains compound across iterations remains untested. The routing signal itself could also be sharpened: right now it organizes curriculum order but does not yet train the router or produce calibrated difficulty estimates.

What NeoHorse-1 demonstrates is that the infrastructure already present in a production routing system contains most of what recursive self-improvement requires. The harness logs capability demand; the logs become curriculum; the curriculum improves the model; the improved model goes back into the harness. The loop is closed in design. Whether it stays closed under repeated iteration is the open question.