News · 2026-09-09

The day's most-upvoted paper turns a model's own routing logs into its next training set

A group publishing as the NeoHorse Team released NeoHorse-1, a system that records which model in a pool handled each user request and how the interaction went, then turns those logs into the training data for the next round. It became the most-upvoted paper of the day on Hugging Face by more than double the runner-up, and its framing is unusually direct: the authors present it as a prototype path toward recursive self-improvement. The most concrete result is that their post-trained 4-billion-parameter model closed most of the gap to an untrained model more than twice its size.

Key facts

• The headline result: a post-trained 4B model substantially closes the gap to the 9B base model across eleven evaluations.
• When: submitted to arXiv 8 September 2026.
• Who: the NeoHorse Team, a 37-author group including Yunhe Wang, Kai Han, Yu Wang and Guohao Dai.
• Primary source: arXiv 2609.08183; code at github.com/TokenRhythm/NeoHorse.

Start with the phrase that made it travel. Recursive self-improvement usually conjures a model editing its own weights in a runaway loop. That is not what this is, and the paper is honest about it. The loop here is offline and operated by people: serve traffic, log it, filter it, train on it, serve again. What makes it interesting is where the training signal comes from.

Most post-training data is either written by humans or generated on purpose. NeoHorse-1 instead harvests the exhaust of its own serving infrastructure. The system runs what the authors call a routing harness in front of a heterogeneous pool of models — a router that guesses how much capability an incoming request needs and dispatches it to a model of the right size. For every user turn it records three things: the predicted capability demand, which service tier it chose, and everything that happened next.

That last part is the trick. Because the router had to make a prediction and the interaction then either went well or badly, ordinary production traffic arrives pre-labelled. The system knows what it expected and what it got.

Those records become training examples that, in the authors' words, "preserve interleaved reasoning, tool calls, and harness context" — not just the final answer but the whole working trace, including the scaffolding the agent operated inside. They are then filtered through structural validation, a six-dimensional semantic evaluation and subscene-level labelling before anything is trained on.

The routing signal does double duty. It also organises the curriculum: supervised fine-tuning runs in three stages ordered by the routing labels, and the same progression carries into what the authors call routing-guided on-policy distillation, where a larger teacher supervises responses the student generated itself. Finally, capability-guided allocation reads the evaluation results and decides what the next training mixture should contain. The authors describe the result as "closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next."

The analogy that fits is a hospital that keeps records of which cases got sent to which specialist and how each turned out, then uses that ledger to redesign its own residency programme — and then uses the next cohort's outcomes to redesign it again. No single step is exotic. The compounding is the claim.

Results span eleven evaluations covering harness-based agents, tool use, coding and instruction following. The meaningful framing is not the raw scores but the shape: post-training lifted the 4B model far enough that it closed most of the aggregate gap to the 9B base model. Getting most of the way to double your parameter count without adding parameters is the kind of efficiency result that matters commercially, whatever one thinks of the self-improvement framing.

Why it matters this week is partly accident and partly not. The same day this paper topped the charts, an Anthropic researcher resigned saying labs are "racing straight to self-improving superintelligence". The loudest news story and the loudest paper of the day were the same subject approached from opposite ends — one describing the destination as a danger, the other publishing a working prototype of the first step. Neither cites the other.

The honest caveats are real and worth stating. The authorship is thinly attributed: "NeoHorse Team" fronts a 37-person list whose institutional affiliations are not spelled out in the listing, and a paper claiming progress toward recursive self-improvement deserves more scrutiny of provenance than a typical release, not less. The demonstrated loop is one or two rounds of an offline pipeline, not a sustained ascent, and the paper does not establish where the returns stop — which is the question that actually determines whether "recursive" is the right word. Training a model on traces its own router selected also carries a well-known risk of narrowing: the system may get better at the traffic it already sees while quietly losing ground elsewhere, a failure mode related to reward hacking and to the general fragility of synthetic data loops. The Hugging Face discussion is worth reading alongside the paper.

Primary source, verified: read the paper → (arXiv 2609.08183)

Ground Truth. (2026, September 9). The day's most-upvoted paper turns a model's own routing logs into its next training set. Ground Truth. https://groundtruth.day/news/neohorse-1-feeds-its-own-routing-logs-back-into-training.html

@misc{groundtruth:neohorse-1-feeds-its-own-routing-logs-back-into-training,
  title  = {The day's most-upvoted paper turns a model's own routing logs into its next training set},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/neohorse-1-feeds-its-own-routing-logs-back-into-training.html}
}

Topics: research · self-improvement · post-training · agents · routing