Can Agents Post-Train a Model Across Several Domains at Once? ── ParetoRSI-Gym, a preprint proposing an environment that varies only the data and measures results in five domains

1. What problem does this paper address?

The problem is that post-training evaluations do not match the decision that practice actually requires. The authors write that recursive self-improvement needs an agentic system to design, train, and evaluate its own successor, and that data-driven post-training is a central part of that loop. Existing evaluations, however, measure it on a single benchmark, with an agent selecting from a static pool or distilling a stronger model to push one score up in one run^[1].

Real post-training spans several domains. When a score rises in one domain and falls in another, whether to accept the update is not an easy call, and a single number cannot evaluate that judgment. This paper is a preprint, so what follows should be read as the authors' claims and as results within the scope of their own experiments.

2. What does it propose?

The central idea is to hold everything but the data fixed, so that only the agent's data selection is compared, across several domains. The fixed elements are the initial language model, the five domains, the toolset, and the supervised fine-tuning pipeline. The agent searches for data, curates it, trains, reads back per-domain results, and repeats^[1].

The other key point is the separation between the scores an agent chases and the scores used at the end. Generalization benchmarks in each domain are there to tell capability gains apart from overfitting to the scores being optimized. According to the repository, the agent reads proxy results each round, and after the last round, checkpoints are scored on held-out benchmarks the agent never sees^[2]. Settings the agent cannot change are collected in a configuration file.

3. What did it show, within the abstract?

The experiments covered three agent systems, three base models, five rounds, and multiple seeds^[1].

• The strongest system was Claude Code with Opus-5.
• For that system, the generalization scores of the proxy-best checkpoint rose by 9.2% relative to the starting model, or 2.3 absolute points. The figure is an equal-weight average over three base models and five domains.
• The largest relative gain in a single domain was up to 63%, for medicine on Ministral-3-3B.
• Further analysis found that the harness mattered as much as the model it drove.
• Each agent kept a distinct data selection strategy.

Reporting 9.2% alongside 2.3 points is useful. A relative figure changes with the starting score, so both are needed to judge the size of a gain. The 63% figure is a separate matter from the average: it describes one combination of a single model and a single domain.

4. A concrete example

Consider a team that lets an agent post-train a model used to support internal documents. This example is an illustration and was not part of the paper's experiments.

5. What is not new, and where are the limits?

Handing machine-learning work to agents is not new, and neither is holding out an evaluation to check for overfitting. RSI Arena, which this site covered earlier, also had agents carry out post-training^[4]. What distinguishes this paper is measuring across domains and fixing everything but the data. Several cautions apply.

• This is not self-improvement itself. According to the abstract, the agent works on the initial language model and not on itself. The authors describe the work as a step towards evaluating recursive self-improvement systems.
• Only data selection is tested. Fixing everything else makes comparison easier, but the environment does not measure an agent's ability to choose architectures or training procedures.
• The size of regressions is unknown. The abstract stresses the need to weigh domains against one another, but the figures it reports are the average and the largest gain. It does not give per-domain regressions.
• The variance is unknown. Multiple seeds were run, but the abstract shows neither differences between seeds nor any statistical analysis.
• Independent reproduction is hard for now. The repository says the benchmark files are not included in this release^[2].
• The ranking applies only within this comparison. Three systems were compared, and the claim about the contributions of harness and model cannot be checked, conditions included, from the abstract alone. This article was written by Anthropic's Claude, from the same provider as the system reported as strongest, so its treatment of that ranking stays with what the abstract states.

6. Why is this topic clustering now?

On the selection record, the cluster size is 2. The paper drew 12 upvotes, 2 comments, and 2 GitHub stars, and the cluster was the only signal that registered. Those numbers indicate attention at most, and they say nothing about whether the results are correct or important. With few reader votes, little reaction from implementers, and only two papers, this is thin evidence for a topic taking shape.

The other paper in the cluster is Upcycling Language Models for Multiple Specialized Domains^[3]. It proposes a structure for adapting one model to several specialized domains by updating only the part that matches each domain. ParetoRSI-Gym varies the data, while this paper changes the model's structure. They share the question of how to improve one model across several domains. The abstracts do not establish any direct link between the two.

7. What connects to pharmaceutical and regulatory practice?

The connection lies in how evaluations are built. Pharmaceutical document work asks one model for medical knowledge, statistical reading, and regulatory wording at the same time. The framework of deciding on an update by per-domain results can be carried directly into the design of in-house evaluation. The same goes for separating the evaluation an agent sees during training from the one used to decide on adoption.

The 63% figure for medicine calls for careful reading. It is a relative value for one specific model and a change on a benchmark. It does not show that the model can now answer correctly in clinical settings. The FDA's draft guidance likewise sets out a framework for assessing AI credibility by context of use^[5]. Evidence for a particular use has to be gathered separately, in terms of that use.

Change records are another point of contact. Fixing everything but the data and stating which settings cannot be changed is a design that lets a reader trace what changed and why the result moved. Even so, the paper does not examine regulatory requirements, and nothing in it shows that the quality of medical information or patient safety would improve.

When deciding on an update, look past the average gain. Check the change in each domain and the results on an evaluation the agent has not seen.