The expansion of large language models beyond assistance roles into autonomous system development represents one of the most discussed shifts in contemporary artificial intelligence. As these models grow in capability, a natural progression emerges: from supporting human researchers to independently conceiving, executing, and refining machine learning experiments. This trajectory, often framed as AI for AI or recursive self-improvement, depends on systems that can manage experimentation over extended timescales. Workflow-driven machine learning engineering platforms provide the orchestration layer necessary for such long-horizon runs, yet they typically require significant human specification. Recent efforts have moved toward model-directed interaction, where the model itself organizes experiments and adapts strategies based on feedback. However, each experiment's outcome is the cumulated result of prior decisions, and a single training run may encompass dozens or hundreds of such steps. Isolating useful supervision from these elongated interaction trajectories, and learning from intermediate outcomes rather than only final results, remains an open problem. The paper under review introduces MLExpert, a 35-billion-parameter Mixture-of-Experts agentic model trained via a novel post-training pipeline designed to address this very challenge.
Standard supervised fine-tuning of language models assumes that labeled data arrives in independent, identically distributed chunks. In the context of autonomous machine learning engineering, however, the model generates its own training signals through interaction with an environment. These signals are sparse, delayed, and entangled with the sequence of actions that preceded them. Simply treating each experiment's final metric as a monolithic reward signal discards the granular information embedded in intermediate states. The authors argue that effective learning in this setting requires two complementary changes: a data construction process that produces executable, well-scoped ML tasks, and a training objective that can assign credit across multiple decision points within a single interaction window.
To meet the data construction challenge, the authors design the Kaggle-MLE Factory. This system generates machine learning engineering tasks modeled after competitions on platforms like Kaggle. Each task includes a well-defined dataset, a modeling objective (such as classification or regression), and automatically generated evaluation metrics. The Factory wraps these tasks in an interaction harness that records the model's actions—such as data preprocessing choices, feature engineering steps, model selection, and hyperparameter adjustments—as a sequential trace. Critically, the harness also captures the intermediate outputs of each step, enabling downstream analysis of which actions contributed most to the final result. Tasks are designed to be executable within a bounded compute budget, and their difficulty spans common ML engineering scenarios: limited training data, imbalanced classes, noisy features, and feature selection under high dimensionality. By automating task generation and execution, the Kaggle-MLE Factory eliminates the need for manual task design while providing a consistent testbed for evaluating model-directed MLE behavior.
Once tasks and interaction harnesses are in place, the next requirement is a training method that can extract useful supervision from the resulting trajectories. The authors propose hierarchical agentic SFT as the first stage. This step transfers general tool-use and coding capabilities from a base model to the MLE domain. The key insight is that many of the individual actions within an ML engineering trace—reading a file, writing a Python script, invoking a pipeline function—are generic tool-use skills already present in strong base models. Hierarchical SFT organizes supervision outcome-guided: instead of treating every token in a trace as equally informative, the method identifies segments of the trajectory that led to favorable intermediate outcomes (such as a decrease in validation loss after a particular feature-engineering step) and uses those segments as positive examples. Less successful segments are downweighted or excluded. This outcome-guided approach ensures that the model learns to reproduce the patterns of effective ML engineering without being flooded by noise from unsuccessful attempts.
The second training stage introduces Windowed Interaction RL (WIRL). While SFT can instill baseline tool-use proficiency, it cannot adequately address the credit-assignment problem across multi-turn decisions. WIRL treats the decisions between successive experiment results as a single execution window and assigns a shared outcome-based reward to the entire window. Concretely, suppose a model makes a sequence of choices—selecting a model architecture, tuning a learning rate, and deciding on a regularization strategy—before observing a validation metric. WIRL Backpropagates the final metric (or a smoothed version of it) through all actions within the window, allowing the policy to update based on the collective impact of those actions. Asynchronous collection and policy correction further support learning: while some windows are being executed and evaluated, other parallel windows generate data that can be used for immediate policy updates. This means the model does not need to wait for a complete trajectory to terminate before its parameters change; learning can proceed continuously as data becomes available.
On MLE-Bench Lite, the combined pipeline yields dramatic improvement. The base model, before any post-training, achieves an average medal rate of 37.9% across the benchmark's tasks. After hierarchical SFT and WIRL, MLExpert's average medal rate climbs to 72.7%. This places the model in the performance range of larger frontier models such as Kimi-K3 and GLM-5.3, despite operating at a fraction of their parameter count. The improvement is consistent across task subtypes within MLE-Bench Lite, including tasks focused on data preprocessing, model selection, and hyperparameter tuning. Notably, the gain is not uniform across all tasks; those requiring longer sequences of decisions tend to show larger relative improvements, suggesting that the WIRL component is particularly effective when the decision horizon is extended.
Beyond MLE-Bench Lite, the authors evaluate MLExpert on several supplementary benchmarks. NatureBench Lite, which focuses on neural architecture search and related optimization problems, sees gains over the base model, though the absolute medal rates are lower than those on MLE-Bench Lite. This is perhaps unsurprising, given the distinct nature of architecture-search tasks versus traditional ML engineering workflows. The authors also report results on agent scaffold optimization and algorithm optimization tasks, where MLExpert again outperforms the base model. Across all these settings, the pattern is consistent: the post-training pipeline imparts a measurable optimization capability that generalizes beyond the specific training tasks. The authors interpret this as evidence that the model learns not just shortcuts for particular competition formats but broader principles of experimental design and iterative improvement.
One technical detail worth highlighting is the role of the Mixture-of-Experts architecture in MLExpert. The base model is a 35B parameter MoE, which activates a subset of experts per token. This sparsity structure has two consequences for the training pipeline. First, the routing decisions made by the model— which experts are activated for a given input token— become part of the trace recorded by the interaction harness. Second, the expert activations themselves can be interpreted as a form of adaptive computation: the model routes different input patterns to different specialized computation paths. The authors note that this structure may contribute to the model's ability to specialize different experts for different sub-tasks within an ML engineering workflow, though they do not isolate this effect experimentally. The MoE gating mechanism is trained jointly with the post-training objectives, and the authors observe that expert diversity, as measured by the entropy of activation patterns, increases after WIRL training.
The paper also acknowledges limitations and open directions. The improvements, while substantial, are still evaluated on a finite set of benchmarks; broader applicability to real-world ML pipelines, where data quality, domain-specific constraints, and regulatory considerations play a role, remains to be seen. The Kaggle-MLE Factory, while effective for generating standardized tasks, abstracts away many of the messy details of real ML engineering projects. Additionally, the WIRL reward is based on final outcome metrics; intermediate rewards shaping—such as rewarding decreases in loss after each mini-batch—is an area the authors flag for future work. The computational cost of the post-training pipeline, while not excessive relative to the size of the model, is non-trivial: each training step requires running full experiment trajectories, which can be expensive even with parallelization. Finally, the authors note that the current approach assumes a fixed horizon per task; adapting the method to settings where the number of permissible decisions varies dynamically is an open problem.
From a practical standpoint, MLExpert demonstrates that post-training methods can meaningfully extend the autonomous capabilities of large Mixture-of-Experts models. The Kaggle-MLE Factory provides a reproducible framework for task generation and trajectory collection, while the combination of hierarchical SFT and WIRL offers a concrete path for learning from long-horizon interaction data. For researchers or organizations looking to build agentic ML systems, the paper suggests a blueprint: start with a base model—ideally one with an MoE architecture for efficiency—design a task factory that produces executable ML engineering challenges, and then apply outcome-guided supervised fine-tuning followed by windowed interaction reinforcement learning. The reported results indicate that even a single cycle of this pipeline can move the needle from near-random performance on MLE tasks to a level competitive with much larger models.
Mirroring broader trends in the field, the work also points toward the growing importance of reinforcement learning from human or environment feedback in the training of large models. The transition from pure next-token prediction to models that can actively plan, execute, and evaluate sequences of actions is a key frontier, and MLExpert contributes a concrete instantiation of this shift. The windowed interaction paradigm, in particular, offers a flexible template: any setting where an agent executes a sequence of actions and receives a delayed reward can potentially benefit from the same credit-assignment strategy. Whether the target is autonomous robotics, code generation, or data science workflow automation, the core idea—treating a series of decisions as a window and assigning shared outcome-based credit—may prove widely applicable.
In summary, the paper addresses a pressing gap in the pursuit of autonomous machine learning engineering: how to train models on long, sparse, and entangled interaction trajectories. The proposed solution—a pipeline combining a task factory, hierarchical outcome-guided SFT, and windowed interaction RL—produces MLExpert, a model that significantly outperforms its base on standard MLE benchmarks and generalizes to related problem domains. The Mixture-of-Experts architecture, already chosen for its parameter efficiency, also appears to align well with the routing and expert-diversity dynamics induced by the training method. While limitations remain, the empirical results are compelling enough that the approach is likely to influence subsequent work on agentic models for ML and beyond. The authors have made the code, data, and trained checkpoints publicly available, facilitating further experimentation and extension.