Agents that build models — automated data engineering, training runs, distillation, and synthetic data. The agent-era successor to AutoML.
English · 简体中文
🤖 Daily arXiv radar: docs/radar.md — auto-refreshed every day.
This is one of three sibling lists, split by the thing that gets improved:
| Product | List |
|---|---|
| Code (software engineering an agent can verify against tests) | awesome-self-improving-coding-agents |
| Skills & organizations (how firms absorb agents) | awesome-ai-native |
| Models & data — this list | awesome-agentic-ml |
In scope: systems where the deliverable is a model or a dataset — agents running training experiments, cleaning and evaluating data, distilling teachers into students, generating synthetic training data. When an agent edits train.py so the val loss drops, that's here; when it edits application code so tests pass, that's the sibling list.
Out of scope: classic MLOps (deploying/monitoring models — see the incumbents in Related lists), and pre-agent AutoML papers (NAS era — same).
flowchart LR
D[Data: clean / eval / synthesize] --> T[Training run]
T --> E[Evaluate]
E -->|agent reads results| A[Agent decides the next change]
A --> D
A --> T
E --> S[Student model / distilled dataset]
| # | Work | Why start here |
|---|---|---|
| 1 | autoresearch — Karpathy | The whole thesis in three files: the agent edits training code overnight, the human writes program.md, a fixed 5-minute budget keeps experiments comparable. |
| 2 | MLE-bench | The yardstick for this entire list: can an agent do ML engineering end-to-end (Kaggle-style)? |
| 3 | Agent-R1 | The training substrate: step-level MDP RL for tool-using agents. |
| 4 | EasyDistill | The data view: black-box teachers → SFT/DPO datasets, including agent-trajectory distillation. |
| System | What it automates |
|---|---|
| autoresearch — Karpathy | Single-GPU overnight training experiments (~100/night): agent edits train.py, keeps what improves val_bpb, discards the rest. |
| Deep Researcher Agent · paper | 24/7 DL experimentation with explicit discipline — narrow search, forced verification, human control. |
| karpathy (the agent, not the person) | An agentic Machine Learning Engineer — the job title as a system. |
| Optim-Agent | LLM agents as your hyperparameter optimizer — the AutoML core loop, agent-first. |
| autogluon-assistant | Multi-agent system on top of AutoGluon for end-to-end multimodal AutoML. |
| System | What it is |
|---|---|
| Agent-R1 · paper | Agentic RL with step-level MDP transitions; Online Policy Distillation branch bridges RL training and distillation. |
| NeoHorse · report | Open-weight 4B/9B agent models built via agentic post-training with a routing harness; the Jev decision primitives. |
| System | What it is |
|---|---|
| dash | A self-learning data agent built with systems-engineering primitives. |
"Distillation" now means three different things; this list tracks the agent-facing end of all three:
| System | Which distillation |
|---|---|
| torchdistill | Classical KD (logits/features) — 26+ methods, maintained since 2019; the elder statesman. |
| EasyDistill | LLM-era data distillation: compose generate/eval/filter/rewrite operators into SFT/DPO pipelines, agent trajectories included. |
| distilly | Cognitive distillation: people → skills. (The org-side story lives in awesome-ai-native.) |
The training loop's fuel is data the model (or an agent) generates itself:
| Year | Work | What it established |
|---|---|---|
| 2022 | STaR | Bootstrap reasoning by training on self-generated rationales that lead to correct answers. |
| 2022 | Self-Instruct | The model writes its own instruction-tuning data — seed of every synthetic-data loop. |
| 2023 | ReST^EM | Generate → filter → fine-tune, scaled past human data. |
| 2024 | SPIN | Self-play fine-tuning: beat your own past iterations. |
| 2024 | Self-Rewarding Language Models | The model judges and revises its own outputs to create its next preference data. |
| 2025 | Absolute Zero | Zero external data: propose tasks, verify with a code executor, train on the result. |
| 2025 | SEAL | The model edits its own fine-tuning data — self-modification at the training-recipe level. |
The counterweight: model collapse (Nature) — recursive training on self-generated data degrades distribution tails. Every pipeline above should be read against this result.
| Benchmark | Measures | Link |
|---|---|---|
| MLE-bench | End-to-end ML engineering, Kaggle-style | paper · code |
- Data agents lack a benchmark. MLE-bench covers training runs; nobody has the "SWE-bench of data engineering" (clean → evaluate → pipeline agents on real warehouses).
- Experiment comparability at scale. autoresearch's fixed-budget trick works on one GPU; there is no standard for comparing agent-run experiments across platforms.
- License clarity. Several fast-moving systems in this space ship non-OSI licenses (source-available, NOASSERTION) — check before commercial use.
- Distillation governance. When agents distill teachers into students (or colleagues into skills), provenance and permission trails don't exist yet.
docs/radar.md is regenerated daily by scripts/radar.py (GitHub Actions, workflow) from arXiv new-submission listings, keyword-filtered for agentic-ML / data-agent topics. Precision-first keywords — tune via PR.
- awesome-self-improving-coding-agents — sibling: agents improving code (SWE-verifiable work)
- awesome-ai-native — sibling: the organizational layer
- KaiWU5/Awesome-AI4AI — the paper-survey companion to this space (223 papers)
- wasiahmad/Awesome-LLM-Synthetic-Data — synthetic-data generation reading list
- hibayesian/awesome-automl-papers — the pre-agent AutoML canon
- EthicalML/awesome-production-machine-learning — the production/MLOps incumbents
PRs welcome — see CONTRIBUTING.md. The bar: primary-source links, one line on what it automates, and respect for the product boundary (code → sibling list; org → sibling list).
CC0 1.0 — public domain.