Skip to content

Data and training signal

Training data

Models that generate, filter, label or curate the data their successors learn from.

Nemotron-4 340B

Aligned model generates training data
Reward model judges and filters it
Data trains next intermediate instruct model
Next model generates better alignment data
repeat

Training dataNVIDIA · 2024

Tongyi DeepResearch

Agent synthesizes agentic pretraining data
Pipeline adjusts training set in real time
Data feeds back into model training
repeat

Training dataTongyi Lab, Alibaba · 2025

Language model generates a self-edit
Edit is applied as weight update
RL rewards self-edit by downstream performance
repeat

Training dataMIT (Zweiger, Pari et al.) · 2025

Datasetswesmith.com

o3-mini injects bugs and writes issues
SWE-agent solves tasks for expert trajectories
Trajectories fine-tune Qwen 2.5 Coder
Open coding agent trained on agent data
repeat

Training dataStanford, Princeton, Alibaba Qwen (Yang et al.) · 2025

DatasetHugging Face

Llama-3-70B scores pages for educational value
Regressor trained on labels filters corpus
Filtered corpus pretrains new language models
LLM judgment sets next training diet
repeat

Training dataHugging Face · 2024

FoundationarXiv

Language model writes step-by-step rationales
Correct rationales are kept as dataset
Model is fine-tuned on self-generated set
Improved model generates next rationale dataset
repeat

Training dataZelikman et al. (Stanford, Google Research) · 2022

Recently on the radar

PaperarXiv·30 Sep 2026·Training data

False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

The loop: Self-evolving search agents generate their own training data, which often suffers from co-cheating where the proposer and solver agree on shared errors. This data is filtered using multi-sample verification and cross-fitting to reliably train the next generation of agents.

loop fit 9/10Meijia Chen, Hao Li, Zheng Lu et al.via arXiv

PaperarXiv·1 Oct 2026·Training data

No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning Collapse

The loop: A model generates synthetic text which is then filtered using a non-parametric entropy rate estimator to preserve diversity. This filtered data is used to iteratively fine-tune the model, preventing model collapse across successive generations.

loop fit 9/10Lewis Mitchellvia arXiv

Other loop types

Self-modifying agents Tools and harness Self-reward and self-play Kernels and compilers Hardware and chips Inference efficiency Architecture and optimizer search Research automation AI R&D evals Interpretability and oversight