Michael Y. Hu, Jane Pan, Ayush Rajesh Jhaveri, Nicholas Lourie, Kyunghyun Cho
Neural scaling laws predict how language model performance improves with increased training inputs. While aggregate metrics like validation loss can follow smooth power-law curves, individual downstream tasks exhibit diverse scaling behaviors: some improve monotonically, others plateau, and some even degrade with scale. We argue that predicting downstream performance from validation loss suffers from two limitations: averaging token-level losses obscures signal, and no simple parametric family can capture the full spectrum of scaling behaviors. To address this, we propose Neural Neural Scaling Laws (NeuNeu), a neural network that frames scaling law prediction as time-series extrapolation. NeuNeu combines temporal context from observed accuracy trajectories with token-level validation losses, learning to predict future performance without the limitations inherent in assuming a specific functional form. Trained entirely on open-source model checkpoints from HuggingFace, NeuNeu achieves 1.99% mean absolute error in predicting model accuracy on 66 downstream tasks -- a 44% reduction compared to logistic scaling laws (3.56% MAE). Furthermore, NeuNeu generalizes zero-shot to unseen model families, architectures, parameter counts, and downstream tasks. Our work suggests that predicting downstream scaling directly from data outperforms parametric alternatives.
Michael Y. Hu, Apurva Gandhi, Kyunghyun Cho, Tal Linzen, Pratyusha Sharma
Data mixing decides how to combine different sources or types of data and is a consequential problem throughout language model training. In pretraining, data composition is a key determinant of model quality; in continual learning and adaptation, it governs what is retained and acquired. Yet existing data mixing methods address only one phase of this lifecycle at a time: some require smaller proxy models tied to a single training phase, others assume a fixed domain set, and continual learning lacks principled guidance altogether. We argue that data mixing is fundamentally an online decision making problem---one that recurs throughout training and demands a single, unified solution. We introduce OP-Mix (On-Policy Mix), a data mixing algorithm that operates across pretraining, midtraining, and instruction tuning. Our main insight is that candidate data mixtures can be cheaply simulated by interpolating between low-rank adapters trained directly on the current model, eliminating separate proxy models and ensuring the search is always grounded in the model's actual learning dynamics. Across the language model training lifecycle, OP-Mix consistently finds near-optimal mixtures while using a fraction of the compute of the baselines. In pretraining, OP-Mix improves upon training without mixing by 6.3% in average perplexity. For continual learning, OP-Mix matches the performance of both retraining and on-policy distillation while using 66% and 95% less overall compute, respectively. OP-Mix suggests a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.
Bing Yan, Anuroop Sriram, Zachary Ward Ulissi, Kyunghyun Cho, Ricky T. Q. Chen, Benjamin Kurt Miller
Molecular 3D structures underpin simulations of thermodynamics, kinetics, crystallization, and nucleation, making their accurate prediction central to drug discovery and materials design. However, molecular geometry can change dramatically across environments. Existing models either ignore environmental effects or train separate models for each environment. We hypothesize that transfer learning across contexts is possible for conformer generation and that a general model will perform better when training data are limited, as in ligand binding and crystallization. We introduce EVA-Flow, a unified framework for environment-aware molecular conformation generation that learns shared chemical structure and adapts it through environment-conditioned generative modeling. Across vacuum, protein-ligand docking, solvation, and crystal packing, pretraining and cross-environment finetuning provide their largest gains in the three data-scarce settings. Evaluation against multiple reference conformations shows that EVA-Flow covers a broad range of solvation conformations. Analyses of molecules shared across environments support its ability to generate different conformations for the same molecule in different environments.
Many high-throughput scientific experiments produce distributional outputs, requiring conditional generative models to interpret the data, and to predict outcomes at new inputs. Acquiring experimental data at new conditions (e.g., for a new antigen or a new cell type) is often expensive, motivating the need for automated experimental design methods. However, existing active learning frameworks are limited as they assume scalar or vector-valued outputs. We first show that the Wasserstein test risk of a conditional generative task decomposes into a mean term and a shape term, demonstrating why mean-based active learning is blind to distributional shape. To capture both, we propose the transport Neural Tangent Kernel (tNTK), a tractable kernel that measures how sensitive a generative model's predicted distribution is with respect to its parameters. The tNTK recovers the standard NTK in the deterministic limit and upper-bounds the posterior Fréchet variance of the prediction under small parameter perturbations. It admits a closed-form expression for affine Gaussian transport and tractable Monte Carlo estimators otherwise. Treating the tNTK as the covariance function of a Gaussian process, we greedily select inputs that minimize the total GP posterior variance across the pool, yielding a distribution-aware active learning strategy. We evaluate this strategy across four classes of conditional generative models, including conditional flow matching and conditional Monge gap, on synthetic distributions, intent-labeled sentences, and single-cell perturbation responses. In these experiments, the tNTK outperforms the baseline NTK computed from the conditional mean, demonstrating the value of distribution-aware active learning for generative-modeling tasks.