One trainer can replace pretraining, midtraining, SFT and RL as separate phases, and a 124M-parameter model trained that way beat staged training on the same compute
AspireAI community · arXiv
A new paper argues the split between pretraining, midtraining, supervised fine-tuning and reinforcement learning is a convenience rather than a necessity. The authors recast suitable RL objectives as maximum-likelihood fitting to model-generated target distributions, so static data and RL generations can share one weighted likelihood loss, one data format and one trainer. Staged training then becomes the special case where all mixture weight sits on one source at a time. In controlled experiments the team trained a 124M-parameter model from scratch on a fixed 20B-token pretraining corpus and a 510M-token midtraining corpus, mixing midtraining data into the tail of pretraining and through RL instead of running the phases in sequence; the best unified configurations beat staged training on task accuracy and on generalization to larger unseen puzzles, and the paper reports that unified training also holds general capabilities better and gives a stronger starting point for continual learning. The limitation is scale: the headline controlled result is a 124M-parameter model, and the real-world experiment uses a 0.76B model trained on Math12K, so this is a framework and a direction, not yet a demonstrated win on frontier-scale training.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.