Predicting how well a large language model will reason no longer requires actually training one.
Researchers have introduced rBridge, a method that uses models with fewer than 1 billion parameters to predict the reasoning performance of models up to 32 billion parameters. The technique works by weighting a standard training signal - negative log-likelihood - with how closely a dataset aligns to the target reasoning task, then using outputs from frontier models as reference labels. In experiments across six reasoning benchmarks, rBridge produced the strongest correlation between small-proxy scores and large-model results of any tested approach. It also transferred those predictive relationships to new pre-training datasets without additional fine-tuning.
The practical implication is significant: teams evaluating which datasets to use before a large training run can now get a reliable signal at a fraction of the cost. The paper claims a cost reduction of over 100x compared to the best prior baseline, which matters because dataset selection at scale is one of the least visible but most expensive parts of building a capable model.
Reasoning has long been the awkward exception to the proxy-model playbook. Emergent capabilities - skills that appear suddenly past a certain parameter threshold, often above 7 billion - made small models unreliable predictors of large-model behavior. rBridge does not eliminate that threshold so much as work around it by aligning the proxy's objective more tightly to the task. Whether it holds up outside controlled benchmark conditions, and across the messier data mixtures real training runs use, is the question the paper leaves open.