TEMPO: Scaling Test-time Training for Large Reasoning Models

The buried problem here is that test-time training on reasoning models tends to degrade because the reward signal itself drifts as the policy improves, creating a feedback loop where the model optimizes against an increasingly stale critic. TEMPO's alternating schedule is essentially a stabilization mechanism, not just a performance trick.

This connects directly to the fixed-point analysis covered in 'Stability and Generalization in Looped Transformers' from April 16, which showed that stable, reachable fixed points require specific architectural conditions during test-time compute scaling. TEMPO approaches the same stability problem from the training dynamics side rather than the architectural side, making the two papers complementary lenses on the same underlying challenge. The 'Evaluation-driven Scaling for Scientific Discovery' piece from the same day is also relevant: SimpleTES uses parallel exploration with feedback loops, and TEMPO's critic recalibration serves a structurally similar corrective function, though the domains and mechanisms differ.

Watch whether TEMPO's stability gains hold when the critic recalibration interval is varied across significantly different task distributions, particularly on multi-step mathematical reasoning benchmarks like AIME 2025. If performance degrades sharply with longer recalibration gaps, the framework's practical utility narrows considerably.