T2 Scaling Optimizes Train & Inference Compute

🔑 Enhanced Key Takeaways

• •T2 scaling addresses the 'inference-time compute' gap by formalizing the trade-off between pretraining compute and test-time compute (e.g., chain-of-thought, self-consistency, or tree-of-thought search).
• •The research demonstrates that for a fixed total compute budget (pretraining + inference), the optimal model size is significantly smaller than predicted by Chinchilla, which only considered pretraining compute.
• •Empirical validation shows that T2-optimized models achieve higher accuracy on reasoning-heavy benchmarks like GSM8K and MATH by allocating more compute to inference-time search rather than static model parameters.

🛠️ Technical Deep Dive

• •The T2 scaling law is defined by a joint optimization function: C_total = C_train + N * C_inference, where N is the number of test-time samples.
• •The framework utilizes a power-law relationship between test-time compute (number of samples) and performance, similar to the power-law relationship between pretraining compute and loss.
• •Implementation involves shifting the 'compute frontier' by reducing parameter count (N) and increasing training tokens (D) to maximize the performance-per-token-per-sample ratio.
• •The approach specifically targets agentic workflows where inference-time search (e.g., Best-of-N sampling) provides diminishing returns for larger models but significant gains for smaller, heavily overtrained models.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline