When to Ponder: Adaptive Compute Allocation for Code Generation via Test-Time Training
This addresses the challenge of adaptive compute allocation for code generation, offering a training-free method to improve efficiency, though it is incremental as it builds on existing test-time training techniques.
The paper tackles the problem of uniform computation in large language models by proposing PonderTTT, a gating strategy that uses self-supervised reconstruction loss to selectively trigger test-time training updates, achieving 82-89% Oracle Recovery and up to 16% lower loss on out-of-distribution languages compared to baselines.
Large language models apply uniform computation to all inputs, regardless of difficulty. We propose PonderTTT, a gating strategy using the TTT layer's self-supervised reconstruction loss to selectively trigger Test-Time Training (TTT) updates. The gating decision itself is training-free--requiring no learned classifier or auxiliary networks; only a single scalar threshold is initially calibrated on unlabeled data and continuously adapted via EMA to maintain target update rates. Our experiments with GPT-2 models (124M to 1.5B) on code language modeling (The Stack v2, teacher-forced perplexity) demonstrate that this signal is inference-compatible, requiring no ground-truth labels. Our Reconstruction Gating achieves 82-89% Oracle Recovery while being fully training-free, significantly outperforming Random Skip baselines (up to 16% lower loss on OOD languages).