Adaptive Depth in Looped Transformers: Diagnosing Learned Halting Gates and Trajectory Readouts

The Entanglement Problem in Adaptive Computation

Looped Transformers represent an innovative approach to increasing computational depth in neural networks by repeatedly applying a single, shared recurrent block. This design aims to achieve adaptive computation, allowing the model to perform more steps for complex inputs and fewer for simpler ones, thereby optimizing inference time. However, the current paradigm for learned halting objectives in these architectures presents a significant challenge. These objectives typically use a single exit distribution that serves a dual purpose: it acts as the stopping rule during inference and simultaneously dictates how strongly each intermediate state is supervised during training. This inherent entanglement means that the mechanism deciding when to halt also shapes the very sequence of internal states, or "trajectory," that the model generates.

The consequence of this entanglement is that suboptimal adaptive-compute performance can arise from several intertwined sources. It might be due to a poorly learned readout mechanism that fails to accurately assess when to stop, or from an inefficiently formed trajectory of intermediate states, or even from a complex interaction between both. This makes diagnosing the root cause of inefficiency particularly difficult, as improvements in one aspect might inadvertently affect the other. The paper argues that this conflation of roles has often been left implicit in prior work on learned halting, obscuring a clearer path to optimization.

A Diagnostic Lens for Looped Transformers

To systematically investigate this issue, the researchers adopted a "trajectory-readout lens," applying it across both controlled synthetic tasks and larger, more realistic models. The synthetic tasks, such as modular arithmetic and binary parity, allowed for precise control over the problem space, enabling clear observation of how different halting strategies influenced model behavior. This controlled environment was crucial for isolating variables and understanding fundamental mechanisms before scaling up.

A key methodological innovation was the use of "fixed-prior depth supervision." This technique focuses on shaping the model's internal trajectory without simultaneously imposing an input-dependent halting policy. By doing so, the study could observe whether the intermediate states generated by the model inherently contained useful stopping signals. The findings were significant: this approach produced "difficulty-aware trajectories," meaning the model's internal states naturally evolved in a way that reflected the complexity of the input, exposing clear opportunities for effective halting.

Furthermore, the research explored the efficacy of "simple post-hoc confidence readouts." These are straightforward mechanisms applied after the trajectory is formed to decide when to halt, rather than being learned jointly with the trajectory. Surprisingly, these simple readouts often matched or even outperformed more complex, learned linear and multi-layer perceptron (MLP) gates. This suggests that the problem might not lie in the expressivity of the halting gate itself, but rather in how the trajectory is formed when the gate is trained concurrently.

Implications for Efficient AI Models

The most critical insight from this diagnostic evaluation is that fitting gates on "frozen trajectories" — trajectories that were shaped independently of the halting gate — localized the failure point. It became evident that the primary impediment to optimal adaptive compute performance stemmed mainly from the trajectory induced by joint gate training, rather than from any inherent limitation in the gate's ability to express a halting policy. This reframes the problem: it's not just about teaching the model when to stop, but also about ensuring the internal thought process (the trajectory) is conducive to an efficient stop.

This pattern was consistently observed in evaluations of large-scale Ouro-1.4B and 2.6B checkpoints, which are substantial pre-trained models. The study found that while existing "ponder gates" (a type of learned halting gate) were competitive, they were not uniformly Pareto-optimal across all metrics. Crucially, the measured latency confirmed that the reductions in average exit depth achieved through these disentangled approaches translated directly into practical, real-world inference-time savings. This validates the theoretical findings with empirical evidence from production-scale models.

In essence, this systematic diagnostic evaluation fundamentally reframes adaptive depth in Looped Transformers. It shifts the focus from merely "gate learning" to a more holistic understanding of adaptive computation as a joint problem involving both "trajectory formation" and "exit readout." By highlighting this distinction, which prior learned-halting work often left implicit, the research provides a clearer roadmap for developing more efficient, dynamically adaptive, and high-performing Transformer architectures, ultimately contributing to the broader goal of more resource-aware artificial intelligence.