Inference in large language models (LLMs) is usually performed by a fixed‑depth, fixed‑order forward pass: every input traverses all layers sequentially. The human brain, however, routes information flexibly via the thalamus to different cortical areas according to demand. Li et al. (2026) introduced program‑of‑layers (PoLar), treating transformer layers as a library of functions rather than a rigid chain, enabling a brain‑like routing scheme. PoLar dynamically selects contiguous blocks of layers to skip or repeat, constructing an adaptive program for each input.
We reconstructed PoLar’s diagnostic Monte‑Carlo Tree Search (MCTS) in greater detail and evaluated it on five models. Our reproduction confirmed several of the paper’s main findings:
- Skipping blocks outperforms the standard forward pass;
- Repeating blocks outperforms pure skipping;
- Combining skipping and repeating yields the best performance.
Shorter programs suffice for easy questions, while harder ones require more layer repetitions. However, we could not replicate the authors’ claim about a learned router for single‑shot inference: its top‑ranked prediction consistently collapsed back to the standard pass, even though the top‑k predicted programs collectively provided a measurable accuracy gain.
Beyond reproduction, we discovered that a small set of generic programs solves most queries. A deeper analysis of program structure and robustness revealed that error‑correcting programs are highly brittle—undoing a single edit inside a program typically breaks the correction. This mirrors thalamo‑cortical coordination in the brain, suggesting PoLar captures key principles of flexible information routing.
The codebase is publicly released at https://datexis.github.io/RE-PoLar/.
Review