Alex Open Research Wiki

PoLar / Program-of-Layers

2 min read

PoLar / Program-of-Layers

Summary

PoLar is the input-dependent execution controller introduced by Skip a Layer or Loop It?. It treats a frozen Transformer’s layers as callable functions, partitions depth into short contiguous segments, and predicts whether each segment should be skipped, executed once, or executed twice.

How It Works

PoLar has two phases:

  1. Offline discovery and learning. MCTS searches for layer-index paths that produce correct answers. A predictor is trained on those valid paths.
  2. Inference-time routing. A frozen Qwen3-Embedding-0.6B encodes the input; learned layer queries and a small cross-layer Transformer predict segmentation and operation logits; beam search returns one or more candidate programs for execution by the frozen base LLM.

The paper reports about 2.1M learned predictor parameters. That count excludes the frozen embedding model loaded by the released pipeline, so it should be read as trainable-controller size rather than total auxiliary deployment footprint.

What It Is Not

  • It is not a training-free system end to end: the base LLM stays frozen, but the controller is supervised on offline MCTS outputs.
  • It is not ordinary early exit: a program may skip internal segments and repeat others.
  • It is not a parameter-shared looped backbone: it reuses selected layers from an already trained stack rather than training one shared block for all depths.
  • It is not yet a time-series or world-model method: the paper does not use temporal numeric inputs, actions/control inputs, interventions, persistent latent state, or closed-loop rollouts.
  • Pass@ for is proposal-set coverage over multiple executed programs; the release does not provide a ground-truth-free selector for the winning answer.

Relation To Foundation TSFM Agenda

Use the source-level mapping in polar-2026 for evidence and caveats. At the entity level, PoLar is an adjacent adaptive-compute design: a future multivariate time-series model could predict a depth program from observation windows, exogenous variables, and actions or control inputs. The open problem is whether that routing preserves dense numeric state and rare events while moving a real quality/latency frontier under a declared compute budget.

Official Artifacts

Graph View

Backlinks