LoopFormer: Elastic-Depth Looped Transformers for Latent Reasoning via Shortcut Modulation

Source

Core Claim

LoopFormer trains looped Transformers on variable-length trajectories using shortcut consistency so representations remain useful under different compute budgets.

Relevance To This Wiki

It turns recurrent depth into an explicit elastic-depth interface, where loop count can be conditioned by the available budget rather than fixed at training and inference.

Limitations

The current evidence is language modeling and reasoning. Budget conditioning needs careful comparison against fixed-depth baselines under the same latency and memory targets.

Foundation TSFM Relevance

Adjacent to fixed-FLOPs hierarchy and dynamic compute, especially for systems that need controllable inference effort.

Open Questions

  • What matched-budget baseline should this source be compared against: unique-depth Transformer layers, recurrent state, explicit memory, or extra inference steps?
  • Which claims transfer from token-sequence reasoning to multivariate time-series state tracking, event streams, or action-conditioned world models?
  • Does shortcut consistency remain useful under hard serving latency budgets, or does trajectory training add overhead that only pays off offline?