Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach

Source

• Raw Markdown: paper_huginn-2025.md
• PDF: paper_huginn-2025.pdf
• Preprint: arXiv 2502.05171
• Official model: tomg-group-umd/huginn-0125
• Official code and data: seal-rg/recurrent-pretraining

Core Claim

Huginn scales a recurrent-depth language model that can spend additional test-time compute by iterating a recurrent block in latent space instead of emitting more reasoning tokens.

Relevance To This Wiki

This is a scale proof for looped Transformers after the original UT idea: the model can be pretrained at billions of parameters and use more loops at inference.

Limitations

Reasoning gains are language-benchmark evidence; they do not directly establish better state tracking for numeric time series.

Foundation TSFM Relevance

Adjacent to dynamic compute: loop count becomes a serving-time budget knob, potentially useful for uncertain windows or planning rollouts if transferred carefully.

Links Into The Wiki

• Huginn
• Looped Transformers And Test-Time Memory
• Efficient Recurrent Sequence Models
• Time-Series Scaling And Efficiency
• Efficient Parallel Samplers for Recurrent-Depth Models
• Foundation Time-Series Model Research Agenda

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?
• How much of the gain comes from recurrent depth versus data mixture, tokenization, prelude/coda design, or other training choices?