ATLAS: Learning to Optimally Memorize the Context at Test Time

Source

• Raw Markdown: paper_atlas-2025.md
• PDF: paper_atlas-2025.pdf
• Preprint: arXiv 2505.23735

Core Claim

ATLAS proposes a higher-capacity long-term memory module that optimizes memory from current and past tokens, then uses it to define a broader DeepTransformers family.

Relevance To This Wiki

ATLAS is a direct continuation of the Titans memory line. It is useful for tracking whether test-time memory improves because of larger memory capacity, better update objectives, or more expressive memory management.

Limitations

It is still mostly upstream sequence-model evidence; the paper should not be treated as proof that test-time memory solves multivariate time-series state tracking without direct domain tests.

Foundation TSFM Relevance

Adjacent to streaming state, long context, and dynamic compute. The open question is whether optimized context memory can carry durable latent state for numeric systems rather than only token recall.

Links Into The Wiki

• ATLAS
• Looped Transformers And Test-Time Memory
• Efficient Recurrent Sequence Models
• Time-Series Scaling And Efficiency
• 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?