Alex Open Research Wiki

EIDOS: Latent-Space Predictive Learning For Time Series Foundation Models

3 min read

  • Important
  • Skimmed

EIDOS: Latent-Space Predictive Learning For Time Series Foundation Models

Source

Core Claim

EIDOS shifts time-series foundation-model pretraining from direct future-value prediction to latent-space predictive learning with observation-space grounding.

Benchmarked Model Entry

  • Model: EIDOS
  • Family: latent-space predictive time-series foundation models
  • Parameters: 12.7M for the main small model; scaling experiments also report 23.3M base and 91.7M large variants.
  • Primary task surface: zero-shot probabilistic forecasting over univariate time series.
  • Evaluation surface: GIFT-Eval as the primary benchmark, with additional fev-bench results.
  • Local artifact status: the expected official code drop external/EIDOS/ and checkpoint external/EIDOS/eidos 1.pt are not present in this checkout.

Key Contributions

  • Trains a causal Transformer to predict the evolution of latent representations.
  • Maps each univariate scalar sample through a SiGLU point-wise tokenizer, using sine basis responses plus data-dependent gating.
  • Uses a lightweight aggregation branch to construct stable target representations.
  • Combines latent alignment, grounding, and forecasting supervision in one objective.
  • Reports robust performance and improved latent organization on GIFT-Eval-style benchmarks.

Method Notes

EIDOS is the main source for Latent-Space Predictive Learning in the time-series cluster and is also linked to Time-Series Foundation Models.

Its tokenizer is important for Number Tokenization: EIDOS keeps the prediction unit aligned with each raw numeric time-series sample instead of aggregating patches first. The paper’s SiGLU tokenizer is a sine-activated gated linear unit, not a standard SwiGLU block. The sine activation supplies bounded Fourier-like scalar responses, while the gate selects useful responses for the latent token.

EIDOS also simplifies JEPA-style target construction. It removes the separate auxiliary target encoder and instead aggregates future point-wise embeddings with a lightweight target branch plus stop-gradient. The grounding head reconstructs future observations from latent targets, so it is best treated as an observation anchor and collapse-prevention mechanism rather than as a general auxiliary numeric-value decoder.

Evidence And Results

The source emphasizes reduced structural fragmentation, noise robustness, feature probing, latent steering, and competitive zero-shot forecasting.

Limitations

The current evidence is forecasting-centered and univariate-first. It should be compared with reasoning-focused TimeOmni-1 and generation-focused TimeOmni-VL. The paper does not yet establish how the SiGLU point-wise tokenizer should extend to multivariate time series, known future exogenous variables, control inputs, interventions, or auxiliary metadata.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Patch size, dynamic tokenization, and point-wise numeric embeddingspartially closesUses SiGLU point-wise scalar tokenization so each raw sample becomes a latent token rather than being compressed into fixed patches.Demonstrated for univariate forecasting; no adaptive token budget or multivariate/channel metadata handling.
Representation quality: semantic state vs dense detailpartially closesPredicts latent future representations while grounding them with an observation-space quantile head; layer probes and latent steering show trend/periodicity directions.Steering is diagnostic, not a general editing API; dense reconstruction/control tradeoffs are not user-selectable.
Anti-collapse regularizationpartially closesReports that grounding loss prevents near-constant latent solutions in the next-embedding objective.Collapse tests are internal ablations, not broad long-tail, cross-channel, or intervention stress tests.
Control and counterfactualsinsufficient evidenceLatent steering changes forecast attributes in representation space.No action, control-input, intervention, or counterfactual rollout interface is evaluated.

Open Questions

  • Can EIDOS-style latent predictive learning support causal or language-based reasoning?
  • How should observation grounding be balanced against latent abstraction?

Graph View

Backlinks