Research map / Time series / Foundation models for time series

TiRex-2: Generalizing TiRex to Multivariate Data and Streaming

NeurIPS 2026 · Time series › Foundation models for time series · approach: LLM-based time series forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces TiRex-2, a recurrent xLSTM-based time series foundation model that handles multivariate targets with past and future covariates, with target-causal bidirectional covariate processing and streaming inference at constant cost per patch.

Problem

Existing multivariate time series foundation models are mostly Transformer-based, so their inference cost grows with context and they must reprocess full history when streaming; univariate recurrent TSFMs ignore cross-variate dependencies.

Approach

Combine a recurrent xLSTM time mixer (forward-only for targets and past covariates, weight-tied reverse pass for future covariates) with an asymmetric grouped-attention variate mixer that blocks covariate-to-target flow, plus an on-the-fly synthetic multivariate coupling pipeline for pretraining.

Key results

• On fev-bench, TiRex-2 attains the lowest mean MASE (ablation full model 1.527 MASE) and is reported as state-of-the-art zero-shot on both fev-bench and GIFT-Eval.
• Uses about 2.2x fewer mean active parameters than Chronos-2 and about 4.3x fewer than TimesFM-2.5 (expected active ~53.84M).
• Streaming evaluation remains stable up to 32M steps, i.e. about 4000x beyond the 8k post-training context.

Surprising findings

• Removing the variate mixer's group attention causes the largest ablation degradation (+0.220 MASE on fev-bench), while removing the binary-aware scaler or sLSTM blocks changes results only marginally.

Limitations

• Future-known covariates must be recomputed when they change online, so streaming is not fully free for them.
• Streaming operates at patch granularity (32 steps).

Experimental setup

Datasets: univariate time series corpora (synthetic coupling) fev-bench GIFT-Eval dysts

Benchmarks: fev-bench GIFT-Eval dysts

Metrics: MASE SQL CRPS pairwise win rate skill score quantile loss

Baselines: Chronos-Bolt Moirai-2 Toto-1.0 TiRex TimesFM-2.5 Chronos-2 Chronos-2-Synth PatchTST-FM-r1 FlowState-r1.1 Moirai Moirai-MoE GTT

Builds on: TiRex xLSTM Chronos-2 TimeXer Timer-XL CITRAS reversible instance normalization

Closest papers by meaning

• TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning (NeurIPS 2025 · similarity 0.92)
• xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories (NeurIPS 2025 · similarity 0.89)
• TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables (NeurIPS 2024 · similarity 0.88)
• Timer-XL: Long-Context Transformers for Unified Time Series Forecasting (ICLR 2025 · similarity 0.88)
• Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.88)
• TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.87)