Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2610.39246
IMRouter: Imbalance-aware Temporal Pattern Routing for LLM-based Time Series Forecasting
IMRouter addresses temporal pattern imbalance in LLM-based time-series forecasting, where frequent local dynamics dominate optimization and less frequent patterns receive poorer modeling. Its Pattern-Adaptive Temporal Representation Learning component routes patches to specialized experts using gradient conflict and local gradient density, while Routing-Guided Semantic Alignment transfers routing information into cross-modal alignment with mutual-information-based distillation. Across seven long-term benchmarks, M4 short-term forecasting, and 10%-data few-shot experiments, IMRouter reports lower forecasting errors than the compared baselines, with particularly clear gains on less frequent Traffic patterns.
How it works
The forecasting backbone normalizes each channel, divides it into sliding-window patches, and maps the patches into temporal tokens. Temporal tokens are aligned with compressed text prototypes through cross-attention and then processed by a frozen LLM. The paper's key motivation is expressed by Eq. (5): the aggregate loss weights each pattern-specific loss by $\frac{|D_j|}{|D|}$, and Eq. (7) consequently weights its gradient contribution by the same empirical frequency.
What the paper evaluates
For long-term forecasting, the paper evaluates ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity (ECL), and Traffic with input length 96 and prediction horizons 96, 192, 336, and 720. It reports MSE and MAE, where lower values are better, against TimeSup, FSCA, CALF, Time-LLM, GPT4TS, PatchTST, iTransformer, and TimesNet. The reported aggregate comparison is provided in tab:long_term_forecasting: IMRouter has 0.171 MSE and 0.263 MAE on ECL, and 0.423 MSE and 0.285 MAE on Traffic. For short-term forecasting, the M4 benchmark contains 100,000 univariate series across Yearly, Quarterly, Monthly, Weekly, Daily, and Hourly frequencies.
Limit to keep in view
The pattern diagnostics and pattern-wise evaluation use TS2Vec representations clustered with HDBSCAN; the paper itself notes that clustering can be sensitive to algorithmic assumptions and hyperparameters.
arXiv:2611.00562
TabAbax: A Tabular Foundation Model Technical Report
TabAbax is a tabular foundation model built on TabICLv2 for predicting on unseen datasets from labeled in-context examples. It keeps the 12-layer in-context transformer and changes two upstream components: it improves the synthetic structural causal model (SCM) pretraining prior and adds learned Fourier scalar features followed by a second column–row encoding pass before row compression. On TabArena’s 13-dataset regression suite, it reports 2,024 Elo, 336 above the TabICLv2 baseline and 21 above TabFM. A prior-only ablation reaches 1,916 Elo, suggesting that much of the regression improvement comes from the changed training prior.
How it works
For a table with n rows and m columns, TabAbax preserves TabICLv2’s column embedding, row interaction, compression, and ICL stages, with reported runtime complexity O(n² + nm²). The grouped embedding forms feature tokens from columns at circular offsets (j, j+1, j+3) mod m and applies a shared linear projection.
What the paper evaluates
The authors evaluate on BeyondArena and TabArena. BeyondArena contains 142 datasets (98 classification and 44 regression), with IID, grouped, and temporal splits and strata covering dataset scale, dimensionality, and feature type. Its size distribution is 80 datasets below 10K rows, 35 between 10K and 100K, and 27 above 100K. TabArena contains 51 datasets (38 classification and 13 regression), with 29 below 10K rows, 20 between 10K and 100K, and two above 100K. Its Elo rating is anchored at default Random Forest = 1000; the reported leaderboards include 99 evaluated entries. The paper also reports average rank, oracle improvability (lower is better), and win rate alongside Elo.
Limit to keep in view
The authors state that predictability filtering reduces the effective synthetic corpus and may leave weak-signal, high-noise real tasks underrepresented.
arXiv:2610.41956
Chronos-2: A foundation model for multivariate time series forecasting
Chronos-2 is a 120M-parameter encoder-only Transformer for zero-shot forecasting across univariate and multivariate time series with heterogeneous covariates. It alternates temporal attention within each series with group attention across related targets and covariates, uses patch-based tokenization and quantile prediction, and is pretrained on synthetic multivariate systems containing instantaneous and lagged dependencies. On fev-bench, Chronos-2 improves skill scores by 5.2 percentage points over the strongest time-series foundation-model baseline and increases win rate by 10 percentage points through in-context group attention, while forecasting 100 series in under one second on an NVIDIA A10G GPU.
How it works
Chronos-2 uses a T5-derived, encoder-only Transformer with 120M parameters. Time attention aggregates information across patches within one series, while group attention aggregates across all series assigned to the same group at a shared patch index. A group can represent one series, multiple related targets, variates of a multivariate series, or targets together with covariates.
What the paper evaluates
The main benchmark suite includes fev-bench with 100 tasks, 68 of which involve multivariate dependencies; Chronos-Bench-II with 27 primarily univariate tasks; and GIFT-Eval with 97 primarily univariate tasks. Baselines include AutoETS, AutoARIMA, AutoTheta, statistical ensembles, LightGBM, CatBoost, DeepAR, TFT, PatchTST, TiRex, TimesFM-2.5, Toto-1.0, Moirai-1.1, Moirai-2.0, Granite-TTM-R2, TimePFN, and TabPFN-TS. Point forecasts use MASE, probabilistic forecasts use SQL or WQL, and the paper also reports win rate, skill score, and improvability.
Limit to keep in view
Zero-shot hydrology remains below the specialized LSTM ensemble, with median NSE of 0.68 versus 0.90, although fine-tuning closes the gap.
arXiv:2611.07870
FARR: Forecast-Aware Residual Refinement with Matched Endpoint Evidence for Long-Horizon Time-Series Forecasting
FARR improves long-horizon time-series forecasting by exposing how auxiliary conditioning changes a frozen forecaster's output. It generates a valid forecast and a matched null-condition forecast, uses their difference together with historical statistics to predict a bounded residual, and applies horizon-wise gating before adding that residual to the effective valid forecast. Across ETTh1, Weather, Traffic, and Electricity (ECL), FARR reports the lowest dataset-average MSE and MAE among the listed methods, while adding the cost of a second backbone inference.
How it works
For each forecast origin, the Endpoint Proposal Module runs the same frozen forecaster twice: once with standard conditioning and once with a fixed null operation. The two passes share the checkpoint, historical input, preprocessing, output settings, and sample order, so their difference isolates the conditioning intervention. The method then forms an 18-dimensional feature vector per horizon and channel from 168 recent observations, endpoint values, disagreement statistics, and normalized horizon position.
What the paper evaluates
The evaluation uses ETTh1, Weather, Traffic, and Electricity (ECL) with input length \(L=672\) and horizons \(H\in\{96,192,336,720\}\). A StandardScaler is fitted on the training split and applied unchanged to validation and test data. Locally evaluated MSE and MAE are computed in standardized space, and dataset averages weight the four horizons equally. Endpoint forecasts use a frozen Qwen2.5-0.5B-Instruct checkpoint with label length 576 and token length 96. Baselines include SE-LLM, TimeMixer++, Time-MoE, TQNet, TimeCMA, LLM-TS, AutoTimes, and Time-LLM.
Limit to keep in view
The method requires two frozen-backbone passes at inference; on ETTh1, full FARR takes 85.12 ms versus 63.03 ms for the single-backbone configuration.
Read the full reviews
- IMRouter: Imbalance-aware Temporal Pattern Routing for LLM-based Time Series Forecasting · arXiv:2610.39246
- TabAbax: A Tabular Foundation Model Technical Report · arXiv:2611.00562
- Chronos-2: A foundation model for multivariate time series forecasting · arXiv:2610.41956
- FARR: Forecast-Aware Residual Refinement with Matched Endpoint Evidence for Long-Horizon Time-Series Forecasting · arXiv:2611.07870
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.