TIC-FM

A split-masked latent-memory framework for training-free time series in-context classification.

Using TIC-FM for time series classification

A split-masked latent-memory framework for training-free time series in-context classification.

In this benchmark, TIC-FM is evaluated using training-free in-context classification. Checkpoint information: TSEncoder_orion_icl_fullv1.0.pt. Check the results below for the settings and dataset scopes with published scores.

Evaluated configurations

Native in-context inference

Classifier heads are grouped under their shared pretrained model here. They remain separate leaderboard entries because they are different experimental configurations.

Organization: GDUT · Huawei Noah’s Ark Lab · Université Paris Cité · MBZUAI
Parameters: Not reported by authors
Checkpoint used: TSEncoder_orion_icl_fullv1.0.pt
Input Type: Univariate time series
Inference: Training-free in-context classification

TIC-FM classification benchmark results

Compare accuracy and average rank only within the same evaluation setting and dataset scope. Standard uses the official training splits; 5-shot and 10-shot use different budget-feasible subsets. Rank and average rank include other models in the full comparison set. Scores are not averaged across configurations.

For comparisons with other backbones, open the foundation model classification leaderboard and use the same setting and scope. The few-shot evaluation protocol explains the label budgets.

Standard · Univariate

Comparison set: 19 configurations

This Standard time series classification comparison reports 1 TIC-FM configuration for the univariate dataset scope. Each row shows its own coverage.

The highest reported average accuracy among this model's configurations in this setting is 80.01%, reported by TIC-FM with coverage 128 / 128. This configuration ranks 3 of 19 in this comparison set. The result describes this configuration and setting, rather than every possible use of the pretrained model.

TIC-FM — Standard Univariate time series classification benchmark results
Configuration	Univariate Rank	Univariate Accuracy (%)	Avg. Rank	Coverage
TIC-FM	3	80.01	7.11	128 / 128

5-shot · Univariate

Comparison set: 19 configurations · Budget-feasible Low-shot subset

This 5-shot time series classification comparison reports 1 TIC-FM configuration for the univariate dataset scope. Each row shows its own coverage.

The highest reported average accuracy among this model's configurations in this setting is 68.08%, reported by TIC-FM with coverage 112 / 112. This configuration ranks 1 of 19 in this comparison set. The result describes this configuration and setting, rather than every possible use of the pretrained model.

TIC-FM — 5-shot Univariate time series classification benchmark results
Configuration	Univariate Rank	Univariate Accuracy (%)	Avg. Rank	Coverage
TIC-FM	1	68.08	7.57	112 / 112

10-shot · Univariate

Comparison set: 19 configurations · Budget-feasible Low-shot subset

This 10-shot time series classification comparison reports 1 TIC-FM configuration for the univariate dataset scope. Each row shows its own coverage.

The highest reported average accuracy among this model's configurations in this setting is 74.33%, reported by TIC-FM with coverage 97 / 97. This configuration ranks 1 of 19 in this comparison set. The result describes this configuration and setting, rather than every possible use of the pretrained model.

TIC-FM — 10-shot Univariate time series classification benchmark results
Configuration	Univariate Rank	Univariate Accuracy (%)	Avg. Rank	Coverage
TIC-FM	1	74.33	6.70	97 / 97

TIC-FM paper, code and checkpoint

Reading this model's evaluation

The method catalog describes the pretrained model; leaderboard eligibility depends on the actual coverage of each experimental configuration. Catalog inclusion does not imply complete coverage of all 198 datasets.

Compare evaluated configurations Read the evaluation and ranking rules Explore aggregate capability analysis