2026 Feature Update 3

[mdekstrand]

This is the 3rd feature update release for the 2026 release series. See the Release Notes for highlights, but major improvements include:

• Much faster SLIM
• Bug fixes for FlexMF
• Steam dataset import
• Direct Optuna tuning works (and is best-supported method)

And much more.

What's Changed

• Add support for importing Steam datasets by @mdekstrand in Add support for importing Steam datasets #1111
• Update locked max dependency versions by @mdekstrand in Update locked max dependency versions #1118
• Fix type-checking and KNN types for NumPy 2.5 / SciPy 1.18 by @mdekstrand in Fix type-checking and KNN types for NumPy 2.5 / SciPy 1.18 #1121
• flexmf: rewrite misranked (for WARP) sampling loop by @mdekstrand in flexmf: rewrite misranked (for WARP) sampling loop #1120
• Save more search results in search backend by @mdekstrand in Save more search results in search backend #1122
• Add TrainingOptions.torch_compilation_enabled and disable by default on free-threaded Python by @mdekstrand in Add TrainingOptions.torch_compilation_enabled and disable by default on free-threaded Python #1123
• Add nested thread pool support and use it in hyperparameter tuning by @mdekstrand in Add nested thread pool support and use it in hyperparameter tuning #1125
• tuning: fix Optuna search point usage by @mdekstrand in tuning: fix Optuna search point usage #1130
• Don't panic when ALS solve fails by @mdekstrand in Don't panic when ALS solve fails #1131
• Fix inverted logic on Torch compilation by @mdekstrand in Fix inverted logic on Torch compilation #1133
• Fix Optuna iterative tuning epoch extraction by @mdekstrand in Fix Optuna iterative tuning epoch extraction #1132
• Improve Optuna tune running by @mdekstrand in Improve Optuna tune running #1134
• Add support for boolean and categorical hyperparameters to tuning by @mdekstrand in Add support for boolean and categorical hyperparameters to tuning #1139
• Speed up Slim recommendation with sparse matrices by @mdekstrand in Speed up Slim recommendation with sparse matrices #1140
• Add tests for Optuna pow2 search space by @mdekstrand in Add tests for Optuna pow2 search space #1144
• Include component defaults in Optuna search space by @mdekstrand in Include component defaults in Optuna search space #1145
• Change from pre-commit to hk and update dev setup instructions by @mdekstrand in Change from pre-commit to hk and update dev setup instructions #1135
• Update dependency versions and locks by @mdekstrand in Update dependency versions and locks #1147
• Add reconfigure() function to test configuration responsiveness by @mdekstrand in Add reconfigure() function to test configuration responsiveness #1146
• Refactor pipeline configuration generation to be in nodes by @mdekstrand in Refactor pipeline configuration generation to be in nodes #1149
• Expose batch results as a row-by-row iterator by @mdekstrand in Expose batch results as a row-by-row iterator #1128
• Move config file schemas to new lenskit.schemas package by @mdekstrand in Move config file schemas to new lenskit.schemas package #1148
• Lock EASE solving to work around parallel bug by @mdekstrand in Lock EASE solving to work around parallel bug #1152
• Add top-level search defaults and make stopping behavior configurable by @mdekstrand in Add top-level search defaults and make stopping behavior configurable #1151
• Nest parallel inference inside tuning by @mdekstrand in Nest parallel inference inside tuning #1153
• Refactor and further split tests by @mdekstrand in Refactor and further split tests #1154
• Refactor Optuna tuner into a package by @mdekstrand in Refactor Optuna tuner into a package #1156
• Implement search point abstraction and fix pow2 space by @mdekstrand in Implement search point abstraction and fix pow2 space #1157
• Reduce time spent on eval tests by @mdekstrand in Reduce time spent on eval tests #1158
• Make summary Markdown optional in data import by @mdekstrand in Make summary Markdown optional in data import #1160
• Update to Ruff 0.16, enable default lints, fix lint errors by @mdekstrand in Update to Ruff 0.16, enable default lints, fix lint errors #1161
• Add and fix Clippy lints by @mdekstrand in Add and fix Clippy lints #1163
• Cast item lists to common schema when converting to Arrow by @mdekstrand in Cast item lists to common schema when converting to Arrow #1167
• Implement collectible warnings and warning collection for batch inference by @mdekstrand in Implement collectible warnings and warning collection for batch inference #1168
• Add UniqueItemCount metric by @mdekstrand in Add UniqueItemCount metric #1169
• Update locked + cargo dependencies for 2026.3 release by @mdekstrand in Update locked + cargo dependencies for 2026.3 release #1171
• Reformat TOML with Tombi by @mdekstrand in Reformat TOML with Tombi #1172
• Add interaction="default" to add_relationship_class by @mdekstrand in Add interaction="default" to add_relationship_class #1174
• Run dataset tests on Woodpecker with full data access by @mdekstrand in Run dataset tests on Woodpecker with full data access #1181
• Import Steam interactions (Australia subset) by @mdekstrand in Import Steam interactions (Australia subset) #1179

Full Changelog: v2026.2.1...v2026.3.0

---

This discussion was created from the release 2026 Feature Update 3.