1.3.1

Compatibility

• Requires BoTorch 0.18.1 (#5226)
  • Makes JAX / fully-Bayesian generator backend an optional dependency.
  • Users who want to leverage the fully-Bayesian generators (method = "quality" in ax/api), should use pip install "ax-platform[fully_bayesian]" to install with NumPyro and other related dependencies.

1.3.0

Compatibility

• Require BoTorch==0.18.0 (#5216)
  • Requires PyTorch>=2.4.0
  • Utilizes NumPyro for faster fully-Bayesian model fitting, requiring JAX as a dependency
• Storage: Add step_size to SQAParameter class, and as a column to the parameter_v2 table in the DB (#5212).

API Changes (ax/api)

• Client.attach_data (and therefore Client.complete_trial) now auto-registers metrics in raw_data that are not part of
  the optimization config as tracking metrics, matching the documented contract. Previously this raised UserInputError (#5135)
• Client.configure_optimization / objective and outcome-constraint strings now support linear equality constraints (w^T x == b) (#5174)

Docs

• New example starting an Ax experiment from dataframe data (#5098)

Core (ax/core)

• Objective and OutcomeConstraint are now expression-based and no longer hold Metric instances;
  metrics live only on the Experiment. Construct them with an expression string, e.g.
  Objective(expression="ne1 + 2*ne2") or OutcomeConstraint(expression="qps >= 7000")
  (single, scalarized, and multi-objective are all expressed this way). Old-style constructors still
  work but discard the Metric after init. Objective.metric is removed -- get a metric's name from
  the expression and look it up via Experiment.get_metric(name). This is a step toward deprecating
  ScalarizedObjective, MultiObjective, ScalarizedOutcomeConstraint, and ObjectiveThreshold (#5000, #5122, #5045, #5041)
  Note: We are aware of some edge-case bugs in this setup and we may be iterating on this in future releases.
• ParameterConstraint gains a new equality= kwarg for linear equality constraints (w^T x == b); equality constraints are honored in SearchSpace membership checks (#5173, #5175)
• Make all mark_* methods on BaseTrial no-op when the status is unchanged, and avoid rewriting timestamps when the status is unchanged (#5097, #5074)
• Runner lifecycle for search space editing: new Runner.update and Runner.on_search_space_update methods, a typed RunnerConfig infrastructure, and an UNSET sentinel (#5131, #5132, #5133, #5130)
• Support for Language-in-the-Loop labeling trials, with data-freshness hashing/stamping and stale-data handling (#4986, #4992, #5144)
• AddExecutionViability transform (lives in adapter; relies on core trial/metric plumbing) (#4547)

Generation (ax/generators, ax/adapter, ax/generation_strategy)

• Replace Pyro with NumPyro (JAX-backed) for fully Bayesian NUTS inference -- roughly 25x faster fit on CPU with equivalent model quality (#5087)
• Threaded equality constraints from ParameterConstraint all the way down to BoTorch (#5176, #5177, #5179, #5178, #5182)
• Open-sourced the Transfer Learning stack: TransferLearningAdapter (registered as Generators.BOTL) (#5052, #5048, #5096, #5047)
• Heterogeneous TL now defaults to MultiTaskGP + LearnedFeatureImputation (ImputedMultiTaskGP) (#5106, #5183, #5193, #5192)
• Dispatch qEUBO for preferential BO (#5093) and support PairwiseGP in ModelList pipelines (#5092)
• GenerationStrategy.fit(experiment, data) public method encapsulating the transition-then-fit pattern (#4922)
• New InSampleUniformGenerator for model-free in-sample candidate selection (#4987)
• FreshLILOLabelCheck transition criterion and hash-based filtering of stale LILO data in the adapter (#4994, #4993)
• Expand the model space for numeric ordered ChoiceParameters (#5210)

Breaking

• Removed the deprecated steps= kwarg from GenerationStrategy.__init__ (use nodes=) (#5142)
• Removed the deprecated use_update argument to GenerationStep (#5187)
• Removed the deprecated generated_points argument to RandomGenerator (#5186)
• Re-index TorchOptConfig.objective_thresholds from (n_outcomes,) to (n_objectives,) (#5018)

Analysis (ax/analysis, ax/plot)

• New TransferLearningAnalysis with paste/diff comparison links (#4918, #4980)
• New UtilityRankingPlot for preference metrics; preference metrics handled across the analysis layer (#5166, #5165)
• New NotApplicableStateAnalysisCard (#5035, #5036, #5051, #5163)
• SensitivityAnalysisPlot aesthetic improvements; use total-order sensitivity for high-dimensional experiments (#5015, #5115)

Storage (ax/storage)

• Add step_size to SQAParameter class, and as a column to the parameter_v2 table in the DB (#5212). This will be leveraged in the next version.
• SQLAlchemy 2.0 migration: migrate SQA declarative classes to SA 2.0 Mapped[T], make the SQA store SA 2.0-compatible (#5205, #5201, #5169, #5203, #5206)
• Support equality constraints in JSON and SQA storage (#5181)

Legacy API (ax/service)

Note: These are deprecated and may be removed in a future release.

• optimize() is rewritten as a thin wrapper over ax.api.client.Client; the OptimizationLoop class is removed. The optimize() signature and return type are preserved (#5143)
• Removed AxClient.get_optimization_trace -- superseded by UtilityProgressionAnalysis (#5168)
• Search space editing on AxClient: new add_parameters, update_parameters, and disable_parameters methods to mutate the search space after trials have started; parameter constraints can be added alongside new parameters (#4945, #5023)

Orchestration (ax/orchestration)

• Removed Scheduler, SchedulerOptions, and SchedulerInternalError -- deprecated since Ax 1.0.0; use the Orchestrator instead (#5188)

Ax 1.2.4

Bug Fixes

• Fix incorrect feasibility computation when using qLogProbabilityOfFeasibility for MOO — objective weights were applied twice via both the posterior transform and the constraint matrix, leading to incorrect results when only objective thresholds (no outcome constraints) were present (#4935)
• Add defensive issubclass guard for acquisition function dispatch to prevent silent fallthrough for future subclasses of qLogProbabilityOfFeasibility (#4938)
• Require only opt_config metrics for prepare_arm_data to fix ArmEffectsPlot failures when tracking metrics are missing (#4957)

Other changes

• Bumped pinned botorch version to 0.17.2 (#4959). This picks up the following changes from botorch 0.17.2:
  • Support post_processing_func in optimize_with_nsgaii for post-processing optimization results, e.g., to round discrete dimensions to valid values
• Remove unused objective_thresholds parameter from Acquisition.get_botorch_objective_and_transform — the parameter was silently discarded (#4939)
• Add Self type annotations to clone methods for better type inference in subclasses (#4907)
• Heterogeneous search space utilities for transfer learning benchmarks (#4767)
• Migrate benchmarking state dict files for GPyTorch compatibility (#4916)
• Move merge_multiple_curves to Advanced tier in complexity classification (#4949)
• Move infer_reference_point_from_experiment and get_tensor_converter_adapter to ax/service/utils/best_point.py (#4940)
• Replace disclosure triangle with info icon in Bento notebooks for analysis cards (#4956)

V1.2.3 - Feb 19 2026

V1.2.3 -- Feb 19, 2026

⚠️ Breaking Changes

Requirements

• Python 3.11+ required (#4810)
• Pandas 3.0 upgrade (#4838)
• BoTorch 0.17.0 (#4911)

Method Removals

• transition_to now required on TransitionCriterion (#4848) — Users must explicitly specify transition targets
• Removed callable serialization (#4806) — Encoding callables now raises an exception
• Removed legacy classes: TData (#4771), MinimumTrialsInStatus (#4786), completion criteria (#4850), arms_per_node override (#4822)

---

🚀 New Features

Experiment Lifecycle Tracking

New ExperimentStatus enum (DRAFT, INITIALIZATION, OPTIMIZATION, COMPLETED) with automatic status updates from the Scheduler based on generation strategy phase. (#4737, #4738, #4891)

BOPE (Preference Learning)

• Utility-based traces via PairwiseGP preference models (#4792)
• UtilityProgressionAnalysis support with "User Preference Score" UI (#4793)

BONSAI (Pruning)

• New pruning_target_parameterization API parameter (#4775)
• Tutorial and documentation added (#4865, #4871)

New methods

• add_tracking_metrics() method (#4858)
• TorchAdapter.botorch_model convenience property (#4827)
• DerivedParameter now supports bool and str types (#4847)
• LLM integration: LLMProvider/LLMMessage abstractions and llm_messages on Experiment (#4826, #4904)

Visualization

• Improved slice/contour plots: uses status_quo → best trial → center hierarchy (#4841)
• Sensitivity analysis excludes 'step' by default (#4777)
• ScalarizedOutcomeConstraint support in feasibility analysis (#4856)

---

⚡ Performance

• Data handling: DataRow-backed Data class with itertuples — 3.6x faster tensor creation (#4773, #4774, #4798)
• Generation strategy caching: Significant speedup in high trial count regimes (#4830)
• Optimization complete logic: O(nodes × TC) → O(TC on current node) (#4828)

---

🐛 Bug Fixes

• Fix GeneratorRun.clone() not copying metadata (mutations affected original) (#4892)
• Fix OneHot transform not updating hierarchical parameter dependents (#4825)
• Fix Float parameters loaded as ints from SQA (#4853)
• Fix scikit-learn 1.8.0 compatibility with XGBoost (#4816)
• Trials now marked ABANDONED (not FAILED) on metric fetch failure (#4779)
• Baseline improvement healthcheck now shows WARNING instead of FAIL (#4883)

---

🧹 Modernization

Codebase updated to Python 3.11+ idioms:

• typing.Self instead of typing_extensions.Self (#4867)
• StrEnum instead of (str, Enum) (#4868)
• ExceptionGroup (PEP 654) for structured errors (#4877)
• asyncio.TaskGroup for structured concurrency (#4878)
• PEP 604 type annotations (X | None) (#4912)

---

Full Changelog

For the complete list of 90+ PRs, see the GitHub releases page.

Ax 1.2.2

NOTE: this will be the last Ax release before SQLAlchemy becomes a required dependency

Deprecations

• Add deprecation warning to AxClient (#4749)
• Add deprecation warning to 'optimize' loop API (#4697)
• Deprecate Trial.runner (#4460)
• Deprecate TensorboardMetric's percentile in favor of quantile (#4676)
• Deprecate default_data_type argument to Experiment (#4698)

New Features

• Efficient leave-one-out cross-validation for Gaussian processes (#4631)
• Add patience parameter to PercentileEarlyStoppingStrategy (#4595)
• Log-scale support for ChoiceParameter (#4591)
• Support ChoiceParameter in Log transform (#4592)
• Add robust trial status polling to Orchestrator (#4756)
• Expose validation for TL experiments and fetching of candidate TL sources through AxService (#4615)
• Add PreferenceOptimizationConfig with storage layer support (#4638)
• Add PLBO transform and metric ordering validation (#4633)
• Add expect_relativized_outcomes flag to PreferenceOptimizationConfig (#4632)
• Add kendall tau rank correlation diagnostic (#4617)
• Vectorize SearchSpace membership check for performance (#4762)

Analyses

• New UtilityProgression Analysis for tracking optimization progress over time (#4535)
• New Best Trials Analysis for identifying top-performing trials (#4545)
• New Early Stopping Healthcheck analysis (#4569)
• New Predictable Metrics Healthcheck analysis (#4598)
• New Baseline Improvement Healthcheck analysis (#4673)
• New Complexity Rating Healthcheck for assessing optimization difficulty (#4556)
• Analysis to visualize experiment generation strategy (#4759)
• Add Pareto frontier display on MOO objective scatter plots (#4708)
• Add Progression Plots for MapMetric experiments to ResultsAnalysis (#4705)
• Add SEM display option to ContourPlot (#4690)
• Add markers to ProgressionPlot line charts (#4693)
• GraphvizAnalysisCard and HierarchicalSearchSpaceGraph visualization (#4616)
• IndividualConstraintsFeasibilityAnalysis replaces ConstraintsFeasibilityAnalysis (#4527)

Bug Fixes

• Fix tied trial bug in PercentileESS: Use rank() for n_best_trial protection (#4587)
• Fix StandardizeY not updating weights in ScalarizedObjective (#4619)
• Fix StratifiedStandardizeY behavior with ScalarizedObjective & ScalarizedOutcomeConstraint (#4621)
• Fix floating point precision issue in step_size validation (#4604)
• Fix progression normalization logic in early-stopping strategies (#4525)
• Fix dependent parameter handling in Log transform (#4679)
• Drop NaN values in MAP_KEY column before align_partial_results (#4634)
• Filter failed trials from plots (#4725)
• Allow single progression early stopping checks when patience > 0 (#4635)
• Update SOBOL transition criterion to exclude ABANDONED and FAILED trials (#4776)

Other

• Speed up MapDataReplayMetric (#4654)
• Fast MapData.df implementation (#4487)
• Validate patience <= min_progression in PercentileEarlyStoppingStrategy (#4639)
• Enforce smoothing in [0, 1) for TensorBoardMetric (#4661)
• Enforce sort_values=True for numeric ordered ChoiceParameter (#4597)
• Add error if PowerTransformY is used with ScalarizedObjective (#4622)
• Support ScalarizedObjective in get_best_parameters with model predictions (#4594)
• Rename model_kwargs -> generator_kwargs (#4668)
• Rename model_gen_kwargs -> generator_gen_kwargs (#4667)
• Rename model_cv_kwargs -> cv_kwargs (#4669)

Ax 1.2.1

Bug fixes

• Improved error messaging for client.compute_analyses when certain analyses are
  not yet available (#4441)
• Fix tooltip mismatch bug in ArmEffectsPlot (#4479)

Other changes

• Bumped pinned botorch version to 0.16.1 (#4570)
• Removed deprecated robust optimization functionality (#4493)
• Allow HierarchicalSearchSpace to be constructed with multiple root nodes (#4560)

Ax 1.2.0

New features

• DerivedParameterConfig allows users to specify parameters which are not tuned,
  instead taking the value of some expression of other tunable parameters (#4454)
• New argument simplify_parameter_changes in client.configure_generation_strategy
  (defaulted to False) which when True informs Ax to change as few parameters as
  possible relative to the status quo without degrading optimization performance. Has
  a near-zero effect on generation time (#4409)
• Default to qLogNParEgo acquisition function for multi-objective optimization in
  multi-objective optimization when the number of objectives is > 4, leading to
  improved walltime performance (#4347)

Bug fixes

• Fix issue during candidate generation involving MapMetrics providing progressions
  at different scales i.e. one progression goes up to 10^9 and the other goes up to
  10^6 by normalizing to [0, 1] (#4458)

Other changes

• Improve visual clarity in ArmEffectsPlot by removing certain elements including
  red "infeasibility" halos and optional cumulative best line (#4397, #4398)
• Instructions on citing Ax included in README.md and ax.dev (#4317, #4357)
• New "Using external methods for candidate generation in Ax" tutorial on website (#4298)

Ax 1.1.2

Bug fixes

• Fixed rendering issue in ArmEffectsPlot when the number of arms displayed is greater than 20 (#4273)

Other changes

• Enabled Winsorization transform by default, improving surrogate performance in the presence of outliers (#4277)

Ax 1.1.1

Bug fixes

• Correctly filter out observations from Abandoned trials/arms during candidate
  generation (#4155)
• Handle scalarized objectives in ResultsAnalysis (#4193)
• Fix bug in polytope sampler when generating more than one candidate in a batch (#4244)

Other changes

• Transition from setup.py to pyproject.toml for builds, modernizing Ax's build
  configuration and bringing it in compliance with PEP 518 and PEP 621 (#4100)
• Add py.typed file, which allows typecheckers like Pyre, mypy, etc. to see Ax's types
  and avoid a TypeStub error when importing Ax (#4139)
• Improve legibility of ArmEffectPlot by modifying legend and x-axis labels (#4220,
  #4243)
• Address logspew in OneHotEncoder transform (#4232)

Ax 1.1.0

New Features

• New option for the method parameter in client.configure_generation_strategy:
  quality -- allows uers to indicate they would like Ax to generate the highest
  quality candidates it is able to at the expense of slower runtime (#4042)
• New logic for deciding which analyses to produce by default in
  client.compute_analyses (#4013)
• New parameters in client.summarize allow users to filter their summary by trial
  index and/or trial status (#4012, #4118)

Bug Fixes

• Allow client.summarize to be called without a GenerationStrategy being set
  (i.e. before client.configure_generation_strategy or client.get_next_trails
  has been called.) (#3801)
• Fixed incorrect grouping in TopSurfacesAnalysis (#4095)
• Fixed ContourPlot failing to compute in certain search spaces with parameter
  constraints (#4124)
• Misc. plotting fixes and improvements

Other changes

• Bumped pinned botorch version to 0.15.1
• Performance improvements in SensitivityAnalysis (#3891)
• Improved optimization performance in constrained optimization settings (#3585)
• Augmented logging in Client, early stopping module (#4044, #4108)