v2.0.0-beta.1

What's Changed

Core PySR / Backend

PySR v2.0.0-beta.1 is the first beta release of PySR 2. It upgrades the Julia backend from SymbolicRegression.jl v2.0.0-alpha.9 to v2.0.0-beta.3 and introduces first-class mutation and plugin configuration.

Breaking changes

• Remove ParametricExpressionSpec in favor of TemplateExpressionSpec by @MilesCranmerBot in #1277
• Switch Slurm allocations to SlurmClusterManager.jl by @MilesCranmer in #794
• Enable simulated annealing by default, matching SymbolicRegression.jl, by @MilesCranmerBot in #1283
• Increase the default crossover probability to 0.2 through the updated backend defaults
• Enable adaptive mutation weights and automatic batching through the updated backend defaults

Saved search states using ParametricExpressionSpec cannot be loaded by PySR 2. Use TemplateExpressionSpec for structured expressions.

Mutations and plugins

• Expose first-class mutation configuration through mutations and default_mutations by @MilesCranmerBot in #1282
• Expose plugin configuration through plugins and default_plugins, including simulated annealing, adaptive parsimony, adaptive mutation weights, and mutation bursts
• Add configurable mutation objects, including BacksolveMutation
• Preserve the legacy weight_* parameters while allowing mutation-specific settings through the new object interface

Search behavior and performance

• Delegate automatic batching to SymbolicRegression.jl while preserving explicit batching and batch_size settings by @MilesCranmerBot in #1285
• Skip the unused Float64 SymbolicRegression.jl precompile workload for PySR's default Float32 configuration by @MilesCranmerBot in #1279
• Reuse evaluation buffers during search through the updated backend
• Improve backend precompilation time and compiled-cache size

Bug fixes

• Respect tempdir when writing temporary equation files by @MilesCranmerBot in #1207
• Convert TemplateExpressionSpec.num_features dictionary keys to Julia symbols by @MilesCranmerBot in #1209

Backend changes from SymbolicRegression.jl v2.0.0-alpha.9 to v2.0.0-beta.3

View the SymbolicRegression.jl changes included in this release

PySR v2.0.0-beta.1 upgrades SymbolicRegression.jl from v2.0.0-alpha.9 to v2.0.0-beta.3.

Backend release: https://github.com/astroautomata/SymbolicRegression.jl/releases/tag/v2.0.0-beta.3

Full backend changelog: astroautomata/SymbolicRegression.jl@v2.0.0-alpha.9...v2.0.0-beta.3

Mutation, crossover, and plugin interfaces

• Add first-class AbstractMutation types and mutation-specific configuration by @MilesCranmer in astroautomata/SymbolicRegression.jl#610
• Add a composable plugin interface with simulated annealing, adaptive mutation weights, adaptive parsimony, and mutation-loop plugins by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#645
• Add the backsolve mutation by @ayagh19 in astroautomata/SymbolicRegression.jl#573
• Add customizable crossovers through AbstractCrossover by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#664
• Allow plugins to contribute default mutations and crossovers by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#663

Defaults and search behavior

• Increase the default crossover probability by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#643
• Enable adaptive mutation weights by default by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#678
• Enable automatic batching for large datasets by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#676
• Preserve the simulated-annealing temperature schedule by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#652

Template expressions and custom value types

• Add generic optimizable parameters for template expressions by @adil-soubki in astroautomata/SymbolicRegression.jl#644
• Remove the dedicated ParametricExpression API in favor of template expressions by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#656
• Adopt DynamicExpressions.jl 2.9 and its EvalContext interface by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#668
• Preserve natural type promotion in ValidVector operations by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#625
• Support SubArray inputs in eval_grad_tree_array by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#566

Performance

• Reuse evaluation buffers during search by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#654
• Narrow the default precompile workload by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#642
• Add mutation function barriers and avoid unnecessary deep copies in the search loop

Correctness and reliability

• Recompute costs after simplification by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#550
• Fix constant-optimization restarts for zero-valued constants by @singhharsh1708 in astroautomata/SymbolicRegression.jl#637
• Prevent multiprocessing teardown hangs by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#641
• Finalize logging only after every search output completes by @adil-soubki in astroautomata/SymbolicRegression.jl#646
• Report mismatched X and y sample counts early by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#660
• Cap crossover constraint retries correctly by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#666
• Make stdin quit monitoring non-blocking by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#562
• Disable simplification for expression-level losses by @MilesCranmerBot in astroautomata/SymbolicRegression.jl#674
• Mark Enzyme support as stable by @wsmoses in astroautomata/SymbolicRegression.jl#632

Packaging / Release / Dependencies

• Move PyPI publication to trusted publishing by @MilesCranmerBot in #1256
• Update repository links following the move to the astroautomata organization by @MilesCranmerBot in #1272
• Update Python, documentation, and GitHub Actions dependencies

CI / Testing / Automation

• Handle skipped scikit-learn estimator checks by @MilesCranmer in #1201
• Reduce low-signal GitHub Actions usage while preserving targeted pull-request coverage by @MilesCranmerBot in #1241
• Restore the large Docker nightly workflow by @MilesCranmerBot in #1259

Documentation / Website / Papers

• Add an agent skill describing effective PySR usage by @MilesCranmer in #1264
• Use Float32 literals in the custom-loss example by @MilesCranmerBot in #1276
• Expand the research-paper showcase with work on microbial growth, human mobility, wind turbines, skin friction, dark energy, biomass pyrolysis, angular coefficients, and stellar chaos
• Update the contributor list by @MilesCranmerBot in #1286

Full Changelog: v2.0.0a2...v2.0.0-beta.1