Uploaded October 2026 | Updated November 2026, 3 hours ago
What’s new with Herb.jl: Teaching Programs how to Program with Program Synthesis by Tilman Hinnerichs, Reuben Gardos Reid
PreTalx: pretalx.com/juliacon-2026/talk/RXG7AD
flyer-RXG7AD: pretalx.com/media/juliacon-2026/submissions/RXG7AD/resources/RXG7AD_XFuhxxV.png
Wouldn’t it be great if Julia could program itself?
You simply tell it what you want, Julia magic happens, and you get correct-by-construction code.
In this talk, we introduce `Herb.jl`, a unifying program synthesis library written in Julia, that gets us closer to this goal.
While we are not fully there yet, we have significantly progressed since our last talk at JuliaCon 2024.
`Herb.jl` is a toolbox and framework for program synthesis: Automatically generating (not just) programs from specifications.
Since our last talk at JuliaCon 2024, `Herb.jl` has grown substantially, combining both reasoning about programs and LLM-guided code generation.
First, we introduce new constraint-solving techniques to `Herb.jl` and program synthesis:
Constraints allow us to prune large regions of possible programs, e.g., by removing symmetries and useless programs, leading to significantly faster solving times.
Second, we introduce provable LLM guidance:
In `Herb.jl`, large language models (LLMs) can propose candidates or search heuristics, while `Herb.jl` retains formal guarantees and reasoning capabilities.
In this talk, we give a brief overview of our architecture, which unifies these ideas into a single extendable pipeline.
We demonstrate how Julia’s metaprogramming and composability enable implementing new and old synthesizer ideas, and show how `Herb.jl` helps to synthesize not just programs.
What’s new with Herb.jl: Teaching Programs how to Program with Program Synthesis by Tilman Hinnerichs, Reuben Gardos Reid
PreTalx: pretalx.com/juliacon-2026/talk/RXG7AD
flyer-RXG7AD: pretalx.com/media/juliacon-2026/submissions/RXG7AD/resources/RXG7AD_XFuhxxV.png
Wouldn’t it be great if Julia could program itself?
You simply tell it what you want, Julia magic happens, and you get correct-by-construction code.
In this talk, we introduce `Herb.jl`, a unifying program synthesis library written in Julia, that gets us closer to this goal.
While we are not fully there yet, we have significantly progressed since our last talk at JuliaCon 2024.
`Herb.jl` is a toolbox and framework for program synthesis: Automatically generating (not just) programs from specifications.
Since our last talk at JuliaCon 2024, `Herb.jl` has grown substantially, combining both reasoning about programs and LLM-guided code generation.
First, we introduce new constraint-solving techniques to `Herb.jl` and program synthesis:
Constraints allow us to prune large regions of possible programs, e.g., by removing symmetries and useless programs, leading to significantly faster solving times.
Second, we introduce provable LLM guidance:
In `Herb.jl`, large language models (LLMs) can propose candidates or search heuristics, while `Herb.jl` retains formal guarantees and reasoning capabilities.
In this talk, we give a brief overview of our architecture, which unifies these ideas into a single extendable pipeline.
We demonstrate how Julia’s metaprogramming and composability enable implementing new and old synthesizer ideas, and show how `Herb.jl` helps to synthesize not just programs.









- Provides a growing collection of benchmark combinatorial decision problems.
- Includes all necessary components to build and train a DFL policy for each problem.
- Offers a general interface for defining custom problems.
- Facilitates reproducibility and experimentation by allowing the same algorithm to be applied across multiple benchmarks with minimal changes.
- [`DecisionFocusedLearningAlgorithms.jl`](https://github.com/JuliaDecisionFocusedLearning/DecisionFocusedLearningAlgorithms.jl)
- Implements generic versions of state-of-the-art DFL training algorithms.
- Provides compatibility with problems defined through the benchmarks interface.
- While some classical DFL approaches can be implemented directly using the lower-level tools provided by InferOpt.jl, recent state-of-the-art methods have become increasingly complex. This package introduces higher-level implementations to make these advanced algorithms accessible without requiring users to master the technical details.
The goal of the talk is to demonstrate how users can leverage these packages to define a problem, select a training strategy, and train a policy for their own problem in a few lines of code, without necessarily needing deep expertise in DFL. JuliaDecisionFocusedLearning: A Practical Introduction to Decision-Focused Learning in Julia | Baty](https://i.ytimg.com/vi/g2XFyvBmsz0/mqdefault.jpg)

 is a fully Julia-based idealized atmospheric flow solver, primarily developed for gravity wave research. The model is highly efficient and integrates the Boussinesq, pseudo-incompressible, and compressible equations over arbitrary topography. Either high-resolution, wave-resolving simulations can be conducted, or simulations with parameterized gravity waves using the parameterization scheme MS-GWaM, where both transience and horizontal propagation of gravity waves are accounted for.
[PinCFlow.jl](https://github.com/Atmospheric-Dynamics-GUF/PinCFlow.jl) (Pseudo-inCompressible Flow solver), developed for conducting idealized atmospheric simulations, integrates the Boussinesq, pseudo-incompressible, and compressible equations using a finite volume approach and a semi-implicit time-stepping scheme. Due to MPI parallelization in all spatial dimensions, high-resolution, wave-resolving simulations can be conducted efficiently. The use of a terrain-following coordinate system allows for arbitrary surface topography.
A unique feature of [PinCFlow.jl](https://github.com/Atmospheric-Dynamics-GUF/PinCFlow.jl) is its numerical realization of multiple-scale WKB theory for the parameterization of gravity-wave impacts on mean-flow dynamics using MS-GWaM (Multi-Scale Gravity-Wave Model). Here, gravity waves are approximated using a ray-tracing technique, where so-called ray volumes propagate through a six-dimensional phase-space. Unlike conventional single-column, steady-state parameterizations, MS-GWaM accounts for wave transience and horizontal propagation.
Current research and development focuses on the generation of orographic gravity waves, the interaction between gravity waves and turbulence, and their impact on ice cloud formation and tracer transport.
Originally developed in Fortran 90, the solver has been fully translated and optimized as a Julia package. Now officially registered and accompanied by extensive documentation, [PinCFlow.jl](https://github.com/Atmospheric-Dynamics-GUF/PinCFlow.jl) provides a modern, high-performance framework for research in computational fluid dynamics. Idealized Atmospheric Flow and Gravity-Wave Modeling with PinCFlow.jl | Steiger](https://i.ytimg.com/vi/h6SKj25kL6w/mqdefault.jpg)