§ 2.1
Distributed orchestration
AHISTER, DEBOHAC, and the cluster lifecycle — distributing evolutionary search across heterogeneous compute under hierarchical capacity and budget.
Read moreResearch at Ufinq
Ufinq is the research arm behind Diafunc — a focused effort to make Symbolic AI scale. Where deep learning compresses data into millions of opaque weights, Ufinq evolves code: small, readable programs that describe how data behaves, deployable on any machine and auditable line by line.
Our work is organized around seven research areas, each a thread of active investigation. The corresponding publications are listed under Papers; short-form notes on individual problems live under Notes.
Symbolic AI is not a smaller deep learning. It is a different category of model — and a different research program.
Research areas
§ 2.2
Symbolic refinement
Algebraic simplification, structural normalization, and the e-graph rewriter — reducing every candidate to its canonical form before it enters the gene pool.
Read more§ 2.3
Constant optimization
CMA-ES, L-BFGS, Nelder-Mead, Coordinate Descent — numeric methods specialized for the constraints of symbolic regression candidates.
Read more§ 2.4
Approximate evaluation
CRESCENT and successors — cheap-then-full evaluation gating that redirects compute toward the candidates that deserve a full-fidelity score.
Read more§ 2.5
Genetic operators
How candidates are generated and recombined — the mutation chain, crossover families, and the structural invariants they must preserve.
Read more§ 2.6
Search strategies
Navigating the universal function space — exploration vs exploitation, adaptive search spaces, and the dual diversity gates that hold populations open.
Read more§ 2.7
Compositional search
Building large models by composing smaller ones — divide-and-conquer in symbolic regression, sub-model assembly, and the seams between them.
Read moreThe Pareto frontier
Internally, the search does not commit to a single model too early. It maintains a frontier — a set of candidate solutions that trade accuracy against complexity. A small, blunt model and a large, accurate model can both sit on the frontier; nothing strictly dominated by another candidate gets to stay. From that frontier the system selects the one model it returns, applying a principled complexity-versus-accuracy criterion rather than leaving the tradeoff to the reader.
Research principles
Ufinq research is closed-source by default, but open in its ideas. Algorithms, designs, and empirical results are described in papers and technical reports; the productized implementation is part of the Diafunc platform. The asymmetry between reading a paper and reproducing a full distributed system is deliberate — and the direction is reversible: closed today, with the option to open components later if and when it serves the research community.
Three commitments shape what gets prioritized:
Readable modelsEvery output is auditable code, not opaque weights — the result must remain readable as it scales.
Distributed by constructionAlgorithms are designed for cluster execution from day one — not single-machine prototypes retrofitted later.
Honest benchmarksEvery claim ships with reproducible SRBench-schema results — no cherry-picked workloads, no contaminated splits.
Where to read more
The full output of the research effort lives on the Papers page. Empirical evaluations are tracked under Benchmarks. For the short version of how the system works end-to-end, see Technology.