LLMs Automate SAT Solver Heuristics in New AutoSAT Framework

A new paper published on July 24 2026 introduces AutoSAT, a framework that leverages large language models (LLMs) to generate novel heuristics for Conflict‑Driven Clause Learning (CDCL) SAT solvers such as EasySAT.¹

Conflict‑driven clause learning has become the dominant approach for solving Boolean satisfiability problems, but its performance hinges heavily on hand‑crafted heuristic choices. Modern CDCL solvers like MiniSat and Kissat bundle multiple heuristics, requiring substantial manual tuning and deep domain expertise. This challenge motivates AutoSAT’s goal: to automate the discovery of effective heuristic designs beyond conventional parameter tuning.¹

Unlike prior automated algorithm design efforts that mainly focus on hyperparameter optimization or operator selection, AutoSAT directly generates new efficient heuristics inside an existing solver. By feeding prompts into an LLM trained on vast text corpora, the system proposes candidate heuristic rules which are then evaluated within the solver environment. The authors report that this first attempt can produce strategies not considered during traditional manual design, potentially improving solver speed on benchmark instances.¹

Expert commentary notes previous attempts using evolutionary algorithms—such as those by Smith et al. in 2018—yielded only modest gains after long search times. In contrast, AutoSAT’s LLM‑based generation bypasses exhaustive search loops, offering a more direct path to novel heuristics. However, it also faces challenges; modern SAT solvers already embed sophisticated machine‑learning components, such as neural network–based clause classification trained offline. Therefore any improvement from LLM‑generated heuristics must surpass these established baselines to be meaningful.¹

The paper acknowledges practical limitations: generating a full CDCL solver from scratch with an LLM is impractical due to the sheer complexity and large context required for contemporary solvers. Instead, AutoSAT focuses on optimizing heuristics within an existing framework (EasySAT), keeping the core algorithm intact while allowing the LLM to propose rule modifications. This modular approach keeps computational costs manageable and facilitates integration into current solver pipelines.¹

AutoSAT’s experimental results show promising performance improvements across several standard benchmarks, suggesting that LLMs can indeed contribute useful heuristic insights for combinatorial optimization problems beyond SAT solving alone. The authors argue that this technique could generalize to other domains where heuristic design remains a bottleneck. Yet they caution that scaling the method will require careful handling of model inversion and gradient estimation costs associated with flow‑based generative models used in related HRL research.

AutoSAT represents a significant step toward automating heuristic selection in SAT solvers by using language models. Future work will determine if its advantages exceed those of traditional machine-learning or evolutionary methods, enabling faster solver improvements without extensive manual tuning.