Large language models (LLMs) have shown remarkable ability to generate code, yet their outputs often violate syntactic or semantic constraints when guided only through natural language prompts. We introduce TreeCoder, the most general and flexible framework to date for exploring decoding strategies, constraints, and hyperparameters in LLMs, and use it in code generation to enforce correctness and structure during decoding rather than relying on prompt engineering. TreeCoder represents decoding as a tree search over candidate programs, where both decoding strategies and constraint functions–such as style, syntax, execution–are treated as first-class, optimisable components. This design enables systematic exploration and automatic tuning of decoding configurations using standard optimisation techniques. Experiments on Python, SQL and Rust show that TreeCoder consistently improves accuracy across open-source models such as CodeLlama, Mistral, DeepSeek and Qwen, often significantly outperforming their unconstrained baselines.

We propose a neurosymbolic architecture aimed at boosting the performance of any Language Model (LM) for SQL query generation. This approach leverages symbolic reasoning to guide the LM's exploration of the search space by considering multiple paths, symbolically evaluating choices at each decision point to choose the next step, with the added novel ability to backtrack. A key innovation is the use of symbolic checks on both partially and fully generated SQL queries, enabling early truncation of unsuccessful search paths. Input consists of textual requirements on the desired query, along with optional example tuples to be selected by the query. Experiments on Xander, our open-source implementation, show it both reduces runtime and increases accuracy of the generated SQL. A specific result is an LM using Xander outperforming a four-times-larger LM.

Traditionally, explainable artificial intelligence seeks to provide explanation and interpretability of high-performing black-box models such as deep neural networks. Interpretation of such models remains difficult, because of their high complexity. An alternative method is to instead force a deep-neural network to use human-intelligible features as the basis for its decisions. We tested this approach using the natural category domain of rock types. We compared the performance of a black-box implementation of transfer-learning using Resnet50 to that of a network first trained to predict expert-identified features and then forced to use these features to categorise rock images. The performance of this feature-constrained network was virtually identical to that of the unconstrained network. Further, a partially constrained network forced to condense down to a small number of features that was not trained with expert features did not result in these abstracted features being intelligible; nevertheless, an affine transformation of these features could be found that aligned well with expert-intelligible features. These findings show that making an AI intrinsically intelligible need not be at the cost of performance.