Massimo Rondelli

Generating accurate and executable code using Large Language Models (LLMs) remains a significant challenge for underrepresented programming languages, such as Prolog and Lisp, due to the scarcity of public training data compared to high-resource languages like Python. This paper introduces a generalizable Reinforcement Learning (RL) approach that combines small-scale versions of the Qwen2.5-Coder model with Group Relative Policy Optimization (GRPO) to enable effective code generation through reasoning. To address the limitations of sparse datasets, we integrate execution-driven feedback directly into the RL loop, utilizing a reward system that exploits both logical correctness and structural formatting. Experimental results on GSM8K dataset demonstrate significant improvements in reasoning quality and code accuracy across underrepresented languages. These findings underscore the potential of our approach to benefit a wide range of programming languages lacking extensive training resources by leveraging symbolic reasoning and interpreter-based feedback.

Large Language Models (LLMs) generate fluent code in mainstream languages but fail systematically when asked to produce Prolog: they hallucinate built-ins, impose imperative control flow on a declarative paradigm, and mishandle the interplay of unification and backtracking. The cause is distributional – logic-programming code is scarce in modern pre-training corpora – and no amount of scale will fix data that does not exist. We show that this gap can be closed by training small open-weight models with reinforcement learning where the reward signal comes from a SWI-Prolog interpreter embedded in the training loop. Using Group Relative Policy Optimization (GRPO) on Qwen2.5-Coder-Instruct (0.5B–7B), we identify a failure mode – reasoning contraction – in which the KL penalty anchors the policy too tightly to the supervised baseline; the model collapses to short memorised patterns and accuracy degrades over training. Removing the KL term and adding a length-aware reward expands reasoning chains and lifts one-shot pass@4 on GSM8K-Prolog from 0.814 to 0.893, outperforming both PROPER (0.72) and substantially larger general-purpose models such as Llama 3.3 70B (0.839) and Qwen2.5-Coder-32B (0.873). Gains transfer to 20 Rosetta Code Prolog tasks and to GSM-Symbolic (zero-shot p2 more than doubles, from 0.298 to 0.673), and the same recipe applied to Lisp lifts zero-shot pass@4 from 0.572 to 0.876. Taken together, the results indicate that, for formal languages with a deterministic semantics, the interpreter itself is sufficient to act as the primary training supervisor.

Automatic generation of executable Blender code from natural language remains challenging, with state-of-the-art LLMs producing frequent syntactic errors and geometrically inconsistent objects. We present BlenderRAG, a retrieval-augmented generation system that operates on a curated multimodal dataset of 500 expert-validated examples (text, code, image) across 50 object categories. By retrieving semantically similar examples during generation, BlenderRAG improves compilation success rates from 40.8% to 70.0% and semantic normalized alignment from 0.41 to 0.77 (CLIP similarity) across four state-of-the-art LLMs, without requiring fine-tuning or specialized hardware, making it immediately accessible for deployment. The dataset and code will be available at https://github.com/MaxRondelli/BlenderRAG.