Large language models (LLMs) have shown impressive capabilities in code generation across many programming languages but even state-of-the-art LLMs generate programs that contain syntactic errors and fail to complete the given tasks, especially for low-resource programming languages (LRPLs). In addition, the high cost of training makes finetuning LLMs unaffordable for those with constrained computational resources, further weakening the effectiveness of LLMs for code generation. In this work, we propose SLMFix, a novel code generation pipeline that leverages a small language model (SLM) finetuned using reinforcement learning (RL) techniques to fix syntactic errors in LLM-generated programs for domain-specific languages (DSLs) based on interpreter feedback. Our experimental results demonstrate the effectiveness and generalizability of our approach across multiple DSLs, improving the validator pass rates by 40% on LRPLs and eliminating more than 50% of syntactic errors for high-resource DSLs. Notably, SLMFix brings substantial performance improvement to the base model and outperforms supervised finetuning approach even for 7B models on LRPLs including Ansible and Lean, showing the potential of our approach in improving the quality of LLM-generated programs.

Adapting visual programming or prompting large language models (LLMs) to generate executable code for visual tasks like visual question answering (VQA) for specialized tasks or domains remains challenging due to high annotation and inference costs. We propose a low-cost visual program distillation method that can be used for models with at most 1 billion parameters and requires no human-generated program annotations. We achieve this through synthetic data augmentation based on decoupling programs into higher-level skills, called templates, and their corresponding arguments. Experimental results show that, with a relatively small amount of question/answer data, small language models can generate high-quality specialized visual programs with the added benefit of much faster inference.

Large Multimodal Models (LMM) are built across modalities and the misalignment between two modalities can result in “hallucination”, generating textual outputs that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the vision-language alignment, where human annotators are asked to compare two responses and pinpoint the more hallucinated one, and the vision-language model is trained to maximize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image-text pairs to improve the general capabilities of our model. To evaluate the proposed approach in real-world scenarios, we develop a new evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaVA-Bench dataset with the 96% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improvement of 60% on MMHAL-BENCH over other baselines.