Razan Baltaji

Razan Baltaji, Saurabh Pujar, Louis Mandel et al. · ibm-research

Large language models (LLMs) have achieved state-of-the-art performance in various software engineering tasks, including error detection, clone detection, and code translation, primarily leveraging high-resource programming languages like Python and Java. However, many critical languages, such as COBOL, as well as emerging languages, such as Rust and Swift, remain low-resource due to limited openly available code. This scarcity hampers the training and effectiveness of LLMs for these languages, increasing software maintenance costs and stifling innovation. Addressing this gap, we investigate the potential of transfer learning to enhance LLM performance on low-resource programming languages by leveraging data from high-resource counterparts. Our extensive empirical study evaluates transferability across 10 to 41 programming languages and five key tasks: code generation, clone detection, code repair, solution domain classification, and error detection. Additionally, we develop a performance prediction model to guess the best source languages for a given target and task, and analyze the features that influence transfer performance. We further replicate a representative subset of experiments with a larger model to test the generalizability of our conclusions to contemporary large-scale LLMs. Our findings demonstrate that cross-lingual transfer significantly outperforms zero-shot learning, with effectiveness varying based on both source and target languages. Furthermore, our model reliably predicts successful transfer sources by considering linguistic and dataset-specific features, offering practical guidance for data acquisition and model training. This work contributes to the development of LLM-driven tools for low-resource programming languages and provides insights into the characteristics that facilitate transfer across language pairs.

Babak Hemmatian, Razan Baltaji, Lav R. Varshney

Abid et al. (2021) showed a tendency in GPT-3 to generate mostly violent completions when prompted about Muslims, compared with other religions. Two pre-registered replication attempts found few violent completions and only a weak anti-Muslim bias in the more recent InstructGPT, fine-tuned to eliminate biased and toxic outputs. However, more pre-registered experiments showed that using common names associated with the religions in prompts increases several-fold the rate of violent completions, revealing a significant second-order anti-Muslim bias. ChatGPT showed a bias many times stronger regardless of prompt format, suggesting that the effects of debiasing were reduced with continued model development. Our content analysis revealed religion-specific themes containing offensive stereotypes across all experiments. Our results show the need for continual de-biasing of models in ways that address both explicit and higher-order associations.

Razan Baltaji, Sourya Basu, Lav R. Varshney

Constructing model-agnostic group equivariant networks, such as equitune (Basu et al., 2023b) and its generalizations (Kim et al., 2023), can be computationally expensive for large product groups. We address this problem by providing efficient model-agnostic equivariant designs for two related problems: one where the network has multiple inputs each with potentially different groups acting on them, and another where there is a single input but the group acting on it is a large product group. For the first design, we initially consider a linear model and characterize the entire equivariant space that satisfies this constraint. This characterization gives rise to a novel fusion layer between different channels that satisfies an invariance-symmetry (IS) constraint, which we call an IS layer. We then extend this design beyond linear models, similar to equitune, consisting of equivariant and IS layers. We also show that the IS layer is a universal approximator of invariant-symmetric functions. Inspired by the first design, we use the notion of the IS property to design a second efficient model-agnostic equivariant design for large product groups acting on a single input. For the first design, we provide experiments on multi-image classification where each view is transformed independently with transformations such as rotations. We find equivariant models are robust to such transformations and perform competitively otherwise. For the second design, we consider three applications: language compositionality on the SCAN dataset to product groups; fairness in natural language generation from GPT-2 to address intersectionality; and robust zero-shot image classification with CLIP. Overall, our methods are simple and general, competitive with equitune and its variants, while also being computationally more efficient.

Razan Baltaji, Babak Hemmatian, Lav R. Varshney

Multi-agent AI systems can be used for simulating collective decision-making in scientific and practical applications. They can also be used to introduce a diverse group discussion step in chatbot pipelines, enhancing the cultural sensitivity of the chatbot's responses. These applications, however, are predicated on the ability of AI agents to reliably adopt assigned personas and mimic human interactions. To see whether LLM agents satisfy these requirements, we examine AI agent ensembles engaged in cross-national collaboration and debate by analyzing their private responses and chat transcripts. Our findings suggest that multi-agent discussions can support collective AI decisions that more often reflect diverse perspectives, yet this effect is tempered by the agents' susceptibility to conformity due to perceived peer pressure and occasional challenges in maintaining consistent personas and opinions. Instructions that encourage debate in support of one's opinions rather than collaboration increase the rate of inconstancy. Without addressing the factors we identify, the full potential of multi-agent frameworks for producing more culturally diverse AI outputs or more realistic simulations of group decision-making may remain untapped.