Chu Fei Luo

Chu Fei Luo, Samuel Dahan, Xiaodan Zhu

Test-time reasoning has become a significant field of study since the introduction of chain-of-thought reasoning in large language models (LLMs). However, the mechanisms of this reasoning process are still under-explored -- from the same input prompt, and even the same partial solution, LLMs can produce varied answers if sampled multiple times. We propose to leverage question-asking as an inference-time intervention that articulates information about the model's hidden state. To achieve that, we present a student-teacher setting where a student asks questions to a teacher. We train a probe on the student's hidden state before and after asking a question and find it is predictive of the trajectory's final correctness, even before generating the teacher's answer. This suggests there is a meaningful signal from the self-diagnosis that occurs during question generation rather than information transfer from the teacher. We then frame question-asking as a sequential decision problem, using this probe as a quality score, and define a gating policy to ask questions that maximize likelihood of correctness. We find that the success of question-asking as an intervention is largely dependent on the model's self-consistency. Our empirical results show a gap between detection and recovery; while our gating policy captures model correctness and uncertainty, interventions are equally likely to harm correct trajectories as they are to recover incorrect ones. This gap between diagnosis and correction has broader implications on language models' capacity for self-refinement under uncertainty.

Chu Fei Luo, Ahmad Ghawanmeh, Bharat Bhimshetty et al.

Ensuring the safe deployment of AI systems is critical in industry settings where biased outputs can lead to significant operational, reputational, and regulatory risks. Thorough evaluation before deployment is essential to prevent these hazards. Red-teaming addresses this need by employing adversarial attacks to develop guardrails that detect and reject biased or harmful queries, enabling models to be retrained or steered away from harmful outputs. However, most red-teaming efforts focus on harmful or unethical instructions rather than addressing social bias, leaving this critical area under-explored despite its significant real-world impact, especially in customer-facing systems. We propose two bias-specific red-teaming methods, Emotional Bias Probe (EBP) and BiasKG, to evaluate how standard safety measures for harmful content affect bias. For BiasKG, we refactor natural language stereotypes into a knowledge graph. We use these attacking strategies to induce biased responses from several open- and closed-source language models. Unlike prior work, these methods specifically target social bias. We find our method increases bias in all models, even those trained with safety guardrails. Our work emphasizes uncovering societal bias in LLMs through rigorous evaluation, and recommends measures ensure AI safety in high-stakes industry deployments.

Chu Fei Luo, Samuel Dahan, Xiaodan Zhu

As language models have a greater impact on society, it is important to ensure they are aligned to a diverse range of perspectives and are able to reflect nuance in human values. However, the most popular training paradigms for modern language models often assume there is one optimal answer for every query, leading to generic responses and poor alignment. In this work, we aim to enhance pluralistic alignment of language models in a low-resource setting with two methods: pluralistic decoding and model steering. We empirically demonstrate that model steering offers consistent improvement over zero-shot and few-shot baselines with only 50 annotated samples. Our proposed methods decrease false positives in several high-stakes tasks such as hate speech detection and misinformation detection, and improves the distributional alignment to human values in GlobalOpinionQA. We hope our work highlights the importance of diversity and how language models can be adapted to consider nuanced perspectives.

Radin Shayanfar, Chu Fei Luo, Rohan Bhambhoria et al.

Building Task-Oriented Dialogue (TOD) systems that generalize across different tasks remains a challenging problem. Data-driven approaches often struggle to transfer effectively to unseen tasks. While recent schema-based TOD frameworks improve generalization by decoupling task logic from language understanding, their reliance on neural or generative models often obscures how task schemas influence behaviour and hence impair interpretability. In this work, we introduce a novel framework, CoDial (Code for Dialogue), which converts a TOD task schema, represented as a novel structured heterogeneous graph, to programmatic LLM guardrailing code, such as NVIDIA's Colang, enabling interpretable and efficient alignment of dialogue policies during inference. We introduce two paradigms, $\text{CoDial}_{\text{free}}$ and $\text{CoDial}_{\text{structured}}$ for generating LLM guardrails, and propose a feedback mechanism that integrates human feedback to iteratively improve the generated code. Empirically, CoDial achieves state-of-the-art (SOTA) performance on the widely used STAR dataset and is on par with SOTA on the MultiWOZ dataset, while also providing interpretability. We additionally demonstrate CoDial's iterative improvement via manual and LLM-aided feedback, making it a practical tool for expert-guided alignment of LLMs in high-stakes domains.

Chu Fei Luo, Rohan Bhambhoria, Samuel Dahan et al.

Deep learning has made significant progress in the past decade, and demonstrates potential to solve problems with extensive social impact. In high-stakes decision making areas such as law, experts often require interpretability for automatic systems to be utilized in practical settings. In this work, we attempt to address these requirements applied to the important problem of legal citation prediction (LCP). We design the task with parallels to the thought-process of lawyers, i.e., with reference to both precedents and legislative provisions. After initial experimental results, we refine the target citation predictions with the feedback of legal experts. Additionally, we introduce a prototype architecture to add interpretability, achieving strong performance while adhering to decision parameters used by lawyers. Our study builds on and leverages the state-of-the-art language processing models for law, while addressing vital considerations for high-stakes tasks with practical societal impact.

Chu Fei Luo, Rohan Bhambhoria, Xiaodan Zhu et al.

Hate speech causes widespread and deep-seated societal issues. Proper enforcement of hate speech laws is key for protecting groups of people against harmful and discriminatory language. However, determining what constitutes hate speech is a complex task that is highly open to subjective interpretations. Existing works do not align their systems with enforceable definitions of hate speech, which can make their outputs inconsistent with the goals of regulators. This research introduces a new perspective and task for enforceable hate speech detection centred around legal definitions, and a dataset annotated on violations of eleven possible definitions by legal experts. Given the challenge of identifying clear, legally enforceable instances of hate speech, we augment the dataset with expert-generated samples and an automatically mined challenge set. We experiment with grounding the model decision in these definitions using zero-shot and few-shot prompting. We then report results on several large language models (LLMs). With this task definition, automatic hate speech detection can be more closely aligned to enforceable laws, and hence assist in more rigorous enforcement of legal protections against harmful speech in public forums.