Faiza Khan Khattak

Stephen Obadinma, Faiza Khan Khattak, Shirley Wang et al. · utoronto

Building Agent Assistants that can help improve customer service support requires inputs from industry users and their customers, as well as knowledge about state-of-the-art Natural Language Processing (NLP) technology. We combine expertise from academia and industry to bridge the gap and build task/domain-specific Neural Agent Assistants (NAA) with three high-level components for: (1) Intent Identification, (2) Context Retrieval, and (3) Response Generation. In this paper, we outline the pipeline of the NAA's core system and also present three case studies in which three industry partners successfully adapt the framework to find solutions to their unique challenges. Our findings suggest that a collaborative process is instrumental in spurring the development of emerging NLP models for Conversational AI tasks in industry. The full reference implementation code and results are available at \url{https://github.com/VectorInstitute/NAA}

Jacob-Junqi Tian, David Emerson, Sevil Zanjani Miyandoab et al.

Prompting large language models has gained immense popularity in recent years due to the advantage of producing good results even without the need for labelled data. However, this requires prompt tuning to get optimal prompts that lead to better model performances. In this paper, we explore the use of soft-prompt tuning on sentiment classification task to quantify the biases of large language models (LLMs) such as Open Pre-trained Transformers (OPT) and Galactica language model. Since these models are trained on real-world data that could be prone to bias toward certain groups of populations, it is important to identify these underlying issues. Using soft-prompts to evaluate bias gives us the extra advantage of avoiding the human-bias injection that can be caused by manually designed prompts. We check the model biases on different sensitive attributes using the group fairness (bias) and find interesting bias patterns. Since LLMs have been used in the industry in various applications, it is crucial to identify the biases before deploying these models in practice. We open-source our pipeline and encourage industry researchers to adapt our work to their use cases.

Jacob-Junqi Tian, Omkar Dige, D. B. Emerson et al.

Large language models (LLMs) are trained on vast, uncurated datasets that contain various forms of biases and language reinforcing harmful stereotypes that may be subsequently inherited by the models themselves. Therefore, it is essential to examine and address biases in language models, integrating fairness into their development to ensure that these models do not perpetuate social biases. In this work, we demonstrate the importance of reasoning in zero-shot stereotype identification across several open-source LLMs. Accurate identification of stereotypical language is a complex task requiring a nuanced understanding of social structures, biases, and existing unfair generalizations about particular groups. While improved accuracy is observed through model scaling, the use of reasoning, especially multi-step reasoning, is crucial to consistent performance. Additionally, through a qualitative analysis of select reasoning traces, we highlight how reasoning improves not just accuracy, but also the interpretability of model decisions. This work firmly establishes reasoning as a critical component in automatic stereotype detection and is a first step towards stronger stereotype mitigation pipelines for LLMs.

Huan Wu, Ali Emami, Muhammad Furquan Hassan et al.

African American English (AAE), a rule-governed dialect spoken by over 30 million people, is routinely misinterpreted and "corrected" by large language models (LLMs). Across six instruction-tuned LLMs (14B to 70B), we show that state-of-the-art models systematically prefer Standard American English (SAE) continuations even when the preceding context is in AAE, effectively rewriting AAE into SAE. We present an end-to-end framework to audit and mitigate this bias. For auditing, we introduce conditional Dialect Group Invariance (cDGI), which isolates true model bias from translator-induced artifacts, and a feature-level localization analysis that identifies which AAE markers most strongly trigger bias; we find that syntactic constructions, especially negative concord (e.g., "ain't nobody"), are universal triggers across all models. For mitigation, we introduce, to our knowledge, the first application of activation steering to dialect bias: a training-free, test-time method that extracts dialect directions via causal tracing and injects them into bias-relevant layers. Activation steering reduces bias 5 to 20 times more than prompting while preserving SAE fluency. To enable this work, we release REAL-AAE , the largest real-AAE parallel corpus to date: 17,479 AAE/SAE/ AAE_back triplets from natural tweets (2 to 6 times larger than prior real-AAE resources), validated automatically (BERTScore F1 = 0.95) and by three native AAE speakers (83.0% semantic agreement).

Omkar Dige, Jacob-Junqi Tian, David Emerson et al.

As the breadth and depth of language model applications continue to expand rapidly, it is increasingly important to build efficient frameworks for measuring and mitigating the learned or inherited social biases of these models. In this paper, we present our work on evaluating instruction fine-tuned language models' ability to identify bias through zero-shot prompting, including Chain-of-Thought (CoT) prompts. Across LLaMA and its two instruction fine-tuned versions, Alpaca 7B performs best on the bias identification task with an accuracy of 56.7%. We also demonstrate that scaling up LLM size and data diversity could lead to further performance gain. This is a work-in-progress presenting the first component of our bias mitigation framework. We will keep updating this work as we get more results.

Alicia Parrish, Rajat Shinde, Sanket Badhe et al.

Current AI safety evaluation and benchmarking frameworks predominantly rely on Western-centric culture-agnostic defaults that mask critical regional laws, socio-linguistic nuances, and cultural taboos, leaving Vision-Language Models (VLMs) vulnerable in global deployments. We introduce Pluralis v0.1: a novel multimodal, multi-regional, and multilingual dataset built from a culture-first perspective. Spanning 6,448 prompts across six Asia-Pacific countries (Bangladesh, India, Korea, Pakistan, Singapore, Taiwan) and eight languages, Pluralis diverges from prior work by natively sourcing localized safety hazards rather than adapting Western datasets. Crucially, it introduces a multimodal evaluation paradigm: user text (e.g., "Should I gift this?") and an image referring to "this" (e.g., a clock) - both innocuous in isolation, but synergistically triggering specific legal or cultural violations. Pluralis disentangles universal safety violations from localized cultural appropriateness, establishing the latter as a first-class evaluation axis. To operationalize this, we present Judge-Pluralis, an agreement-gated LLM-as-a-Judge ensemble trained on examples classified in an empirically derived cultural taxonomy. Observing VLM behavior on a subset of the Pluralis surfaces recurring, locale-specific failure modes such as image misidentifications with downstream harm, missed item-context-locale interactions, and inadequate refusals. These failure modes vary systematically across locales and languages, exposing blind spots that globally averaged metrics conceal. Ultimately, Pluralis is not presented as a solved evaluation framework for cultural alignment, but rather as a first step and catalyst for future innovation. We call upon the research community to utilize this foundation to advance the science of multilingual, multicultural evaluation to better support AI cultural alignment globally.

Omkar Dige, Diljot Singh, Tsz Fung Yau et al.

Mitigating bias in language models (LMs) has become a critical problem due to the widespread deployment of LMs. Numerous approaches revolve around data pre-processing and fine-tuning of language models, tasks that can be both time-consuming and computationally demanding. Consequently, there is a growing interest in machine unlearning techniques given their capacity to induce the forgetting of undesired behaviors of the existing pre-trained or fine-tuned models with lower computational cost. In this work, we explore two unlearning methods, (1) Partitioned Contrastive Gradient Unlearning (PCGU) applied on decoder models and (2) Negation via Task Vector, to reduce social biases in state-of-the-art and open-source LMs such as LLaMA-2 and OPT. We also implement distributed PCGU for large models. It is empirically shown, through quantitative and qualitative analyses, that negation via Task Vector method outperforms PCGU in debiasing with minimum deterioration in performance and perplexity of the models. On LLaMA-27B, negation via Task Vector reduces the bias score by 11.8%

Farnaz Kohankhaki, D. B. Emerson, Jacob-Junqi Tian et al.

Bias in large language models (LLMs) has many forms, from overt discrimination to implicit stereotypes. Counterfactual bias evaluation is a widely used approach to quantifying bias and often relies on template-based probes that explicitly state group membership. It measures whether the outcome of a task performed by an LLM is invariant to a change in group membership. In this work, we find that template-based probes can lead to unrealistic bias measurements. For example, LLMs appear to mistakenly cast text associated with White race as negative at higher rates than other groups. We hypothesize that this arises artificially via a mismatch between commonly unstated norms, in the form of markedness, in the pretraining text of LLMs (e.g., Black president vs. president) and templates used for bias measurement (e.g., Black president vs. White president). The findings highlight the potential misleading impact of varying group membership through explicit mention in counterfactual bias quantification.

Faiza Khan Khattak, Vallijah Subasri, Amrit Krishnan et al.

Machine Learning Health Operations (MLHOps) is the combination of processes for reliable, efficient, usable, and ethical deployment and maintenance of machine learning models in healthcare settings. This paper provides both a survey of work in this area and guidelines for developers and clinicians to deploy and maintain their own models in clinical practice. We cover the foundational concepts of general machine learning operations, describe the initial setup of MLHOps pipelines (including data sources, preparation, engineering, and tools). We then describe long-term monitoring and updating (including data distribution shifts and model updating) and ethical considerations (including bias, fairness, interpretability, and privacy). This work therefore provides guidance across the full pipeline of MLHOps from conception to initial and ongoing deployment.

Paul Grouchy, Shobhit Jain, Michael Liu et al.

With the growing amount of text in health data, there have been rapid advances in large pre-trained models that can be applied to a wide variety of biomedical tasks with minimal task-specific modifications. Emphasizing the cost of these models, which renders technical replication challenging, this paper summarizes experiments conducted in replicating BioBERT and further pre-training and careful fine-tuning in the biomedical domain. We also investigate the effectiveness of domain-specific and domain-agnostic pre-trained models across downstream biomedical NLP tasks. Our finding confirms that pre-trained models can be impactful in some downstream NLP tasks (QA and NER) in the biomedical domain; however, this improvement may not justify the high cost of domain-specific pre-training.

Faiza Khan Khattak, Ansaf Salleb-Aouissi

Crowd-labeling emerged from the need to label large-scale and complex data, a tedious, expensive, and time-consuming task. One of the main challenges in the crowd-labeling task is to control for or determine in advance the proportion of low-quality/malicious labelers. If that proportion grows too high, there is often a phase transition leading to a steep, non-linear drop in labeling accuracy as noted by Karger et al. [2014]. To address these challenges, we propose a new framework called Expert Label Injected Crowd Estimation (ELICE) and extend it to different versions and variants that delay phase transition leading to a better labeling accuracy. ELICE automatically combines and boosts bulk crowd labels supported by labels from experts for limited number of instances from the dataset. The expert-labels help to estimate the individual ability of crowd labelers and difficulty of each instance, both of which are used to aggregate the labels. Empirical evaluation shows the superiority of ELICE as compared to other state-of-the-art methods. We also derive a lower bound on the number of expert-labeled instances needed to estimate the crowd ability and dataset difficulty as well as to get better quality labels.