Marcelo Grave

Tiago Machado, Maysa Malfiza Garcia de Macedo, Rogerio Abreu de Paula et al.

This work aims to investigate how different Large Language Models (LLMs) alignment methods affect the models' responses to prompt attacks. We selected open source models based on the most common alignment methods, namely, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning with Human Feedback (RLHF). We conducted a systematic analysis using statistical methods to verify how sensitive the Attack Success Rate (ASR) is when we apply variations to prompts designed to elicit inappropriate content from LLMs. Our results show that even small prompt modifications can significantly change the Attack Success Rate (ASR) according to the statistical tests we run, making the models more or less susceptible to types of attack. Critically, our results demonstrate that running existing 'attack benchmarks' alone may not be sufficient to elicit all possible vulnerabilities of both models and alignment methods. This paper thus contributes to ongoing efforts on model attack evaluation by means of systematic and statistically-based analyses of the different alignment methods and how sensitive their ASR is to prompt variation.

Heloisa Candello, Muneeza Azmat, Uma Sushmitha Gunturi et al.

With the rapid uptake of generative AI, investigating human perceptions of generated responses has become crucial. A major challenge is their `aptitude' for hallucinating and generating harmful contents. Despite major efforts for implementing guardrails, human perceptions of these mitigation strategies are largely unknown. We conducted a mixed-method experiment for evaluating the responses of a mitigation strategy across multiple-dimensions: faithfulness, fairness, harm-removal capacity, and relevance. In a within-subject study design, 57 participants assessed the responses under two conditions: harmful response plus its mitigation and solely mitigated response. Results revealed that participants' native language, AI work experience, and annotation familiarity significantly influenced evaluations. Participants showed high sensitivity to linguistic and contextual attributes, penalizing minor grammar errors while rewarding preserved semantic contexts. This contrasts with how language is often treated in the quantitative evaluation of LLMs. We also introduced new metrics for training and evaluating mitigation strategies and insights for human-AI evaluation studies.

Claudio Pinhanez, Paulo Cavalin, Cassia Sanctos et al.

This work explores the consistency of small LLMs (2B-8B parameters) in answering multiple times the same question. We present a study on known, open-source LLMs responding to 10 repetitions of questions from the multiple-choice benchmarks MMLU-Redux and MedQA, considering different inference temperatures, small vs. medium models (50B-80B), finetuned vs. base models, and other parameters. We also look into the effects of requiring multi-trial answer consistency on accuracy and the trade-offs involved in deciding which model best provides both of them. To support those studies, we propose some new analytical and graphical tools. Results show that the number of questions which can be answered consistently vary considerably among models but are typically in the 50%-80% range for small models at low inference temperatures. Also, accuracy among consistent answers seems to reasonably correlate with overall accuracy. Results for medium-sized models seem to indicate much higher levels of answer consistency.

Claudio Pinhanez, Raul Fernandez, Marcelo Grave et al. · ibm-research

Representations of AI agents in user interfaces and robotics are predominantly White, not only in terms of facial and skin features, but also in the synthetic voices they use. In this paper we explore some unexpected challenges in the representation of race we found in the process of developing an U.S. English Text-to-Speech (TTS) system aimed to sound like an educated, professional, regional accent-free African American woman. The paper starts by presenting the results of focus groups with African American IT professionals where guidelines and challenges for the creation of a representative and appropriate TTS system were discussed and gathered, followed by a discussion about some of the technical difficulties faced by the TTS system developers. We then describe two studies with U.S. English speakers where the participants were not able to attribute the correct race to the African American TTS voice while overwhelmingly correctly recognizing the race of a White TTS system of similar quality. A focus group with African American IT workers not only confirmed the representativeness of the African American voice we built, but also suggested that the surprising recognition results may have been caused by the inability or the latent prejudice from non-African Americans to associate educated, non-vernacular, professionally-sounding voices to African American people.

Paulo Cavalin, Cassia Sanctos, Marcelo Grave et al.

In this work we present the Consistency-Rebalanced Accuracy (CoRA) metric, improving the reliability of Large Language Model (LLM) scores computed on multiple choice (MC) benchmarks. Our metric explores the response consistency of the LLMs, taking advantage of synthetically-generated questions with altered answer choices. With two intermediate scores, i.e. Bare-Minimum-Consistency Accuracy (BMCA) and Consistency Index (CI), CoRA is computed by adjusting the multiple-choice question answering (MCQA) scores to better reflect the level of consistency of the LLM. We present evaluations in different benchmarks using diverse LLMs, and not only demonstrate that LLMs can present low response consistency even when they present high MCQA scores, but also that CoRA can successfully scale down the scores of inconsistent models.

Paulo Cavalin, Cassia Sanctos, Marcelo Grave et al.

We introduce \textsc{CAT}, a framework designed to evaluate and visualize the \emph{interplay} of \emph{accuracy} and \emph{response consistency} of Large Language Models (LLMs) under controllable input variations, using multiple-choice (MC) benchmarks as a case study. Current evaluation practices primarily focus on model capabilities such as accuracy or benchmark scores and, more recently, measuring consistency is being considered an essential property for deploying LLMs in high-stake, real-world applications. We argue in this paper that although both dimensions should still be evaluated independently, their inter-dependency also need to be considered for a more nuanced evaluation of LLMs. At the core of \textsc{CAT} are the \emph{Consistency-Accuracy Relation (CAR)} curves, which visualize how model accuracy varies with increasing consistency requirements, as defined by the \emph{Minimum-Consistency Accuracy (MCA)} metric. We further propose the \emph{Consistency-Oriented Robustness Estimate (CORE)} index, a global metric that combines the area and shape of the CAR curve to quantify the trade-off between accuracy and consistency. We present a practical demonstration of our framework across a diverse set of generalist and domain-specific LLMs, evaluated on multiple MC benchmarks. We also outline how \textsc{CAT} can be extended beyond MC tasks to support long-form, open-ended evaluations through adaptable scoring functions.