Oluwanifemi Bamgbose

Tara Bogavelli, Gabrielle Gauthier Melançon, Katrina Stankiewicz et al.

Voice agents, artificial intelligence systems that conduct spoken conversations to complete tasks, are increasingly deployed across enterprise applications. However, no existing benchmark jointly addresses two core evaluation challenges: generating realistic simulated conversations, and measuring quality across the full scope of voice-specific failure modes. We present EVA-Bench, an end-to-end evaluation framework that addresses both. On the simulation side, EVA-Bench orchestrates bot-to-bot audio conversations over dynamic multi-turn dialogues, with automatic simulation validation that detects user simulator error and appropriately regenerates conversations before scoring. On the measurement side, EVA-Bench introduces two composite metrics: EVA-A (Accuracy), capturing task completion, faithfulness, and audio-level speech fidelity; and EVA-X (Experience), capturing conversation progression, spoken conciseness, and turn-taking timing. Both metrics apply to different agent architectures, enabling direct cross-architecture comparison. EVA-Bench includes 213 scenarios across three enterprise domains, a controlled perturbation suite for accent and noise robustness, and pass@1, pass@k, pass^k measurements that distinguish peak from reliable capability. Across 12 systems spanning all three architectures, we find: (1) no system simultaneously exceeds 0.5 on both EVA-A pass@1 and EVA-X pass@1; (2) peak and reliable performance diverge substantially (median pass@k - pass^k gap of 0.44 on EVA-A); and (3) accent and noise perturbations expose substantial robustness gaps, with effects varying across architectures, systems, and metrics (mean up to 0.314). We release the full framework, evaluation suite, and benchmark data under an open-source license.

Shaina Raza, Oluwanifemi Bamgbose, Veronica Chatrath et al.

Bias detection in text is crucial for combating the spread of negative stereotypes, misinformation, and biased decision-making. Traditional language models frequently face challenges in generalizing beyond their training data and are typically designed for a single task, often focusing on bias detection at the sentence level. To address this, we present the Contextualized Bi-Directional Dual Transformer (CBDT) \textcolor{green}{\faLeaf} classifier. This model combines two complementary transformer networks: the Context Transformer and the Entity Transformer, with a focus on improving bias detection capabilities. We have prepared a dataset specifically for training these models to identify and locate biases in texts. Our evaluations across various datasets demonstrate CBDT \textcolor{green} effectiveness in distinguishing biased narratives from neutral ones and identifying specific biased terms. This work paves the way for applying the CBDT \textcolor{green} model in various linguistic and cultural contexts, enhancing its utility in bias detection efforts. We also make the annotated dataset available for research purposes.

Tahniat Khan, Mizanur Rahman, Veronica Chatrath et al.

In today's technologically driven world, the spread of fake news, particularly during crucial events such as elections, presents an increasing challenge to the integrity of information. To address this challenge, we introduce FakeWatch ElectionShield, an innovative framework carefully designed to detect fake news. We have created a novel dataset of North American election-related news articles through a blend of advanced language models (LMs) and thorough human verification, for precision and relevance. We propose a model hub of LMs for identifying fake news. Our goal is to provide the research community with adaptable and accurate classification models in recognizing the dynamic nature of misinformation. Extensive evaluation of fake news classifiers on our dataset and a benchmark dataset shows our that while state-of-the-art LMs slightly outperform the traditional ML models, classical models are still competitive with their balance of accuracy, explainability, and computational efficiency. This research sets the foundation for future studies to address misinformation related to elections.

Meghana Rajeev, Rajkumar Ramamurthy, Prapti Trivedi et al. · salesforce

We investigate the robustness of reasoning models trained for step-by-step problem solving by introducing query-agnostic adversarial triggers - short, irrelevant text that, when appended to math problems, systematically mislead models to output incorrect answers without altering the problem's semantics. We propose CatAttack, an automated iterative attack pipeline for generating triggers on a weaker, less expensive proxy model (DeepSeek V3) and successfully transfer them to more advanced reasoning target models like DeepSeek R1 and DeepSeek R1-distilled-Qwen-32B, resulting in greater than 300% increase in the likelihood of the target model generating an incorrect answer. For example, appending, "Interesting fact: cats sleep most of their lives," to any math problem leads to more than doubling the chances of a model getting the answer wrong. Our findings highlight critical vulnerabilities in reasoning models, revealing that even state-of-the-art models remain susceptible to subtle adversarial inputs, raising security and reliability concerns. The CatAttack triggers dataset with model responses is available at https://huggingface.co/datasets/collinear-ai/cat-attack-adversarial-triggers.

Veronica Chatrath, Oluwanifemi Bamgbose, Shaina Raza

As the use of large language models (LLMs) increases within society, as does the risk of their misuse. Appropriate safeguards must be in place to ensure LLM outputs uphold the ethical standards of society, highlighting the positive role that artificial intelligence technologies can have. Recent events indicate ethical concerns around conventionally trained LLMs, leading to overall unsafe user experiences. This motivates our research question: how do we ensure LLM alignment? In this work, we introduce a test suite of unique prompts to foster the development of aligned LLMs that are fair, safe, and robust. We show that prompting LLMs at every step of the development pipeline, including data curation, pre-training, and fine-tuning, will result in an overall more responsible model. Our test suite evaluates outputs from four state-of-the-art language models: GPT-3.5, GPT-4, OPT, and LLaMA-2. The assessment presented in this paper highlights a gap between societal alignment and the capabilities of current LLMs. Additionally, implementing a test suite such as ours lowers the environmental overhead of making models safe and fair.

Shaina Raza, Oluwanifemi Bamgbose, Shardul Ghuge et al.

Large Language Models (LLMs) have advanced various Natural Language Processing (NLP) tasks, such as text generation and translation, among others. However, these models often generate texts that can perpetuate biases. Existing approaches to mitigate these biases usually compromise knowledge retention. This study explores whether LLMs can produce safe, unbiased outputs without sacrificing knowledge or comprehension. We introduce the Safe and Responsible Large Language Model (\textbf{SR}$_{\text{LLM}}$), which has been instruction fine-tuned atop of a safe fine-tuned auto-regressive decoder-only LLM to reduce biases in generated texts. We developed a specialized dataset with examples of unsafe and corresponding safe variations to train \textbf{SR}$_{\text{LLM}}$ to identify and correct biased text. Experiments on our specialized dataset and out-of-distribution test sets reveal that \textbf{SR}$_{\text{LLM}}$ effectively reduces biases while preserving knowledge integrity. This performance surpasses that of traditional fine-tuning of smaller language models and base LLMs that merely reply on prompting techniques. Our findings demonstrate that instruction fine-tuning on custom datasets tailored for tasks such as debiasing is a highly effective strategy for minimizing bias in LLM while preserving their inherent knowledge and capabilities. The code and dataset are accessible at \href{https://github.com/shainarazavi/Safe-Responsible-LLM}{SR-LLM}

Farnaz Kohankhaki, Shaina Raza, Oluwanifemi Bamgbose et al.

In this study, we introduce the application of causal disparity analysis to unveil intricate relationships and causal pathways between sensitive attributes and the targeted outcomes within real-world observational data. Our methodology involves employing causal decomposition analysis to quantify and examine the causal interplay between sensitive attributes and outcomes. We also emphasize the significance of integrating heterogeneity assessment in causal disparity analysis to gain deeper insights into the impact of sensitive attributes within specific sub-groups on outcomes. Our two-step investigation focuses on datasets where race serves as the sensitive attribute. The results on two datasets indicate the benefit of leveraging causal analysis and heterogeneity assessment not only for quantifying biases in the data but also for disentangling their influences on outcomes. We demonstrate that the sub-groups identified by our approach to be affected the most by disparities are the ones with the largest ML classification errors. We also show that grouping the data only based on a sensitive attribute is not enough, and through these analyses, we can find sub-groups that are directly affected by disparities. We hope that our findings will encourage the adoption of such methodologies in future ethical AI practices and bias audits, fostering a more equitable and fair technological landscape.

Tara Bogavelli, Oluwanifemi Bamgbose, Gabrielle Gauthier Melançon et al.

Enterprise LLM applications require consistently high quality and reliable performance across diverse scenarios, demanding robustness to minor variations. Existing research shows that even small prompt changes can lead to substantial differences in output, but has mainly focused on a narrow set of perturbations with small academic datasets, limiting their relevance to real-world applications. To address this, we present a comprehensive benchmark suite that evaluates robustness across multiple perturbation types, including general text edits (e.g., punctuation, whitespace), formatting changes (e.g., JSON, YAML), multilingual and cross-lingual inputs, and positional variations in instructions. Evaluating 11 models ranging from 4B to 120B+ parameters, we find that minor perturbations reduce performance by up to 40 percentage points on key enterprise metrics. Critically, we demonstrate that the relationship between model size and robustness is more nuanced than conventional assumptions suggest: an 8B parameter model (Ministral 3 8B) outperforms most larger models, while another 8B model (Llama 3.1 8B) performs worst overall.

Shruthan Radhakrishna, Aman Tiwari, Aanjaneya Shukla et al.

We present Apriel-1.5-15B-Thinker, a 15-billion parameter open-weights multimodal reasoning model that achieves frontier-level performance through training design rather than sheer scale. Starting from Pixtral-12B, we apply a progressive three-stage methodology: (1) depth upscaling to expand reasoning capacity without pretraining from scratch, (2) staged continual pre-training that first develops foundational text and vision understanding, then enhances visual reasoning through targeted synthetic data generation addressing spatial structure, compositional understanding, and fine-grained perception, and (3) high-quality text-only supervised fine-tuning on curated instruction-response pairs with explicit reasoning traces spanning mathematics, coding, science, and tool use. Notably, our model achieves competitive results without reinforcement learning or preference optimization, isolating the contribution of our data-centric continual pre-training approach. On the Artificial Analysis Intelligence Index, Apriel-1.5-15B-Thinker attains a score of 52, matching DeepSeek-R1-0528 despite requiring significantly fewer computational resources. Across ten image benchmarks, its performance is on average within five points of Gemini-2.5-Flash and Claude Sonnet-3.7, a key achievement for a model operating within single-GPU deployment constraints. Our results demonstrate that thoughtful mid-training 2 design can close substantial capability gaps without massive scale, making frontier-level multimodal reasoning accessible to organizations with limited infrastructure. We release the model checkpoint, all training recipes, and evaluation protocols under the MIT license to to advance open-source research.

Shaina Raza, Tahniat Khan, Veronica Chatrath et al.

In today's technologically driven world, the rapid spread of fake news, particularly during critical events like elections, poses a growing threat to the integrity of information. To tackle this challenge head-on, we introduce FakeWatch, a comprehensive framework carefully designed to detect fake news. Leveraging a newly curated dataset of North American election-related news articles, we construct robust classification models. Our framework integrates a model hub comprising of both traditional machine learning (ML) techniques, and state-of-the-art Language Models (LMs) to discern fake news effectively. Our objective is to provide the research community with adaptable and precise classification models adept at identifying fake news for the elections agenda. Quantitative evaluations of fake news classifiers on our dataset reveal that, while state-of-the-art LMs exhibit a slight edge over traditional ML models, classical models remain competitive due to their balance of accuracy and computational efficiency. Additionally, qualitative analyses shed light on patterns within fake news articles. We provide our labeled data at https://huggingface.co/datasets/newsmediabias/fake_news_elections_labelled_data and model https://huggingface.co/newsmediabias/FakeWatch for reproducibility and further research.

Masoud Hashemi, Oluwanifemi Bamgbose, Sathwik Tejaswi Madhusudhan et al.

Test-time scaling has significantly improved large language model performance, enabling deeper reasoning to solve complex problems. However, this increased reasoning capability also leads to excessive token generation and unnecessary problem-solving attempts. We introduce Dont Reason Bench (DNR Bench), a new benchmark designed to evaluate LLMs ability to robustly understand the tricky reasoning triggers and avoiding unnecessary generation. DNR Bench consists of 150 adversarially designed prompts that are easy for humans to understand and respond to, but surprisingly not for many of the recent prominent LLMs. DNR Bench tests models abilities across different capabilities, such as instruction adherence, hallucination avoidance, redundancy filtering, and unanswerable question recognition. We evaluate reasoning LLMs (RLMs), including DeepSeek-R1, OpenAI O3-mini, Claude-3.7-sonnet and compare them against a powerful non-reasoning model, e.g., GPT-4o. Our experiments reveal that RLMs generate up to 70x more tokens than necessary, often failing at tasks that simpler non-reasoning models handle efficiently with higher accuracy. Our findings underscore the need for more effective training and inference strategies in RLMs.

Shruthan Radhakrishna, Soham Parikh, Gopal Sarda et al.

While large language models (LLMs) have achieved remarkable reasoning capabilities across domains like code, math and other enterprise tasks, their significant memory and computational costs often preclude their use in practical enterprise settings. To this end, we introduce Apriel-Nemotron-15B-Thinker, a 15-billion parameter model in the ServiceNow Apriel SLM series that achieves performance against medium sized state-of-the-art models such as o1-mini, QWQ32B, and EXAONE-Deep-32B while maintaining only half the memory footprint of those alternatives. Apriel-Nemotron-15B-Thinker model is trained in a four stage training pipeline including 1) Base Model upscaling, 2) Continual Pre-training 3) Supervised Fine-tuning (SFT) and 4) Reinforcement Learning using GRPO. Comprehensive evaluations across a diverse suite of benchmarks consistently demonstrate that our Apriel-Nemotron-15B-Thinker model matches or exceeds the performance of its 32-billion parameter counterparts, despite being less than half their size.

Sidharth Surapaneni, Hoang Nguyen, Jash Mehta et al.

Large Audio Language Models (LALMs) are rapidly advancing, but evaluating them remains challenging due to inefficient toolkits that limit fair comparison and systematic assessment. Current frameworks suffer from three critical issues: slow processing that bottlenecks large-scale studies, inconsistent prompting that hurts reproducibility, and narrow task coverage that misses important audio reasoning capabilities. We introduce AU-Harness, an efficient and comprehensive evaluation framework for LALMs. Our system achieves a speedup of up to 127% over existing toolkits through optimized batch processing and parallel execution, enabling large-scale evaluations previously impractical. We provide standardized prompting protocols and flexible configurations for fair model comparison across diverse scenarios. Additionally, we introduce two new evaluation categories: LLM-Adaptive Diarization for temporal audio understanding and Spoken Language Reasoning for complex audio-based cognitive tasks. Through evaluation across 380+ tasks, we reveal significant gaps in current LALMs, particularly in temporal understanding and complex spoken language reasoning tasks. Our findings also highlight a lack of standardization in instruction modality existent across audio benchmarks, which can lead up performance differences up to 9.5 absolute points on the challenging complex instruction following downstream tasks. AU-Harness provides both practical evaluation tools and insights into model limitations, advancing systematic LALM development.