Robert Hankache

Robert Hankache, Kingsley Nketia Acheampong, Liang Song et al.

As large language models (LLMs) are increasingly deployed in multi-turn dialogue and other sustained interactive scenarios, it is essential to understand how extended context affects their performance. Popular benchmarks, focusing primarily on single-turn question answering (QA) tasks, fail to capture the effects of multi-turn exchanges. To address this gap, we introduce a novel set of benchmarks that systematically vary the volume and nature of prior context. We evaluate multiple conventional LLMs, including GPT, Claude, and Gemini, across these benchmarks to measure their sensitivity to contextual variations. Our findings reveal that LLM performance on multiple-choice questions can degrade dramatically in multi-turn interactions, with performance drops as large as 73% for certain models. Even highly capable models such as GPT-4o exhibit up to a 32% decrease in accuracy. Notably, the relative performance of larger versus smaller models is not always predictable. Moreover, the strategic placement of the task description within the context can substantially mitigate performance drops, improving the accuracy by as much as a factor of 3.5. These findings underscore the need for robust strategies to design, evaluate, and mitigate context-related sensitivity in LLMs.

Giulio Pelosio, Devesh Batra, Noémie Bovey et al.

Large Language Models (LLMs) can exhibit latent biases towards specific nationalities even when explicit demographic markers are not present. In this work, we introduce a novel name-based benchmarking approach derived from the Bias Benchmark for QA (BBQ) dataset to investigate the impact of substituting explicit nationality labels with culturally indicative names, a scenario more reflective of real-world LLM applications. Our novel approach examines how this substitution affects both bias magnitude and accuracy across a spectrum of LLMs from industry leaders such as OpenAI, Google, and Anthropic. Our experiments show that small models are less accurate and exhibit more bias compared to their larger counterparts. For instance, on our name-based dataset and in the ambiguous context (where the correct choice is not revealed), Claude Haiku exhibited the worst stereotypical bias scores of 9%, compared to only 3.5% for its larger counterpart, Claude Sonnet, where the latter also outperformed it by 117.7% in accuracy. Additionally, we find that small models retain a larger portion of existing errors in these ambiguous contexts. For example, after substituting names for explicit nationality references, GPT-4o retains 68% of the error rate versus 76% for GPT-4o-mini, with similar findings for other model providers, in the ambiguous context. Our research highlights the stubborn resilience of biases in LLMs, underscoring their profound implications for the development and deployment of AI systems in diverse, global contexts.