GroundCocoa: A Benchmark for Evaluating Compositional & Conditional Reasoning in Language Models
Harsh Kohli, Sachin Kumar, Huan Sun
TL;DR
GroundCocoa introduces a controllable data-generation pipeline to create grounded, conditional, and compositional reasoning tasks in a flight-booking domain. It encodes user requirements as a Product-of-Sums over flight attributes and evaluates multiple LLMs under direct, CoT, and L2M prompting, highlighting substantial model variation and a ceiling around 67% accuracy for GPT-4 Turbo. Entropy-based analysis and POS-derived dependency graphs reveal that higher complexity and conditional branching degrade performance, underscoring gaps in current models’ grounding abilities. The work provides a scalable benchmark and releaseable tooling to extend evaluation to other domains and more complex reasoning forms, guiding future improvements in robust, grounded reasoning in LLMs.
Abstract
The rapid progress of large language models (LLMs) has seen them excel and frequently surpass human performance on standard benchmarks. This has enabled many downstream applications, such as LLM agents, to rely on their reasoning to address complex task requirements. However, LLMs are known to unexpectedly falter in simple tasks and under seemingly straightforward circumstances - underscoring the need for better and more diverse evaluation setups to measure their true capabilities. To this end, we choose to study compositional and conditional reasoning, two aspects that are central to human cognition, and introduce GroundCocoa - a lexically diverse benchmark connecting these reasoning skills to the real-world problem of flight booking. Our task involves aligning detailed user preferences with available flight options presented in a multiple-choice format. Results indicate a significant disparity in performance among current state-of-the-art LLMs with even the best performing model, GPT-4 Turbo, not exceeding 67% accuracy despite advanced prompting techniques.