Answers are grounded in this paper and cite their sources.
Large language models are increasingly deployed as reasoning and decision support components in intelligent systems, yet aggregate benchmark scores provide limited evidence about whether they can reliably access primitive operations, combine them compositionally, adapt to reformulated contexts, and preserve successful behavior under perturbation. We introduce PPC-Bench, a controlled behavioral benchmark comprising 5000 deterministic samples across mathematical computation, Python code execution, and rule-based reasoning. Nine open-source models are evaluated along five dimensions: Direct Accessibility, Compositional Utilization, Prompt Elicitation, Contextual Adaptation, and Robustness Consistency. To distinguish inaccessible primitive capability from failed composition, we further construct a separate paired set of 1800 samples organized into 240 matched task families. Each family aligns atomic probes with composition, prompted, contextual, and perturbed variants of the same task. Qwen3.5-27B correctly answers 839 of 840 atomic probes but successfully composes only 74 of 239 atom-accessible families. In contrast, Llama-3.1-8B-Instruct correctly composes 132 of 164 atom-accessible families despite lower atomic accuracy. Gemma-3-1B-IT exhibits limitations in both atomic accessibility and compositional utilization. Step-by-step prompting recovers approximately half of the accessible-but-uncomposed families for Qwen3.5-27B and Llama-3.1-8B-Instruct, while successful composition is preserved under perturbation in more than 91% of eligible families for both models. These results provide condition-specific behavioral profiles that distinguish initial composition success, recovery of composition failures, and correctness preservation within successful subsets. PPC-Bench combines broad model-level evaluation with matched-family conditional analysis, providing a basis for investigating model and prompt choices without inferring internal mechanisms from task accuracy.
No takes yet. Share an insight, caveat, or question.
Eryang Huan (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: