Value crucible for evaluating robustness of value attributed LLM response profiles via agent adversarial debates

Static benchmarks reveal what large language models (LLMs) output in isolated prompts but provide limited insight into whether value-attributed response profiles remain consistent during extended interactions. We introduce Value Crucible, a three-stage agent-based framework for evaluating the robustness of value-attributed responses under role conditioning and adversarial conversational pressure. Guided by Schwartz's refined value theory as a standardized coding space and informed by Bardi and Goodwin's dual-route account of value change as a design heuristic for conversational influence, we evaluate 11 mainstream LLMs across 10 social roles and 57 value scenarios. Value Crucible first elicits default assistant-persona and human-norm estimation profiles, then examines how role prompts reshape value-attributed responses, and finally uses self-confrontation debates to assess stance preservation after sustained challenge using the Stance Preservation Index (SPI). Across models, roles, and value dimensions, we observe differences in response robustness. Initially moderate ratings show greater post-debate movement than strongly endorsed or rejected ratings, a pattern that remains robust under mixed-effects and boundary-corrected analyses. Response shifts also exhibit circumplex-consistent trade-offs within the Schwartz-coded measurement space. By extending evaluation from static outputs to dynamic interactions, Value Crucible provides a scalable framework for identifying when value-attributed responses remain robust or become susceptible to conversational reframing.