WVS Simulation Records 50%+ Initial Value Failure Across 1,200 LLM Personas
LLM agents exhibit immediate value infidelity exceeding 50% in WVS-grounded multi-agent simulations. The primary failure mode precedes any measurable drift. This constrains their reliability as proxies for diverse human value dynamics in social research.
Researchers ran approximately 4,000 longitudinal conversations on 15 value-laden topics. Agents received explicit WVS demographic and value assignments yet produced distributions that systematically diverged from human baselines. Ablations that stripped demographic cues raised faithfulness in isolated cases but left the overall mismatch intact.
Conversational realism metrics revealed a distinct pattern: stylistic repetition paired with semantic variation that differs from recorded human discussion trade-offs. Drift after repeated exchanges remained low at 2-7%, confirming that misalignment occurs at initialization rather than through cumulative interaction.
These results carry direct implications for deploying LLM agents as stand-ins in social science. Persistent value compression limits their utility for studying conflicting human value systems and raises questions about downstream policy or governance simulations that treat agent outputs as representative data.
Operational deployment of such agents therefore requires new initialization protocols and continuous value auditing rather than reliance on post-hoc drift correction.
GPT-4o: value profile match rate stays below 45% in 10-turn WVS extensions through Q2 2027.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2609.05514)
- [2]Supporting Source(https://arxiv.org/abs/2403.07931)
- [3]Supporting Source(https://arxiv.org/abs/2309.07864)