arXiv:2603.17218v2 Announce Type: replace Abstract: Post-training alignment optimizes language models to match human preference signals, but this objective is not equivalent to modeling observed human behavior. We compare 120 base-aligned model pairs on more than 10,000 real human decisions in multi-round strategic games - bargaining, persuasion, negotiation, and repeated matrix games. In these settings, base models outperform their aligned counterparts in predicting human choices by nearly 10:1, robustly across model families, prompt formulations, and game configurations. This pattern reverse
The proliferation of aligned language models has reached a point where their practical application and underlying assumptions can be rigorously tested against human behavior. This research utilizes a significant number of models and human interactions to draw robust conclusions.
This finding fundamentally challenges the assumption that current alignment methods lead to models accurately reflecting human behavior, redirecting focus towards more descriptive, rather than normative, AI development for certain applications. For strategic readers, this suggests new pathways for model evaluation and improvement, especially in sensitive decision-making contexts.
The understanding of 'alignment' shifts from a direct path to human-like behavior prediction to an optimization for human preference, opening a gap in predictive accuracy that requires new model architectures or alignment techniques. This also has implications for the use of AI in strategic decision-making and forecasting.
- · AI ethicists
- · Academics researching human behavior modeling
- · Developers of 'base' foundational models
- · Researchers focused on descriptive AI
- · Companies over-relying on current aligned models for behavior prediction
- · Advocates of simple preference alignment as a universal solution
- · Applications requiring precise human choice forecasting
The immediate first-order effect will be a re-evaluation of alignment strategies and metrics to differentiate between normative alignment and descriptive accuracy.
A plausible second-order consequence is a push for new benchmarks and datasets focused on real-world, multi-agent strategic interactions to better assess AI's predictive capabilities.
A speculative third-order consequence could be the emergence of dual AI systems: one aligned for normative tasks and another optimized for descriptive accuracy, leading to more nuanced AI deployments.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL