A new decoding method makes automated human value detection match the theory it was built on.
Researchers benchmarked a DeBERTa-v3-base classifier against the 19 Schwartz basic values — a framework describing human motivation as a circular continuum where neighboring values align and opposing ones conflict. Standard classifiers treat each label as independent, ignoring that structure entirely. The team tested two fixes: geometry-aware training objectives and a post-hoc energy decoder that scores full label sets jointly. Training-time approaches produced only modest gains, and performed no better with the real Schwartz ordering than with a random one. The decoder, however, reliably shifted outputs toward label sets that respect the continuum — and it held Macro-F1 and Micro-F1 flat by design.
The finding matters because value detection is increasingly used to audit AI outputs and measure alignment. If a classifier labels text as endorsing both "power" and "universalism" — values the Schwartz model places in direct tension — it is producing outputs that contradict the theoretical grounding the task claims. A decoder that enforces soft geometric consistency is a cheap way to close that gap without retraining. The gain also disappeared when the decoder used a random permutation or an empirical co-occurrence graph instead of the true Schwartz ordering, which argues against the result being a statistical artifact.
As a supplementary check, the team ran a diagnostic using Qwen2.5-72B-Instruct — a publicly documented open-weight model — and found that feeding the continuum structure at inference shifted its behavior but fell short of supervised structured prediction, suggesting the decoder trick does not transfer for free to large language models. The method is lightweight and interpretable, which is more than most alignment tooling can claim, though it remains one benchmark away from a real endorsement.