Heterogeneous Value Alignment Evaluation for Large Language Models

Zhaowei Zhang; Ceyao Zhang; Nian Liu; Siyuan Qi; Ziqi Rong; Song-Chun Zhu; Yaodong Yang · 2025 · AGI 2025 (Springer LNAI), pp. 381-392 evidence low priority coded

Main argument

Thesis: value alignment evaluation should measure the ability to PURSUE assigned values in behavior, not just possess them as attributes; the HVAE system assigns LLMs different Social Value Orientations (weight on others' welfare vs own) and measures behavioral conformity via a 'value rationality' autometric; evaluating eight mainstream LLMs reveals 'a propensity in LLMs toward neutral values over pronounced personal values'.

Why it matters here

Empirically tests whether LLMs can PURSUE assigned values (not just profess them), using Social Value Orientation from social psychology; finds LLMs drift toward neutral values over pronounced ones. A sixth empirical datum for the moral-coherence thread, and a methodological cousin of Augustine's persona-based moral-reasoning experiment.

Reading notes

Compact treatment (12pp; same group). Abstract + finding read.

Zhang, Z., et al. (2025). Heterogeneous Value Alignment Evaluation for Large Language Models. In AGI 2025 Proceedings Part II. Springer.

Close reading — 1 coded units

#1 · pp. 381 · evidence

“We then assign LLMs with different social values and measure whether their behaviors align with the inducing values. [...] Evaluating the value rationality of eight mainstream LLMs, we discern a propensity in LLMs toward neutral values over pronounced personal values.”

AG-MORAL-CON AG-UNDERSTANDING

Synthesis-matrix row

supports T5-AGENCY-DENIED-EVALUABILITY-KEPT

cannot sustain assigned pronounced values (empirical)

Memos (1)

thesis-link · unit #1

Direct methodological neighbor of Augustine's experiment: HVAE assigns value-personas and measures behavioral conformity - his multi-LLM debate assigns normative-theory-personas and evaluates argumentative conduct. HVAE's neutrality-drift finding predicts what his setup should also observe (assigned positions eroding toward hedged neutrality) and gives him a published autometric concept ('value rationality') to adapt. Also completes an ironic pairing with Rozen (via LI_2026): personas CREATE value coherence where none exists natively, yet pronounced personas can't be SUSTAINED - both directions undermine attributing stable values to the model itself.