AI Alignment: A Comprehensive Survey

Jiaming Ji; Tianyi Qiu; Boyuan Chen; Borong Zhang; et al. (Yaodong Yang group) · 2025 · arXiv:2310.19852v6 (Peking/Cambridge/Oxford/CMU/HKUST/USC) background high priority coded

Main argument

Thesis (survey): alignment's objectives are RICE - Robustness, Interpretability, Controllability, Ethicality; the field decomposes into FORWARD alignment (training: learning from feedback incl. preference modeling/RLHF/scalable oversight; learning under distribution shift) and BACKWARD alignment (evidence + governance: assurance incl. safety evaluation, interpretability, human values verification; and governance incl. multi-stakeholder approaches and open-source policy) forming an alignment CYCLE. Human-values verification is treated via formal machine ethics and game-theoretic cooperation frameworks. The survey closes by 'Rethinking AI Alignment from a Socio-technical Perspective' - conceding that alignment extends beyond technical training into governance, multi-stakeholder legitimacy, and deployment context.

Why it matters here

The current canonical technical survey (105pp): RICE objectives (Robustness, Interpretability, Controllability, Ethicality), the forward/backward alignment cycle, and - crucially for the dissertation - its own concluding turn to a 'socio-technical perspective' with multi-stakeholder governance. Use as THE citation for what technical alignment comprises, and for showing the technical field's own arc bends toward the governance/normative questions it cannot internally answer.

Reading notes

Targeted treatment: structure + governance/human-values-verification sections read (105pp; same Yaodong Yang group as ZHANG_2025 chapters). The 'Human Values Verification' section's formal-machine-ethics + game-theory-for-cooperation taxonomy is the technical field's thin slot where the dissertation's entire subject sits.

Ji, J., Qiu, T., Chen, B., Zhang, B., et al. (2025). AI Alignment: A Comprehensive Survey. arXiv:2310.19852v6.

Close reading — 3 coded units

#1 · pp. 1 · definition

“we identify four principles as the key objectives of AI alignment: Robustness, Interpretability, Controllability, and Ethicality (RICE). [...] we outline the landscape of current alignment research and decompose them into two key components: forward alignment and backward alignment.”

TU-LITREV

#2 · pp. 52 · definition

“Human Values Alignment refers to the expectation that AI systems should adhere to the community's social and moral norms. [...] if these systems fail to grasp the inherent complexity and adaptability of human values, their decisions could result in negative social [consequences].”

VC-ROLE

#3 · pp. 58 · argument

“[Concluding section: 'Rethinking AI Alignment from a Socio-technical Perspective' - governance, multi-stakeholder approaches, and open-source policy as integral to the alignment cycle, with European policymakers requiring 'performance, predictability, interpretability, corrigibility, security' across the lifecycle.]”

RL-INST

Synthesis-matrix row

complicates T7-AGENTIC-BREAKS-FRAMES

agents in scope; ethicality pillar remains thin

Memos (1)

thesis-link · unit #2

The survey's 'Ethicality' pillar and 'Human Values Verification' subsection are strikingly thin relative to the other 100 pages - formal machine ethics + game theory, with community norms invoked but never theorized (unit 2). This is the measured version of the McKinlay finding at the technical field's own summit: the E in RICE is a placeholder. Lit-review use: cite Ji as the definitive map of what technical alignment CAN do, then locate the dissertation precisely in the placeholder - the verification of value alignment (unit 2's 'adherence to community's social and moral norms') requires exactly what the xphi corpus + convergentist metaethics provide: an empirically grounded, normatively defensible account of which community norms, whose, and why.