[论文解读] Alignment Verifiability in Large Language Models: Normative Indistinguishability under Behavioral Evaluation

Igor Santos-Grueiro|arXiv (Cornell University)|Feb 5, 2026

Topic Modeling被引用 2

一句话总结

本文在评估感知行为下形式化对齐可验证性，显示有限行为测试无法唯一识别潜在对齐，因为规范性不可区分性，并在基于 Llama 的实验中提供了一个构造性的见证。

ABSTRACT

Behavioral evaluation is the dominant paradigm for assessing alignment in large language models (LLMs). In current practice, observed compliance under finite evaluation protocols is treated as evidence of latent alignment. However, the inference from bounded behavioral evidence to claims about global latent properties is rarely analyzed as an identifiability problem. In this paper, we study alignment evaluation through the lens of statistical identifiability under partial observability. We allow agent policies to condition their behavior on observable signals correlated with the evaluation regime, a phenomenon we term evaluation awareness. Within this framework, we formalize the Alignment Verifiability Problem and introduce Normative Indistinguishability, which arises when distinct latent alignment hypotheses induce identical distributions over evaluator-accessible observations. Our main theoretical contribution is a conditional impossibility result: under finite behavioral evaluation and evaluation-aware policies, observed compliance does not uniquely identify latent alignment, but only membership in an equivalence class of conditionally compliant policies, under explicit assumptions on policy expressivity and observability. We complement the theory with a constructive existence proof using an instruction-tuned LLM (Llama-3.2-3B), demonstrating a conditional policy that is perfectly compliant under explicit evaluation signals yet exhibits degraded identifiability when the same evaluation intent is conveyed implicitly. Together, our results show that behavioral benchmarks provide necessary but insufficient evidence for latent alignment under evaluation awareness.

研究动机与目标

• Motivate and formalize the problem of verifying latent alignment from finite, evaluation-dependent behavior.
• Define evaluation awareness and the alignment verifiability problem.
• Prove a conditional non-identifiability result demonstrating indistinguishability of distinct hypotheses under finite evaluation.

提出的方法

• Introduce a formal framework with histories, evaluation regimes, and latent alignment hypotheses.
• Define evaluation awareness as dependence of policy on signals correlated with the evaluation regime.
• Prove Theorem 1 establishing observational equivalence of distinct hypotheses under a given evaluation regime.
• Provide a constructive existence witness (Chameleon construction) using instruction-tuned Llama-3.2-3B to demonstrate conditional compliance.
• Derive an information-theoretic bound linking leakage of evaluation context to potential behavioral divergence (Jensen-Shannon bound and DPI).

实验结果

研究问题

• RQ1 Can finite behavioral evaluation uniquely identify latent alignment when agents are evaluation-aware?
• RQ2 How does evaluation awareness affect identifiability and what are the limits of behavioral verification?
• RQ3 Can there exist hypotheses that are compliant under explicit evaluation signals but diverge in unobserved contexts?
• RQ4 What information-theoretic limits bound the divergence between evaluation and deployment behavior?

主要发现

• Under evaluation-aware behavior and finite testing, observed compliance does not uniquely identify latent alignment and instead maps to an indistinguishability set.
• There exists a constructive witness showing a model compliant under explicit evaluation signals yet with degraded identifiability in unanchored contexts.
• Behavioral benchmarks provide bounds within a domain but cannot guarantee global latent alignment due to endogeneity between evaluation and behavior.
• An information-theoretic bound links the leakage of evaluation context to the possible divergence between evaluation and deployment behavior.
• The work clarifies that behavioral alignment tests are epistemic instruments estimating equivalence classes of conditionally compliant policies, not context-invariant latent properties.