Relational Preference Encoding in Looped Transformer Internal States

In many applications, human and LLM evaluators use assessments of relevant criteria to create an overall evaluation for an item or individual. For example, in admissions, committees assess candidates on attributes such as test scores, GPA, and research experience to evaluate their overall fit for the program. Another example arises in medical care where clinicians use patient reports of symptoms to consider preliminary diagnoses and assess risks. Each setting involves mapping multiple criteria to an overall evaluation -- a process that reflects the evaluator's underlying preferences. We focus on the fundamental question of learning these preferences. Many applications of this problem make specific modeling assumptions on evaluator preferences that may be substantially violated in the real world. We make the minimal assumption that the preference function is coordinate-wise non-decreasing, which is reasonable in a large number of evaluation settings. We theoretically characterize the severity of model mismatch for many common assumptions and show that it can lead to significant issues for learning evaluator preferences and other important downstream tasks. We then present an algorithm for learning evaluators' preferences that is robust to model mismatch. We prove theoretically that our algorithm can learn any preference function without sacrificing performance when the linearity assumption holds. Evaluations of our algorithm with synthetic simulations and real-world data confirm its ability to learn preferences robustly and illustrate key aspects of LLM and human preferences.

In RLHF, each training example contains a prompt $x$ and two candidate responses $y,y'$, and annotators provide pairwise preferences between these responses. The learning problem is to convert these heterogeneous pairwise judgments into a single scalar reward $r(x,y)$ that measures response quality for each prompt. Classical social choice implies an impossibility because heterogeneous annotator samples can induce pooled preferences with Condorcet cycles, so no scalar reward can evaluate all compared response pairs consistently. A growing literature analyzes RLHF as a social-choice problem, but usually assumes a fixed finite set of alternatives, i.e., a pre-enumerated finite set of candidate responses for each prompt. Modern pipelines instead score responses through a learned representation $φ(x,y)$ before a scalar head, so $φ$ determines which responses are treated as distinguishable alternatives and which comparisons are visible to the reward model. Once this embedding is part of the problem, the impossibility results from social choice theory become a tradeoff. We show that the excess cross-entropy loss of any reward built on $φ$ decomposes exactly into a representational term, which a richer $φ$ shrinks, and an aggregation term, which a richer $φ$ enlarges by exposing more comparisons that no scalar can rank consistently. The same results extend to direct preference optimization (DPO), and jointly training the embedding and the reward cannot guarantee to recover the sweet spot of this tradeoff. Experiments on synthetic data and real preference datasets corroborate our results.

Pairwise human comparisons are a primary interface through which modern AI systems learn human preferences. RLHF and related alignment pipelines typically model such comparisons with Bradley--Terry log-odds, where choice probabilities are governed by latent reward differences. This paper examines what this assumption misses through a reduced-form model motivated by rational inattention, in which each label is generated by a low-capacity evaluation channel. The model separates two forms of ambiguity that standard reward modeling tends to conflate: a comparison may be difficult because the two candidates are genuinely close in value, or because the relevant distinction is hard to detect under limited attention. We show that limited attention can fundamentally distort what pairwise comparisons reveal. In particular, passive comparison data cannot generally distinguish reward, attention, and default tendencies, and heterogeneous attention can make standard Bradley--Terry reward modeling recover misleading rankings. Our analysis shows that learning is governed not by the raw number of labels, but by the amount of attended information each label carries. A case study on human votes over language-model pairs from Chatbot Arena exhibits the predicted signature, a cyclic component of the comparison data that exceeds sampling noise and that no scalar reward can represent; a second case study on perceptual comparisons shows that response times and gaze carry gap information that the labels do not. This perspective suggests that human feedback should be treated not as direct revealed preference, but as an attention-limited measurement process: a weak preference signal may reflect hidden evaluation difficulty rather than genuine indifference.