cs.GTNov 9, 2026

Enforcing Axioms for AI Alignment under Loss-Based Rules

Authors: Alexandros Hollender, Sonja Kraiczy

Organizations: Technical University of Munich · University of Oxford, Principles of Intelligence

Abstract

Many alignment methods, most notably reinforcement learning from human feedback (RLHF), train reward models by minimizing a loss on pairwise comparisons of model responses. Inspired by principle-guided methods such as Constitutional AI, we treat each principle as a voter that compares responses according to that principle, for example, preferring the more helpful or less toxic response. We represent these voters as linear directions in a fixed embedding space, a simplifying assumption motivated by the Linear Representation Hypothesis. In this linear social choice model, Ge et al. (2024) showed that an optimal linear reward model can violate Pareto optimality (PO): From the principles-as-voters lens, this means a response A can be less helpful and more toxic than B, yet still receive a higher reward. We ask whether richer reward models or changes to comparison data can recover PO within loss-based training. We show that PO can still fail for polynomials of any fixed degree and one-hidden-layer ReLU multilayer perceptrons (MLPs) of any fixed width d≥2d\ge 2 with suitable finite parameter bounds. We then recover PO for loss-based training of linear rewards under suitable assumptions in an idealized setting with uniform comparison directions and equal comparison lengths. These results motivate dataset design as an approach to recovering PO in constitutional-style alignment while retaining loss-based training.

Figures & tables

Explore similar work

Sep 26, 2026

Axiom Satisfiability of Linear Rewards in Alignment

Learning from human preference data is the dominant route to aligning language models with human values. In linear social choice, where rewards are linear in a fixed feature representation of prompt-response pairs, Ge et al.[2024] show that fitting such a reward by minimizing any non-decreasing convex loss, including BTL, fails PO and PMC. Moreover, no rule that reads only the majority relation can satisfy PO once the output is required to be linearly induced. We ask what it costs to enforce these axioms anyway. To this end, we relax the linear model to allow per-candidate slack. We compute the relaxed linear reward with the smallest total slack that satisfies the axioms with a margin ηη, the minimum required difference between two reward values. Our solution satisfies the axioms under no assumptions about the voters or how comparisons were collected. We bound the optimal total slack by O(1)O(1) when ηη is at most O(1m2)O(\frac{1}{m^2}) for mm candidates. Furthermore, we exhibit an instance that forces this bound, concluding that the rate is tight up to constants. In practice, the no. of candidates far exceeds the feature dimension, and only a linear reward can be evaluated on unseen responses. We therefore introduce a new method that charges the linear part for each comparison it gets wrong. It simultaneously minimizes the total slack and the no. of violations, with a parameter λλ trading off between them. We show that the total slack is monotone but saturating in λλ: raising it reduces the violations of the deployed linear reward and increases the slack, yet the slack stays below (η+Δd)⌊m2/4⌋(η+Δ\sqrt{d})\lfloor m^2/4\rfloor, where ΔΔ and dd are the diameter and dimension of the features, respectively. Experiments on both synthetic and real-life preference data corroborate our theory and show that the linear reward output by our method beats linear BTL.
Sep 27, 2026

From Constitutions to Control: Interpretable Rewards for Aligning Language Models

Current approaches to aligning language models often make it hard to know what behavior is being rewarded or to change that reward in a targeted way. In particular, standard preference-based methods collapse multiple considerations into aggregate human judgments, obscuring what drives the resulting reward, while principle-based methods specify high-level values without fully operationalizing them. To address this gap, we develop a rubric-based framework to transform a general-purpose constitution into an interpretable and tunable reward model, using constitution-guided AI feedback to estimate initial weights for the constituent rubric items. We then reweight those dimensions to construct modified rewards for training. Across experiments on political alignment and safety-helpfulness tradeoffs, reweighting individual dimensions predictably changes targeted behaviors largely independently while navigating tradeoffs between conflicting alignment objectives. We show that the same framework can mitigate label bias encoded in preference judgments -- including sycophancy and demographic bias -- by reducing their influence on the training reward. Our results demonstrate that constitution-derived, interpretable rewards can translate high-level alignment principles into more transparent and controllable model behavior.
Mar 17, 2026

Via Negativa for AI Alignment: Why Negative Constraints Are Structurally Superior to Positive Preferences

Recent empirical results have demonstrated that training large language models (LLMs) with negative-only feedback can match or exceed standard reinforcement learning from human feedback (RLHF). Negative Sample Reinforcement achieves parity with PPO on mathematical reasoning; Distributional Dispreference Optimization trains effectively using only dispreferred samples; and Constitutional AI outperforms pure RLHF on harmlessness benchmarks. Yet no unified theoretical account explains why negative signals are so effective. This paper proposes such an account: positive preferences and negative constraints are structurally asymmetric. Positive preferences ("which is better") encode continuously coupled, context-dependent human values that cannot be exhaustively specified -- leading models to learn surface correlates such as agreement with the user (sycophancy). Negative constraints ("what is wrong") encode discrete, finite, independently verifiable prohibitions that can converge to a stable boundary. This asymmetry -- rooted in Popper's falsification logic and the epistemology of negative knowledge -- explains both the sycophancy failure of preference-based RLHF and the surprising effectiveness of negative-signal methods. We argue that alignment research should shift its center of gravity from "learning what humans prefer" to "learning what humans reject," and offer testable predictions for this framework.