arXiv subjects

Suvadip Sana

Publications and source records attributed to Suvadip Sana.

5 recordsLinked to original sources

arXiv2026

Mixing Times of Glauber Dynamics on Masked Language Models

Masked language models (MLMs) define local conditional distributions over tokens but do not, in general, correspond to any consistent joint distribution over sequences. This raises a fundamental question: what global distributional behavior is induced when such conditionals are used iteratively for generation? We address this question by modeling iterative masked-token resampling as a Glauber dynamics Markov chain on the discrete space of token sequences. We first show that MLM conditionals are intrinsically incompatible: we introduce a rectangle test that certifies this incompatibility and empirically verify its prevalence across modern MLMs. We then provide a theoretical analysis of the induced Markov chain. Under bounded cross-token influence, we establish a high-temperature contraction result implying $O(n\log n)$ mixing time where $n$ is the sequence length. In contrast, we prove that under a uniform local margin condition, the chain exhibits metastability, with exponentially slow escape from semantic basins at low temperatures. Empirically, we demonstrate a phase transition in mixing behavior as a function of temperature and sequence length, consistent with the theoretical predictions. We further characterize the induced stationary behavior through semantic trajectories, identifying persistent structures such as long-lived traps and recurrent semantic basins, with political content serving as a measurable case study.

cs.LG↗

arXiv2026

EigenBench: A Comparative Behavioral Measure of Value Alignment

Aligning AI with human values is a pressing unsolved problem. To address the lack of quantitative metrics for value alignment, we propose EigenBench: a black-box method for comparatively benchmarking language models' values. Given an ensemble of models, a constitution describing a value system, and a dataset of scenarios, our method returns a vector of scores quantifying each model's alignment to the given constitution. To produce these scores, each model judges the outputs of other models across many scenarios, and these judgments are aggregated with EigenTrust (Kamvar et al., 2003), yielding scores that reflect a weighted consensus judgment of the whole ensemble. EigenBench uses no ground truth labels, as it is designed to quantify subjective traits for which reasonable judges may disagree on the correct label. Hence, to validate our method, we collect human judgments on the same ensemble of models and show that EigenBench's judgments align closely with those of human evaluators. We further demonstrate that EigenBench can recover model rankings on the GPQA benchmark without access to objective labels, supporting its viability as a framework for evaluating subjective values for which no ground truths exist. The code is available at https://github.com/jchang153/EigenBench.

cs.AI↗

arXiv2026

Democratic Preference Alignment via Sortition-Weighted RLHF

Whose values should AI systems learn? Preference based alignment methods like RLHF derive their training signal from human raters, yet these rater pools are typically convenience samples that systematically over represent some demographics and under represent others. We introduce Democratic Preference Optimization, or DemPO, a framework that applies algorithmic sortition, the same mechanism used to construct citizen assemblies, to preference based fine tuning. DemPO offers two training schemes. Hard Panel trains exclusively on preferences from a quota satisfying mini public sampled via sortition. Soft Panel retains all data but reweights each rater by their inclusion probability under the sortition lottery. We prove that Soft Panel weighting recovers the expected Hard Panel objective in closed form. Using a public preference dataset that pairs human judgments with rater demographics and a seventy five clause constitution independently elicited from a representative United States panel, we evaluate Llama models from one billion to eight billion parameters fine tuned under each scheme. Across six aggregation methods, the Hard Panel consistently ranks first and the Soft Panel consistently outperforms the unweighted baseline, with effect sizes growing as model capacity increases. These results demonstrate that enforcing demographic representativeness at the preference collection stage, rather than post hoc correction, yields models whose behavior better reflects values elicited from representative publics.

cs.AI↗

arXiv2025

Quantitative Relaxations of Arrow's Axioms

In this paper we develop a novel approach to relaxing Arrow's axioms for voting rules, addressing a long-standing critique in social choice theory. Classical axioms (often styled as fairness axioms or fairness criteria) are assessed in a binary manner, so that a voting rule fails the axiom if it fails in even one corner case. Many authors have proposed a probabilistic framework to soften the axiomatic approach. Instead of immediately passing to random preference profiles, we begin by measuring the degree to which an axiom is upheld or violated on a given profile. We focus on two foundational axioms-Independence of Irrelevant Alternatives (IIA) and Unanimity (U)-and extend them to take values in $[0,1]$. Our $σ_{IIA}$ measures the stability of a voting rule when candidates are removed from consideration, while $σ_{U}$ captures the degree to which the outcome respects majority preferences. Together, these metrics quantify how a voting rule navigates the fundamental trade-off highlighted by Arrow's Theorem. We show that $σ_{IIA}\equiv 1$ recovers classical IIA, and $σ_{U}>0$ recovers classical Unanimity, allowing a quantitative restatement of Arrow's Theorem. In the empirical part of the paper, we test these metrics on two kinds of data: a set of over 1000 ranked choice preference profiles from Scottish local elections, and a batch of synthetic preference profiles generated with a Bradley-Terry-type model. We use those to investigate four positional voting rules-Plurality, 2-Approval, 3-Approval, and the Borda rule-as well as the iterative rule known as Single Transferable Vote (STV). The Borda rule consistently receives the highest $σ_{IIA}$ and $σ_{U}$ scores across observed and synthetic elections. This compares interestingly with a recent result of Maskin showing that weakening IIA to include voter preference intensity uniquely selects Borda.

cs.GT↗

arXiv2024

A scalar matching factor on the Birkhoff polytope characterizing permutation and uniform matrices

Birkhoff polytope is the set of all bistochastic matrices (also known as doubly stochastic matrices). Bistochastic matrices form a special class of stochastic matrices where each row and column sums up to one. Permutation matrices and uniform matrices are special extreme cases of bistochastic matrices. In this paper, we define a scalar quantity called the matching factor on the Birkhoff polytope. Given a bistochastic matrix, we define the matching factor by taking the product of the squares of the Euclidean norms of each row and column and show that permutation matrices and uniform matrices maximize and minimize the matching factor, respectively. We also extend this definition of scalar matching factor to a larger class of matrices and show similar maximization and minimization properties.

math.RA↗