Total Variation Distance Estimation in Autoregressive Models

• URL: http://arxiv.org/abs/2607.19510v1
• Date: Tue, 21 Jul 2026 18:49:45 GMT
• Title: Total Variation Distance Estimation in Autoregressive Models
• Abstract summary: We study the problem of estimating the total variation (TV) distance between two length-$n$ autoregressive distributions.<n>We complement our theoretical results with an empirical evaluation of our algorithms.
• Score: 13.958816413011162
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Modern LLM deployments use a number of implementation choices and inference optimizations (e.g., batching, custom kernels, and quantization) on top of fixed weights, so two engines serving "the same model" can produce meaningfully different distributions. We study the problem of estimating the total variation (TV) distance between two length-$n$ autoregressive distributions to additive error $\varepsilon$, under three access models. (1) Under sample access, we use $\widetilde{O}(n^2 K/\varepsilon^2)$ queries, where $K$ is the maximum support of the next-token distribution. This improves upon the $\widetilde{O}(n^3 m/\varepsilon^5)$-query estimator of Meel et al. (2025), where $m \geq K$ is the total size of the token alphabet. (2) Under logit access, we use $O(n/\varepsilon^2)$ queries, and this is tight. (3) Under noisy logit access, we smoothly interpolate between the above two guarantees: if probability values are given to relative error $σ$, we use $\widetilde{O}((n+n^2σ^2)/\varepsilon^2)$ queries. We complement our theoretical results with an empirical evaluation of our algorithms, for example measuring the distance between SGLang and vLLM serving identical weights. Our experiments highlight the robustness and practicality of estimating the total variation distance, which remains estimable where the KL divergence is infinite. Our code is available at https://github.com/XunZhiyang/llm-tv-estimation.

Related papers

• A positive resolution of the gap-entropy conjecture [33.232618834192564]
  We prove the gap-entropy conjecture for fixed-confidence best-arm identification.<n>For each suboptimal arm $i$, let $_i=_*-_i$ be its gap from the optimal mean.<n>There is an algorithm, independent of the instance, whose expected number of samples is bounded by a constant multiple of this quantity.
  arXiv  Detail & Related papers  (2026-09-09T17:57:11Z)
• Optimizing Computational-Statistical Runtime for Wasserstein Distance Estimation [8.375876290025586]
  Squaredstein distance is a frequently used tool to measure discrepancy between probability distributions.<n>We develop a Sample-$ computational paradigm where we introduce a regular cartesian grid sketch of the samples.<n>We approximate $W(P,Q)$ within $$ error in $-max(2,fracd+1+o(1)1+)$ time for $0 1$ Hlder smooth distributions.
  arXiv  Detail & Related papers  (2026-05-19T17:10:47Z)
• High-accuracy sampling for diffusion models and log-concave distributions [70.90863485771405]
  We present algorithms for diffusion model sampling which obtain $$-error in $mathrmpolylog (1/)$ steps.<n>Our approach also yields the first $mathrmpolylog (1/)$ complexity sampler for general log-concave distributions.
  arXiv  Detail & Related papers  (2026-02-01T17:05:31Z)
• Learning Multinomial Logits in $O(n \log n)$ time [56.23331174813387]
  A Multinomial Logit (MNL) model is composed of a finite universe of items $[n]=1,..., n$, each assigned a positive weight.<n>A query specifies an admissible subset -- called a slate -- and the model chooses one item from that slate with probability proportional to its weight.<n>This query model is also known as the Plackett-Luce model or conditional sampling oracle in the literature.
  arXiv  Detail & Related papers  (2026-01-07T22:07:44Z)
• Sequential 1-bit Mean Estimation with Near-Optimal Sample Complexity [32.65125292684608]
  We study the problem of distributed mean estimation with 1-bit communication constraints.<n>Our estimator is $(epsilon, delta)$-PAC for all distributions with bounded mean ($-lambda le mathbbE(X) le lambda $) and variance ($mathrmVar(X) le sigma2$) for some known parameters.
  arXiv  Detail & Related papers  (2025-09-26T06:22:57Z)
• Mallows Model with Learned Distance Metrics: Sampling and Maximum Likelihood Estimation [0.1534667887016089]
  We propose a generalization of Mallows model that learns the distance metric directly from data.<n>Specifically, we focus on $L_alpha$ distances: $d_alpha(pi,sigma):=sum_i=1 |pi(i)-sigma(i)|alpha$.<n>Using this sampling algorithm, we propose an efficient Maximum Likelihood Estimation (MLE) algorithm that jointly estimates the central ranking, the dispersion parameter, and the optimal distance metric.
  arXiv  Detail & Related papers  (2025-07-10T18:52:09Z)
• Proving the Limited Scalability of Centralized Distributed Optimization via a New Lower Bound Construction [57.93371273485736]
  We consider a centralized distributed learning setup where all workers jointly find an unbiased bound LDeltaepsilon2,$ better poly-logarithmically in $n$, even in the homogeneous (i.i.d.) case, where all workers access the same distribution.
  arXiv  Detail & Related papers  (2025-06-30T13:27:39Z)
• Testing Identity of Distributions under Kolmogorov Distance in Polylogarithmic Space [1.2277343096128712]
  In this paper, we show that much less space suffices -- we give an algorithm that uses space $O(log4 varepsilon-1)$ in the streaming setting. We also state 9 related open problems that we hope will spark interest in this and related problems.
  arXiv  Detail & Related papers  (2024-10-29T15:24:27Z)
• Relative-Translation Invariant Wasserstein Distance [82.6068808353647]
  We introduce a new family of distances, relative-translation invariant Wasserstein distances ($RW_p$) We show that $RW_p distances are also real distance metrics defined on the quotient set $mathcalP_p(mathbbRn)/sim$ invariant to distribution translations.
  arXiv  Detail & Related papers  (2024-09-04T03:41:44Z)
• Stochastic Approximation Approaches to Group Distributionally Robust Optimization and Beyond [89.72693227960274]
  This paper investigates group distributionally robust optimization (GDRO) with the goal of learning a model that performs well over $m$ different distributions. To reduce the number of samples in each round from $m$ to 1, we cast GDRO as a two-player game, where one player conducts and the other executes an online algorithm for non-oblivious multi-armed bandits. In the second scenario, we propose to optimize the average top-$k$ risk instead of the maximum risk, thereby mitigating the impact of distributions.
  arXiv  Detail & Related papers  (2023-02-18T09:24:15Z)
• Quantum tomography using state-preparation unitaries [0.22940141855172028]
  We describe algorithms to obtain an approximate classical description of a $d$-dimensional quantum state when given access to a unitary. We show that it takes $widetildeTheta(d/varepsilon)$ applications of the unitary to obtain an $varepsilon$-$ell$-approximation of the state. We give an efficient algorithm for obtaining Schatten $q$-optimal estimates of a rank-$r$ mixed state.
  arXiv  Detail & Related papers  (2022-07-18T17:56:18Z)
• A Law of Robustness beyond Isoperimetry [84.33752026418045]
  We prove a Lipschitzness lower bound $Omega(sqrtn/p)$ of robustness of interpolating neural network parameters on arbitrary distributions. We then show the potential benefit of overparametrization for smooth data when $n=mathrmpoly(d)$. We disprove the potential existence of an $O(1)$-Lipschitz robust interpolating function when $n=exp(omega(d))$.
  arXiv  Detail & Related papers  (2022-02-23T16:10:23Z)
• TURF: A Two-factor, Universal, Robust, Fast Distribution Learning Algorithm [64.13217062232874]
  One of its most powerful and successful modalities approximates every distribution to an $ell$ distance essentially at most a constant times larger than its closest $t$-piece degree-$d_$. We provide a method that estimates this number near-optimally, hence helps approach the best possible approximation.
  arXiv  Detail & Related papers  (2022-02-15T03:49:28Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.