New methods correct bias in LLM-as-a-Judge evaluations, but reliability depends on judge quality and model calibration.
problem Systematic bias in LLM-as-a-Judge evaluations using naive estimators.
method Analytical results, simulations, and real-data case study to diagnose reliability of corrected estimates.
result Corrected estimates, especially shared-calibration comparisons, can be unreliable under certain conditions.
VLM judges rank well but score poorly; task difficulty and annotation quality affect interval width.
problem VLMs as judges lack reliability indicators in multimodal evaluations.
method Conformal prediction using score-token log-probabilities.
result Evaluation uncertainty is task-dependent, affecting interval width and reliability.
A new framework evaluates LLMs by considering judge reliability.
problem Evaluating LLMs without ground truth labels can lead to biased results.
method Introduces judge-specific discrimination parameters and estimates model quality and judge reliability.
result Improves agreement with human preferences and produces calibrated uncertainty quantification.
CARE improves LLM aggregation by accounting for shared confounders.
problem LLM judges' correlated errors due to shared latent confounders.
method CARE explicitly models judges' scores as true quality and confounders.
result CARE reduces aggregation error by up to 26.8% across various benchmarks.
Researchers use LLMs to judge other LLMs, but this study provides a new geometric perspective to understand when it works.
problem The challenge of evaluating LLMs using other LLMs as judges, considering both aleatoric and epistemic uncertainties.
method A geometric perspective on ranking LLM candidates using probability simplices, analyzing conditions for identifiable rankings and designing Bayesian priors.
result Geometric analysis reveals that rankings based on LLM judges are robust in many but not all datasets, emphasizing the importance of modeling epistemic uncertainty.
LLM evaluation suffers from systematic biases and lacks reliable positive judgments.
problem LLM evaluation suffers from systematic biases and lacks reliable positive judgments.
method Formulate LLM evaluation as a positive-unlabelled learning problem and propose a geometric auditing framework based on Partial Optimal Transport.
result Improved alignment with human preferences, increased robustness to presentation biases, and interpretable confidence estimates.
New models can't beat existing ones, so debiasing methods only slightly reduce needed labels.
problem Limiting scalability in model evaluation due to self-preferencing biases.
method Study of debiasing methods using a few high-quality labels to reduce model judgments.
result Debiasing methods can't decrease required ground truth labels by more than half when the judge is no more accurate than the model.
MultiwayPAM clusters LLM-as-a-Judge scores to reveal evaluator bias.
problem Revealing bias in LLM evaluations of text quality.
method Tensor clustering with MultiwayPAM method.
result Medoids reveal cluster membership of questions, answerers, and evaluators.
Paper improves uncertainty estimation in LLM-as-a-judge systems.
problem Improving uncertainty estimation in LLM-as-a-judge frameworks.
method Generalised probabilistic modelling and improved uncertainty estimates.
result Proposed uncertainty estimates significantly improve system efficiency.
ValueBlindBench tests LLM-generated investment rationales for validity before returns are known.
problem Delayed-ground-truth evaluation of LLM-generated investment rationales.
method Agreement-gated stress testing protocol to validate LLM-judged rationales.
result ValueBlindBench prevents overclaims and identifies flawed financial constructs.
Unified model combines scores and rankings for grant panel review.
problem Combining scores and rankings for quality assessment in panel review.
method Mallows-Binomial model with tree-search algorithm for exact MLE.
result Model combines scores and rankings to quantify object quality and measure consensus.
Traditionally, the vision community has devised algorithms to estimate the distance between an original image and images that have been subject to perturbations. Inspiration was usually taken from the human visual perceptual system and how the system processes different perturbations in order to replicate to what exten…
This paper proposes a method to use LLMs as auxiliary evaluators in place of human judges.
problem The need for cost-effective and scalable evaluation of AI systems.
method Formulates a two-stage sampling design with LLM evaluations and human ratings, using a doubly robust estimator.
result Proposes a method to determine optimal sample sizes for human and LLM ratings.
PDO optimizes LLM prompts without labels, improving performance.
problem Optimizing prompts for LLMs without access to labeled data.
method Pairwise preference feedback, dueling bandits, Thompson Sampling, mutation.
result PDO identifies stronger prompts than label-free methods.
No-knowledge alarms detect misaligned LLM judges without trusting them.
problem Monitoring LLM judges for consistency when ground truth is unknown.
method Formal logic and Linear Programming to detect disagreements.
result No false positives in detecting misaligned judges.
AURA: Adaptive Uncertainty-aware Refinement for LLM-as-a-Judge Auditing
problem Auditing LLM-as-a-Judge decisions
method Adaptive uncertainty-aware refinement
result Human-consistent signal learning and evidence propagation
Feature Learning aims to extract relevant information contained in data sets in an automated fashion. It is driving force behind the current deep learning trend, a set of methods that have had widespread empirical success. What is lacking is a theoretical understanding of different feature learning schemes. This work p…
Algorithm reduces audit costs by identifying best service configurations from biased textual evidence.
problem Designing service systems from textual evidence requires accurate selection despite biased automated scoring.
method Developed PP-LUCB algorithm combining LLM scores and selective audits to minimize costs.
result Correctly identified the best model in 40/40 trials with 90% cost reduction.
New method improves speech synthesis quality.
problem Efficiently train parallel speech synthesis models.
method Spectral energy distance for implicit generative models.
result State-of-the-art generation quality achieved.
RACER optimizes LLM-as-judge accuracy with dynamic reasoning selection.
problem Balancing reasoning accuracy with computational cost in LLM-as-judge settings.
method Formulates routing as a constrained distributionally robust optimization problem, accounting for distribution shift via KL-divergence uncertainty set.
result RACER achieves superior accuracy-cost trade-offs under distribution shift.
Optimizes budgeted evaluations of LLMs by allocating queries to judges efficiently.
problem Evaluating LLMs with heterogeneous judges and varying costs and reliability.
method Formalizes and analyzes budgeted heteroskedastic multi-judge estimation, proposing EST-IVWE for practical implementation.
result EST-IVWE matches the oracle IVWE rate up to lower-order terms in the budget and is instance-optimal.
We conduct a large-scale, systematic study to evaluate the existing evaluation methods for natural language generation in the context of generating online product reviews. We compare human-based evaluators with a variety of automated evaluation procedures, including discriminative evaluators that measure how well machi…
Normalizing flows improve ptychography reconstruction quality and uncertainty quantification.
problem Challenges in ptychography due to large-scale nonlinear and non-convex inverse problems and photon statistics.
method Use of normalizing flows to model the posterior distribution and quantify reconstruction uncertainty.
result Normalizing flows enable better characterization and uncertainty quantification in ptychography reconstructions.
Discovering novel materials can be greatly accelerated by iterative machine learning-informed proposal of candidates---active learning. However, standard \emph{global-scope error} metrics for model quality are not predictive of discovery performance, and can be misleading. We introduce the notion of \emph{Pareto shell-…
New method improves LLM judge accuracy by accounting for dependencies in aggregated binary labels.
problem Classical label aggregation methods fail to account for dependencies among LLM judges, leading to miscalibrated predictions.
method Dependence-aware models based on Ising graphical models and latent factors.
result The proposed method outperforms classical methods on real-world datasets, reducing excess risk.
Study identifies and measures biases in legal case data.
problem Addressing representation biases and sentencing disparities in legal case data.
method Utilizes two regression models: a baseline and a fair judge model.
result Quantifies biases across demographic groups in criminal data from Cook County (Illinois).
CJE calibrates cheap LLM judges against an oracle, achieving high accuracy at a fraction of the cost.
problem Inexpensive LLM judges can produce biased rankings, leading to unreliable outcomes.
method CJE uses a small oracle to calibrate cheap scores, then evaluates at scale with valid uncertainty.
result CJE achieves 99% pairwise ranking accuracy at 14x lower cost compared to a 16x oracle/judge cost ratio.
Experimental evaluation is a major research methodology for investigating clustering algorithms and many other machine learning algorithms. For this purpose, a number of benchmark datasets have been widely used in the literature and their quality plays a key role on the value of the research work. However, in most of t…
Corrects bias in LLM-as-a-judge evaluations using adaptive calibration.
problem Bias in LLM evaluations due to imperfect sensitivity and specificity.
method Plug-in framework with confidence intervals accounting for test and calibration dataset uncertainties.
result LML-based evaluation yields more reliable estimates than human-only evaluation.
Unified framework to bridge human and LLM judgments.
problem Systematic discrepancies between human and LLM evaluations.
method Latent human preference score and linear transformations of covariates.
result Higher agreement with human ratings and exposure of systematic gaps.
To make AI systems broadly useful for challenging real-world tasks, we need them to learn complex human goals and preferences. One approach to specifying complex goals asks humans to judge during training which agent behaviors are safe and useful, but this approach can fail if the task is too complicated for a human to…
Meta-Router optimizes LLM selection using gold-standard and preference-based data.
problem Training a high-quality LLM router with combined data sources is challenging due to bias and scarcity.
method Developed an integrative causal router training framework to correct bias and improve routing accuracy.
result Our approach delivers more accurate routing and improves the trade-off between cost and quality.
FUSE improves verification quality without ground truth labels.
problem Verification of model outputs using imperfect judges and reward models.
method Ensembling verifiers without ground truth labels using spectral algorithms.
result FUSE matches or improves upon semi-supervised alternatives in test-time scaling experiments.
Using the Minority Game model we study a broad spectrum of problems of market mechanism. We study the role of different types of agents: producers, speculators as well as noise traders. The central issue here is the information flow : producers feed in the information whereas speculators make it away. How well each age…
We study a bad arm existing checking problem in which a player's task is to judge whether a positive arm exists or not among given K arms by drawing as small number of arms as possible. Here, an arm is positive if its expected loss suffered by drawing the arm is at least a given threshold. This problem is a formalizati…
The paper deals with the adaptation of a new measure for the unsupervised feature selection problems. The proposed measure is based on space filling concept and is called the coverage measure. This measure was used for judging the quality of an experimental space filling design. In the present work, the coverage measur…
This study examines how the size and alignment of pretraining data affect the performance of large language models on downstream tasks.
problem Understanding how the size and alignment of pretraining data impact the performance of large language models on downstream tasks.
method Investigated the scaling behavior of large language models in a transfer learning setting, focusing on machine translation tasks.
result The size of the finetuning dataset and the distribution alignment between pretraining and downstream data significantly influence the scaling behavior of downstream performance.
This paper revisits the problem of analyzing multiple ratings given by different judges. Different from previous work that focuses on distilling the true labels from noisy crowdsourcing ratings, we emphasize gaining diagnostic insights into our in-house well-trained judges. We generalize the well-known DawidSkene model…
When dealing with subjective, noisy, or otherwise nebulous features, the "wisdom of crowds" suggests that one may benefit from multiple judgments of the same feature on the same object. We give theoretically-motivated `feature multi-selection' algorithms that choose, among a large set of candidate features, not only wh…
Gen AI improves document understanding but not data analysis in public sector tasks.
problem Understanding the impact of Gen AI on public sector tasks.
method Pre-registered field experiment comparing Gen AI to control group performance.
result Mixed results: Gen AI improves document understanding but not data analysis.
Decisions by Machine Learning (ML) models have become ubiquitous. Trusting these decisions requires understanding how algorithms take them. Hence interpretability methods for ML are an active focus of research. A central problem in this context is that both the quality of interpretability methods as well as trust in ML…
In this paper, we study the problem of safe online learning to re-rank, where user feedback is used to improve the quality of displayed lists. Learning to rank has traditionally been studied in two settings. In the offline setting, rankers are typically learned from relevance labels created by judges. This approach has…
New metric evaluates generative models across domains, diagnosing fidelity, diversity, and generalization.
problem Evaluating generative models in diverse domains with limited metrics.
method Introduces a 3D evaluation metric (α-Precision, β-Recall, Authenticity) for domain-agnostic diagnostics. result Unified metric characterizes fidelity, diversity, and generalization, diagnosing model performance.
FinReflectKG builds a comprehensive financial knowledge graph from SEC filings, improving extraction quality.
problem Lack of large-scale, open-source financial knowledge graph datasets.
method Intelligent document parsing, table-aware chunking, schema-guided iterative extraction, reflection-driven feedback loop.
result Reflection-agent-based mode achieves best balance of efficiency, accuracy, and reliability.
Detects unusual inputs to neural networks to prevent flawed predictions.
problem Erratic predictions from neural networks on unexpected inputs.
method Evaluates input unusualness by comparing its content to learned parameters.
result Simple, effective method for comparing input metrics across different scales.
From doctors diagnosing patients to judges setting bail, experts often base their decisions on experience and intuition rather than on statistical models. While understandable, relying on intuition over models has often been found to result in inferior outcomes. Here we present a new method, select-regress-and-round, f…
SAGE improves memory efficiency by selectively adding, merging, or ignoring new facts.
problem Efficiently managing new facts in agentic LLMs to avoid costly write-time reasoning.
method SAGE uses a von Mises-Fisher-based density estimator to score and route candidate facts.
result SAGE achieves the best average token-F1 on LoCoMo and reduces add-phase API cost by 3.4x on GPT-4o-mini.
We give several new criteria to judge whether a simple convex polytope in a Euclidean space is combinatorially equivalent to a product of simplices. These criteria are mixtures of combinatorial, geometrical and topological conditions that are inspired by the ideas from toric topology.