Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Inference-Time Personalized Alignment with a Few User Preference Queries

Authors: Victor-Alexandru Pădurean, Parameswaran Kamalaruban, Nachiket Kotalwar, Alkis Gotovos, Adish Singla

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results across several tasks, involving personalized text and image generation, showcase the effectiveness of USERALIGN in achieving personalized alignment.
Researcher Affiliation Collaboration Victor-Alexandru P adurean MPI-SWS EMAIL Parameswaran Kamalaruban Featurespace Innovation Lab, Visa EMAIL Nachiket Kotalwar Carnegie Mellon University EMAIL Alkis Gotovos MPI-SWS EMAIL Adish Singla MPI-SWS EMAIL
Pseudocode Yes Algorithm 1 System-User Interaction Algorithm 2 USERALIGN: Eliciting User Preferences Algorithm 3 USERALIGN: SOLVE Subroutine
Open Source Code Yes We release our implementation and datasets to support further research.1 (Sections 5 and 6) 1Github repo: https://github.com/machine-teaching-group/neurips2025-useralign.
Open Datasets Yes We release our implementation and datasets to support further research.1 (Sections 5 and 6) Domain dsp64d. For this domain, we similarly define Θ with dimensionality based on pretrained embeddings and S = 3. First, we define X as 100 questions sampled from the Domain Specific Preference (DSP) dataset [58, 35].
Dataset Splits No The paper describes experimental setups such as simulating user preferences or assigning participants for human evaluation, but it does not specify traditional training/test/validation dataset splits for model training in the conventional sense.
Hardware Specification Yes All experiments ran on a compute node with dual AMD EPYC 7702 64-core processors (128 cores total) and 2TB DDR4 ECC memory (2933MHz).
Software Dependencies No All optimization problems are solved with the cvxpy Python package using the CLARABEL conic solver with default settings. If CLARABEL is numerically unstable, the implementation falls back to ECOS, and then to SCS with eps=1e-6 and a 50,000 iteration cap.
Experiment Setup Yes Both methods are parameterized by (ϵ, δ) which determines their stopping condition (see lines 1 and 10 in Algorithm 2). We set δ = 0.05 for all the experiments. When reporting results, we will vary the value of ϵ in [0, S] to get a trade-off between interaction cost and win-rate.