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..

Tractable Multinomial Logit Contextual Bandits with Non-Linear Utilities

Authors: Taehyun Hwang, Dahngoon Kim, Min-hwan Oh

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive numerical experiments validate the effectiveness of our approach, showing robust performance not only in realizable settings but also in scenarios with model misspecification.
Researcher Affiliation Academia Taehyun Hwang Seoul National University EMAIL Dahngoon Kim Seoul National University EMAIL Min-hwan Oh Seoul National University EMAIL
Pseudocode Yes Algorithm 1 ONL-MNL (Optimistic Non-Linear Utility for Contextual MNL Bandit)
Open Source Code Yes We have attached the data and code with sufficient instructions to reproduce the main experimental results in the supplementary material.
Open Datasets Yes Semi-synthetic experiment with real-world dataset. In this section, we provide an additional semi-synthetic experiment leveraging a real-world dataset. We used the IMDB Large Movie Review dataset [26]
Dataset Splits Yes We used 40,000 of these samples as training data to fit a binary classification model.
Hardware Specification Yes All experiments are run on a computing cluster with Intel Xeon Gold 6526R (16-core, 2.8 GHz, 37.5 MB cache, 3 UPI, 195 W).
Software Dependencies No We approximate the offline regression oracle using the Adam optimizer with a learning rate of 10 4 over 2000 iterations. The ε parameter starts at 0.1 and decays multiplicatively by a factor of 0.995 after each step, with a minimum threshold of 0.001. Implementation of ONL-MNL. After Phase I, the pilot estimator ˆw0 is approximated by minimizing the negative log-likelihood using the Adam optimizer with a learning rate of 10 4 for 2000 iterations. The regularization parameter λ is set as λ = cλ κ 5/2dw T, and the confidence radius βt is set as βt = cβ κ 4dw t T . Both scaling constants, cλ and cβ, are selected via grid search.
Experiment Setup Yes We approximate the offline regression oracle using the Adam optimizer with a learning rate of 10 4 over 2000 iterations. The ε parameter starts at 0.1 and decays multiplicatively by a factor of 0.995 after each step, with a minimum threshold of 0.001. After Phase I, the pilot estimator ˆw0 is approximated by minimizing the negative log-likelihood using the Adam optimizer with a learning rate of 10 4 for 2000 iterations. The regularization parameter λ is set as λ = cλ κ 5/2dw T, and the confidence radius βt is set as βt = cβ κ 4dw t T . Both scaling constants, cλ and cβ, are selected via grid search.