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..

Foundations of Top-$k$ Decoding for Language Models

Authors: Georgy Noarov, Soham Mallick, Tao Wang, Sunay Joshi, Yan Sun, Yangxinyu Xie, Mengxin Yu, Edgar Dobriban

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, in Section 5, we study the empirical performance of some of the novel decoding schemes on open-ended text generation and mathematical problem solving tasks with LLMs, and find that they perform competitively with top-k decoding. ... We conduct experiments using the GPT-2 Large [43] and Llama 3.1 8B [25] models. We evaluate on two benchmarks: (1) open-ended text generation using the Web Text test set from the GPT-2 output dataset [40], and (2) grade school math reasoning using the GSM8K Chainof-Thought benchmark [13].
Researcher Affiliation Academia 1 University of Pennsylvania 2 New Jersey Institute of Technology 3 Washington University in St. Louis
Pseudocode Yes F.2 Pseudocode for algorithms See Algorithm 3 and Algorithm 4 for pseudocode for sparse primal (resp. dual) Bregman decoding.
Open Source Code Yes Code is included in the supplementary materials, and an open-source Git Hub repository will be released upon publication.
Open Datasets Yes We evaluate on two benchmarks: (1) open-ended text generation using the Web Text test set from the GPT-2 output dataset [40], and (2) grade school math reasoning using the GSM8K Chain-of-Thought benchmark [13]. ... The GPT-2 output dataset and GSM8K dataset used are publicly available.
Dataset Splits Yes We evaluate on two benchmarks: (1) open-ended text generation using the Web Text test set from the GPT-2 output dataset [40], and (2) grade school math reasoning using the GSM8K Chain-of-Thought benchmark [13]. ... For open-ended text generation, following Chen et al. [12], we use the first 35 tokens of each Web Text test sample as a prompt and generate up to 256 tokens. ... Using the full decoding strategy, we evaluate the LLa MA 3.1 8B model using 8-shot Co T prompting.
Hardware Specification Yes The experiments were conducted on a system running Rocky Linux 8.10, with 64 CPU cores of Intel(R) Xeon(R) Gold 6448Y processors at 2.10 GHz, 1 TB of RAM, and 8 NVIDIA L40S GPUs with 46 GB of memory each.
Software Dependencies Yes The software environment used Python 3.11.11, Py Torch 2.5.1, and CUDA 12.4.
Experiment Setup Yes We test various temperatures, regularization strengths λ {0.01, 0.0001} and primal decoding parameters α {1.5, 2.0}. ... Using the full decoding strategy, we evaluate the LLa MA 3.1 8B model using 8-shot Co T prompting.