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..
On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation
Authors: Liyao Tang, Zhe Chen, Dacheng Tao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that GEM achieves performance comparable to or sometimes even exceeding full fine-tuning, while only updating 1.6% of the model s parameters, fewer than other PEFT methods. Empirically, we validate our approach on large-scale 3D scene datasets, including both indoor [12, 61, 2, 84] and outdoor [5] scenes. Our results indicate that models equipped with GEM consistently achieve performance matching or sometimes surpassing full fine-tuning methods... |
| Researcher Affiliation | Academia | Liyao Tang The University of Sydney EMAIL Zhe Chen1 La Trobe University EMAIL Dacheng Tao2 Nanyang Technological University EMAIL |
| Pseudocode | No | The paper describes the methodology using equations and textual descriptions, but it does not include explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps in a code-like format. |
| Open Source Code | Yes | Code is available at https://github.com/Liyao Tang/GEM. |
| Open Datasets | Yes | Empirically, we validate our approach on large-scale 3D scene datasets, including both indoor [12, 61, 2, 84] and outdoor [5] scenes. |
| Dataset Splits | Yes | We primarily follow the comprehensive protocols proposed in Sonata [76], covering Scan Net [12], Scan Net200 [61], Scan Net++ [84] and S3DIS [2]. |
| Hardware Specification | Yes | Leveraging the parameter efficiency of our method, we train on a single 4090 GPU for much fewer epochs to obtain the reported performance. |
| Software Dependencies | No | Our implementation is based on the open-source codebase Pointcept (here) and follows the official implementations for Sonata [76], PPT [79], as well as PTv3 [77]. |
| Experiment Setup | Yes | For training, we follow the widely accepted fine-tuning setups to update only the inserted or selected weights with the pre-trained backbone weights remaining frozen. All PEFT baselines follow the implementations from released code, adopt the suggested common practice [47, 52], and are tuned to their best validation setting in Fig. 1(c). Specifically, we set the default rank to be r = 32 and the number of learnable tokens to be m = 4. More details are given in the supplementary. |