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..
TreeGen: A Bayesian Generative Model for Hierarchies
Authors: Marcel Kollovieh, Nils Fleischmann, Filippo Guerranti, Bertrand Charpentier, Stephan Günnemann
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically evaluate Tree Gen on the jet-clustering task in high-energy physics, demonstrating that it consistently generates valid trees that adhere to physical constraints and closely align with ground-truth log-likelihoods. We compare Tree Gen against various baselines and test how well the generated hierarchies approximate the ground-truth posterior distributions. |
| Researcher Affiliation | Collaboration | 1 School of Computation, Information and Technology, Technical University of Munich 2 Munich Data Science Institute 3 Munich Center for Machine Learning 4 Pruna AI |
| Pseudocode | Yes | Algorithm 1 Sampling with Tree Gen input: Neural network fθ, entropy schedule αt for t [0, 1], sampling steps N output: Discrete hierarchy T1 Algorithm 2 Training Tree Gen input: Dataset D, neural network fθ, entropy schedule αt for t [0, 1], training steps N |
| Open Source Code | Yes | Project page: cs.cit.tum.de/daml/treegen Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The code is available on the project page and contains instructions to reproduce the main experimental results. |
| Open Datasets | Yes | Our evaluation uses five datasets obtained with the GINKGO jet shower generator [12]: QCD jets: QCD-S, QCD-M, QCD-L, and W-Boson jets: W-S, W-M. [12] Kyle Cranmer, Sebastian Macaluso, and Duccio Pappadopulo. Toy generative model for jets. URL https://github.com/Sebastian Macaluso/ginkgo/blob/master/notes/ toyshower_v4.pdf. |
| Dataset Splits | Yes | Each dataset consists of 100,000 hierarchies. Split into 98,000 hierarchies for training and 1,000 for validation and testing. |
| Hardware Specification | Yes | All experiments are conducted on A100 GPUs. |
| Software Dependencies | No | The paper mentions software like scikit-learn [36] and Adam [28] but does not provide specific version numbers for the software dependencies critical for reproducing the experiments. |
| Experiment Setup | Yes | We train all models using Adam [28] with a learning rate of 0.0001 and gradient clipping set to 0.5 for 50 epochs. For generation, all models use 1000 steps, except for the ablation in Fig. 7, where all use 100. We provide an overview of all hyperparameters in App. B.1. Table 3: Hyperparameters of Tree Gen. Training (LR 10^-4, Epochs 50, Grad. Clip 0.5, EMA 0.9999), Sampling (Steps 1000), Graph-Transformer (Layers 4, Dim (h) 64, Heads 2), Upscaler (Node Layers 1, Edge Layers 3), Prediction Head (Layers 2). |