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..
Resounding Acoustic Fields with Reciprocity
Authors: Zitong Lan, Yiduo Hao, Mingmin Zhao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We comprehensively evaluate Versa on the resounding task using both simulated and real-world datasets. Versa-ELE is model-agnostic and can be applied to existing neural acoustic field models, yielding average improvements of 34% on C50 and 31% on STFT. On top of the AVR [34], Versa-SSL further improves performance by 24% on C50 and 48% on STFT, demonstrating its effectiveness in scenarios with asymmetric gain patterns. |
| Researcher Affiliation | Academia | Zitong Lan University of Pennsylvania EMAIL Yiduo Hao University of Pennsylvania EMAIL Mingmin Zhao University of Pennsylvania EMAIL |
| Pseudocode | No | The paper describes the methods narratively and illustrates concepts with figures, but does not contain a specific section or block labeled "Pseudocode" or "Algorithm". |
| Open Source Code | No | Code, dataset and demo videos are available on the project website. 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: [No] Justification: Our code and dataset will be open-sourced after the review process. |
| Open Datasets | Yes | We use both simulated and real-world datasets to comprehensively evaluate our methods. We use Mesh RIR [29] and RAF [18] as our real datasets, where Mesh RIR shows similar emitter/listener gain patterns, and RAF features different emitter/listener patterns. We use the Acousti X [34] simulator to create synthetic datasets because it provides flexibility to customize emitter and listener gain patterns in the simulation. |
| Dataset Splits | Yes | Acousti X. For each scene, there are five emitter positions in the training set, and each with around 5K emitter pairs, resulting 25K training samples. We randomly sample 10 novel emitter locations to construct the testing set, resulting in a total of 15K samples. Mesh RIR. We subdivide the S32-441 variant to use 7 emitter positions (3K emitter/listener pairs) as training samples, while reserving 25 emitter positions (11K total emitter/listener pairs) for testing. RAF. We randomly select 4 emitter positions in both Furnished Room and Empty Room split, resulting in a total of 1.5K training samples, and we resample 6K testing samples in each setting. |
| Hardware Specification | Yes | We train all the models on NVIDIA L40s GPUs for 200 epochs. |
| Software Dependencies | No | For NAF, INRAS, AV-Ne RF, and AVR experiments, we use the Adam W optimizer [28] with a cosine learning rate scheduler that starts from 10 3 and decays to 10 4, with a batch size of 64 for NAF, INRAS, and AV-Ne RF, and a batch size of 4 for AVR. To implement Versa-SSL method on AVR, we use a fourth-order spherical harmonics [27] to estimate the emitter gain pattern. For AV-Ne RF, we use nerfacto [55] to render RGB and depth images following [37, 18]. The paper mentions software components like "Adam W optimizer" and "nerfacto", but does not provide specific version numbers for these or other libraries/frameworks. |
| Experiment Setup | Yes | For NAF, INRAS, AV-Ne RF, and AVR experiments, we use the Adam W optimizer [28] with a cosine learning rate scheduler that starts from 10 3 and decays to 10 4, with a batch size of 64 for NAF, INRAS, and AV-Ne RF, and a batch size of 4 for AVR. To implement Versa-SSL method on AVR, we use a fourth-order spherical harmonics [27] to estimate the emitter gain pattern. We initiate the second stage of Versa-SSL on AVR halfway through the total training epochs, gradually adding position variation to a standard deviation of 0.3 m. We use a loss weight ΜΈ=0.8. In the Diff RIR experiments, we use an Adam W optimizer with a learning rate 10 3 and a batch size of 4. We train all the models on NVIDIA L40s GPUs for 200 epochs. |