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..

Multi-scale Temporal Prediction via Incremental Generation and Multi-agent Collaboration

Authors: Zhitao Zeng, Guojian Yuan, Junyuan Mao, Yuxuan Wang, Xiaoshuang Jia, Yueming Jin

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on the MSTP Benchmark in general and surgical scene show that IG-MC is a generalizable plug-and-play method for MSTP, demonstrating the effectiveness of incremental generation and the stability of decision-driven multi-agent collaboration.
Researcher Affiliation Collaboration Zhitao Zeng1 Guojian Yuan1 Junyuan Mao1 Yuxuan Wang2 Xiaoshuang Jia3 Yueming Jin1 1National University of Singapore 2Alibaba Group 3Renmin University of China
Pseudocode No The paper describes the methodology in Section 3, but there are no explicitly labeled "Pseudocode" or "Algorithm" blocks.
Open Source Code Yes The code, model weights, benchmark can be found in https://github.com/jinlab-imvr/MSTP.
Open Datasets Yes Our MSTP Benchmark in general scene (MSTP-General) is built on top of the Action Genome (AG) dataset [76]... Our MSTP Benchmark in Surgery (MSTP-Surgery) is built on top of the Gra SP dataset [77]...
Dataset Splits Yes From Gra SP s 32 h / 13video source, we generate scale-aware future-prediction clips at four horizons (1 s, 5 s, 30 s, 60 s) and split them 10:1 into train/test, yielding 40k training and 4k test clips in total. At each scale there are 10k training and 1k test clips (Table 3).
Hardware Specification Yes We profile computational efficiency and latency on a single NVIDIA H200 GPU... Our Generation Agent... fine-tuned on 4 NVIDIA H100 GPUs... Decision-making Agent... finetuned on our MSTP decision dataset using 4 NVIDIA H100 GPUs...
Software Dependencies Yes Our Generation Agent is implemented via the Stable Diffusion 3.5 Large architecture [79]...Decision-making Agent leverages VLMs including Qwen2.5-VL-7B-Instruct [80]...
Experiment Setup Yes Our Generation Agent... trained for one epoch with a per-GPU batch size of 34... Adam W with bf16 precision, a cosine learning-rate schedule and a 10% warm-up phase, weight decay set to 0... At inference, we used 30 denoising steps, a CFG scale of 7.5... Decision-making Agent... employed a per-device batch size of 32, no gradient accumulation, and the Adam W optimizer (initial learning rate 2 10 5, cosine decay, 10% warm-up, weight decay 1 10 2)... At inference, greedy decoding was applied (maximum length 128 tokens) with top-p sampling (p = 0.9).