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..
Learn and Ensemble Bridge Adapters for Multi-domain Task Incremental Learning
Authors: Ziqi Gu, Chunyan Xu, Wenxuan Fang, Xin Liu, Yide Qiu, Zhen Cui
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across multiple benchmarks validate the effectiveness and superiority of LEBA over state-of-the-art methods. |
| Researcher Affiliation | Collaboration | Ziqi Gu1, Chunyan Xu1, , Wenxuan Fang1, Xin Liu2, Yide Qiu1, Zhen Cui3, 1School of Computer Science and Engineering, Nanjing University of Science and Technology 2Nanjing Seetacloud Technology 3School of Artificial Intelligence, Beijing Normal University |
| Pseudocode | Yes | Algorithm 1 LEBA Training Procedure Input: Task sequence St = {(Dt, Ct) | t = 1, . . . , T}; diffusion-based generator G; continuous-domain bridge adapter Γ; adaptive weight ω; incremental model Θt=1 Output: Incremental model ΘT , adaptive weights ω, and continuous-domain bridge adapter Γ 1: Initialize model Θt=1, generator G, weight ω, and adapter Γ 2: for t = 1 to T do 3: # Supervised update with current task data 4: Train Θt=1 on St=1 = (Dt=1, Ct=1) 5: if t > 1 then 6: # Progressive Knowledge Ensemble 7: Construct adapter pool PK from previous adapters {Θl}j K l=j1 with jk {1, , t 1} 8: Generate replay samples bxt i from generator G conditioned on semantic concepts {Cj}t 1 j=1 9: Compute adaptive weights ω for each replay sample bxt i 10: Evaluate similarity η of current adapter Θt and update adapter pool PK 11: # Continuous-Domain Bridge Adaptation 12: Construct continuous-domain bridge adapter Γ with replay data bxt i and adaptive weight ω 13: # Joint Optimization 14: Update Θt, ω, and Γ by minimizing the total loss ζtotal (Eqn. 12) 15: end if 16: # Update the incremental model for the next domain 17: Θt+1 Θt 18: end for |
| Open Source Code | No | Answer: [No] Justification: No codes are included in this submission, but the codes will be provide when the paper is accected. |
| Open Datasets | Yes | Datasets: We evaluate our LEBA in the multi-domain task incremental learning(MTIL) [16]. In this configuration, tasks are sourced from multiple domains, each necessitating unique domain knowledge to achieve high accuracy. The MTIL benchmark comprises 11 tasks and contains a total of 1,201 classes. We evaluate the method using two different task orders: the first follows an alphabetical order (Order-I): Aircraft [41], Caltech101 [42], CIFAR100 [43], DTD [44], Euro SAT [45], Flowers [46], Food [47], MNIST [48], Oxford Pet [49], Stanford Cars [50], and SUN397 [51]. |
| Dataset Splits | No | The paper states: "To evaluate LEBA in the multi-task incremental learning (MTIL) setting, we follow the protocol introduced in ZSCL [16]". While this implies using splits from the referenced protocol, the paper itself does not explicitly provide specific dataset split percentages, sample counts, or direct citations for the splits within its text. |
| Hardware Specification | Yes | All experiments are conducted using Py Torch on NVIDIA Ge Force RTX 4090 GPUs. |
| Software Dependencies | Yes | Following previous work [16], we adopt CLIP with Vi T-B/16 [52] as the backbone for all experiments. Each task s adapter is composed using Lo RA [53]. For generative replay, we employ the Stable Diffusion-V1.4 model [54]... |
| Experiment Setup | Yes | The continuous-domain bridge adapter Γ is implemented as a four-layer MLP. We set the balancing factors γ = 0.1 and β = 0.4, and use a step size of m = 20 and an adapter selection threshold of η = 0.3, and an adapter pool containing K = 2 adapters. Optimization is performed using the Adam W optimizer [55], with label smoothing [56] applied to improve baseline performance. For the MTIL benchmark, we use a batch size of 64 and search the learning rate α within {1 10 3, . . . , 1 10 5}. |