AI Paper Atlas

Research map / Vision-language and multimodal / Cross-modal learning

Multimodal Fusion via Self-Consistent Task-Gradient Fields

Jiayu Xiong, Jing Wang, Jun Xue, Wanlong Wang, Jianlong Kwan, Xiaosen Lyu, Zhouqiang Jiang

ICML 2026 · Vision-language and multimodal › Cross-modal learning · approach: Multimodal fusion with missing modalities · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add reconstruction loss on top of expansion structure (ablation)SCFAE · ActivityNet (128-128)mAP@10: 0.284 -> 0.363Improved
add expansion (n>1) instead of single linear projection (ablation)SCFAE · ActivityNet (128-128)mAP@10: 0.245 -> 0.363Improved
use n=2 expansion instead of n=1 (ablation)SCFAE · ActivityNet (128-128)mAP@10: 0.335 -> 0.363Improved
use reconstruction loss instead of contrastive loss on shared features (ablation)SCFAE · ActivityNet (128-128)mAP@10: 0.331 -> 0.363Improved
use cosine similarity instead of MSE for reconstruction (ablation)SCFAE · ActivityNet (128-128)mAP@10: 0.358 -> 0.363Improved
keep original long video features (unequal length) rather than compressing to 128SCFAE · ActivityNetmAP@10: 0.363 -> 0.367Improved
increase expansion factor from n=2 to n=4 (ablation)SCFAE · CMU-MOSEI {a,v}accuracy: 72.0 -> 72.1No clear effect
vary boundary ratio b from 0.5 to 0.25 or 0.75 (ablation)SCFAE · CMU-MOSEI {a,v}accuracy: 72.0 -> 71.2 (b=0.25 at n=2), 71.5 (b=0.75)Hurt
increase reconstruction weight lambda from 1.0 to 5.0 (ablation)SCFAE · CMU-MOSEI {a,t,v}accuracy: 87.8 -> 87.2Hurt
decrease reconstruction weight lambda from 1.0 to 0.1 (ablation)SCFAE · CMU-MOSEI {a,t,v}accuracy: 87.8 -> 86.9Hurt
permute modality order in cyclic shared-reference path (ablation)SCFAE · CMU-MOSEIaverage accuracy: 80.3 -> 80.1No clear effect
use SCFAE fusion instead of 4-layer cross-attentionjoint audio-visual training · FakeAVCeleb visual ACC delta (VideoMAE V2-S + WavLM-B)delta ACC (%): -7.54 -> -0.88Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy