Research map / Vision-language and multimodal / Human and embodied modeling
Uni-X: Mitigating Modality Conflict with a Two-End-Separated Architecture for Unified Multimodal Models
Jitai Hao, Hao Liu, Xinyan Xiao, Qiang Huang, Jun Yu
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 changed | Where | Measured | Effect |
|---|---|---|---|
| separate shallow and deep layers by modality (Uni-X 9:5) vs fully shared transformer | Shared Transformer · identical training (28B tokens), Qwen2.5-1.5B | average of MMLU/GenEval/MMB: 38.0 -> 41.6 | Improved |
| Uni-X vs hard-routed vision MoE | HardMoE · identical training (28B tokens), Qwen2.5-1.5B | average: 41.3 -> 41.6 | Improved |
| Uni-X vs MoT and UniFork baselines | MoT, UniFork · identical training (28B tokens), Qwen2.5-1.5B | average: MoT 34.6, UniFork 29.5 -> 41.6 | Improved |
| add loss masking on instruction tokens in pretraining (ablation) | Uni-X (9:5) · identical training (28B tokens), Qwen2.5-1.5B | GenEval: 34.8 -> 43.3 | Improved |
| increase total separated layers from 6 (3:3) to 14 (7:7) (ablation) | Uni-X · Qwen2.5-1.5B | average: 38.9 -> 40.1 | Improved |
| increase separated layers to 22 (11:11) (ablation) | Uni-X · Qwen2.5-1.5B | average: 40.1 -> 39.8 | Hurt |
| use more shallow than deep separated layers (9:5 vs 5:9) (ablation) | Uni-X · Qwen2.5-1.5B | average: 39.1 -> 41.6 | Improved |
| use 11:3 split (too few deep separated layers) (ablation) | Uni-X · Qwen2.5-1.5B | GenEval: 43.3 -> 25.1 | Hurt |
| increase vision-specific separated layers (14:8 to 14:20) (ablation) | Uni-X · Qwen2.5-1.5B | average: 37.4 -> 41.2 | Improved |
| scale Uni-X from 140B to 240B training tokens | Uni-X · Qwen2.5-3B | GenEval: 82 -> 83 | Improved |
| use Uni-X architecture vs shared transformer at inference | Uni-X · H800 inference | tokens/s: 975.2 -> 910.2 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Liquid (7B) — Liquid: Language Models are Scalable and Unified Multi-Modal Generators (IJCV 2026) · better on 2 of 2 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
- UNIALIGN: Scaling Multimodal Alignment within One Unified Model (CVPR 2025 · similarity 0.91)
- UniM: A Unified Any-to-Any Interleaved Multimodal Benchmark (CVPR 2026 · similarity 0.89)
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.88)
- MANZANO: A Simple and Scalable Unified Multimodal Model with a Hybrid Vision Tokenizer (ICLR 2026 · similarity 0.88)
- TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs (NeurIPS 2026 · similarity 0.88)
- UniGen: Enhanced Training & Test-Time Strategies for Unified Multimodal Understanding and Generation (NeurIPS 2025 · similarity 0.88)
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.
