Research map / Vision-language and multimodal / Multimodal LLMs
OrderChain: Towards General Instruct-Tuning for Stimulating the Ordinal Understanding Ability of MLLM
Jinhong Wang, Shuo Tong, Jian Liu, Dongqi Tang, Weiqiang Wang, Wentong Li, Hongxia Xu, Danny Z. Chen, Jintai Chen, Jian Wu
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 |
|---|---|---|---|
| add Domain Knowledge Prompts (ablation) | LLaVA baseline · Adience | accuracy: 47.5 -> 58.0 | Improved |
| add Category Feature Prompts alone (ablation) | LLaVA baseline · Adience | accuracy: 47.5 -> 32.5 | Hurt |
| add Category Feature Prompts on top of Domain Knowledge Prompts, no RO-CoT (ablation) | LLaVA + DKP · Adience | accuracy: 58.0 -> 38.7 | Hurt |
| add RO-CoT to Category Feature Prompts (ablation) | LLaVA + CFP · Adience | accuracy: 32.5 -> 84.6 | Improved |
| add Domain Knowledge Prompts to Category Feature Prompts + RO-CoT (full OrderChain) (ablation) | LLaVA + CFP + RO-CoT · Adience | accuracy: 84.6 -> 93.2 | Improved |
| add Domain Knowledge Prompts to Category Feature Prompts + RO-CoT (full OrderChain) (ablation) | LLaVA + CFP + RO-CoT · Adience | MAE: 0.18 -> 0.12 | Improved |
| add OrderChain to LLaVA-1.5 | LLaVA-1.5 baseline · HCI | accuracy: 61.4 -> 73.0 | Improved |
| add OrderChain to LLaVA-1.5 | LLaVA-1.5 baseline · DR | accuracy: 30.0 -> 85.7 | Improved |
| add OrderChain to LLaVA-1.5 | LLaVA-1.5 baseline · Image Aesthetics (overall) | accuracy: 59.56 -> 73.83 | Improved |
| compare OrderChain with NumCLIP on subjective aesthetics labels | LLaVA-1.5 + OrderChain · Image Aesthetics (overall) | accuracy: 76.53 -> 73.83 | Hurt |
| use OrderChain-trained LLaVA for zero-shot transfer to APTOS | LLaVA baseline · APTOS (zero-shot transfer) | accuracy: 23.2 -> 70.0 | Improved |
| use OrderChain-trained LLaVA for zero-shot transfer to KonIQ | LLaVA baseline · KonIQ (zero-shot transfer) | accuracy: 36.3 -> 51.5 | Improved |
Closest papers by meaning
- Scaling Prompt Synthesis for Large Language Model Reasoning (ICML 2026 · similarity 0.86)
- Learning Hierarchical Prompt with Structured Linguistic Knowledge for Vision-Language Models (AAAI 2024 · similarity 0.86)
- MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark (ICML 2024 · similarity 0.86)
- LLaVA-MoD: Making LLaVA Tiny via MoE-Knowledge Distillation (ICLR 2025 · similarity 0.84)
- HelpSteer3-Preference: Open Human-Annotated Preference Data across Diverse Tasks and Languages (NeurIPS 2025 · similarity 0.84)
- StrategyLLM: Large Language Models as Strategy Generators, Executors, Optimizers, and Evaluators for Problem Solving (NeurIPS 2024 · similarity 0.84)
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.
