Research map / Vision-language and multimodal / Human and embodied modeling
Hyper-ICL: Attention Calibration with Hyperbolic Anchor Distillation for Multimodal In-Context Learning
Niloufar Alipour Talemi, Hossein Kashiani, Fatemeh Afghah
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 |
|---|---|---|---|
| replace Euclidean anchor with hyperbolic anchor distillation (ablation) | Hyper-ICL (Idefics-9B) · VQAv2 | accuracy: 59.64 -> 62.08 | Improved |
| replace Euclidean anchor with hyperbolic anchor distillation (ablation) | Hyper-ICL (Idefics-9B) · OK-VQA | accuracy: 53.76 -> 55.31 | Improved |
| add anchor distillation to supervised loss (ablation) | Hyper-ICL (Idefics-9B) · VQAv2 | accuracy: 57.11 -> 62.08 | Improved |
| add supervised loss to hyperbolic anchor loss (ablation) | Hyper-ICL (Idefics-9B) · OK-VQA | accuracy: 53.78 -> 55.31 | Improved |
| apply module at attention logits instead of post-attention output (ablation) | Hyper-ICL (Idefics-9B) · VQAv2 | accuracy: 58.42 -> 62.08 | Improved |
| apply module at attention logits instead of post-attention output (ablation) | Hyper-ICL (Idefics-9B) · OK-VQA | accuracy: 50.18 -> 55.31 | Improved |
| increase demonstrations from 32 to 64 shots (ablation) | Hyper-ICL (Idefics-9B) · OK-VQA | accuracy: 55.31 -> 55.26 | No clear effect |
| increase demonstrations from 4 to 32 shots (ablation) | Hyper-ICL (Idefics-9B) · VQAv2 | accuracy: 57.87 -> 62.08 | Improved |
| increase adapter rank beyond 4 (ablation) | Hyper-ICL (Idefics-9B) · VQAv2 | accuracy: 62.1 -> 62.2 | No clear effect |
| set curvature to 0.1 rather than 0.01 or 1.0 (ablation) | hyperbolic anchor distillation · VQAv2 (Idefics-9B) | accuracy | Improved |
| replace 32-shot ICL with Hyper-ICL at inference | Idefics-9B · inference FLOPs and runtime | FLOPs (T) / runtime (ms): 23.364 -> 0.955; 468.55 -> 59.32 | Improved |
| replace standard ICL with Hyper-ICL | Idefics-9B · COCO caption hallucination | CHAIRs: 16.78 -> 8.29 | Improved |
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.
- LIVE — LIVE: Learnable In-Context Vector for Visual Question Answering (NeurIPS 2024) · better on 4 of 4 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
- AIM: Let Any Multimodal Large Language Models Embrace Efficient In-Context Learning (AAAI 2025 · similarity 0.91)
- Make LVLMs Focus: Context-Aware Attention Modulation for Better Multimodal In-Context Learning (AAAI 2026 · similarity 0.89)
- VL-ICL Bench: The Devil in the Details of Multimodal In-Context Learning (ICLR 2025 · similarity 0.89)
- Capturing In-Context Learning Dynamics with Task Operators (NeurIPS 2026 · similarity 0.88)
- Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning (ICLR 2024 · similarity 0.88)
- Revisiting In-context Learning Inference Circuit in Large Language Models (ICLR 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.
