Research map / Vision-language and multimodal / Visual question answering
Attention-Logit Steering to Compositional Generalization for Continual VQA
Suyoung Yang
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 factorized old/new expert control (ablation) | Q-STEER · Continual VQA v2 | AP: 48.73 -> 53.68 | Improved |
| add ICR attention-logit steering (ablation) | Q-STEER · Continual VQA v2 | AF: 5.90 -> 4.51 | Improved |
| add probe-based expert expansion (ablation) | Q-STEER · Continual VQA v2 | AP: 50.96 -> 53.68 | Improved |
| use KL drift instead of constant d(x) (ablation) | Q-STEER · Continual VQA v2 | AP: 52.87 -> 53.68 | Improved |
| use KL drift instead of cosine drift (ablation) | Q-STEER · Continual VQA v2 | AP: 53.14 -> 53.68 | Improved |
| add active pre-softmax mask (ablation) | Q-STEER · Continual VQA v2 | AP: 51.28 -> 53.68 | Improved |
| add sparse top-K backward updates (ablation) | Q-STEER · Continual VQA v2 | AF: 6.89 -> 4.51 | Improved |
| add ICR steering on Evidence-Mass (ablation) | Q-STEER · Continual VQA v2 grounding | Evidence-Mass: 0.421 -> 0.487 | Improved |
| add ICR steering on attention entropy (ablation) | Q-STEER · Continual VQA v2 grounding | attention entropy: 2.31 -> 2.07 | Improved |
| vary ICR scale s_icr from 0.3 to 0.1 or 0.5 (ablation) | Q-STEER · Continual VQA v2 | AP: 53.68 -> 52.76 (0.1), 52.98 (0.5) | Hurt |
| extend controlled late layers from 6 to 8 (ablation) | Q-STEER · Continual VQA v2 | AP: 53.68 -> 52.94 | Hurt |
| increase controlled late layers from 1 to 6 (ablation) | Q-STEER · Continual VQA v2 | AP: 50.72 -> 53.68 | Improved |
Closest papers by meaning
- Ask and Remember: A Questions-Only Replay Strategy for Continual Visual Question Answering (ICCV 2025 · similarity 0.94)
- SCoT: Similarity-guided Conflict-aware Task Consolidation for Continual VQA (ECCV 2026 · similarity 0.91)
- AVQACL: A Novel Benchmark for Audio-Visual Question Answering Continual Learning (CVPR 2025 · similarity 0.89)
- Dynamic Cross-Modal Prompt Generation for Multimodal Continual Instruction Tuning (NeurIPS 2026 · similarity 0.88)
- Question-Aware Gaussian Experts for Audio-Visual Question Answering (CVPR 2025 · similarity 0.87)
- CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025 · similarity 0.87)
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.
