Research map / Vision-language and multimodal / Multimodal LLMs
VISOR++ : VISUAL INPUT BASED STEERING FOR LARGE VISION LANGUAGE MODELS
Ravikumar Balakrishnan, Mansi Phute
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 |
|---|---|---|---|
| use ensemble VISOR++ image instead of system prompt | LLaVA-1.5-7B · Refusal (negative steering) | BAS: 0.698 -> 0.353 | Improved |
| use ensemble VISOR++ image instead of system prompt | IDEFICS2-8B · Sycophancy (negative steering) | BAS: 0.759 -> 0.581 | Improved |
| optimize per-model VISOR++ image instead of steering vector | IDEFICS2-8B · Refusal (positive steering) | BAS: 0.817 -> 0.94 | Improved |
| jointly optimize one ensemble image instead of per-model images | LLaVA-1.5-7B · Sycophancy (negative steering) | BAS: 0.393 -> 0.623 | Hurt |
| jointly optimize one ensemble image instead of per-model images | IDEFICS2-8B · Refusal (negative steering) | BAS: 0.231 -> 0.29 | No clear effect |
| insert ensemble VISOR++ image (trained on LLaVA and IDEFICS2) on unseen model | GPT-4-Turbo · Survival instinct (negative steering) | behavior fraction: 0.388 -> 0.312 | Improved |
| insert ensemble VISOR++ image on unseen model | GPT-4-Turbo · Sycophancy (negative steering) | behavior fraction: 0.460 -> 0.390 | Improved |
| insert ensemble VISOR++ image on unseen model | Claude Sonnet 3.5 · Refusal (negative steering) | behavior fraction: 0.609 -> 0.609 | No clear effect |
| insert ensemble VISOR++ image on unseen model | Qwen2-vl-7b · Refusal (negative steering) | BAS: 0.866 -> 0.859 | No clear effect |
| insert ensemble VISOR++ image on unseen model | Llama-3.2-11B · Sycophancy (negative steering) | BAS: 0.496 -> 0.518 | Hurt |
| insert ensemble VISOR++ image on unseen model | LLaVA-NeXT · Refusal (negative steering) | BAS: 0.879 -> 0.852 | Improved |
| add ensemble VISOR++ image versus random image | LLaVA-1.5-7B · MMLU (14,000 samples) | mean score: 0.491 -> 0.492 | No clear effect |
Closest papers by meaning
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.89)
- Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- Prune2Drive: A Plug-and-Play Framework for Accelerating Vision-Language Models in Autonomous Driving (CVPR 2026 · similarity 0.88)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
- VLsI: Verbalized Layers-to-Interactions from Large to Small Vision Language Models (CVPR 2025 · similarity 0.88)
- Less Language, More Latents: Annotation-Efficient VLAs for Driving (NeurIPS 2026 · 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.
