Research map / Vision-language and multimodal / Multimodal LLMs
Stop Mixing Things Up! BISCUIT Teaches Vision-Language Models to Learn New Concepts from Images on the Spot
Jiahua Bao, Siyao Cheng, Jiaxing Du, Yuhang Jia, Boyang Niu, Zeming Lang, Changjiang He, Hao Zhang, Jie Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BISCUIT, a method that teaches vision-language models to learn and use newly introduced visual concepts from images at inference time, addressing the "Mixing Things Up" failure where concept names get confused.
Problem
VLMs often fail to apply newly introduced visual concepts during inference, confusing concept names and giving vague descriptions and poor grounding; prior work mainly handled person-related concepts through text prompts.
Approach
Teaches VLMs to learn new visual concepts directly from images on the spot, rather than relying on text prompts designed mainly for person-related concepts (details beyond the abstract not available).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add BISCUIT training | Qwen2.5-VL-7B · BiscuitVQA overall | SANLS summed: 339.9 -> 483.9 | Improved |
| add BISCUIT training | LLaVA-1.5-7B · BiscuitVQA overall | SANLS summed: 274.0 -> 325.2 | Improved |
| add dual-stream (RGB + edge) encoder (ablation) | BISCUIT on Qwen2.5-VL-7B · BiscuitVQA overall | SGPT summed: 731.4 -> 799.6 | Improved |
| add SVD-based initialization of edge-stream modules (ablation) | BISCUIT on Qwen2.5-VL-7B · BiscuitVQA overall | SGPT summed: 738.6 -> 799.6 | Improved |
| add identification-aware loss instead of standard DPO (ablation) | BISCUIT on Qwen2.5-VL-7B · BiscuitVQA overall | SGPT summed: 721.7 -> 799.6 | Improved |
| add identification-aware loss instead of standard DPO (ablation) | BISCUIT on Qwen2.5-VL-7B · Generalization retention | retention (%): 94.7% -> 94.1% | No clear effect |
| use Canny edges instead of no edge detection (ablation) | BISCUIT on Qwen2.5-VL-7B · IICL | IICL: 608.0 -> 641.8 | Improved |
| use Sobel edges instead of Canny (ablation) | BISCUIT on Qwen2.5-VL-7B · IICL | IICL: 641.8 -> 587.4 | Hurt |
| use QR basis fusion instead of direct sum (ablation) | BISCUIT on Qwen2.5-VL-7B · IICL | IICL: 621.5 -> 641.8 | Improved |
| set threshold tau=0.0 instead of 0.1 (ablation) | identification-aware loss · Generalization, Qwen2.5-VL-3B | generalization | Hurt |
| use BISCUIT vs other concept-learning methods on generalization | Qwen2.5-VL-7B · Generalization retention across 11 benchmarks | retention (%): 91.09% -> 94.07% | 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.
- Yo'Chameleon — Yo'Chameleon: Personalized Vision and Language Generation (CVPR 2025) · better on 3 of 3 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
- Bridging the gap to real-world language-grounded visual concept learning (NeurIPS 2025 · similarity 0.89)
- Same or Not? Enhancing Visual Perception in Vision-Language Models (CVPR 2026 · similarity 0.88)
- VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set (NeurIPS 2025 · similarity 0.88)
- V2C-CBM: Building Concept Bottlenecks with Vision-to-Concept Tokenizer (AAAI 2025 · similarity 0.88)
- Language-Informed Visual Concept Learning (ICLR 2024 · similarity 0.87)
- The Neglected Tails in Vision-Language Models (CVPR 2024 · 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.
