Research map / Vision-language and multimodal / Multimodal LLMs
Falcon: Functional Assembly and Language for Compositional Reasoning in X-ray
Yonathan Michael, Mohamad Alansari, Natnael Takele, Andreas Henschel, Naoufel Werghi
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 the structured safety adapter (risk, presence, link heads) to the region-grounded VLM (ablation) | Falcon · Falcon-X RFG | cIoU: 35.82 -> 50.45 | Improved |
| add presence head on top of risk head (ablation) | SSA · Falcon-X RFG | cIoU: 39.33 -> 45.52 | Improved |
| add the link head (ablation) | SSA · Falcon-X RFG | mIoU: 58.60 -> 69.58 | Improved |
| add SSA structured supervision on top of LoRA adaptation (ablation) | Falcon · Falcon-X RFG | cIoU: 44.63 -> 50.45 | Improved |
| swap Vicuna-7B for Qwen3.5-VL backbone (ablation) | Falcon · Falcon-X RFG | cIoU: 50.45 -> 51.07 (with SSA); 35.82 -> 36.91 (without) | No clear effect |
| use Falcon instead of fine-tuned baselines for referring functional grounding | Falcon · Falcon-X RFG | cIoU: 20.07 -> 50.45 | Improved |
| structured safety bottleneck effect on panoptic segmentation | Falcon · Falcon-X PS | cIoU: 36.73 -> 14.37 | Hurt |
| relax the structured bottleneck (ablation) | Falcon · Falcon-X PS/RPS vs RFG | cIoU: +5.27 PS cIoU, +3.99 RPS cIoU, -7.39 RFG cIoU | Mixed |
| increase proposal recall and unfreeze visual encoder (ablation) | Falcon · Falcon-X PS/RPS | cIoU: recall 92.6% -> 97.8%, change under 1 cIoU | No clear effect |
| replace zero-fill removal with mask-guided inpainting for counterfactuals (ablation) | Falcon-X counterfactual images · DINO-FID realism check | DINO-FID: 32.74 -> 6.72 | Improved |
| replace Groma with Falcon on missing-component identification | Falcon · Falcon-X MCI | accuracy: 97.1 -> 94.75 | Hurt |
| Falcon versus Sting-Bee on cross-dataset referring localization | Falcon · STCray/PIDray/SIXray | cIoU: 10.11 -> 16.60 (STCray) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Sting-Bee — STING-BEE: Towards Vision-Language Model for Real-World X-ray Baggage Security Inspection (CVPR 2025) · better on 1 of 1 shared result
- Groma — Groma: Localized Visual Tokenization for Grounding Multimodal Large Language Models (ECCV 2024) · better on 1 of 1 shared result
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
- STING-BEE: Towards Vision-Language Model for Real-World X-ray Baggage Security Inspection (CVPR 2025 · similarity 0.85)
- SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization (CVPR 2026 · similarity 0.84)
- VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety (ICLR 2026 · similarity 0.84)
- Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models (ICLR 2026 · similarity 0.84)
- CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays (NeurIPS 2025 · similarity 0.84)
- Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models (NeurIPS 2026 · similarity 0.83)
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.
