Research map / Vision-language and multimodal / Multimodal LLMs
ReasonCLIP-58M: Visually Grounded Commonsense Reasoning Supervision for CLIP
Sicheng Zhang, Muhammad Muzammal Naseer, Binzhu Xie, Naufal Suryanto, Shi Qiu, Jamal Bentahar, NAVEED AKHTAR, Shah Mubarak
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 Stage 1 reasoning-aware alignment continual pretraining on ReasonLite-42M (ablation) | CLIP ViT-L/14 336 · COCO-5K I->T | R@1: 58.0 -> 65.1 | Improved |
| add Stage 1 continual pretraining (ablation) | CLIP ViT-L/14 336 · Urban-1K I->T | R@1: 73.0 -> 83.0 | Improved |
| add Stage 2 explicit reasoning supervision after Stage 1 (CLIP-L) (ablation) | ReasonCLIP L/14 224 · COCO-5K I->T | R@1: 64.5 -> 59.9 | Hurt |
| add Stage 2 explicit reasoning supervision after Stage 1 (ablation) | ReasonCLIP L/14 336 · RCLIP-V3-5K T->I | R@1: 38.5 -> 42.8 | Improved |
| add Stage 2 for large SigLIP backbone (ablation) | ReasonSigLIP So400m · RCLIP-V3-5K T->I | R@1: 43.4 -> 49.5 | Improved |
| reasoning continual pretraining on compositional reasoning benchmarks | CLIP B/32 · compositional reasoning avg | average accuracy: 52.1 -> 57.7 | Improved |
| reasoning continual pretraining on SigLIP2 G/16 | SigLIP2 G/16 · SugarCrepe | accuracy: 83.0 -> 76.8 | Hurt |
| apply READ-CLIP fine-tuning on top of ReasonCLIP-B/32 | ReasonCLIP B/32 · compositional reasoning avg | average accuracy: 64.1 -> 65.7 | Improved |
| replace CLIP with ReasonCLIP as LLaVA-NeXT visual tower | LLaVA-NeXT (Qwen3-1.7B) · MMVP | accuracy: 58.7 -> 62.3 | Improved |
| replace CLIP with ReasonCLIP as visual tower | LLaVA-NeXT (Qwen3-1.7B) · SciQA | accuracy: 73.1 -> 73.0 | No clear effect |
| stage 1 loss weighting: descriptive-heavy or reasoning-heavy instead of default schedule (ablation) | ReasonCLIP Stage 1 · ablation on training configurations | overall | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Long-CLIP — Long-CLIP: Unlocking the Long-Text Capability of CLIP (ECCV 2024) · better on 1 of 2 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
- Enhancing Compositional Reasoning in CLIP via Reconstruction and Alignment of Text Descriptions (NeurIPS 2025 · similarity 0.89)
- SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.89)
- Point-RFT: Improving Multimodal Reasoning with Visually Grounded Reinforcement Finetuning (NeurIPS 2025 · similarity 0.89)
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.89)
- VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
- Activating Visual Context and Commonsense Reasoning Through Masked Prediction in VLMs (AAAI 2026 · similarity 0.89)
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.
