Research map / Vision-language and multimodal / Multimodal LLMs
Learning Chain of Counterfactual Thought for Bias-Robust Vision-Language Reasoning
Yifeng Zhang, Ming Jiang, Qi Zhao
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 TLM reasoning-process predictor to CoCT (ablation) | CoCT on LLaVA-1.5 · Bingo | accuracy: 18.8 -> 29.8 | Improved |
| add bias-robust loss L_br to TLM training (ablation) | CoCT on LLaVA-1.5 · CoBRa counterfactual split | accuracy: 34.2 -> 38.3 | Improved |
| include both original and counterfactual in-context examples instead of one type (ablation) | CoCT on LLaVA-1.5 · CoBRa counterfactual split | accuracy: 33.7 -> 38.3 | Improved |
| select in-context examples by reasoning similarity rather than knowledge similarity (ablation) | CoCT on LLaVA-1.5 · CoBRa counterfactual split | accuracy: 31.2 -> 32.1 | Improved |
| knowledge-similarity selection instead of reasoning similarity (ablation) | CoCT on LLaVA-1.5 · ScienceQA-IMG | accuracy: 70.8 vs 70.1 (random) | Mixed |
| add reasoning processes to the prompt (ablation) | CoCT on LLaVA-1.5 · CoBRa gap (delta) | delta (lower better): 44.2 -> 38.1 | Improved |
| add knowledge graphs to the prompt (ablation) | CoCT on LLaVA-1.5 · CoBRa counterfactual split | accuracy: 29.9 -> 38.3 | Improved |
| add in-context examples versus zero-shot (ablation) | CoCT on LLaVA-1.5 · HallusionBench | accuracy: 29.4 -> 30.7 (random) | Improved |
| apply CoCT (CoBRa examples) over plain LVLM | LLaVA-1.5 · Bingo | accuracy: 12.5 -> 29.8 | Improved |
| use CoCT over DDCoT on ScienceQA | LLaVA-1.5 and GPT-4V · ScienceQA-IMG | accuracy: 74.6 -> 71.3 (LLaVA); 80.2 -> 75.8 (GPT-4V) | Hurt |
| use CoBRa instead of GQA as example source | CoCT · GQA-OOD / VQA-CP (LLaVA-1.5) | accuracy: VQA-CP 56.2 -> 56.0 | Mixed |
Closest papers by meaning
- What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models (CVPR 2024 · similarity 0.90)
- Scaling Test-Time Robustness of Vision-Language Models via Self-Critical Inference Framework (CVPR 2026 · similarity 0.90)
- CF-VLM:CounterFactual Vision-Language Fine-tuning (NeurIPS 2025 · similarity 0.89)
- Latent Chain-of-Thought for Visual Reasoning (NeurIPS 2025 · similarity 0.88)
- CFPO: Counterfactual Policy Optimization for Multimodal Reasoning (ICML 2026 · similarity 0.88)
- When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.88)
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.
