Research map / Vision-language and multimodal / Multimodal LLMs

Harnessing Chain-of-Thought Reasoning in Multimodal Large Language Models for Face Anti-Spoofing

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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.

Introduces FaceCoT, a 1.08M-sample chain-of-thought VQA dataset for face anti-spoofing, and a two-stage CoT-Enhanced Progressive Learning (CEPL) strategy fine-tuning MiniCPM-V-2.6 for spoof detection with interpretable reasoning.

Problem

Face anti-spoofing relies on single visual modality and binary labels, giving poor cross-domain generalization and no explanations; MLLMs for FAS lack vision-language CoT training data.

Approach

Build a six-part hierarchical CoT annotation format (caption, facial description, facial attributes, reasoning, spoofing description, conclusion) generated by GPT-4o with human refinement, expand it with an RL-tuned caption model, and train an MLLM in two stages: CoT-only vision-enhancement SFT followed by LoRA-based joint CoT and binary classification.

Key results

• Trained on FaceCoT-All, average AUC 97.77% and HTER 6.30% across 11 benchmarks, versus I-FAS at 93.71% AUC and 11.30% HTER.
• On the widely used leave-one-out protocol, average HTER 1.06% and AUC 99.85%, best among compared methods.
• CEPL outperforms single-stage joint training with HTER 7.65% vs 8.84% and AUC 96.59% vs 95.91% on the Gold100K ablation.

Surprising findings

• Adding RL after CEPL stage 2 gives nearly no change in the main metrics (HTER 7.80% vs 7.65%), which the authors attribute to extensive SFT data.
• Zero-shot MiniCPM-V-2.6 and Qwen2.5-VL-7B both underperform the I-FAS baseline substantially, indicating off-the-shelf VLMs are weak for FAS.

Limitations

• Does not capture uncommon real-world device or environment variations.
• Potential as a standardized evaluation benchmark is not fully explored.
• Gains from RL after CEPL are marginal and not clearly explained.

Experimental setup

Datasets: CelebA-Spoof Wild-FAS (WFAS) FaceCoT-Gold100K FaceCoT-Silver982K

Benchmarks: CASIA-MFSD CASIA-SURF-3DMask HKBU-MARs-V1+ HiFiMask MSU-MFSD OULU-NPU Replay-Attack Rose-Youtu SIW SIW-M-V2 WMCA O&C&I→M protocol cross-dataset

Metrics: HTER AUC accuracy

Baselines: ViTAF ViT-B ViT-L FLIP I-FAS FGHV GDA PatchNet SSAN IADG UDG-FAS TTDG

Builds on: LLaVA-CoT VRFT LoRA MiniCPM-V

Closest papers by meaning

• Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models (AAAI 2025 · similarity 0.92)
• FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.92)
• PA-FAS: Towards Interpretable and Generalizable Multimodal Face Anti-Spoofing via Path-Augmented Reinforcement Learning (AAAI 2026 · similarity 0.90)
• From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.89)
• mmFAS: Multimodal Face Anti-Spoofing Using Multi-Level Alignment and Switch-Attention Fusion (AAAI 2025 · similarity 0.87)
• SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining (AAAI 2025 · similarity 0.87)