Research map / Vision-language and multimodal / Multimodal LLMs

PA-FAS: Towards Interpretable and Generalizable Multimodal Face Anti-Spoofing via Path-Augmented Reinforcement Learning

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 1 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.

Proposes PA-FAS, which augments reasoning paths from a label tree (PNRPS sampling) and uses answer shuffling during SFT before GRPO, to make SFT+RL work for multimodal domain-generalized face anti-spoofing with interpretable reasoning.

Problem

SFT+RL for multimodal face anti-spoofing is hampered by scarce, binary-label annotations: SFT yields overconfident models with polarized rewards, limited RL exploration, and reasoning shortcuts where models map images directly to answers while ignoring the chain-of-thought.

Approach

Build a classification tree over labels and sample positive-negative random paths to generate many CoT variants per annotated sample (reasoning path augmentation), then shuffle final answers across samples during SFT so the model must rely on the reasoning text, followed by GRPO without a KL term.

Key results

• PA-FAS reaches average HTER 15.21% and AUC 89.13% under fixed-modal cross-dataset protocol, versus 23.25% for Qwen2.5-VL-3B-SFT and 34.37% for SFT+GRPO.
• Under missing-modality protocol, PA-FAS average HTER is 15.85% versus 33.46% for zero-shot Qwen2.5-VL-3B.
• Answer shuffling reduces HTER from 24.45% to 15.21% relative to no shuffling in the ablation.

Surprising findings

• Standard SFT+GRPO raised average HTER from 23.25% (SFT alone) to 34.37%, indicating RL collapse under scarce data.
• Path-level shuffling gave no gain (HTER 24.12%) while answer shuffling gave large gain, suggesting models ignore reasoning text when answers are predictable.

Limitations

• Labeled data is small (800 samples), making training less stable and efficient, as the authors note.
• Multimodal information utilization remains sub-optimal; mild interference from other modalities observed.
• Reward design is simple (format plus binary correctness), and KL term is dropped without extensive analysis.

Experimental setup

Datasets: WMCA SURF (CASIA-SURF) CeFA (CASIA-CeFA) PADISI

Benchmarks: WMCA SURF CeFA PADISI

Metrics: HTER AUC

Baselines: Qwen2.5-VL-3B Qwen2.5-VL-3B-SFT Qwen2.5-VL-3B-SFT+GRPO MMDG DADM FLIP CLIP SSAN VP-FAS

Builds on: GRPO Qwen2.5-VL MMDG

Closest papers by meaning

• Harnessing Chain-of-Thought Reasoning in Multimodal Large Language Models for Face Anti-Spoofing (CVPR 2026 · similarity 0.90)
• mmFAS: Multimodal Face Anti-Spoofing Using Multi-Level Alignment and Switch-Attention Fusion (AAAI 2025 · similarity 0.89)
• Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models (AAAI 2025 · similarity 0.89)
• From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.88)
• Revisiting the Data Sampling in Multimodal Post-training from a Difficulty-Distinguish View (AAAI 2026 · similarity 0.88)
• FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.87)