Research map / Vision-language and multimodal / Multimodal LLMs

From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Proposes TAR-FAS, which lets an MLLM (InternVL3-8B) adaptively call visual tools (zoom, LBP, FFT, wavelet, HOG, edge) in a chain-of-thought for face anti-spoofing, trained on a new ToolFAS-16K trajectory dataset and a tool-diversity GRPO variant (DT-GRPO).

Problem

Brief textual-description MLLM-based face anti-spoofing captures coarse semantic cues and misses fine-grained spoof traces, limiting cross-domain generalization under the one-to-eleven protocol.

Approach

Reformulate FAS as chain-of-thought with visual tools (CoT-VT), where the MLLM first gives a quick intuitive answer and then invokes classical visual operators (zoom-in, LBP, FFT/wavelet, HOG/edge) over multiple turns, with expert-classifier guidance for annotation and a tool-diversity reward in GRPO to encourage varied tool use.

Key results

• Under one-to-eleven protocol, TAR-FAS reaches 7.54% average HTER and 96.67 AUC, versus 11.30% HTER for I-FAS.
• On HiFiMask, TAR-FAS gets 17.97% HTER vs 28.23% for I-FAS; on HKBU-MARs-V1+ 3.48% vs 18.64%.
• Under ICMO leave-one-out protocol, average HTER is 1.30%, slightly better than I-FAS (1.33%); on MSU-MFSD and SIW and SIW-M-V2 I-FAS is better.

Surprising findings

• Without the format-injection stage, RL alone reaches only 20.79% HTER, showing RL from scratch fails to learn tool use effectively.
• On MSU-MFSD, SIW and SIW-M-V2 the proposed method is slightly worse than I-FAS, and under ICMO it is only marginally better than I-FAS.

Limitations

• Relies on Gemini-2.5 Pro for annotation and on classical tools only, with a limited tool set (6 operators).
• Gains over I-FAS are marginal on several benchmarks (e.g., MSU-MFSD, SIW, SIW-M-V2, OULU-NPU) and under ICMO protocol.
• Expert-model guidance adds dependence on tool-specific classifiers trained on CelebA-Spoof.

Experimental setup

Datasets: CelebA-Spoof ToolFAS-16K

Benchmarks: MSU-MFSD CASIA-MFSD Replay-Attack OULU-NPU SIW Rose-Youtu HKBU-MARs-V1+ WMCA SIW-M-V2 CASIA-SURF-3DMask HiFiMask

Metrics: HTER AUC

Baselines: ViT-L ViTAF FLIP I-FAS CFPL-FAS ST-GRPO

Builds on: I-FAS DeepEyes GRPO Qwen2.5-VL InternVL3

Closest papers by meaning

• FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.91)
• Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models (AAAI 2025 · similarity 0.90)
• SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining (AAAI 2025 · similarity 0.89)
• Harnessing Chain-of-Thought Reasoning in Multimodal Large Language Models for Face Anti-Spoofing (CVPR 2026 · similarity 0.89)
• mmFAS: Multimodal Face Anti-Spoofing Using Multi-Level Alignment and Switch-Attention Fusion (AAAI 2025 · similarity 0.88)
• Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.88)