Cognitive Inception: Agentic Reasoning against Visual Deceptions by Injecting Skepticism

TL;DR

This work tackles the vulnerability of multi-modal LLMs to AIGC visual deceptions by introducing Inception, a fully reasoning-based agentic framework that injects skepticism into the inference process. It models skepticism as iterative, dual-agent reasoning between External and Internal Skeptics, coupled with a logic-verification mechanism that treats valid skeptical logics as evidence of AI-generated content when sufficient conditions are found. The approach achieves state-of-the-art results on the AEGIS benchmark and strong gains on Forensics-Bench, demonstrating robust generalization across video and image modalities without feature-based detectors. By formalizing skeptical logics, their verification, and iterative expansion, the method improves authenticity verification while providing interpretable reasoning traces, signaling practical impact for trustworthy AI perception. The framework’s reliance on reasoning dynamics rather than training on new detectors suggests broad applicability to emerging generative visual threats and evolving data distributions.

Abstract

As the development of AI-generated contents (AIGC), multi-modal Large Language Models (LLM) struggle to identify generated visual inputs from real ones. Such shortcoming causes vulnerability against visual deceptions, where the models are deceived by generated contents, and the reliability of reasoning processes is jeopardized. Therefore, facing rapidly emerging generative models and diverse data distribution, it is of vital importance to improve LLMs' generalizable reasoning to verify the authenticity of visual inputs against potential deceptions. Inspired by human cognitive processes, we discovered that LLMs exhibit tendency of over-trusting the visual inputs, while injecting skepticism could significantly improve the models visual cognitive capability against visual deceptions. Based on this discovery, we propose \textbf{Inception}, a fully reasoning-based agentic reasoning framework to conduct generalizable authenticity verification by injecting skepticism, where LLMs' reasoning logic is iteratively enhanced between External Skeptic and Internal Skeptic agents. To the best of our knowledge, this is the first fully reasoning-based framework against AIGC visual deceptions. Our approach achieved a large margin of performance improvement over the strongest existing LLM baselines and SOTA performance on AEGIS benchmark.

Figures (5)

• Figure 1: An inception of skepticism can improve human cognitive robustness. (a) Humans assume information to be authentic at a trust default state trust_default_6, and are vulnerable to deceptions. (b) When a trigger of skepticism is given, trust default could be alleviated, and humans exhibit significantly stronger reasoning and cognitive capabilities against visual deception. We evaluated LLMs' reasoning on AIGC visual deceptions, and observed similar reasoning shortcomings and trust default behavior. Inspired by this, we propose an agentic reasoning framework by explicitly injecting skepticism into LLMs' inference process to improve reasoning reliability against visual deceptions.
• Figure 2: Skepticism largely benefits LLMs' reasoning against deceptions. We discovered that LLMs tend to be trust their visual inputs by default and struggle to discover useful visual information against visual deceptions. To explore the trust default tendency of LLMs, we evaluated LLMs' reasoning on AI-generated videos from AEGIS benchmark benchmark_1 under 3 different triggers: Default, Neutral and Skeptic. Significant performance differences were observed. (a): Skeptic triggers lead to more effortful reasoning processes of LLMs and produce the most visual elements denoted by the reasoning processes. (b) & (c): Explicitly triggering skepticism contributes to higher reasoning quality, with significantly higher precision and recall rates of visual information retrieved by the LLMs.
• Figure 3: Overall framework. We propose Inception to iteratively refine and enhance the skeptical logics given by LLMs. An External Skeptic agent with access to the visual input skeptically produces a set of Initial Skeptic Logics $r_{ex}^i$ over it. Then the skeptic logics are evaluated by an Internal Skeptic, which skeptically question any logic given to it. The Internal Skeptic returns the Logic Verification flags with corresponding its reasoning processes. Verification flags represent 3 different types of logic judgments: Valid, Invalid and Epochē (Suspension of Judgment). The Epochē Logics lead to the next round of External Skeptic reasoning to iteratively retrieve more visual information to enhance the logics. The valid logics are kept for decision making and invalid logics are discarded.
• Figure 4: Iterative logic expansion and verification. $r_{\textrm{ex}}^{(\cdot)}$ is an independent and indivisible skeptic logic statement given by the External Skeptic, and $v^{(\cdot)}$ is the verification flag returned by the Internal Skeptic. To verify the Initial Skeptical Logics, Inception iteratively expands the epochē logic into child logic nodes, until the epochē logic is proved to be valid or invalid. In this reasoning tree, logic chains are formed to verify a skeptic logic by back-tracing the valid logics along such reasoning tree expansion process.
• Figure 5: Accuracy scatter plot against number of verified skeptical logics. The area of each marker is proportional to the population size of data samples. In this experiment, the contribution of skepticism on the reasoning process is evaluated. Verified skeptical logics refer to the skeptical logics eventually proved "valid" or "invalid". The charts titled "Skeptical" are produced by the complete Inception framework, and the charts titled "Neutral" are produced by an ablation setting, replacing the External Skeptic with a neutral reasoning agent. It is observed that skeptical reasoning's accuracy is positively correlated with the total number of logics verified by the framework. Vice versa, the reasoning process's accuracy is damaged when more logics are verified.