Perception, Verdict, and Evolution: Hindsight-Driven Self-Refining Forensics Agent for AI-Generated Image Detection¶
🕒 Published (v1): 2026-06-25 02:59 UTC · Source: Arxiv · link
Why this paper was selected
DAMO/Alibaba; self-refining forensics agent with hindsight-driven evolution; novel refinement loop
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
ForeAgent is an agentic forensics framework for AI-generated image (AIGI) detection that combines multi-view perceptual cues (semantic, frequency-domain, spatial) with an MLLM verdict module, and continuously self-improves via a Hindsight-Driven Self-Refining loop without relying on external frontier model annotations. It achieves 82.18% accuracy on the Chameleon benchmark (+16.41% over AIDE) and 93.3% mean accuracy across 16 generators on AIGCDetectBenchmark.
Problem¶
Existing AIGI detectors face two critical gaps: (1) feature-based methods lack high-level semantic reasoning, while MLLM-based methods lack sensitivity to fine-grained forensic artifacts; (2) current MLLM-based approaches depend on static synthetic supervision generated by expensive frontier models (GPT-4o, Pixtral-124B), leaving iterative self-refinement on misclassified samples unexplored.
Method¶
ForeAgent is built on Qwen3-VL-8B-Instruct without architectural modification and operates in two phases:
Perception–Verdict Architecture: The agent receives dual-image input — the original image \(x\) and its wavelet frequency-domain counterpart \(x_{freq}\) — for semantic and spectral perception. Simultaneously, it invokes an external NPR (Neighboring Pixel Relationships) detector as a callable tool to obtain spatial-domain artifact probabilities \(y_2 = \{Pr_{real}, Pr_{fake}\}\). A verdict module synthesizes \(y_1 = F_{MLLM}(x, x_{freq})\) and \(y_2\) into a confidence-calibrated, logically grounded final decision \(\hat{y}\).
Hindsight-Driven Self-Refining (Sampling–Reflection–Evolution): The training set is partitioned into \(N\) disjoint shards. At each iteration \(t\), the current agent \(\pi_{\theta_{t-1}}\) runs inference rollouts and categorizes trajectories as (1) incorrect predictions, (2) correct predictions with low-quality reasoning, or (3) high-quality. Failure cases trigger reflection: given ground-truth label \(v_{gt}\) as hindsight, the agent regenerates two candidate traces (\(tr_1\) for perception, \(tr_2\) for verdict). A Dual-Expert Quality Gating (DEG) module — using both ForeAgent and Qwen3-VL-Plus as independent judges — admits samples under a "lenient entry, strict exit" principle. High-quality samples \(D_{high}\) receive full reasoning supervision; correct-but-weak samples \(D_{label}\) receive label-only supervision. The agent is LoRA fine-tuned each iteration with performance-gated checkpoint selection (revert if \(S_t \leq S_{best}\)). Adaptive quality thresholds \(\tau_t\) start low and tighten across iterations.
Key Contributions¶
- Perception–Verdict architecture: unifies semantic (original image), frequency-domain (wavelet diagonal maps at 256×256), and spatial (NPR tool-call output) cues in a single MLLM reasoning engine without backbone modification.
- Hindsight-Driven Self-Refining: a closed-loop Sampling–Reflection–Evolution paradigm that converts failure cases into high-value training signals using only ground-truth labels — no frontier model annotation required.
- Dual-Expert Quality Gating: two independent MLLM judges with "lenient entry, strict exit" logic to filter synthesized reasoning traces, mitigating single-model evaluation bias.
- State-of-the-art results on Chameleon (82.18%) and AIGCDetectBenchmark (93.3% mean) with superior reasoning quality over GPT-5.
Results¶
- Chameleon benchmark: ForeAgent 82.18% overall (96.69% real, 61.86% fake) vs. GPT-5 81.91%, GPT-5-mini 69.65%, AIDE 65.77%, AIGI-Holmes 75.90%.
- AIGCDetectBenchmark (16 generators): ForeAgent 93.3% mean vs. AIGI-Holmes 93.2%, AIDE 92.8%, NPR 82.9%.
- Diffusion-model generators (notable gaps): Midjourney 96.0% vs. AIGI-Holmes 81.6%; SD1.4 98.3% vs. AIGI-Holmes 91.3%; DALLE2 98.9% vs. AIGI-Holmes 85.3%.
- Weak point: WFIR generator 63.4% — CNN-based CNNSpot (91.7%) and AIDE (94.2%) are substantially stronger there.
- Reasoning quality: external evaluation shows ForeAgent produces more consistent and causally grounded reasoning than GPT-5 and GPT-5-mini.
- Trained on 8Ă—A100; backbone Qwen3-VL-8B with LoRA (rank 64, alpha 128).
Limitations¶
- Underperforms CNN-based detectors on WFIR (63.4%), suggesting domain-specific post-processing characteristics remain a blind spot for the frequency-spatial fusion approach.
- Self-refining loop is computationally expensive: requires inference rollouts, dual-expert evaluation, and LoRA fine-tuning per iteration across \(N\) data shards on 8Ă—A100s.
- NPR spatial expert is used without fine-tuning; any distribution mismatch in its predictions propagates uncorrected into the verdict.
- Maximum iterations \(T_{max} = 5\) is a fixed cap; the convergence behavior beyond this is not analyzed.
- Chameleon fake-image accuracy (61.86%) remains well below real-image accuracy (96.69%), indicating residual bias toward classifying images as real.
Relevance to Agentic AI / LLM Agents¶
ForeAgent is a concrete instance of a tool-augmented MLLM agent applied to a forensic perception task: it treats an external NPR detector as a callable tool, uses the MLLM as a reasoning/arbitration engine, and closes the training loop via hindsight-driven self-refinement — a pattern directly analogous to ReAct-style agents with self-improvement. The Sampling–Reflection–Evolution paradigm is a domain-specific instantiation of iterative self-play or STaR-style self-improvement, demonstrating that agent self-curation of training data (with quality gating) can substitute for costly frontier-model supervision. The dual-expert quality gating mechanism is also relevant to researchers building LLM-as-judge pipelines, showing empirically that dual independent judges reduce single-model bias.