TL;DR
The rapid evolution of large language models (LLMs) has made traditional, black-box AI-generated content (AIGC) detectors obsolete. This paper introduces REVEAL, a novel framework that shifts detection from opaque classification to transparent, reasoning-based decision-making using a two-stage training strategy (SFT + RL). Alongside the framework, the authors release AIGC-text-bank, a massive, parallel multi-domain dataset featuring text from state-of-the-art models across pure and hybrid (AI-polished) authorship scenarios.
What Problem Does This Paper Address?
As generative models achieve human-level fluency, distinguishing machine-authored text from human writing has become a critical challenge for academic integrity, fraud prevention, and digital trust. Existing AIGC detectors suffer from two major flaws:
- Opaqueness and Poor Generalization: Statistical and black-box neural classifiers exploit surface-level patterns that fail to generalize as new LLMs emerge.
- The "AI-Polish" Blindspot: Most benchmarks focus on binary classification (pure human vs. pure AI). They fail to address the realistic, nuanced scenario of human-AI collaborative writing, where an AI polishes a human-authored draft.
This paper tackles these gaps by building a detector that articulates a human-readable reasoning process before delivering a verdict, specifically designed to handle both pure AI-generated and hybrid AI-polished text.
Methods
REVEAL adopts a Think-then-Answer paradigm, forcing the model to generate explicit reasoning traces before outputting a final label. The training pipeline consists of two main stages:
- Reasoning Initialization via SFT: To bootstrap reasoning capabilities, the authors use a hindsight analysis strategy. Instead of asking a teacher model to predict a label, they provide OpenAI's o3 model with both the input text and its ground-truth label, instructing it to reconstruct a plausible decision-making process. The target model (Qwen3-8B) is then fine-tuned using an Outcome-Weighted Objective, which applies a higher loss coefficient to the final answer token to ensure the model uses the reasoning path as supporting context rather than losing focus.
- Reasoning Refinement via Reinforcement Learning (RL): To eliminate hallucinations and logical inconsistencies, the model undergoes direct preference alignment via the DAPO algorithm. Crucially, the authors implement variance-based data selection, filtering the training set to focus only on borderline cases where stochastic rollouts yield inconsistent predictions. The policy is guided by a composite reward function balancing final accuracy, strict formatting constraints, and a fine-grained logical consistency score evaluated by GPT-4o.
For practical applications requiring high throughput, the authors also introduce REVEAL-Fast, a variant trained directly on 3-class labels that bypasses reasoning generation to output a well-calibrated "AIGC score" based on normalized logits.
Datasets
The authors introduce AIGC-text-bank, a large-scale parallel corpus containing 1,498,279 samples across 10 diverse domains (e.g., academic papers, social media, literature). The dataset pairs authentic human writing (published strictly before the release of ChatGPT to guarantee human authorship) with synthetic counterparts generated by 12 different LLMs (including GPT-5, DeepSeek-R1, Grok-4.1, and Llama-3.3).
The synthetic data is split into two categories:
- AI-Native: Fully machine-generated text, length-matched to human references, with 20% of samples explicitly prompted to imitate human writing styles.
- AI-Polish: Human-authored texts refined by LLMs to improve fluency while strictly preserving the original semantic intent and logical structure.
For evaluation, the framework was tested across five benchmarks: their own AIGC-bench test set, DetectRL (adversarial attacks), M4 (domain adaptation), Pan (collaborative analysis), and LOKI (real-world scenarios).
Key Insights and Ideas
- Hindsight Curation Prevents Error Amplification: Generating reasoning chains by giving the teacher model the ground-truth answer ensures high-quality, accurate rationales for training, avoiding the compounding errors common in standard zero-shot generation.
- Uncertainty Filtering Powers Generalization: In the RL phase, filtering out samples where the model is already deterministically successful or failing forces the optimization process to focus entirely on ambiguous, boundary cases. This significantly mitigates overfitting to trivial patterns and drives out-of-distribution (OOD) robustness.
- Disentangling the "Machine Surface" from the "Human Core": The linguistic analysis reveals that AI-polished text retains a human core (high information density, specific proper nouns) but possesses a machine surface (unusual smoothness, lack of natural irregularities). REVEAL successfully learns to weigh these conflicting signals to identify subtle machine interventions.
Key Results
- State-of-the-Art Detection: REVEAL achieved consistently high cross-domain performance, averaging 91.15% accuracy across all five benchmarks, outperforming zero-shot statistical detectors and general LLMs.
- Mastering Fine-Grained Discrimination: On the highly challenging 3-class classification task (Human vs. AI-Native vs. AI-Polish), strong proprietary models like GPT-5 and OpenAI o3 performed near chance (~48% accuracy) due to a severe bias toward predicting the AI-Native class. In contrast, REVEAL achieved 70.74% accuracy, demonstrating a highly balanced prediction profile.
- Superior Transfer Learning: When used as a base model for downstream tasks, REVEAL proved to be an excellent parameter initialization, achieving the highest accuracies across domain adaptation (97.33% on M4) and complex collaborative pattern analysis (49.07% on the 6-class Pan dataset).
Limitations and Future Work
The authors openly acknowledge several limitations in their current framework:
- Inference Latency: The Think-then-Answer paradigm inherently introduces higher inference latency due to the generation of lengthy reasoning chains, making it challenging for real-time applications. Future work will explore model distillation and early-exit mechanisms.
- Text-Only Modality: REVEAL currently cannot process multimodal inputs (images, audio, video). Extending the framework to support cross-modal reasoning mechanisms is a key next step.
- The Generative Moving Target: The rapid evolution of LLMs means detectors must continuously adapt. The authors suggest that future research should focus on continual learning strategies to incrementally update detectors with minimal retraining.
Summary updated