MAGA-Bench: Machine-Augment-Generated Text via Alignment Detection Benchmark

TL;DR

This work introduces MAGA, a comprehensive benchmark and generation framework that augments machine-generated text with alignment techniques to stress-test and improve detector robustness. The MAGA dataset spans 20 English and Chinese domains with 12 diverse generators, and employs four alignment methods plus RLDF variants to produce more human-aligned machine text. Experiments show MAGA reliably degrades many detectors on MAGA data, while detectors fine-tuned on MAGA exhibit improved generalization on external datasets, validating the benchmark's utility for advancing detector robustness. The approach provides a practical pathway to evaluate and enhance detector generalization, informing future detector design and content-generation safeguards.

Abstract

Large Language Models (LLMs) alignment is constantly evolving. Machine-Generated Text (MGT) is becoming increasingly difficult to distinguish from Human-Written Text (HWT). This has exacerbated abuse issues such as fake news and online fraud. Fine-tuned detectors' generalization ability is highly dependent on dataset quality, and simply expanding the sources of MGT is insufficient. Further augment of generation process is required. According to HC-Var's theory, enhancing the alignment of generated text can not only facilitate attacks on existing detectors to test their robustness, but also help improve the generalization ability of detectors fine-tuned on it. Therefore, we propose \textbf{M}achine-\textbf{A}ugment-\textbf{G}enerated Text via \textbf{A}lignment (MAGA). MAGA's pipeline achieves comprehensive alignment from prompt construction to reasoning process, among which \textbf{R}einforced \textbf{L}earning from \textbf{D}etectors \textbf{F}eedback (RLDF), systematically proposed by us, serves as a key component. In our experiments, the RoBERTa detector fine-tuned on MAGA training set achieved an average improvement of 4.60\% in generalization detection AUC. MAGA Dataset caused an average decrease of 8.13\% in the AUC of the selected detectors, expecting to provide indicative significance for future research on the generalization detection ability of detectors.

Figures (14)

• Figure 1: More aligned MGT not only evades detection by existing detectors but also facilitates fine-tuning of neural-based detectors, which enhancing their generalization capability for wild detection.
• Figure 2: MAGA-Bench Overview. Our dataset construction contains 20 domains and 12 generators. We evaluate selected detectors on our dataset. We also adopted various decoding strategies, which are not presented here but detailed in §\ref{['sec:decode']}.
• Figure 3: MAGA Pipeline
• Figure 4: RLDF (Reinforce Learning from Detectors Feedback). Fine-tune the LLM via RL using detector D as the RM, resulting in harder-to-detect and better human-aligned generated text. Fine-tune detector D (RoBERTa) on the dataset constructed from HWT and its corresponding better-aligned MGT, to obtain a detector D with stronger generalized detection capability, thus enabling multi-round adversarial fine-tuning. RLDF-CD and RLDF-CM resolve RLDF’s practical infeasibility caused by RoBERTa overfitting by means of cross-reward.
• Figure 5: RLDF-CD matrix analysis for attack and generalization. The horizontal axis represents the RM domain, and the vertical axis represents the LLM domain.