Direct Discrepancy Learning (DDL)

• Direct Discrepancy Learning (DDL) is an optimization strategy that directly aligns the discrepancy score with the separation of human-written and machine-generated text.
• It employs a normalized conditional discrepancy and adaptive reference clustering to improve detectors’ robustness across diverse models and domains.
• Benchmark results show that DDL enhances performance and computational efficiency compared to traditional training-based and zero-shot detection methods.

Direct Discrepancy Learning (DDL) is an optimization strategy for machine-generated text detection introduced as the core training idea of DetectAnyLLM. It is designed for detectors that score a text by a discrepancy induced by a scoring LLM and its perturbations, and it is presented as a remedy for the mismatch between standard training objectives and the actual human-written text (HWT) versus machine-generated text (MGT) decision boundary. In this formulation, DDL does not train the scorer to remain close to a reference LLM; instead, it directly shapes the discrepancy statistic used at inference, driving HWT toward discrepancy $0$ and MGT toward a positive margin $\gamma$, with final decisions made by a reference-clustering rule. The acronym is not uniform across the literature and must be distinguished from unrelated uses such as Differential Description Length and Dynamic Discrepancy Learning (Fu et al., 15 Sep 2025).

1. Discrepancy as the detection primitive

In DetectAnyLLM, the starting point is a discrepancy score derived from a scoring LLM. For a text $x$, the basic discrepancy is

$$d(x,f_\theta,q_\phi) = \log f_\theta(x) - \mathbb E_{\tilde{x}\sim q_\phi(\cdot|x)}[\log f_\theta(\tilde{x})],$$

where $f_\theta$ is the scoring model and $q_\phi$ generates perturbed samples. The operational intuition given for MGTD is that machine-generated text tends to contain high-probability tokens, so its original log-probability stays high while perturbations hurt it, yielding a large positive discrepancy; human text tends to have a smaller discrepancy. This yields the expected ordering

$d(x_m,f_\theta,q_\phi) > d(x_h,f_\theta,q_\phi).$

To make the detector efficient, Fast-DetectGPT replaces full perturbation likelihood with a conditional probability approximation,

$$f_\theta(\tilde{x}) = \prod_i f_\theta(\tilde{x}_i|\tilde{x}_{<i}) \sim \prod_i f_\theta(\tilde{x}_i|x_{<i}) = f_\theta(\tilde{x}|x),$$

and DetectAnyLLM uses the normalized conditional discrepancy

$$d_c(x,f_\theta,q_\phi) = \frac{\log f_\theta(x|x) - \tilde{\mu}}{\tilde{\sigma}},$$

with

$$\tilde{\mu} = \mathbb E_{\tilde{x}\sim q_\phi(\tilde{x}|x)}[\log f_\theta(\tilde{x}|x)].$$

The normalization term $\gamma$0 is explicitly part of the detector score, and the paper later reports that removing this normalization substantially degrades performance. This makes discrepancy not merely a heuristic statistic but the central geometric quantity around which both training and inference are organized (Fu et al., 15 Sep 2025).

2. Objective alignment and the formal DDL loss

The principal claim behind DDL is that prior training-based detectors are misaligned with the MGTD task because they inherit objectives from language-model fine-tuning or preference learning rather than directly optimizing the HWT/MGT separation. The paper specifically criticizes DPO-style training, arguing that its implicit KL regularization keeps the scoring model anchored to the original LLM or reference model. In the MGTD setting, the stated goal is not to preserve language-model behavior but to make the model “learn to be a detector rather than another LLM.”

DDL removes that regularization and directly optimizes the discrepancy score used at test time. The reward is defined as

$\gamma$1

Human-written text is therefore pushed toward discrepancy $\gamma$2, while machine-generated text is pushed toward a positive target margin $\gamma$3. In practice, the paper sets $\gamma$4, following ImBD’s training setup, which yields the actual objective

$\gamma$5

This is the formal DDL objective. Its defining feature is that the detector is optimized on the statistic it will actually use at inference. The authors present this as the central mechanism by which DDL improves robustness and generalization across domains, models, and generation tasks (Fu et al., 15 Sep 2025).

3. Integration into DetectAnyLLM

DetectAnyLLM retains Fast-DetectGPT’s three-stage structure: Re-sampling / perturbation generation, Discrepancy calculation, and Final decision. DDL is used to train the scoring model so that the perturbation-induced discrepancy becomes more discriminative.

The final decision rule is called reference clustering. After training, the method constructs two reference sets, $\gamma$6 for machine-generated texts and $\gamma$7 for human-written texts. For each reference text it computes $\gamma$8, producing empirical distributions $\gamma$9 and $x$0. For a test text $x$1, it then defines an adaptive window

$x$2

Within that window, it counts how many reference discrepancies belong to the machine and human sets:

$x$3

$x$4

The estimated probability that $x$5 is machine-generated is then

$x$6

The paper claims that this adaptive-window procedure is more stable across varying data density than a fixed threshold. In the full framework, DDL thus determines the discrepancy geometry, while reference clustering converts that geometry into a probabilistic decision head (Fu et al., 15 Sep 2025).

4. Benchmark design and reported empirical performance

The empirical evaluation is centered on MIRAGE, described as the most diverse multi-task MGTD benchmark in the paper. MIRAGE samples human-written texts from 10 corpora across 5 text-domains—Academic, Email, Website, News, Comment—and then re-generates or revises them using 17 LLMs total, including 13 proprietary models and 4 open-source models. It evaluates three tasks—Generate, Polish, Rewrite—under two settings: disjoint-input generation (DIG) and shared-input generation (SIG).

On MIRAGE-DIG, DetectAnyLLM reports the following results: Generate: AUROC 0.9525, MCC 0.7975, TPR@5% 0.7770; Polish: AUROC 0.9297, MCC 0.7487, TPR@5% 0.7756; Rewrite: AUROC 0.9234, MCC 0.7447, TPR@5% 0.7778. On MIRAGE-SIG, it reports Generate: AUROC 0.9526, Balanced Accuracy 0.9059, MCC 0.8119, TPR@5% 0.7722; Polish: AUROC 0.9316, Balanced Accuracy 0.8740, MCC 0.7483, TPR@5% 0.7779; Rewrite: AUROC 0.9158, Balanced Accuracy 0.8643, MCC 0.7320, TPR@5% 0.7574.

The abstract summarizes the overall effect as “over a 70% performance improvement under the same training data and base scoring model.” The paper further states that DetectAnyLLM consistently outperforms zero-shot methods such as Likelihood, LogRank, Entropy, LRR, NPR, DetectGPT, and Fast-DetectGPT, as well as training-based baselines including RoBERTa-Base/Large and ImBD. It also reports strong performance across specific generator and domain combinations involving models such as GPT-4o, GPT-4o-mini, GPT-o3-mini, DeepSeek-R1/V3, Claude-3.5/3.7, Gemini-2.0, Qwen2.5, Llama 3.1, Grok2, Moonshot-v1, and Doubao-1.5-pro, and interprets this as evidence for cross-model generalization rather than memorization of a training generator (Fu et al., 15 Sep 2025).

5. Ablations, robustness, and computational profile

The paper includes several analyses intended to isolate DDL’s contribution. A central ablation varies the DDL margin $x$7. On MIRAGE-DIG polish, performance changes from AUROC 0.5481 at $x$8 to 0.8692 at $x$9, 0.9257 at $$d(x,f_\theta,q_\phi) = \log f_\theta(x) - \mathbb E_{\tilde{x}\sim q_\phi(\cdot|x)}[\log f_\theta(\tilde{x})],$$0, 0.9360 at $$d(x,f_\theta,q_\phi) = \log f_\theta(x) - \mathbb E_{\tilde{x}\sim q_\phi(\cdot|x)}[\log f_\theta(\tilde{x})],$$1, and 0.9377 at $$d(x,f_\theta,q_\phi) = \log f_\theta(x) - \mathbb E_{\tilde{x}\sim q_\phi(\cdot|x)}[\log f_\theta(\tilde{x})],$$2; for $$d(x,f_\theta,q_\phi) = \log f_\theta(x) - \mathbb E_{\tilde{x}\sim q_\phi(\cdot|x)}[\log f_\theta(\tilde{x})],$$3, AUROC is reported to stabilize around 0.9259–0.9377 depending on the reporting slice. The authors choose $$d(x,f_\theta,q_\phi) = \log f_\theta(x) - \mathbb E_{\tilde{x}\sim q_\phi(\cdot|x)}[\log f_\theta(\tilde{x})],$$4 in the main experiments because performance has already saturated and becomes insensitive to further increase.

Normalization is reported to matter strongly. On MIRAGE-DIG polish, using normalization yields AUROC 0.9297, MCC 0.7487, TPR@5% 0.7756; removing it yields AUROC 0.8499, MCC 0.5563, TPR@5% 0.5232. This is used to support the claim that $$d(x,f_\theta,q_\phi) = \log f_\theta(x) - \mathbb E_{\tilde{x}\sim q_\phi(\cdot|x)}[\log f_\theta(\tilde{x})],$$5 helps standardize discrepancy values across heterogeneous domains and models.

DDL is also reported to be computationally cheaper than SPO-style training because it does not need a separate reference model during training. On GPT-J-6B, the paper reports SPO: 166s/epoch, 31.45GB memory and DDL: 116s/epoch, 20.16GB memory, corresponding to about a 30.12% reduction in time and 35.90% reduction in memory. A further study of the implicit KL strength in DPO-style training finds that larger $$d(x,f_\theta,q_\phi) = \log f_\theta(x) - \mathbb E_{\tilde{x}\sim q_\phi(\cdot|x)}[\log f_\theta(\tilde{x})],$$6 degrades detection and collapses the discrepancy gap between human and machine text, which is presented as empirical support for removing the KL term. Model-size ablations on MIRAGE-SIG Rewrite show DDL outperforming SPO across Qwen2-0.5B (0.9370 vs 0.8570), GPT-Neo-2.7B (0.9158 vs 0.7694), and GPT-J-6B (0.8909 vs 0.8367) (Fu et al., 15 Sep 2025).

6. Terminological distinctions and neighboring discrepancy-based methods

The literature does not use “DDL” as a stable acronym. In technical writing, the expansion must therefore be stated explicitly.

Term	Paper	Meaning
Direct Discrepancy Learning	DetectAnyLLM (Fu et al., 15 Sep 2025)	MGTD objective that directly optimizes the discrepancy statistic
Differential Description Length	"Differential Description Length for Hyperparameter Selection in Machine Learning" (Abolfazli et al., 2019)	model/hyperparameter selection by differences of description lengths
Dynamic Discrepancy Learning	"Learning Contrastive Self-Distillation for Ultra-Fine-Grained Visual Categorization Targeting Limited Samples" (Fang et al., 2023)	feature-level contrastive learning with a dynamic memory queue
Discrepancy-based Self-supervised LeArning (D-SLA)	"Graph Self-supervised Learning with Accurate Discrepancy Learning" (Kim et al., 2022)	graph SSL that learns discrepancy between original and perturbed graphs
Discrepancy Reduction	"Learning to Adapt by Minimizing Discrepancy" (II et al., 2017)	local learning in a temporal recurrent generative model via prediction–correction mismatch

The distinction from Differential Description Length is especially important. In (Abolfazli et al., 2019), DDL means a model and hyperparameter selection criterion built from differences of description lengths of the data, under a log-loss interpretation of codelength. That method estimates generalization error from the training data alone and is unrelated to MGTD. Likewise, the Dynamic Discrepancy Learning module in CSDNet stores historical intermediate features in a dynamic queue for supervised contrastive learning in ultra-fine-grained visual categorization, and it is used only during training (Fang et al., 2023).

The broader discrepancy-learning theme nevertheless recurs across fields. D-SLA for graphs replaces the contrastive assumption that two perturbations of the same graph should be similar with an objective that distinguishes original from perturbed graphs and calibrates embedding distance to graph edit distance (Kim et al., 2022). Earlier, Discrepancy Reduction in the Temporal Neural Coding Network used local error units and corrected latent targets to update parameters directly from prediction–correction mismatch, avoiding backpropagation through time (II et al., 2017). These methods do not define a single common algorithm, but they share an emphasis on discrepancy as the training signal rather than invariance alone or generic maximum-likelihood optimization.