GigaCheck: Detecting LLM-generated Content via Object-Centric Span Localization¶

Conference: ACL 2026 arXiv: 2410.23728 Code: GitHub Area: Object Detection Keywords: LLM-generated text detection, object detection paradigm, DETR, text span localization, human-machine collaborative text

TL;DR¶

GigaCheck is proposed as a dual-strategy framework: document-level classification via fine-tuned LLM, and span-level detection that innovatively treats AI-generated text spans as "objects," employing a DETR-like architecture for end-to-end character-level localization.

Background & Motivation¶

Background: With the rapid improvement in LLM-generated content quality, AI-generated text has become increasingly indistinguishable from human-written text in many contexts. Detecting AI-generated content has emerged as a critical need for combating misinformation, academic fraud, and spam proliferation.

Limitations of Prior Work: (1) Document-level detection methods are insufficiently reliable on human-machine collaborative text (partially human-written, partially machine-written); (2) Existing span-level detection methods primarily rely on token-level sequence labeling (BIO), requiring manual post-processing to aggregate tokens into contiguous spans, and are constrained by sentence boundaries and fixed granularity; (3) Detection methods lag behind the advancement of generative models.

Key Challenge: There is a need to simultaneously address document-level classification ("Is this article AI-generated?") and span-level localization ("Which specific passages are AI-generated?"), with shared representations between the two tasks to improve efficiency.

Goal: To propose a unified framework capable of both high-accuracy document-level detection and precise localization of AI-generated text spans.

Core Idea: AI-generated text spans are analogized to "objects" in images, leveraging the well-established DETR architecture from visual object detection for end-to-end 1D span detection, transferring the robustness of visual detection to the text domain.

Method¶

Overall Architecture¶

GigaCheck adopts a shared backbone with dual-head architecture: (1) Unified Backbone: LoRA-fine-tuned Mistral-7B provides text embeddings; (2) DETR Head: treats the embedding sequence as a 1D feature map and uses DN-DAB-DETR architecture to detect AI-generated spans; (3) Classification Head: applies an MLP on the hidden state of the final EOS token for document-level binary classification. Both heads can be used independently while sharing the fine-tuned backbone.

Key Designs¶

Object-Centric Span Localization:
- Function: Treats contiguous AI-generated text spans as 1D "objects" for end-to-end detection and localization.
- Mechanism: Token embeddings \(\mathbf{E}\) are obtained from the fine-tuned LLM, then projected linearly and passed through a Transformer encoder to yield contextual features \(\mathbf{R}\). \(N\) learnable anchor queries \((c, w)\) (center and width) are iteratively refined through the Transformer decoder, predicting offsets \((\Delta c, \Delta w)\) at each layer. The final output is a triple \((c, w, p)\)—center position, width, and confidence—all normalized to a character-level interval of \([0,1]\).
- Design Motivation: Sequence labeling methods require manual post-processing to aggregate tokens into spans, whereas DETR directly regresses contiguous intervals, eliminating the need for heuristic post-processing. Character-level localization, as opposed to token-level, is more flexible and tokenizer-agnostic.
Unified Text-Representation Backbone:
- Function: Provides shared high-quality text embeddings for both detection and classification tasks.
- Mechanism: Mistral-7B is fine-tuned with LoRA using a proxy task: three-class classification (human-written / machine-written / collaborative) for frozen feature extraction (used by DETR), and two-class classification (human-written / machine-written) jointly trained with the classification head. LoRA keeps the pre-trained weights frozen, training only the low-rank matrices.
- Design Motivation: Datasets are typically small; LoRA generalizes better on small data and trains faster. The shared backbone validates the generalizability of the learned embeddings.
DN-DAB-DETR Adaptation:
- Function: Stabilizes training and improves localization accuracy.
- Mechanism: Employs DAB-DETR's learnable anchor boxes as positional queries and DN-DETR's denoising training strategy (simultaneously training learnable queries and noise-augmented GT queries). Hungarian matching is used for prediction–GT assignment.
- Design Motivation: DN-DAB-DETR demonstrates superior localization accuracy and training stability in visual detection, outperforming DAB-DETR, Deformable DETR, and CO-DETR.

Loss & Training¶

The training loss is a weighted sum of L1, gIoU, and Focal Loss, computed separately for Hungarian-matched predictions and denoising GT queries. The classification head uses binary cross-entropy. During DETR training, the backbone is frozen; during classification head training, the backbone is trainable.

Key Experimental Results¶

Main Results (Classification)¶

Dataset	GigaCheck (Mistral-7B)	Prev. SOTA	Notes
TuringBench (FAIR)	High accuracy	RoBERTa-based methods	Strong classification with unified backbone alone
TweepFake	High accuracy	—	Validated on tweet domain
MAGE	High accuracy	—	Large-scale validation across multiple generators and domains

Main Results (Span Detection)¶

Dataset	GigaCheck (DETR)	Prior Methods	Notes
RoFT	Strong localization	Sequence labeling methods	Single-boundary scenario
RoFT-ChatGPT	Strong localization	—	ChatGPT generation scenario
TriBERT	Strong localization	—	Multi-boundary (1–3) scenario

Key Findings¶

The DETR architecture can be successfully generalized from the visual domain to the text domain, demonstrating the feasibility of the object detection paradigm in NLP.
The same fine-tuned backbone performs strongly on both classification and detection tasks, validating the strong generalizability of the learned embeddings.
End-to-end span detection eliminates the need for heuristic post-processing in sequence labeling approaches.
LoRA-based parameter-efficient fine-tuning is particularly effective on small datasets.

Highlights & Insights¶

Paradigm Innovation: Reframing text span detection as a 1D object detection problem represents an elegant and effective cross-domain transfer.
Unified Framework: A single fine-tuned backbone serves both detection and classification, achieving efficiency while validating embedding generality.
End-to-End Design: DETR directly outputs character-level intervals, avoiding the cumbersome pipeline of BIO labeling and post-processing.
Model Agnosticism: The backbone can be replaced with any decoder-based LLM, conferring good extensibility to the framework.
Open-Source Contribution: Full code is publicly released, promoting reproducibility.

Limitations & Future Work¶

Evaluation is currently limited to English text; multilingual adaptation is an important future direction.
Generators in the detection datasets are primarily earlier models (GPT-2/3, CTRL); detection performance on the latest LLMs remains unknown.
The number of DETR queries \(N\) must be preset per dataset and cannot adapt dynamically.
Robustness analysis against adversarial attacks (e.g., paraphrasing, watermark removal) is insufficient.
Future work may explore multi-granularity detection (joint paragraph-level, sentence-level, and word-level detection).

vs. Sci-SpanDet: Sci-SpanDet relies on IMRaD document structure for scientific paper detection; GigaCheck is domain-agnostic and applicable to arbitrary text.
vs. Sequence Labeling Methods (BIO): Sequence labeling requires manual aggregation of tokens into spans; GigaCheck directly regresses contiguous intervals.
vs. Statistical Methods (DetectGPT, etc.): Statistical methods require access to the probability distribution of the target LLM; GigaCheck has no such requirement.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ First application of DETR to text span localization; the paradigm innovation carries substantial significance.
Experimental Thoroughness: ⭐⭐⭐⭐ Dual validation across 3 classification and 3 detection benchmarks, though testing on the latest LLMs is lacking.
Writing Quality: ⭐⭐⭐⭐ Architecture diagrams are clear, method descriptions are rigorous, and the cross-modal analogy is apt.
Value: ⭐⭐⭐⭐ Provides a new technical direction for AI-generated text detection; open-source release enhances impact.