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-AI collaborative text

TL;DR

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

Background & Motivation

Background: With the rapid improvement in the quality of LLM-generated content, AI-generated text has become difficult to distinguish from human-written text in many scenarios. Detecting AI-generated content has become a critical requirement for combating misinformation, academic fraud, and the spread of spam.

Limitations of Prior Work: (1) Document-level detection methods lack reliability on human-AI collaborative text (partially human-written and partially AI-generated); (2) Existing span-level detection methods are primarily based on token-level sequence labeling (BIO), requiring manual post-processing to aggregate tokens into continuous spans, and are limited by sentence boundaries and fixed granularity; (3) The development of detection methods lags behind the progress of generative models.

Key Challenge: There is a need to simultaneously solve document-level classification ("Is this article AI-generated?") and span-level localization ("Which specific segments are AI-generated?"), where representations should be shared between both tasks to improve efficiency.

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

Key Insight: Analogizing AI-generated text spans to "objects" in images, the mature DETR architecture from visual object detection is utilized for end-to-end 1D span detection, transferring the robustness of visual detection to the NLP domain.

Method

Overall Architecture

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

Key Designs

1. Object-Centric Span Localization:

   • Function: Performs end-to-end detection and localization of continuous AI-generated text spans as 1D "objects."
   • Mechanism: Token embeddings \(\mathbf{E}\) are obtained from the fine-tuned LLM and processed via linear projection and a Transformer encoder to get contextual features \(\mathbf{R}\). \(N\) learnable anchor queries \((c, w)\) (center and width) are iteratively refined through a Transformer decoder, with each layer predicting offsets \((\Delta c, \Delta w)\). The final output is a triplet \((c, w, p)\)—center position, width, and confidence—where all values are normalized to a character-level interval of \([0,1]\).
   • Design Motivation: Sequence labeling methods requires manual post-processing to aggregate tokens into spans, whereas DETR directly regresses continuous intervals, eliminating the reliance on heuristic post-processing. Character-level rather than token-level localization is more flexible and independent of the tokenizer.

2. Unified Text-Representation Backbone:

   • Function: Provides shared, high-quality text embeddings for both detection and classification tasks.
   • Mechanism: Use LoRA to fine-tune Mistral-7B, training on proxy tasks: a three-class classification (human/AI/collaborative) for frozen feature extraction (for DETR), and binary classification (human/AI) joint-training with the classification head. LoRA keeps pre-trained weights frozen and only trains low-rank matrices.
   • Design Motivation: Datasets are typically small; LoRA generalizes better on small data and trains faster. The shared backbone validates the generalization capability of the embeddings.

3. DN-DAB-DETR Adaption:

   • Function: Stabilizes training and improves localization precision.
   • Mechanism: Adopts DAB-DETR's learnable anchor boxes as positional queries and DN-DETR's denoising training strategy (simultaneously training learnable queries and noisy GT queries). Hungarian matching is used for prediction-GT pairing.
   • Design Motivation: DN-DAB-DETR has demonstrated the best 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 + Focal Loss, calculated for both 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	Description
TuringBench (FAIR)	High Precision	RoBERTa-based methods	Unified backbone achieves strong classification performance
TweepFake	High Precision	-	Validation in the tweet domain
MAGE	High Precision	-	Large-scale validation across multiple generators and domains

Main Results (Span Detection)

Dataset	GigaCheck (DETR)	Previous Methods	Description
RoFT	Strong Localization Accuracy	Sequence labeling methods	Single-boundary scenarios
RoFT-ChatGPT	Strong Localization Accuracy	-	ChatGPT generation scenarios
TriBERT	Strong Localization Accuracy	-	Multi-boundary (1-3) scenarios

Key Findings

• The DETR architecture can be successfully generalized from visual space to text space, proving the feasibility of the object detection paradigm in NLP.
• The same fine-tuned backbone performs excellently on both classification and detection tasks, verifying that the learned embeddings possess strong generalization capabilities.
• End-to-end span detection eliminates the need for heuristic post-processing required in sequence labeling methods.
• LoRA-based parameter-efficient fine-tuning is particularly effective on small datasets.

Highlights & Insights

• Novelty: Redefining text span detection as a 1D object detection problem is an elegant and effective cross-domain transfer.
• Unified Framework: A single fine-tuned backbone serves both detection and classification, which is not only efficient but also validates the universality of the embeddings.
• End-to-End Design: DETR directly outputs character-level intervals, avoiding the cumbersome process of BIO labeling and post-processing.
• Model Agnostic: The backbone can be replaced with any decoder-only LLM, giving the framework good scalability.
• Value: The complete code is publicly released, promoting reproducibility.

Limitations & Future Work

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

Related Work & Insights

• vs Sci-SpanDet: Sci-SpanDet relies on IMRaD document structures for scientific paper detection; GigaCheck is domain-agnostic and applicable to any text.
• vs Sequence Labeling (BIO): Sequence labeling requires manual aggregation of tokens into spans; GigaCheck directly regresses continuous 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; significant paradigm innovation.
• Experimental Thoroughness: ⭐⭐⭐⭐ Double validation across 3 classification and 3 detection benchmarks, though lacks testing on the latest LLMs.
• Writing Quality: ⭐⭐⭐⭐ Clear architecture diagrams, rigorous method descriptions, and appropriate cross-modal analogies.
• Value: ⭐⭐⭐⭐ Provides a new technical route for AI-generated text detection; open-sourcing enhances impact.

Related Papers

• [ACL 2026] Temporal Flattening in LLM-Generated Text: Comparing Human and LLM Writing Trajectories
• [NeurIPS 2025] Can LLMs Write Faithfully? An Agent-Based Evaluation of LLM-generated Islamic Content
• [ACL 2026] Beyond the Final Actor: Modeling the Dual Roles of Creator and Editor for Fine-Grained LLM-Generated Text Detection
• [ACL 2026] DetectRL-X: Towards Reliable Multilingual and Real-World LLM-Generated Text Detection
• [ACL 2026] Who Wrote This Line? Evaluating the Detection of LLM-Generated Classical Chinese Poetry