TraceDet

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

★ 10 Open GitHub ↗

Official implementation of TraceDet (ICLR 2026)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Figure 1: Overview of TraceDet. The framework detects hallucinations
 in D-LLMs via generation traces and entropy-based analysis.

# TraceDet: Tracing Hallucination Detection in Large Language Models

This repository contains the implementation of **TraceDet**, a comprehensive framework for detecting and analyzing hallucinations in diffusion Large Language Models (D-LLMs) through generation trace and entropy-based analysis.

## Project Overview

TraceDet focuses on:
- **Evaluating factuality** on benchmarks like TriviaQA, hotpotQA and CommensenseQA
- **Training classifiers** to predict hallucination probability at different stages
- **Detecting hallucinations** in LLM generation through hidden state analysis

---

## Directory Structure

```
TraceDet_publish/
├── train_main.py                 # Main training script for TimeHalu model
├── train_lstm_baseline.py        # LSTM baseline implementation
├── train_transformer_baseline.py # Transformer baseline
├── train_tuning.py               # Hyperparameter tuning
├── step_entropy_ave.py           # Entropy analysis utility
│
├── models/                       # Model architectures
│   └── TimeHalu.py               # Core TimeHalu model implementation
│
├── trainers/                     # Training utilities
│   └── train_TimeHalu.py         # Training loop for TimeHalu
│
├── DLLM_generate/                # D-LLM generation
│   ├── dream_entropy.py          # Generation from Dream models
│   ├── data_loader.py            # Data loading utilities
│   ├── data_generator.py         # Synthetic data generation
│   ├── generate.py               # Generation pipeline
│   ├── generate_interoutput.py   # Intermediate output generation
│   └── metric_qwen.py            # Qwen model evaluation metrics
│
├── environment.yml               # Conda environment specification
├── run_data.sh                   # Data processing script
├── qwen_eval.sh                  # Qwen model evaluation script
├── tuning.sh                     # Tuning execution script
└── triviaqa_test_resultlist_mask.json  # TruthfulQA test results
```

---

## Setup & Installation

### Prerequisites
- Python 3.10

### Environment Setup

```bash
# Create conda environment
conda env create -f environment.yml
conda activate python310

```

---

## Reproducibility Pipeline

All experiments in the paper follow a **three-step pipeline**.

---

### Step 1: Generate LLM Outputs & Entropy Traces

Run LLM generation and extract token-level entropy and intermediate hidden states.

```bash
bash generate_data.sh
```
This step corresponds to entropy trace signal genration from D-LLMs in the paper.

---

### Step 2: Hallucination Annotation

Automatically label hallucinations using Qwen-based evaluation.

```bash
bash qwen_eval.sh
```

### Step 3: Train TimeHalu Detector

Train the temporal hallucination detection model.

```bash
python train_main.py \
```

This step corresponds to the main structure of TraceDet, following previous work of TimeX++ on the temporal infromation bottelneck modeling

---

## Citation

If you use TraceDet in your research, please cite:

```bibtex
@article{chang2025tracedet,
  title={TraceDet: Hallucination Detection from the Decoding Trace of Diffusion Large Language Models},
  author={Chang, Shenxu and Yu, Junchi and Wang, Weixing and Chen, Yongqiang and Yu, Jialin and Torr, Philip and Gu, Jindong},
  journal={arXiv preprint arXiv:2510.01274},
  year={2025}
}
```
---
## Acknowledgements

This project is inspired by and builds upon prior open-source research efforts.
We sincerely thank the authors and contributors of TimeX++ and LLaDa for making their codebases publicly available, which greatly facilitated the development of this work.

---

## License

This project is provided as-is for research purposes.

---

## Contact

For questions or collaborations, please reach out to the project maintainers. For bug reports and feature requests, please open an issue or contact the maintainers.

TraceDet

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README