Research map / Large language models / Language, safety and interpretability
Zero-Shot Detection of LLM-Generated Text via Implicit Reward Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes IRM, a zero-shot LLM-generated text detector that scores texts by the implicit reward log(pi_instruct/pi_base) from publicly available instruction-tuned and base model pairs, requiring no training.
Problem
Zero-shot detection of LLM-generated text suffers when the source LLM is unknown and proxy models are used; existing reward-based detectors (ReMoDetect) need preference data and task-specific fine-tuning and generalize poorly to unseen LLMs.
Approach
Use the implicit reward from DPO's closed-form relation, log(pi_instruct(y)/pi_base(y)) computed from a public instruction-tuned/base model pair, as a training-free detection score, avoiding preference collection and fine-tuning.
Key results
- With the Llama-3.2-1B family, IRM reaches an average score of 91.77% on DetectRL, versus 87.67% for Binoculars and 85.86% for the supervised ReMoDetect.
- With the Gemma-2-2B family, IRM averages 77.05% on DetectRL, slightly below Binoculars at 77.81%.
- On the length-generalization sub-task, IRM with the Gemma-2-2B family drops significantly (60.30 train / 58.15 test F1), as the optimal threshold shifts with text length.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use implicit reward scores from the Llama-3.2-1B family instead of Binoculars | LLM-generated text detection · DetectRL | average score: 87.67% to 91.77% | Helped |
| Use implicit reward scores from the Gemma-2-2B family instead of Binoculars | LLM-generated text detection · DetectRL | average score: 77.81% to 77.05% | Hurt |
| Use the Llama-3.2-1B family instead of the Gemma-2-2B family (ablation) | IRM · DetectRL | average score: 77.05% to 91.77% | Helped |
| Use Qwen2 instead of Qwen2.5 at 0.5B (ablation) | IRM · DetectRL | average score: 73.24% to 82.79% | Helped |
| Use the 0.5B rather than 1.5B Qwen2.5 model (ablation) | IRM · DetectRL | average score: 69.69% to 73.24% | Helped |
| Use Gemma-2B rather than Gemma-2-2B (ablation) | IRM · DetectRL | average score: 77.05% to 83.72% | Helped |
| Scale Llama from 1B to 3B (ablation) | IRM · DetectRL | average score: 91.77% to 91.08% | Hurt |
| Use the instruction-tuned model instead of the base model for Log-Likelihood (ablation) | Log-Likelihood · DetectRL | average score: 66.44% to 77.46% | Helped |
| Use the instruction-tuned model instead of the base model for Fast-DetectGPT (ablation) | Fast-DetectGPT · DetectRL | average score: +3.98 | Hurt |
| Use the Llama-3.2-1B-family IRM rather than ReMoDetect | LLM-generated text detection · DetectRL | average score: 85.86% to 91.77% | Helped |
| Use Llama-3.2-1B-family IRM rather than Binoculars | LLM-generated text detection · DivScore | F1: 93.96 to 96.61 | Helped |
| Use Llama-3.2-1B-family IRM rather than ReMoDetect | LLM-generated text detection · DivScore | AUROC: 98.47 to 98.42 | Hurt |
Surprising findings
- Instruction-tuned models improved most zero-shot baselines compared to base models (5 of 6 methods), suggesting aligned models are better detectors.
- Reward models from RewardBench (GRM, RM-Deberta) and Skywork-Reward-V2 perform poorly on detection, while the implicit reward from the same family works well.
Limitations
- Only lightweight models (about 0.5B to 3B) were evaluated due to compute constraints.
- Requires an instruction-tuned and base model pair from the same family; not feasible when only the instruction-tuned model is available.
- Length-dependent threshold: the summed reward shifts with text length, hurting cross-length generalization for the Gemma-2-2B family.
Experimental setup
Datasets: DetectRL RAID DivScore
Benchmarks: DetectRL RAID DivScore
Metrics: AUROC F1
Baselines: Log-Likelihood Rank Log-Rank LRR Fast-DetectGPT Lastde Binoculars ReMoDetect RM-Deberta-v3-large-v2 GRM-gemma2-2B GRM-Llama3.2-3B Skywork-Reward-V2
Builds on: DPO ReMoDetect Binoculars Fast-DetectGPT DetectRL
Closest papers by meaning
- Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection (ICLR 2025 · similarity 0.90)
- Training-free LLM-generated Text Detection by Mining Token Probability Sequences (ICLR 2025 · similarity 0.89)
- DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios (NeurIPS 2024 · similarity 0.88)
- ReMoDetect: Reward Models Recognize Aligned LLM's Generations (NeurIPS 2024 · similarity 0.88)
- DALD: Improving Logits-based Detector without Logits from Black-box LLMs (NeurIPS 2024 · similarity 0.88)
- Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.