Research map / Large language models / Alignment and preferences
LLM Evaluators Recognize and Favor Their Own Generations
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.
Shows LLM evaluators (GPT-4, GPT-3.5, Llama 2) can partly recognize their own summaries and that fine-tuning for self-recognition yields a linear correlation with self-preference bias, with controls for confounders.
Problem
Self-preference bias in LLM self-evaluation: whether LLMs favor their own outputs because they recognize them as their own (self-recognition), rather than by coincidence or quality.
Approach
Treats self-recognition as a measurable capability separate from self-preference, then fine-tunes LLMs to raise or lower self-recognition (plus unrelated control tasks) and measures the resulting linear correlation with self-preference on news summarization.
Key results
- GPT-4 reaches 73.5% accuracy distinguishing itself from two other LLMs and humans out of the box in pairwise setting.
- GPT-3.5 and Llama 2 both exceed 90% self-recognition accuracy after fine-tuning on 500 examples.
- Self-preference strength shows a linear correlation with self-recognition across fine-tuned models, and control-task fine-tunes fall on the same trend.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune GPT-3.5 on 500 self-recognition examples | GPT-3.5 evaluator · XSUM | self-preference: 0.582 -> 0.898 | Helped |
| Fine-tune GPT-3.5 on 500 self-recognition examples | GPT-3.5 evaluator · CNN | self-preference: 0.431 -> 0.97 | Helped |
| Fine-tune GPT-3.5 on 500 self-recognition examples from XSUM | GPT-3.5 evaluator · CNN, out-of-domain | self-preference: 0.431 -> 0.75 | Helped |
| Fine-tune Llama-2-7b on 500 self-recognition examples | Llama-2-7b evaluator · XSUM | self-preference: 0.511 -> 0.485 | Hurt |
| Fine-tune Llama-2-7b on 500 self-recognition examples from XSUM | Llama-2-7b evaluator · CNN, out-of-domain | self-preference: 0.505 -> 0.788 | Helped |
| Fine-tune Llama-2-7b on 500 self-recognition examples from CNN | Llama-2-7b evaluator · CNN | self-preference: 0.505 -> 0.5 | No clear effect |
| Fine-tune GPT-3.5 on 10 self-recognition examples from XSUM (ablation) | GPT-3.5 evaluator · XSUM | self-preference: 0.582 -> 0.657 | Helped |
| Fine-tune GPT-3.5 on 10 self-recognition examples from XSUM (ablation) | GPT-3.5 evaluator · CNN, out-of-domain | self-preference: 0.431 -> 0.421 | No clear effect |
| Fine-tune Llama-2-7b on 10 self-recognition examples from XSUM (ablation) | Llama-2-7b evaluator · XSUM | self-preference: 0.511 -> 0.665 | Helped |
| Fine-tune Llama-2-7b on 10 self-recognition examples from XSUM (ablation) | Llama-2-7b evaluator · CNN, out-of-domain | self-preference: 0.505 -> 0.81 | Helped |
| Fine-tune on length, vowel-count, readability, constant-answer, or random tasks instead of self-recognition (ablation) | GPT-3.5 and Llama-2-7b evaluators · XSUM and CNN | self-preference | Mixed |
| Swap the source labels in the pairwise preference prompt (ablation) | GPT-4 and GPT-3.5 evaluators · XSUM | self-preference: 0.73 -> 0.32; 0.61 -> 0.46 | Mixed |
Surprising findings
- Llama 2 is completely incapable of distinguishing itself from GPT-3.5 and GPT-4, and GPT-4 does not distinguish Llama 2 summaries from its own more easily than GPT-3.5 summaries.
- In the individual setting GPT-4 is strongly biased against predicting any text as its own, regardless of the true author.
Limitations
- Causal hypothesis is only supported by correlational evidence and control experiments, not validated mechanistically.
- Limited to text summarization on two datasets and a few models.
- Self-preference not controlled against ground-truth generation quality.
Experimental setup
Datasets: XSUM CNN/DailyMail
Benchmarks: XSUM CNN/DailyMail
Metrics: self-recognition score self-preference score Kendall's tau accuracy
Baselines: Llama-2-7b-chat GPT-3.5 GPT-4
Builds on: LLM-as-a-judge calibration of verbalized uncertainty
Closest papers by meaning
- SELF-[IN]CORRECT: LLMs Struggle with Discriminating Self-Generated Responses (AAAI 2025 · similarity 0.89)
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.89)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.88)
- Evaluating the Evaluator: Measuring LLMs’ Adherence to Task Evaluation Instructions (AAAI 2025 · similarity 0.87)
- On Evaluating LLM Alignment by Evaluating LLMs as Judges (NeurIPS 2025 · similarity 0.87)
- Accelerating Unbiased LLM Evaluation via Synthetic Feedback (ICML 2025 · similarity 0.87)
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.