Official Comment · ICLR 2026

There are two aspects to cross-model generalization: (i) Can we train with different combinations of prompt generator and prompt evaluator models and (ii) after training, can we switch to a different prompt evaluator model without retraining. PRL is robust at both of these cross-model generalizations. Regarding (i), using different generator and evaluator models can be beneficial. In Table 6 in the paper, we have introduced a PRL-LLaMA variant that uses Llama3.1-8B-Instruct as the generator. The results are reproduced below (the first line corresponds to PRL-Llama in the paper, the second line to PRL): | Prompt Generator | Prompt Evaluator (Training) | Prompt Evaluator (Test) | Rouge-1 | Rouge-2 | Rouge-L | |------------------------|-----------------------|-----------------------|--------------|--------------|--------------| | LLama3.1-8B-Instruct | Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct | 43.70±0.02 | 16.66±0.25 | 37.46±0.5 | | Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct | 42.47±0.83 | 16.17±0.24 | 37.73±0.36 | | Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct | LLama3.1-8B-Instruct | 39.38±2.82 | 15.77±1.56 | 30.57±2.80 | These results show that PRL is flexible: different generator models can be used without degrading performance, and cross-model training can even yield improvements. Moreover, for this rebutall we crafted a new experiment in which we train with Qwen2.5-7B-Instruct as prompt generator and Llama3.1-8B-Instruct as evaluator on the summarization task. We present the results in the following table: | Prompt Generator: Qwen2.5-7B-Instruct | Prompt Evaluator: Llama3.1-8B-Instruct (Train/Test) | | ------------------------------------------- | --------------------------------------------------------- | | Method | ROUGE-1 | ROUGE-2 | ROUGE-L | | ------- | ----------------- | ----------------- | ----------------- | | MI | 28.92 | 9.29 | 25.09 | | APE | 38.47±0.66 | 12.96±0.14 | 33.17±0.22 | | GA | 37.48±2.65 | 13.92±1.14 | 33.48±2.54| | **PRL** | **40.06±0.02** | **13.98±0.12** | **34.06±0.01** | | Regarding (ii), we also see at the table above that prompts produced by PRL using Qwen2.5-7B-Instruct (as both generator and evaluator during training) can be directly applied to Llama3.1-8B-Instruct for evaluation during test time. These transferred prompts achieve state-of-the-art results on the summarization task, demonstrating that PRL produces general prompts that remain effective across evaluation models.