The Preference Prediction shared task, part of the ELOQUENT Lab at CLEF-2025, challenges participants to build systems that simulate human judgment in evaluating pairs of LLM-generated responses. This novel task includes two sub-tasks. The first sub-task requires predicting which of two model responses humans prefer, using an accuracy metric. The second sub-task extends the first one by asking systems to provide natural language explanations for their predictions, automatically evaluated through standard NLG metrics and an LLM-as-a-judge approach. This overview paper summarizes the task design, evaluation methods, participation statistics, baseline performance, and lessons learned. We discuss the system performance in both sub-tasks. The results show room for improvement in generating coherent and human-like explanations, despite acceptable performance in the preference prediction.
Side-by-side evaluation has become a widely adopted paradigm for assessing how well large language
models (LLMs) align with human preferences across various natural language tasks. Generally, human
annotators compare two LLM-generated responses to a prompt and indicate which answer is the best
based on criteria such as coherence, helpfulness, factuality, and safety. To mitigate the high cost and
scalability challenges of collecting human judgments, recent research has focused on developing “judge”
models that can automatically predict human preferences with increasing accuracy. Notable eforts
include RewardBench [
Despite progress in preference prediction, the ability of LLMs to explain why a particular response
should be preferred remains underexplored. Recent work has highlighted that LLM judges may exhibit
unfairness or instability when tasked with evaluative reasoning [
The first year of the Preference Prediction 1 tests the capability of systems to predict human preferences for diferent outputs from LLMs and explain their predictions with respect to five criteria: relevance, naturalness, truthfulness, safety, and overall quality. This task ofers two sub-tasks: 2. Preference prediction & explanation generation. Predict human preferences between two
LLM responses with respect to the criteria and explain the system’s predictions.
GPT-4o Mistral-7B-IT Llama-3-70B-IT tulu-2-dpo-70b Claude 3.7 Sonnet
GPT-4 Mistral-7B-v0.3 Llama-3-70B Llama-2-70B N/A
OpenAI
Apache 2.0
LLaMA 3
Apache 2.0
Anthropic
Hurst et al. [
Jiang et al. [
Ivison et al. [8]
Anthropic [9] 1http://eloquent-lab.github.io/task-preference-prediction/
We conducted an in-house annotation to label the response pairs. The annotation team consisted of a team leader, with a near-native proficiency in English and a background in computational linguistics, and 15 native English-speaking annotators who had experience creating data sets for learning from human feedback and for explainable AI methods. The team leader manually checked the annotations and exchanged feedback with the annotators. The average number of tasks completed by an annotator in the final dataset was 87. The average pay rate was $15/hour.
The annotation task consisted of two sub-tasks: • Selecting the better of two LLM responses based on five evaluation criteria; • Providing a written explanation for each judgment.
The annotators were asked to evaluate each criterion independently, selecting the better response for that specific criterion without considering their judgments for other criteria. For each criterion, they could choose among four options: • A is better, • B is better, • Both are good, • Both are bad.
We used the following set of evaluation criteria: • Relevance: Which response better follows the prompt and completes the user’s request? • Naturalness: Which response is more human-like? • Truthfulness: Which response is more truthful? • Safety: Which response is less harmful? • Overall quality: Which response is best overall?
The prompt and responses were displayed side by side in a web-based interface, with radio buttons for each criterion and a text field for explanations.
Dataset and codebase Our shared task dataset is created as part of Primeape,2 a novel benchmark of human-annotated preferences and explanations for evaluating LLM judges. The dataset is served as the development and private test set.3 We ofer a baseline based on meta-llama/Llama-3.1-8B-Instruct,4 which is utilized as a judge LLM in a zero-shot regime. Our baseline and evaluation codebase is available in the ELOQUENT 2025 GitHub repository.5 Subtask 1 was evaluated using accuracy across five criteria: relevance, naturalness, truthfulness, safety, and overall quality. For each instance, systems were required to select the preferred response, and their predictions were compared to human-annotated preferences for each criterion. Subtask 2 extended Subtask 1 by requiring systems to also generate natural language explanations for their predictions. The explanation quality was assessed using multiple standard language generation evaluation metrics: BERTScore and ROUGE-L were used to measure lexical and semantic similarity with human-written reference explanations, while an LLM-as-a-judge framework provided a more interpretive assessment. The latter was implemented via a custom evaluation script that constructs a detailed prompt in which a large language model (LLM) is instructed to assess whether an AI-generated explanation aligns with the human rationale. The prompt includes the user instruction, outputs from two assistants, the AI system’s decision and explanation, and the human judgment. The LLM is asked to ignore irrelevant biases such as response order and length and to output a verdict in the form of [[Yes]] or [[No]]. These outputs are then used to compute alignment scores as the percentage of the system-generated explanations that align with the human-written explanations.
Subtask 1: Preference Prediction The first subtask evaluated systems on their ability to predict human preferences across five criteria: relevance, naturalness, truthfulness, safety, and overall quality. 2This is ongoing work on multilingual human preference prediction and explanation. The English subset is created as part of the Preference Prediction shared task. Data collection and annotation will be documented in detail in an upcoming paper. To be available at https://github.com/Toloka/primeape. 3https://huggingface.co/datasets/Eloquent/preference_prediction 4https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct 5https://github.com/eloquent-lab/eloquent-lab.github.io/tree/main See Table 3 for the subtask results. The top-performing system was submitted by VerbaNexAI [11], achieving the highest average accuracy of 56.99%, with particularly strong performance in truthfulness (75.16%) and safety (94.15%). The UTK team [12] secured second place with an average accuracy of 38.72%, outperforming the baseline in all categories, most notably in relevance and naturalness. The baseline model, based on simple prompting without fine-tuning, achieved an average accuracy of 29.70%. Although the FHS team submitted strong results (average accuracy of 53.89%), their submission was received after the deadline and was not included in the oficial ranking.
Team VerbaNexAI UTK Baseline
Avg. (Acc./ROUGE-L/BERTScore/LLM-Judge) 56.99 / 20.04 / 87.00 / 33.04 38.72 / 9.00 / 83.46 / 18.38 29.70 / 8.40 / 83.13 / 24.27
evaluation to include explanation generation, with systems assessed across four metrics: accuracy, ROUGE-L, BERTScore, and an LLM-as-a-judge evaluation (using GPT-4o). See Table 4 for the subtask results. VerbaNexAI again achieved top performance across all criteria, with an average accuracy of 56.99%, and strong explanation quality as reflected in high BERTScores (87.00%) and favorable LLMas-a-judge ratings (33.04). Their explanations were particularly strong in the safety and truthfulness categories. UTK placed second with a lower average accuracy (38.72%) and explanation metrics slightly behind VerbaNexAI, but ahead of the baseline. The baseline system performed reasonably in BERTScore (83.13%) but struggled in accuracy (29.70%) and LLM-as-a-judge scores (24.27). The final ranking was determined using a Borda count [13, 14], with VerbaNexAI achieving a top performance.
Subtask 1: Preference Prediction VerbaNexAI employed a lightweight in-context learning approach without any fine-tuning. Their system used meta- llama/Llama-3.2-1B-Instruct for generation and facebook/bart-large-cnn for summarization. To construct prompts, they used sentence embeddings generated by all-MiniLM-L6-v2 to retrieve the most similar examples from a training pool. These examples were summarized before being inserted into the final few-shot prompt. A custom system instruction guided the model in generating preference judgments and explanations. Postprocessing involved parsing model outputs with regular expressions, comparing predicted preferences to ground truth, and computing cosine similarity between explanation embeddings. The system was notable for its modular architecture, retrieval-augmented prompting, and compact model size (100M–1B parameters), achieving the best overall performance in both sub-tasks.
Team UTK used a fine-tuning approach based on unsloth/llama-3-8b-Instruct-bnb-4bit, applying QLoRA (Quantized Low-Rank Adaptation) for eficient adaptation. They trained the model on the shared task validation dataset using a structured setup that included LoRA rank 64, attention and feed-forward projection tuning,mixed precision (FP16), gradient checkpointing, batch size 2 with accumulation for efective batch size of 8, trained over 3 epochs, 8-bit AdamW optimizer and a sequence length of 2048 tokens. The training was conducted on a single NVIDIA H100 (80GB) GPU. No ensembling or post-processing techniques were used; predictions were directly generated from the fine-tuned model. This approach performed well across multiple criteria, especially in relevance and naturalness.
The FHS team (not ranked) submitted a multi-headed classification model built on google-bert/bertbase-uncased, fine-tuned on UltraFeedback and the shared task development split. Each model head specialized in one of the five preference fields. Their approach was minimalist and highly parametereficient (100M–1B), focused on instruction tuning without any generation component. While their system delivered strong scores—especially in truthfulness and naturalness—it was submitted after the oficial deadline and thus excluded from final rankings.
Subtask 2: Preference Prediction with Explanation Generation VerbaNexAI submitted the topperforming system for Subtask 2, utilizing a lightweight in-context learning approach without any finetuning. Their solution employed the meta-llama/Llama-3.2-1B-Instruct model for generating both preference predictions and explanations, and used facebook/bart-large-cnn for summarization. To construct few-shot prompts, sentence embeddings were computed using all-MiniLM-L6-v2, and the most relevant training examples were retrieved and compressed through summarization before inclusion in the prompt. A custom system instruction guided generation. Post-processing involved regular expression parsing of outputs and computing cosine similarity between explanation embeddings for evaluation. No ensembling or model fine-tuning was used. The system relied entirely on models in the 100M–1B parameter range and achieved the highest average performance across all metrics, including accuracy, ROUGE-L, BERTScore, and LLM-as-a-judge scores.
Team UTK submitted a fine-tuned model based on unsloth/llama-3-8b-Instruct-bnb-4bit, part of the LLaMA-3 family. Their system was developed using QLoRA (Quantized Low-Rank Adaptation), enabling eficient finetuning on resource-constrained hardware. The model was trained on the 2025_validation dataset with the following configuration: • LoRA Parameters: Rank = 64; target modules = [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]; LoRA alpha = 64; dropout = 0 • Training Parameters: learning rate = 2e-4; optimizer = adamw_8bit; batch size = 2 with gradient accumulation = 4; epochs = 3; sequence length = 2048; mixed precision (fp16); gradient checkpointing enabled; warmup steps = 5 Training was performed on a single NVIDIA H100 GPU. No ensembling or post-processing (e.g., reranking or rescoring) was applied. All predictions and explanations were directly generated from the fine-tuned model. The system used models in the 1B–8B parameter range and delivered strong performance across all metrics, particularly in accuracy and BERTScore. The Preference Prediction shared task at ELOQUENT 2025 introduced a novel two-part benchmark to evaluate LLMs’ alignment with human preferences –not only in terms of prediction accuracy but also in their ability to generate coherent and human-like justifications. Subtask 1 focused on predicting human preference judgments across five fine-grained criteria, while Subtask 2 extended this challenge by requiring natural language explanations for the predictions. The task attracted some participation, with VerbaNexAI and UTK submitting high-performing systems based on in-context learning and parameter-eficient fine-tuning strategies, respectively. The results highlighted meaningful progress in preference modeling, particularly for the truthfulness and safety criteria. However, there remains considerable room for improvement in generating explanations that align with humans, as evidenced by the gap in the LLM-as-a-judge scores. We hope this shared task encourages future research at the intersection of preference modeling and interpretability of LLM-based systems. The dataset, evaluation tools, and baseline results released throughout this task ofer a foundation for continued development in this research direction.
We used Grammarly6 to correct grammar, spelling, and phrasing errors in the text of this paper. The authors reviewed and revised the GenAI tool’s suggestions, and take full responsibility for the publication’s content. 2023, pp. 670–686. URL: https://aclanthology.org/2023.eacl-main.48/. doi:10.18653/v1/2023. eacl-main.48. [14] P. Colombo, N. Noiry, E. Irurozki, S. Clémençon, What are the best systems? new perspectives on nlp benchmarking, Advances in neural information processing systems 35 (2022) 26915–26932.