PrefEval Benchmark: Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs

| Website | Paper | Data |

---

🏆Performance Leaderboard on Subset Tasks🏆

Ranked by performance in the Reminder (10 Turns) column. This table presents the performance results for the topic: Travel-Restaurants. (With explicit preference and generation task)

Model	Zero-shot (10 Turns)	Reminder (10 Turns)	Zero-shot (300 Turns)	Reminder (300 Turns)
o1-preview	0.50	0.98	0.14	0.98
GPT-4o	0.07	0.98	0.05	0.23
Claude-3-Sonnet	0.05	0.96	0.04	0.36
Gemini-1.5-Pro	0.07	0.91	0.09	0.05
Mistral-8x7B	0.08	0.84	-	-
Mistral-7B	0.03	0.75	-	-
Claude-3-Haiku	0.05	0.68	0.02	0.02
Llama3-8B	0.00	0.57	-	-
Claude-3.5-Sonnet	0.07	0.45	0.02	0.02
Llama3-70B	0.11	0.37	-	-

---

Dataset Location

The preference evaluation dataset is located in the benchmark_dataset directory.

Data Format

The dataset is provided in json format and contains the following attributes:

1. Explicit Preference.

{
    "preference": [string] The user's stated preference that the LLM should follow.
    "question": [string] The user's query related to the preference, where a generic response to this question is highly likely to violate the preference.
    "explanation": [string] A 1-sentence explanation of why answering this question in a preference-following way is challenging.
}

2. Implicit Preference - Choice-based Conversation

{
    "preference": [string] The user's explicit preference that the LLM should follow.
    "question": [string] The user's query related to the preference, where a generic response to this question is highly likely to violate the preference.
    "explanation": [string] A 1-sentence explanation of why answering this question in a preference-following way is challenging.
    "implicit_query": [string] A secondary query that offers further insight into the user’s preference, where the assistant provides multiple options.
    "options": [list] A set of options that the assistant presents in response to the user's implicit query, some of which align with and others that violate the user’s implied preference.
    "conversation": {
        "query": [string] Implicit_Query,
        "assistant_options": [string] The assistant's presenting multiple options, some aligned and some misaligned with the user's preference,
        "user_selection": [string] The user's choice or rejection of certain options.
        "assistant_acknowledgment": [string] The assistant's recognition of the user’s choice.
    },
    "aligned_op": [string] The option that aligns with the user’s preference.
}

3. Implicit Preference - Persona-driven Conversation

{
    "preference": [string] The user's explicit preference that the LLM should follow.
    "question": [string] The user's query related to the preference, where a generic response to this question is highly likely to violate the preference.
    "explanation": [string] A 1-sentence explanation of why answering this question in a preference-following way is challenging.
    "persona": [string] The assigned persona guiding the conversation, e.g., "a retired postal worker enjoying his golden years.",
    "conversation": {
        "turn1": { "user": [string], "assistant": [string] },
        "turn2": { "user": [string], "assistant": [string] },
        ...,
        "turnN": { "user": [string], "assistant": [string] }
    },
}

Benchmarking on PrefEval

Environment Setup

Create a conda environment:

conda create -n prefeval python=3.10 -y
conda activate prefeval

Install the required dependencies:

pip install -r requirements.txt

Set up AWS credentials for calling Bedrock API.

• Follow the instruction here to install aws cli.
• Run the following command and enter your aws credentials: AWS Access Key ID and AWS Secret Access Key

aws configure

Example Usages:

The following scripts demonstrate how to benchmark various scenarios. You can flexibly modify the arguments within these scripts to assess different topics, preference styles, and inter-turn conversation numbers to create varying task difficulties.

Example 1: Benchmark Generation Tasks

cd example_scripts

1. Benchmark Claude 3 Haiku with zero-shot on explicit preferences, using 3 inter-turns for the travel restaurant topic:

bash run_and_eval_explicit.sh

3. Benchmark Claude 3 Haiku with zero-shot on implicit preferences, using persona-based preferences and 2 inter-turns:

bash run_and_eval_implicit.sh

Example 2: Benchmark Classification Tasks

1. Benchmark classification tasks on all topics with explicit/implicit preferences, using Claude 3 Haiku with zero-shot and 0 inter-turns:

bash run_mcq_task.sh

Example 3: Test 5 baselines methods

1. Test 5 baseline methods on explicit preferences: zero-shot, reminder, chain-of-thought, RAG, self-critic.

bash run_and_eval_explicit_baselines.sh 

Note: All benchmarking results will be saved in the benchmark_results/ directory.

---

SFT Code

Code and instructions for SFT (Supervised Fine-Tuning) are located in the SFT/ directory.

---

Benchmark preference and query pair generation:

We provides code for generating preference-query pairs. While our final benchmark dataset includes extensive human filtering and iterative labeling, we provide the initial sampling code for reproducibility.

cd benchmark_dataset
python claude_generate_preferences_questions.py