AdamJovine/LISTEN-benchmark

LISTEN Multi-Objective Selection Benchmark

Datasets accompanying the paper "LISTEN to Your Preferences: An LLM Framework for Multi-Objective Selection" (IJCAI-ECAI 2026).

• 📄 Paper: https://huggingface.co/papers/2510.25799
• 💻 Code: https://github.com/AdamJovine/LISTEN

Overview

LISTEN is a benchmark for evaluating how well an LLM can elicit a user's preferences and select a top option from a large candidate set with many numeric and categorical attributes. The benchmark contains four scenarios, each with a candidate set, a documented preference statement, and a small human-curated set of acceptable "winning" rows (human_sol). It also contains five independent human re-rankings of LLM-produced top-20 shortlists, used in the paper to measure agreement between LISTEN and human raters.

Scenarios

Config name	Scenario	n candidates	n ranked (top-K)	Description
exam	Exam Scheduling	4,938	20	Cornell final-exam schedules scored on conflict, clustering, and lateness penalties.
flights_chi_nyc	Flights CHI→NYC	903	20	Filtered Google Flights itineraries with structured layover / duration / price columns.
flights_ithaca_reston	Flights Ithaca→Reston	216	20	Smaller multi-leg search; more varied carrier / stop combinations.
headphones	Headphones	77	20	Amazon product listings with technical specs (driver size, battery, ANC, reviews).

Loading a scenario:

from datasets import load_dataset

exam = load_dataset("AdamJovine/LISTEN-benchmark", "exam", split="train")
flights = load_dataset("AdamJovine/LISTEN-benchmark", "flights_chi_nyc", split="train")
headphones = load_dataset("AdamJovine/LISTEN-benchmark", "headphones", split="train")

Human re-rankings (rerank_h1 … rerank_h5)

Five different humans (h1–h5) were each asked to re-rank the top-20 shortlist produced by LISTEN on a subset of scenarios. Each rerank CSV adds two leading columns to the scenario schema:

• rank — the human's 1-best ordering (1 = preferred).
• _row_id — the original row index in the full candidate CSV.

Coverage per annotator (not every annotator did every scenario):

	exam	flights_chi_nyc	flights_ithaca_reston	headphones
h1		✓	✓	✓
h2	✓		✓	✓
h3	✓	✓	✓
h4	✓	✓	✓	✓
h5	✓	✓		✓

Loading a rerank:

rerank = load_dataset("AdamJovine/LISTEN-benchmark", "rerank_h4", split="train")

Schema notes

• exam_data.csv — numeric penalty/weight columns (conflicts, quints, quads, triple in 24h (no gaps), lateness, avg_max, …) plus the schedule's hyperparameters (alpha, gamma, delta, vega, theta, …).
• Flights CSVs — flight itineraries with origin, destination, departure time, arrival time, duration, stops, price, distance-from-airport columns, and (for flights_chi_nyc) a derived meets_requirements flag.
• headphones_data.csv — Amazon product attributes: asin, product_name, brand (encoded), price, type, connectivity, noise_cancellation, battery_life, bluetooth_version, driver_size, weight, water_resistance, microphone, review_rating, review_count, description.

The complete preference statements used in the paper (and the human_sol IDs for each scenario) live alongside the scenario configs in the GitHub repo under configs/.

How the benchmark is used

In the paper, an LLM is given (a) a preference statement and (b) the candidate set, and is asked to return a top-K shortlist. The primary evaluation metric is:

• NAR (Normalized Average Rank) — the average rank of the LLM-selected items under each human re-ranking, normalized to [0, 1] so that lower is better (rank 1 = the human's most preferred item).

A complementary dataset diagnostic, concordance (fraction of random linear utility functions whose argmax lands in human_sol, paper §4.1), is implemented in post_analysis/concordance_analysis.py in the GitHub repository.

License

MIT. The headphones_data.csv description column contains short product blurbs scraped from public Amazon listings; redistributed under fair-use for academic benchmarking.

Citation

@inproceedings{jovine2026listen,
  title={LISTEN to Your Preferences: An LLM Framework for Multi-Objective Selection},
  author={Jovine, Adam S and Ye, Tinghan and Bahk, Francis and Wang, Jingjing and Ford, Matthew and Shmoys, David B and Frazier, Peter I},
  booktitle={Proceedings of the 35th International Joint Conference on Artificial Intelligence (IJCAI-ECAI 2026)},
  year={2026}
}