cff-version: 1.2.0
message: "If you use GDB in research, please cite as below."
title: "When Empathy Misses the Goal: A Benchmark for Goal Displacement in LLM Advice"
abstract: >-
  GDB is a 230-scenario benchmark for measuring goal displacement in
  large language models that provide advice: the substitution of warmth,
  reassurance, or non-directive encouragement for goal-advancing content
  when the latter is required. The release contains scenarios across nine domains, 4,134
  human-scored model responses spanning 18 systems, and three supplementary
  probes (external rater portability, post-training stage decomposition,
  and a complexity-by-pressure factorial that dissociates displacement from
  social sycophancy).
authors:
  - family-names: "Ariel"
    given-names: "Dean"
    affiliation: "Department of Family Medicine, Clalit Health Services, Tel Aviv District; Faculty of Medicine and Health Sciences, Tel Aviv University; Stein Faculty of Computer and Information Science, Ben-Gurion University of the Negev"
  - family-names: "Laban"
    given-names: "Guy"
    affiliation: "Department of Industrial Engineering and Management, Ben-Gurion University of the Negev"
    email: "laban@bgu.ac.il"
version: "1.0.0"
date-released: "2026-10-16"
doi: "10.57967/hf/11042"
license: "CC-BY-4.0"
keywords:
  - "language models"
  - "evaluation"
  - "benchmark"
  - "goal displacement"
  - "alignment"
  - "RLHF"
  - "sycophancy"
repository-code: "https://github.com/Guylaban/GDB"
url: "https://huggingface.co/datasets/guylaban/gdb-benchmark"
preferred-citation:
  type: conference-paper
  title: "When Empathy Misses the Goal: A Benchmark for Goal Displacement in LLM Advice"
  authors:
    - family-names: "Ariel"
      given-names: "Dean"
    - family-names: "Laban"
      given-names: "Guy"
  year: 2026
  collection-title: "Proceedings of the 40th Conference on Neural Information Processing Systems (NeurIPS 2026), Track on Evaluations and Datasets"