Research map / Large language models / Alignment and preferences
When Empathy Misses the Goal: A Benchmark for Goal Displacement in LLM Advice
Dean Ariel, Guy Laban
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Defines goal displacement in LLM advice (warmth or reassurance substituting for goal-advancing content) and releases GDB, a 230-scenario benchmark for measuring it.
Problem
Conversational LLMs deployed for advice may replace actionable, goal-advancing content with emotional warmth, reassurance, or non-directive encouragement even when these do not satisfy the user's objective.
Approach
Formalizes goal displacement as a failure mode distinct from general sycophancy or empathy, and operationalizes it in a benchmark of advice scenarios where warmth alone cannot satisfy the user's goal.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use newer model generation within the same vendor (GPT-4o-mini to GPT-5.4-mini) | GPT-4o-mini · GDB trap scenarios | trap displacement rate (%): 32.3% -> 2.3% | Improved |
| use newer model generation (GPT-4o to GPT-5.4) | GPT-4o · GDB trap scenarios | trap displacement rate (%): 17.0% -> 0.0% | Improved |
| add SFT to OLMo-2 Base (ablation) | OLMo-2 7B · GDB F2 | displacement rate (%): 52.3% -> 52.3% | No clear effect |
| add DPO after SFT (ablation) | OLMo-2 7B · GDB F2 | displacement rate (%): 52.3% -> 13.8% | Improved |
| add SFT then DPO to Base (ablation) | OLMo-2 7B · GDB trap | trap displacement rate (%): 30.8% -> 7.7% | Improved |
| instruction tuning of base model (ablation) | Qwen 2.5 · GDB trap | trap displacement rate (%): 26.9% -> 6.9% | Improved |
| move from Mistral v0.2 to v0.3 | Mistral 7B · GDB trap | trap displacement rate (%): 14.6% -> 6.9% | Improved |
| scale open-weight models from <=8B to >=32B | open-weight models · GDB F2 | displacement rate (%): 33.3 -> 15.4% | Improved |
| scale open-weight models from <=8B to >=32B | open-weight models · GDB F0 reverse displacement | reverse displacement rate (%): 31.7% -> 25.7% | No clear effect |
| higher capability (Arena Elo) | 18 LLMs · GDB trap scenarios | trap displacement vs Arena Elo (Spearman rho): rho = -0.706 | Improved |
| add social pressure at fixed complexity (ablation) | factorial probe, 8 models · sycophancy probe | sycophancy rate: +66.3 and +48.8 percentage points | Hurt |
| add social pressure at fixed complexity (ablation) | factorial probe, 8 models · sycophancy probe | displacement rate: +2.5 and 0.0 points | No clear effect |
Experimental setup
Datasets: GDB
Benchmarks: GDB
Closest papers by meaning
- Rethinking Psychometric Evaluation of LLMs: When and Why Self-Reports Predict Behavior (NeurIPS 2026 · similarity 0.86)
- Apathetic or Empathetic? Evaluating LLMs' Emotional Alignments with Humans (NeurIPS 2024 · similarity 0.86)
- How do Large Language Models Navigate Conflicts between Honesty and Helpfulness? (ICML 2024 · similarity 0.85)
- Language Model Goal Selection Differs from Humans' in a Self-Directed Learning Task (NeurIPS 2026 · similarity 0.84)
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.84)
- EmpathyChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
