When Empathy Misses the Goal: A Benchmark for Goal Displacement in LLM Advice

Key Findings

• — Trap displacement (validating an infeasible goal, or addressing emotion while ignoring a hard constraint) ranged from 0.0% (Gemini 2.5 Pro, Gemini 3.1 Pro, GPT-5.4, DeepSeek-V3, Claude Haiku 4.5) to 37.7% (Llama 3 8B); GPT-4o reached 17.0% overall and 30.8% on the feasibility-override family, GPT-4o-mini 32.3%.
• — Trap displacement correlated negatively with Chatbot Arena Elo (Spearman rho -0.706, n = 15 models).
• — Reverse displacement (unsolicited problem-solving when emotional support was requested) reached 41.9% (Mistral 7B), including substantial rates in models with negligible trap displacement (DeepSeek-V3 22.9%).
• — In a factorial probe, human-scored responses in the high-complexity, low-pressure condition showed 35.0% sycophancy but 7.5% displacement, with little association between the two constructs.
• — On matched OLMo-2 checkpoints, mixed empathy-with-constraints displacement stayed at 52.3% after supervised fine-tuning and fell to 13.8% after direct preference optimization.
• — By domain, trap displacement was highest in education (20.1%), relationship (14.1%), career (13.0%) and parenting (11.9%), and lowest in legal (2.8%) and financial (3.8%).

Methodology Notes

Scenarios written by the authors (no user logs); models run single-turn at fixed temperature with no system prompt; responses scored by the two authors blind to model identity, with two independent Prolific raters on a 191-item stratified subset (weighted kappa 0.82 for displacement severity against author scores); GPT-4.1 and Kimi-K2 judges validated as cheap screeners. Ethics approval from Ben-Gurion University IE&M departmental committee. Limitations: author-designed scenarios, single-turn, English. Verified from arXiv HTML v1; HF dataset guylaban/gdb-benchmark public (lastModified 2026-10-18). arXiv v1 submitted 2026-10-18 (Tuesday 20 October 2026 mailing). Accepted to the NeurIPS 2026 Evaluations and Datasets Track per the authors. Dataset released on Hugging Face (guylaban/gdb-benchmark, CC BY 4.0, dataset DOI 10.57967/hf/11042).

Sources