PrefEval: Models' Preference-Following Accuracy Falls Below 10% After Just 10 Turns
math_rachel · x · 2026-10-19
From Rachel Thomas' thread: the PrefEval benchmark tests LLMs' ability to infer, memorize and adhere to user preferences in long-context conversations, with 3,000 manually curated preference-query pairs across 20 topics (dietary, travel, entertainment).
Key findings:
- In zero-shot settings, preference-following accuracy falls below 10% at merely 10 turns (3k tokens) across most evaluated models
- Even with advanced prompting and retrieval methods, preference following still deteriorates in long-context conversations
A stark reminder of current models' weakness in long-term personalization.
More from Research
- 1.9 MB extracted (not trained) LoRA fixes vertical color stripes in MOSS-Image SFT — Nynxz · 2026-10-19
- Three ways to execute a AND b AND c: DataFusion filter PR explained — Abhishekcur · 2026-10-19
- Why safety reports can't prove safety: the adversarial logic flaw — davidmanheim · 2026-10-19
- ACT3D trains VLMs to actively explore 3D worlds with RL, hitting 53.2% accuracy — UCSanDiego · 2026-10-19
- OPMD: ShanghaiTech distills teacher intermediate computations beyond output supervision — ShanghaiTech · 2026-10-19
- LexFrame: interpreting LM residual streams in the model's own semantics — Jian Gu · 2026-10-19