Reasoning & knowledge

IdeaTaste

by Shanshan Zhong, Chen Xu, Jingjie Ning, Zhengzhong Liu, Aviral Kumar, Chenyan Xiong

What is IdeaTaste? IdeaTaste is a reasoning & knowledge eval made by Shanshan Zhong, Chen Xu, Jingjie Ning, Zhengzhong Liu, Aviral Kumar, Chenyan Xiong. A research-taste benchmark: LLM agents read a literature corpus frozen at the end of 2025 (about 263K papers across ML systems, code generation, world models and auto-research) and write research proposals, which are then matched against the roughly 11.6K papers those communities actually published in 2026. Scores cover citation matching, idea matching, citation diversity and idea diversity; the authors' run spans 6 models and over 90,000 proposals. Code, dataset (Hugging Face) and a leaderboard are public, but scoring calls paid LLM judges, so reruns cost API spend.

Page last updated 2026-10-18 · Evals Review

Editor note

A research-taste benchmark: LLM agents read a literature corpus frozen at the end of 2025 (about 263K papers across ML systems, code generation, world models and auto-research) and write research proposals, which are then matched against the roughly 11.6K papers those communities actually published in 2026. Scores cover citation matching, idea matching, citation diversity and idea diversity; the authors' run spans 6 models and over 90,000 proposals. Code, dataset (Hugging Face) and a leaderboard are public, but scoring calls paid LLM judges, so reruns cost API spend.

What it tests
Whether an LLM agent proposes the research directions and cites the prior work that a field later pursued.
Original source
github.com/cxcscmu/IdeaTaste
Last updated
Repo last push 2026-11-03
License
MIT

Facts come only from the linked source (checked 15–16 Oct 2026). Credit belongs to the creators. We don't score or rank evals: how we review.

Reviews

No reviews yet.

Reviews are written by people who sign in with GitHub, one per account per eval. We can't confirm that each reviewer used the eval; we remove reviews that break our rules, never just for being negative, and we never add, buy or edit reviews. Averages, when shown, come only from these reviews.

Sign in with GitHub to review

Related reasoning & knowledge evals

  • GPQA — Expert-level science reasoning.
  • Humanity's Last Exam — Frontier academic knowledge and reasoning.
  • LiveBench — Hard, contamination-limited tasks with verifiable ground-truth answers (no LLM judge) in v1; LiveBench 2.0 (14 Oct 2026) adds long-horizon real-work tasks.
  • MMLU — World knowledge and problem solving across many school and professional subjects.

All Reasoning & knowledge evals · Comparison table

Discussion

No one has discussed IdeaTaste yet. Have you run it? Share what it caught, what it missed or how you set it up.

Sign in with GitHub to comment

Be specific and kind. No promotion or affiliate links. Community rules · All discussions