How PertEval Reveals the Real Knowledge Limits of Large Language Models

On December 10, the NeurIPS 2024 conference opened in Vancouver, receiving over 15,000 paper submissions. Alibaba Cloud PAI team’s paper “PertEval: Unveiling Real Knowledge Capacity of LLMs with Knowledge‑Invariant Perturbations” was selected as a Spotlight (3% acceptance).

Innovative Evaluation Method

PertEval introduces “knowledge‑invariant perturbations” that rewrite static benchmark questions without altering the underlying knowledge, mitigating memorization and data‑contamination effects and providing a more reliable measure of LLM knowledge.

Findings on Existing Benchmarks

Re‑evaluating six representative LLMs, including GPT‑4, the study found that performance on static benchmarks such as MMLU is significantly over‑estimated (GPT‑4 by 26%). The inflation stems from models’ hesitation on uncertain knowledge and rote memorization of correct answers.

Impact and Future Work

The results highlight the need for evaluation methods that reflect true model capabilities. PertEval will be integrated into the PAI platform, enabling one‑click real‑knowledge assessment for any model and publishing a leaderboard of “knowledge ability” scores.

Responsible AI Demo at NeurIPS

The team also delivered a keynote on “Core Technical Interpretation and Best Practices of Responsible AI,” showcasing Alibaba Cloud’s enterprise‑grade trustworthy AI solution, which includes a T‑shaped safety architecture, fairness and error analysis, and a suite of demo capabilities such as fine‑tuning Qwen2.5‑Coder, building RAG systems, and AI‑generated content tools.