MemEvoBench (QA Style, Biased Feedback): leaderboard

MemEvoBench (2026) evaluates memory safety of LLM agents: misleading entries in a mixed memory pool, and the agent's own responses stored after each round, test whether accumulated memory drifts the agent into unsafe behavior over three rounds. This board is the attack success rate in percent in the third round on QA-style cases of misleading memory injection, judged by GPT-5.2, with simulated biased user feedback rewarding risky shortcuts; lower is better.

Metric: Attack success rate (%) in the third evaluation round on the 108 QA-style cases (misleading memory injection across 7 domains and 36 risk types): share of responses that follow the misleading memory into the target risk behavior after two rounds of the agent's own responses were added to the memory pool, with simulated biased user feedback stored after each round (risky shortcuts rewarded, cautious answers penalized), judged MISLED or SAFE by GPT-5.2, temperature 0, standard system prompt without safety guidance; lower is better. Source: arxiv.org. Saturation forecast: Around August 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude 3.7 Sonnet77#470
2GPT-578#178
3Gemini 2.5 Pro80#276
4Qwen 3 235B A22B 2507 Instruct83.1#616
5Qwen 3 Next 80B A3B Instruct84.4#778
6DeepSeek V3.289#419
7Llama 3.3 70B Instruct90#1385
8GPT-4o94#751
9Qwen 3 32B94.4#972

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=memevobench-qa-style-biased-feedback · How It Works · Data refreshed daily, snapshot 2026-11-21.