MemSyco-Bench - Memory-Evidence Conflict (Full Dialog): leaderboard
MemSyco-Bench (Xiamen University, 2026) tests memory-induced sycophancy, where remembered user beliefs or preferences improperly steer an answer. This task asks for decisions where verified evidence contradicts a remembered user preference and the evidence should win. The model receives the whole earlier dialogue in context with no memory system; generation accuracy in percent, higher is better.
Metric: Generation accuracy (%; correct under the reference answer and task-specific rubric; decisions where verified evidence conflicts with a remembered user preference; the whole preceding multi-turn dialogue is given in context with no memory system). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Flash | 59.67 | #258 |
| 2 | Llama 3.3 70B Instruct | 29 | #1357 |
| 3 | GPT-4o Mini | 7 | #1623 |
| 4 | Llama 3.1 8B Instruct | 4 | #4246 |
| 5 | Qwen 3 8B | 0.67 | #1830 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=memsyco-bench-memory-evidence-conflict-full-dialog · How It Works · Data refreshed daily, snapshot 2026-11-13.