Skip to main content

Safety

Multi-MoE

Per-question binary correctness (string-match) on MMLU (5-shot, 1000 items) and TruthfulQA (5-shot MC, 812 items) for five Mixtral-8x7B serving configurations (base, Instruct, weight-avg, Multi-MoE with runtime reconfiguration, Multi-MoE no-reconfig) released with the QoS-efficient MoE-serving paper.

1,811items
5subjects
100%observed
unknownlicense
knowledgedomain
reasoningdomain
safetydomain
textmodality

Response matrix

Fit to width. Hover for subject & item; click a cell for details.

Multi-MoE response matrix: AI models (rows) against items (columns)
Correct (1)Incorrect (0)Unobserved

Scale: 1 = correct · 0 = incorrect

Subjects

  1. 1Mixtral-8x7B-v0.1-avg (weight-averaging baseline)0.7257
  2. 2Mixtral-8x7B-v0.1-Instruct0.723
  3. 3Mixtral-8x7B-v0.1 Multi-MoE (proposed, runtime reconfiguration)0.7119
  4. 4Mixtral-8x7B-v0.1 Multi-MoE (no reconfiguration)0.7086
  5. 5Mixtral-8x7B-v0.1-base0.6849