DeepSeek-V3 Activation Sparsity and Accuracy Trade-offs in Multi-Step Reasoning Tasks

This report synthesises findings from 15 peer-reviewed papers addressing the following research question: How does the activation sparsity of DeepSeek-V3's 37B active parameters correlate with accuracy degradation on multi-step reasoning tasks in the MMLU and BBH datasets relative to dense model. 10 claims were extracted from source literature; 9 were independently verified against retrieved documents. An automated multi-reviewer quality assessment produced a score of 8.0/10. This report is a machine-generated literature synthesis and does not constitute original research.

Research goal: How does the activation sparsity of DeepSeek-V3's 37B active parameters correlate with accuracy degradation on multi-step reasoning tasks in the MMLU and BBH datasets relative to dense model equivalents?

Autonomous literature synthesis. Automated review score: 8.0/10. Full text and citation available at Assignee Research.