Qwen3.5-4B-Base Blind Spots: 12 Failure Cases Across 10 Categories

This dataset documents 12 specific failure cases identified in the Qwen3.5-4B-Base multimodal model. Created by PhoenixA and updated in March 2026, it categorizes blind spots across 10 domains including arithmetic, negation, and visual grounding. The data focuses on models utilizing Gated DeltaNet and sparse Mixture of Experts (MoE) architectures.

Use Cases

• Analyzing arithmetic failures in Qwen3.5-4B-Base
• Evaluating visual grounding errors in multimodal outputs
• Investigating multilingual consistency gaps

Strengths

• Covers 10 distinct failure categories
• Targets specific Gated DeltaNet and sparse MoE architecture behaviors
• Expert-identified failure modes

Limitations

• Extremely small sample size of 12 records
• Schema details and column names are not explicitly provided in the source
• Restricted to a single model version

Provenance

Source

PhoenixA

Collection Method

Discovery of model failure cases

Freshness

Last updated March 2026.