Researchers diagnose 'A' bias in AI model fine-tuning
Qwen3-0.6B model stuck predicting 'A' 81% of the time after KDA conversion
In a new arXiv paper, researcher Ronglong Bao uncovered a fascinating failure mode in AI model fine-tuning that reveals how superficial patterns can override actual learning. The study started with a straightforward experiment: converting 21 of 28 full-attention layers in Alibaba's open-source Qwen3-0.6B-Base model to KDA (Kimi Delta Attention) linear-attention layers, all while working within consumer-grade GPU constraints.
The surprising result wasn't that the model performed worse—it matched its teacher's perplexity scores through hidden-state alignment and distillation. Instead, the model developed an 'interface injury' where it became pathologically attached to the letter 'A' in multiple-choice answers. Diagnostics showed the model predicted 'A' 81% of the time, and this behavior persisted even when answer options were rotated (106/161 questions kept the same label across all permutations). Standard evaluation metrics failed to catch this bias, which only became apparent through careful diagnostic testing. A targeted 1,000-step format-focused fine-tuning stage eventually repaired the interface, boosting C-Eval scores by +12.48 points and halving the label-stickiness phenomenon.
- Qwen3-0.6B-Base model developed 81% bias toward answering 'A' after KDA linear-attention conversion
- Standard perplexity metrics failed to detect the 'interface injury'—only rotation diagnostics revealed the bias
- Targeted fine-tuning recovered +12.48 points on C-Eval benchmarks and reduced label-stickiness by ~50%
Why It Matters
Exposes blind spots in AI evaluation metrics and highlights risks of superficial pattern learning in fine-tuning