Transformers' Hard Decision Layer Reveals Abrupt Answer Commitments
A recent investigation has pinpointed the Hard Decision Layer (HDL) in transformer-based language models, where the rankings of answer options suddenly stabilize during inference. Validation through four models (Qwen, Llama, Granite, Mistral) and four benchmark datasets consistently reveals the emergence of the HDL, independent of learned routing policies. This layer remains unchanged even after fine-tuning. Notably, accuracy enhancements at the HDL can reach +0.61 (Qwen on CommonsenseQA), after which performance levels off. Systematic ablations regarding label formats and problem complexity affirm that this phenomenon is intrinsic to model architecture. These results provide insights into transformer inference mechanisms and highlight potential for efficient reasoning and model guidance. All related code and findings are accessible on arXiv.
Key facts
- Hard Decision Layer (HDL) is a natural architectural property where answer option rankings stabilize abruptly.
- Validated across Qwen, Llama, Granite, and Mistral models.
- Tested on four benchmark datasets.
- HDL emerges without learned routing policies.
- HDL is invariant to fine-tuning.
- Accuracy improvements at HDL up to +0.61 (Qwen on CommonsenseQA).
- Phenomenon is fundamental to model architecture.
- All code and results are available on arXiv.
Entities
—