Merge Barriers in BPE Tokenization: How Tokenizer Design Causally Determines Attention Head Specialization. Paper, 23 eval scripts, 86 result files, tokenizer definitions, 4 model checkpoints.
nlp deep-learning reproducible-research transformer llama structured-data attention-mechanism tokenization bpe gqa sentencepiece attention-heads gpt-neox llm mechanistic-interpretability subword-tokenization merge-barriers causal-ablation delimiter-specialization architecture-independence
-
Updated
Jul 6, 2026 - Python