frozen-backbone

通过权重修改，让开源模型讯飞星火Spark-X2.5-4B彻底相信自己是一只傲娇的AI鲸鱼

TapeLM: facts as fingerprints on character ink — not token-id memory, not chunk RAG. One frozen curve encoder for generation and structured slot memory (write, bind, hop, resolve). Noisy recall, lexicon calibration, one-shot edits, clean unlearning — vs fair GPT/RAG; reproducible JSON benchmarks.

DETR-style detection head on a frozen DINOv3 ViT-B/16 backbone, trained end-to-end on COCO 2017. Mixed Query Selection, aspect-ratio bucketing, Hungarian matching with L1 + CIoU. 8.5M trainable params, 0.431 AP on val2017.

Modular knowledge injection into frozen LLM backbones. Measured on Qwen3-4B / Qwen2.5-3B / SmolLM2-1.7B: a 64 to 128 sample phase transition, and a 2,560-parameter module form with no measurable drift on a held task.