Classic BM25 still beats dense retrieval in some low-resource settings, while techniques like KVBoost, RENDER, and compact legal embeddings improve memory use and domain-specific retrieval.
A 0.6 billion parameter legal embedding model shows that distillation, hard negative mining, and binary quantization can move legal retrieval toward self-hosted deployments. The structural issue is not benchmark performance alone, but the ability to ...
GreenLeaf Law Embed Tiny is a 0.6B parameter embedding model for legal retrieval. It achieves 75.11% on MLEB and 64.38% on MTEB Law. Its two-stage pipeline combines distillation and fine-tuning with hard negative mining on 3.4 million query-passage p...
RENDER introduces a control for evaluating memory in LLMs: it fixes the conversation and varies the reader-facing artifact. Across 500 LongMemEval questions and nine models, matched-budget resolved packets beat recency-truncated raw dialogue by 42.4–...
A study on 3,000 Khmer documents shows BM25 outperforms dense retrieval and Qwen2.5-assisted query expansion. The larger model produces better expansions but still introduces noise. A signal for teams building low-cost local search.
KVBoost reuses key-value cache at chunk level and reduces time-to-first-token by 4.49x on Qwen2.5-3B, with no accuracy loss. Unlike prefix caching, it does not require a shared leading prefix: a dual hash separates positional and content identity. As...
A hybrid framework combines lexical signals and SciBERT semantics to identify software and database names in biomedical literature. Trained with a pipeline mixing citation extraction and LLM-assisted distillation, it outperforms specialized methods a...
