Update README.md

README.md

@@ -20,12 +20,26 @@ GLINT-base is a 149M-parameter English late-interaction retriever built from
 128-dimensional token embeddings and uses MaxSim retrieval with 32 query tokens and 300
 document tokens.
 
-Training has two stages:
+## What is new in GLINT?
 
-1. supervised hard-negative fine-tuning;
-2. mixed listwise knowledge distillation on the seven-source hard-negative mixture, with
-   `jinaai/jina-reranker-v3.5` scores, temperature sharpening, false-negative masking, and an
-   InfoNCE anchor.
+GLINT is designed around the mismatch between ordinary dense hard-negative mining and a
+late-interaction retriever. Dense mining selects documents that are difficult under one pooled
+vector; GLINT instead mines negatives under the same token-level MaxSim geometry used at
+retrieval time. This exposes lexical, compositional, and localized token matches that a
+single-vector miner can miss.
+
+The training recipe has two stages:
+
+1. supervised fine-tuning with multi-vector (MaxSim) hard negatives;
+2. mixed listwise knowledge distillation over a diverse seven-source hard-negative mixture.
+
+For the second stage, a frozen listwise teacher (`jinaai/jina-reranker-v3.5`) scores each
+32-document candidate set jointly. GLINT distils that ordering with a sharpened listwise KL
+objective, while a false-negative-masked InfoNCE term preserves a direct retrieval signal.
+The result is still one compact, first-stage late-interaction retriever: it is neither an
+ensemble nor a re-ranker at inference time.
+
+This is a private research release.
 
 ## Usage
 
@@ -63,4 +77,4 @@ computed by summing, over query tokens, the maximum similarity to a document tok
 The corresponding private training artifacts are in
 [GLINT-data](https://huggingface.co/datasets/chungimungi/GLINT-data). It contains the complete
 prepared SFT data, the 1,046,009-row seven-source KD mixture, and Jina teacher-score parquet
-shards. The repository contains no BEIR evaluation corpus or evaluation labels.
+shards. The repository contains no BEIR evaluation corpus or evaluation labels.