Update README.md

README.md

@@ -15,34 +15,18 @@ tags:
 
 # TrimKV-Qwen3-8B-Math
 
-TRIM-KV is an efficient and learnable key–value eviction strategy designed to improve the efficiency of large language models (LLMs) in long-horizon inference.
+> **TRIM-KV** is an efficient and learnable key–value eviction strategy designed to improve the efficiency of large language models (LLMs) in long-horizon inference.
 
-This model is a Qwen3-8B checkpoint trained with TRIM-KV on math reasoning tasks, as presented in the paper [Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction](https://huggingface.co/papers/2605.09649).
+This model is a version of `TrimKV-Qwen3-8B-Math` enhanced with TRIM-KV, as presented in the paper [Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs](https://huggingface.co/papers/2512.03324).
 
-- **Paper:** [https://huggingface.co/papers/2512.03324](https://huggingface.co/papers/2512.03324)
-- **GitHub:** [https://github.com/ngocbh/trimkv](https://github.com/ngocbh/trimkv)
+The core idea behind TRIM-KV is to learn the intrinsic importance of each key–value pair at creation time, which we call *token retention*, and then decay this importance exponentially over time to mimic standard inference running with eviction. 
 
-<a href="https://arxiv.org/pdf/2512.03324"><img src="https://img.shields.io/badge/arxiv-2512.03324-red?style=for-the-badge"></a>
+The retention score is query-agnostic and captures the long-term utility of tokens. This is different from attention scores, which are query-dependent: they capture the short-term utility for predicting the next token and are recomputed at every step.
 
-### About TRIM-KV
+- **Paper:** [Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs](https://huggingface.co/papers/2605.09649)
+- **Code:** [GitHub - ngocbh/trimkv](https://github.com/ngocbh/trimkv)
+- **Arxiv:** [2512.03324](https://arxiv.org/abs/2512.03324)
 
-The core idea behind TRIM-KV is to learn the intrinsic importance of each key–value pair at creation time, which we call *token retention*, and then decay this importance exponentially over time to mimic the standard inference running with eviction. 
-
-The retention score is query-agnostic and captures the long-term utility of tokens. This is different from attention scores, which are query-dependent: they capture the short-term utility for predicting the next token and are recomputed at every step, making them local, myopic, and highly dependent on the transient decoding state.
-
-### Why TRIM-KV?
-
-It's fast, smart, and interpretable.
-
-<div align="center">
-    <img width="1000" alt="speed" src="https://github.com/ngocbh/trimkv/blob/main/assets/speed.png?raw=true"/>
-</div>
-
-<div align="center">
-    <img width="1000" alt="performance" src="https://github.com/ngocbh/trimkv/blob/main/assets/performance.png?raw=true"/>
-</div>
-
----
 
 ## Quick Start
 
@@ -94,6 +78,12 @@ tokenizer = AutoTokenizer.from_pretrained(
 ## Citation
 
 ```bibtex
+@article{bui2025cache,
+  title={Cache what lasts: Token retention for memory-bounded kv cache in llms},
+  author={Bui, Ngoc and Sharma, Shubham and Lamba, Simran and Mishra, Saumitra and Ying, Rex},
+  journal={arXiv preprint arXiv:2512.03324},
+  year={2025}
+}
 @article{bui2025make,
   title={Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction},
   author={Bui, Ngoc and Nguyen, Hieu Trung and Cohan, Arman and Ying, Rex},