---
title: "Chaeyeon Mia Lee (@miainflorence) | Hashnode"
description: "Chaeyeon Mia Lee (@miainflorence) writes on Hashnode. Read their latest articles and follow their work."
url: "https://hashnode.com/@miainflorence"
---

@miainflorence

# Chaeyeon Mia Lee

@miainflorence · Seoul, South Korea · Joined July 2026

### [ResKV: Recovering Evicted Token Contributions via Residual KV Cache — LongBench 32/32](https://miainflorence.hashnode.dev/reskv-recovering-evicted-token-contributions-via-residual-kv-cache-longbench-32-32?utm_source=hashnode&utm_medium=feed)

ResKV adds a residual cache to eviction-based KV compression, restoring dropped tokens' softmax mass. No extra memory. LongBench 32/32, RULER 63/64.

8 min read · [miainflorence.hashnode.dev](https://miainflorence.hashnode.dev/reskv-recovering-evicted-token-contributions-via-residual-kv-cache-longbench-32-32?utm_source=hashnode&utm_medium=feed)

[Aug 6](https://hashnode.com/posts/reskv-recovering-evicted-token-contributions-via-residual-kv-cache-longbench-32-32/6a7443e1926c9a411d768471)

### [Cross-Model KV Cache Transfer: Skip Re-Prefill When Switching LLMs](https://miainflorence.hashnode.dev/cross-model-kv-cache-transfer-skip-re-prefill-when-switching-llms?utm_source=hashnode&utm_medium=feed)

Ridge regression maps KV caches across same-family LLMs, skipping re-prefill on model swaps. 73-98% accuracy retained, up to 25x faster.

6 min read · [miainflorence.hashnode.dev](https://miainflorence.hashnode.dev/cross-model-kv-cache-transfer-skip-re-prefill-when-switching-llms?utm_source=hashnode&utm_medium=feed)

[Aug 6](https://hashnode.com/posts/cross-model-kv-cache-transfer-skip-re-prefill-when-switching-llms/6a73d7eb0f7b259f10e32120)

### [GradCuit: Credit-Assigned Gradient Flow for Robust Test-Time Latent Reasoning in LLMs](https://miainflorence.hashnode.dev/gradcuit-credit-assigned-gradient-flow-for-robust-test-time-latent-reasoning-in-llms?utm_source=hashnode&utm_medium=feed)

GradCuit optimizes latent vectors at intermediate LLM layers at test time, beating CoT by 6.6pp and LatentSeek by 2.4pp.

7 min read · [miainflorence.hashnode.dev](https://miainflorence.hashnode.dev/gradcuit-credit-assigned-gradient-flow-for-robust-test-time-latent-reasoning-in-llms?utm_source=hashnode&utm_medium=feed)

[Aug 5](https://hashnode.com/posts/gradcuit-credit-assigned-gradient-flow-for-robust-test-time-latent-reasoning-in-llms/6a72bdeddf703efde3910bfa)

### [GEPA: How Prompt Evolution Beats Reinforcement Learning Without Touching Model Weights](https://miainflorence.hashnode.dev/gepa-how-prompt-evolution-beats-reinforcement-learning-without-touching-model-weights?utm_source=hashnode&utm_medium=feed)

GEPA is an ICLR 2026 Oral showing LLM reflection-guided prompt evolution outperforms GRPO on compound AI tasks. No GPU, works with any API model.

10 min read · [miainflorence.hashnode.dev](https://miainflorence.hashnode.dev/gepa-how-prompt-evolution-beats-reinforcement-learning-without-touching-model-weights?utm_source=hashnode&utm_medium=feed)

[1](https://hashnode.com/posts/gepa-how-prompt-evolution-beats-reinforcement-learning-without-touching-model-weights/6a7150950bfc7c10ecfedeb4) [Aug 4](https://hashnode.com/posts/gepa-how-prompt-evolution-beats-reinforcement-learning-without-touching-model-weights/6a7150950bfc7c10ecfedeb4)

### [The "Understanding Is Done Early" Insight Behind CoMem: A Deep Technical Breakdown](https://miainflorence.hashnode.dev/the-understanding-is-done-early-insight-behind-comem-a-deep-technical-breakdown?utm_source=hashnode&utm_medium=feed)

CoMem caches transformer mid-layer states to cut 128k-token memory from 89 GB to 18 GB, with 7.83× prefill speedup and RULER score 97.05.

10 min read · [miainflorence.hashnode.dev](https://miainflorence.hashnode.dev/the-understanding-is-done-early-insight-behind-comem-a-deep-technical-breakdown?utm_source=hashnode&utm_medium=feed)

[Aug 4](https://hashnode.com/posts/the-understanding-is-done-early-insight-behind-comem-a-deep-technical-breakdown/6a714c51a552ebfcc4684d9b)
