CMChaeyeon Mia Leeinmiainflorence.hashnode.dev·2d ago · 8 min readResKV: Recovering Evicted Token Contributions via Residual KV Cache — LongBench 32/32Long-context LLM inference has a memory problem that compounds with scale. Every additional token grows the KV cache linearly — a single 32K-token forward pass through an 8B model can exhaust several 00
CMChaeyeon Mia Leeinmiainflorence.hashnode.dev·2d ago · 6 min readCross-Model KV Cache Transfer: Skip Re-Prefill When Switching LLMsThe Problem Nobody Had Solved Every production LLM system eventually confronts the same wall: you want to swap between models — route a cheap query to the 14B, an expensive one to the 32B — but the mo00
CMChaeyeon Mia Leeinmiainflorence.hashnode.dev·3d ago · 7 min readGradCuit: Credit-Assigned Gradient Flow for Robust Test-Time Latent Reasoning in LLMsThe Problem Nobody Had Solved Test-time scaling is one of the most active frontiers in LLM research: spending more compute at inference to get better answers without touching model weights. Chain-of-T00
CMChaeyeon Mia Leeinmiainflorence.hashnode.dev·4d ago · 10 min readGEPA: How Prompt Evolution Beats Reinforcement Learning Without Touching Model WeightsYou've spent 40 GPU-hours fine-tuning a model with GRPO. It works great — until your downstream pipeline changes, or the base model gets upgraded, and you're back to square one. Meanwhile, your engine11K
CMChaeyeon Mia Leeinmiainflorence.hashnode.dev·4d ago · 10 min readThe "Understanding Is Done Early" Insight Behind CoMem: A Deep Technical BreakdownYour GPU runs out of memory at 128k tokens. Your RAG pipeline loses critical context in the middle. You've tried KV cache compression and watched accuracy drop. These aren't configuration problems — t20
