ResKV: Recovering Evicted Token Contributions via Residual KV Cache — LongBench 32/32ResKV 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.devAug 6
Cross-Model KV Cache Transfer: Skip Re-Prefill When Switching LLMsRidge 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.devAug 6
GradCuit: Credit-Assigned Gradient Flow for Robust Test-Time Latent Reasoning in LLMsGradCuit 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.devAug 5
GEPA: How Prompt Evolution Beats Reinforcement Learning Without Touching Model WeightsGEPA 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.dev11KAug 4
The "Understanding Is Done Early" Insight Behind CoMem: A Deep Technical BreakdownCoMem 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.dev2Aug 4