TL;DR: Some of heads should not do KV cache eviction. This paper predicts which heads to do KV cache eviction and which heads to do full KV cache. MDP process trained with RL/GRPO based on task succes

Last layer hidden state already has information related to uncertainty.

Train a smaller model to predict what lines to be deleted for the read action of coding agent. The pruned part is sent back to the agent. The training pipeline is not clear.

Train an agent to spawn an auxiliary agent from the ongoing trajectory to select blocks of KV cache from history and delete at regular intervals.

My class notes on simulation modeling, based on the textbook Simulation Modeling and Analysis (Averill M. Law, 5th edition, McGraw-Hill).

What a discrete-event simulation consists of — entities, events, the event list, the clock, and statistical counters — walked through on a single-server FCFS queue.

A recap of the probability and statistics needed for simulation: probability axioms, random variables, convergence modes, stochastic processes, and estimation of means and variances.

Every random variate in a simulation is built from i.i.d. U(0,1) numbers; these notes cover where those numbers come from and what makes a generator good.

How to turn i.i.d. U(0,1) random numbers into observations from the distributions a model actually needs — the inverse transform and composition methods, and what makes a variate-generation algorithm

How to choose the input distributions a simulation feeds on: assess sample independence, hypothesize a distribution family, estimate its parameters by maximum likelihood, and judge the fit with chi-sq