ICLR Drops Old Reasoning Blocks From Long-Horizon Agent Context, Cutting Cache-Read Tokens 33.3% While Reward Rises Slightly
arXiv·medium signal
arXiv 2609.29875 is a training-free online method. It ranks past reasoning blocks by frozen-proxy entropy and deletes them while keeping every action, tool call and observation. On 260 WorkBuddyBench tasks, average reward went from 0.699 to 0.718, and input, output and cache-read tokens fell 25.5%, 14.4% and 33.3%. Probing suggests old reasoning becomes safe to drop once its derived state is written out to files, code or tool outputs, which argues for agents that write state to disk.