Harness Can Iterate, but Experience Gets Lost: Agent Self-Improvement Still Lacks a Ledger
A paper on Self-Improvement Memory records the reason, result, and applicable environment for each modification. In cross-model tasks, a Harness with memory scores 84.75, while a solution that only inherits code scores 73.42. Code inheritance is only superficial; whether experience can be re-judged is the real difficulty.
Harnesses can keep iterating, and code can be inherited. The trouble is that when the next person or Agent doing optimization opens the repository, they only see the changed prompts, toolchain, and evaluation scripts—not why the change was made at the time, or what problem it solved after the change.
In a tweet, Xudong Han mentioned a paper titled Self-Improvement Memory Enables Continual Recursive Self-Improvement, arXiv:2611.18856, submitted on November 13, 2026, with first author Xin Yu and co-authors including Fei Xia, Jiaqi Pan, and others. The paper is concerned with Continual Recursive Self-Improvement—plainly put, whether existing Harness optimization methods can continue to be used after changing datasets or underlying models, and become more accurate the more they are used.
Let's first clarify what a Harness is. Here it can be understood as an execution framework wrapped around the model; prompts, tool calls, task decomposition, evaluation pipelines, and failure-repair strategies may all be part of it. The model is responsible for generation, and the Harness is responsible for constraining the generation process to be executable and evaluable. The problem is that many designs in a Harness carry contextual information. A certain retry logic was added for a certain type of timeout; a certain prompt order was meant to work around the formatting quirks of a certain model version. Once the underlying model changes, some of these designs are still valid, while others are already outdated. Inheriting only the final code is equivalent to throwing away the on-site records.
The Self-Improvement Memory proposed in the paper, SIM for short, records a modification ledger. Each change is grouped by recurring Harness problems, and the reason for the modification, expected effect, actual result, and evaluation environment are written down. When a new task arrives, the optimizer can refer to this ledger to judge which experiences are worth keeping and which need to be tried again. SIM sits on the Harness optimizer, not on the task-solving policy. This placement is crucial: it records the optimization process itself.
Experimental data are more persuasive than concepts. In cross-model Workspace tasks, the final score using SIM is 84.75, while the approach that only inherits the Harness and continues optimizing scores 73.42. In another 10 groups of Coding optimization experiments, SIM leads in 9 groups. The paper abstract also gives more detailed numbers: in 10 independent source runs for coding, paired continuations with SIM average 2.36 points higher than harness-only continuations, p=0.003. In cross-dataset and paired cross-model comparisons across four domains, SIM's final scores are also higher than those of harness-only and root-cause-only continuations.
These numbers show one thing: memory of the optimization process has independent value. Code inheritance is like handing over a project with only the results and no discussion records. SIM supplies the discussion records, and it requires recording the applicable environment. When the underlying model or task changes, the system cannot unconditionally trust old experience; it needs to re-judge it. This action is harder than simply saving history. Saving history is only logging; re-judging is what makes it memory.
One judgment worth noting is that measuring a Harness's self-improvement capability perhaps should not focus only on how many points it can rack up on a fixed model and fixed benchmark. It should focus more on how much valid experience it retains and how many outdated judgments it proactively corrects as models and tasks keep changing. There are many systems that can inherit code; few can explain why they made a change and overturn old conclusions in a new environment.
Paper link: https://arxiv.org/abs/2611.18856
Tweet link: https://x.com/Xudong07452910/status/2122898087668793390
发布时间: 2026-11-19 01:34