arXiv Machine Learning

The RSI Tax: When Recursive Self-Improvement Is Not Worth the Cost

The paper introduces the concept of the RSI Tax, the token cost incurred during recursive self‑improvement (RSI) of language‑model agents. By evaluating six RSI methods on continual task streams, the authors find that a bounded‑note baseline can match or surpass the performance of more complex methods while saving 65.6–71.0% of total tokens. The study decomposes the RSI Tax into verification, exploration, and writing costs, revealing that verification dominates the expense and that much of the cost is tied to unused states or artifacts, highlighting a practical limitation of costly RSI procedures.

arXiv AI
Oct 31

A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement

A-RSI is a method for zero‑data multimodal self‑improvement that separates the decision of when to refresh training experience from the generation of new tasks. It monitors reward variation and accuracy in training logs to trigger task pool refreshes without relying on validation or test scores. The approach uses a static library of model‑authored programs to generate multimodal questions and verify answers, achieving a performance increase from 57.77 to 61.05 on an eleven‑benchmark suite with 800 fresh tasks.

By Yu Liu, Youxing Li, Yuedong Tan, Zhiwei Yang, Zihan Su, Hanqi Feng, Shaojie Zhang, Libin Liao, Zihao Wan, Chenxu Wu, Cong Cao, Fangfang Yuan, Yanbing Liu, Ziheng Chen, Pei Fu, Radu Timofte, Zongwei Wu, Nicu Sebe, Barnabas Poczos, Jian Luan
arXiv AI
Oct 26

HarnessMM: Recursive Self-Improvement for Harness Token Efficiency

HarnessMM (Harness Minus Minus) is a method that improves harness token efficiency by using guarded decision reuse and recursive self‑improvement (RSI). It refines harness configurations from training feedback, reuses resolved decisions to replace later reasoning callbacks, and stores actions in Executable Decision Graphs. Experiments show significant token savings while maintaining or improving success rates on benchmarks such as HumanEval and DeepSeek Harness.

By Yuhua Jiang, Chang Xu, Feifei Gao, Li Zhao
arXiv Machine Learning
Sep 22

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

The paper introduces Regularized Recursive Self-Improvement of Agent Harnesses (RRSI), a method that applies regularization principles to the iterative editing of an LLM agent’s harness—prompts, control flow, tooling, memory, and context management. RRSI limits the number of edits per candidate, encourages novel trajectories, and uses a critic and pruner to filter out benchmark‑specific or ineffective changes, thereby favoring reusable agent mechanisms. Experiments on eight benchmarks show RRSI improves performance by up to 14.1 points on the training split and 4.7 points on out‑of‑distribution tests, while reducing policy token usage by 30% compared to unregularized evolution.

By Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhang, Yoonho Lee, Chengsong Huang, Han Yu, Zhongying CuiZhu, Yifei Ming, Huaxiu Yao, Burak Gokturk, Tomas Pfister, Chen-Yu Lee
arXiv AI
Oct 26

RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training

RSI-Opt introduces Recursive Self-Improvement Optimizer Agents that evolve their own training policies by accumulating knowledge from previous runs, including prompts, playbooks, optimizer code, and decision cadences. A meta‑agent rewrites and archives these policies, enabling reusable procedures and code that adapt to new tasks. Across multiple generations and tasks, RSI‑Opt improves optimizer performance, achieving higher z‑scores on evolution suites, better cross‑model transfer, and significant error reductions on PDE benchmarks compared to fixed‑prompt baselines.

By Binghang Lu, Yige Sun, Guang Lin
arXiv Computer Vision
Nov 9

Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents

Flex‑RSI is a new benchmark that evaluates how agents can recursively improve their own harness—prompts, tools, memory, and control loop—while solving vision and robotics tasks under a fixed budget. Each round alternates between an executor that refines the task solution and an improver that revises the harness, with only higher‑scoring solutions retained for the next iteration. The benchmark tests both closed‑source and open‑weight models across perception, reasoning, robot action, 3D reconstruction, visual creation, and infrastructure tasks, revealing that self‑improvement is most effective when agents build their own solutions, that simple success/failure feedback limits open‑weight models, and that harness editing itself contributes significantly to performance gains.

By Jingdong Zhang, Jionghao Wang, Zhengming Yu, Yancheng Cai, Wei Chow, Chris Choy, Zhiyang Dou, Hang Hua, Jiachen Li, Zhixuan Liu, Chen Wang, Songlin Wei, Siyuan Yang, Yu Yuan, Xiaohang Zhan, Yucheng Zhang, Shuhong Zheng, Xin Li, Lingjie Liu, Wojciech Matusik, Zhengzhong Tu, Wenping Wang
arXiv AI
Oct 5

VERSE: Verified Self-Evolving Optimizer for Agent Harnesses

The paper introduces VERSE, a Verified Self‑Evolving optimizer that enhances LLM agent harnesses by allowing the optimizer to test edits, replay failures, and perturb steps while tracking fixes and regressions. VERSE builds its own tools for failure analysis, verification, training audits, and workflow control, and uses this feedback to revise the harness’s prompts, skills, tools, hooks, and notes without changing model weights. In experiments across five executors and multiple languages, VERSE improves all evaluated harness optimizers, achieving higher accuracy on held‑out and out‑of‑distribution tasks compared to the strongest baselines.

By Zekai Wang, Yingqiang Ge, Zekun Wang, Hai Wang, Yuhui Xu, Joshua Frandsen, Shancong Fu, Ashia C. Wilson, Chandan K. Reddy
arXiv AI
6d ago

Veritas: A Verifier Generation Harness for Agent Tasks

Veritas is a harness that automatically generates verifiers for agent tasks, adding an evidence header, a library of domain rules, and authoring principles refined through a self‑improvement loop. It was evaluated on six benchmarks, outperforming seven other harnesses on all four execution‑scored tasks and matching or exceeding performance on pairwise comparison tasks. The system also improves agreement with official verdicts across multiple model families and harness‑benchmark pairs.

By Daoan Zhang, Qinbo Bai, Dehai Min, Xingyu Mu, Ziyi Chen, Jianing Yin