TL;DR
Recursive self-improvement (RSI) allows LLM agents to learn from their mistakes, but this "learning" comes with a steep hidden cost in tokens. This paper introduces the RSI Tax to quantify this overhead and demonstrates that complex improvement methods are often less efficient than a simple note-taking baseline, which achieves superior accuracy while saving roughly 70% in total token costs.
What Problem Does This Paper Address?
As LLM agents move from simple chat interfaces to autonomous workers (handling software repos or terminal tasks), researchers are obsessed with making them "self-improving." The idea is that an agent should analyze its failures, update its internal "harness" or skill library, and perform better on the next task.
However, most research focuses solely on accuracy gains, ignoring the massive computational cost of the improvement process itself. In a production environment, every token spent on "thinking about how to improve" is a token you pay for. The authors argue that we need a holistic evaluation that treats self-improvement as a trade-off: does the performance boost justify the RSI Tax?
Methods
The authors define the RSI Tax as the total token cost of the self-improvement process, which they decompose into three functional categories:
- Verify: Evaluating candidate updates on past tasks (often the most expensive part).
- Explore: Searching for improvements, generating proposals, and analyzing trajectories.
- Write: The actual act of updating persistent state (harnesses, skills, or notes).
They compare several state-of-the-art RSI methods—including Meta-Harness, EvoSkill, and RRSI—against a new, intentionally simple baseline called Simple Reflection with Notes (SRN).
Unlike complex methods that run expensive searches and validations, SRN simply maintains a single, bounded (2,048 token) note. When it fails a task, it uses the LLM to revise that note with lessons learned. There is no candidate search and no re-running of old tasks to "verify" the note.
Datasets
The study evaluates these methods across two distinct environments using two backbone models (DeepSeek V4.1 Flash and GPT-6 Luna):
- SWE-Chain-Evo: A subset of 105 software engineering tasks across seven repository chains.
- Terminal-Bench 2.1: A stream of 89 heterogeneous command-line tasks.
These are "streaming" evaluations, meaning the agent encounters tasks one by one, and its state can only be updated using feedback from tasks it has already completed.
Key Insights and Ideas
- Verification is the primary cost driver: For the most advanced RSI methods, verification accounts for a staggering 74.8–93.6% of the total RSI Tax. Re-running old tasks to see if a new skill works is incredibly expensive.
- The "Tax" can outweigh the "Income": In many settings, the RSI Tax actually exceeds the cost of executing the tasks themselves.
- High Waste: A significant portion of the RSI Tax (up to 63.6%) is spent on creating artifacts or state changes that the agent never actually uses again in the task stream.
- Simplicity is a Feature: The SRN baseline proved that you don't need a complex "evolutionary" search to improve. By just keeping a small note of "lessons learned," the agent avoids the search/verify loop entirely, keeping the RSI Tax below 0.4% of total tokens.
Key Results
The results present a sobering reality for complex RSI research. As shown in the accuracy-vs-cost plots, the simple SRN baseline often sits in the "ideal" top-left corner (high accuracy, low cost).
- Accuracy: SRN achieved the highest observed accuracy in three out of four settings and tied in the fourth.
- Efficiency: Compared to Meta-Harness, EvoSkill, and RRSI, SRN reduced total token usage by 65.6–71.0% on average.
- Execution Cost vs. Total Cost: The authors highlight that if you ignore the RSI Tax, some methods look cheaper than they actually are. For example, Meta-Harness might use fewer tokens during execution than SRN, but its massive RSI Tax makes it nearly 4x more expensive in total.
Limitations and Future Work
The authors acknowledge several limitations:
- Scope: The study only covers two models and two benchmarks. The stability of these results across different workloads or model scales remains to be seen.
- Descriptive Attribution: The "subsequent use" analysis is descriptive. It shows that an artifact wasn't used, but it doesn't prove that not creating it would have preserved the same performance.
- Future Direction: The findings motivate the need for cost-aware update policies. Instead of updating every $N$ tasks, agents should learn to forecast whether a potential update is worth the token investment required to find and verify it.
Summary updated