Evolver: Turning Agent Exploration into Evolving Capabilities
Evolver is an experience-driven framework that transforms transient agent exploration into persistent, reusable capabilities without updating foundation model parameters, thereby improving task success rates and execution efficiency across multiple benchmarks.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of artificial intelligence, there is a growing class of programs known as agents. Unlike simple chatbots that answer questions, these agents are designed to take action. They can open files, run code, search the internet, and interact with complex digital environments to solve multi-step problems, much like a human engineer or researcher might. To succeed, these agents must explore their surroundings, try different tools, and learn from mistakes. However, a significant bottleneck has emerged: every time an agent solves a difficult problem, the valuable lessons it learns are often lost. Once the task is finished, the detailed record of how it struggled, what it tried, and what finally worked is discarded. The next time a similar problem arises, the agent starts from scratch, repeating the same exploratory efforts and wasting computational resources. The central question facing researchers is how to capture this fleeting experience and turn it into a lasting resource that helps the agent get better over time, without needing to fundamentally retrain its underlying brain.
A team of researchers has introduced a system called Evolver to solve this problem. Instead of trying to change the core software that powers the agent, which is a difficult and expensive process, Evolver acts as an external memory bank that grows smarter with every task. The system operates on a simple principle: transform the raw, messy history of an agent's exploration into organized, reusable instructions. When an agent finishes a task, Evolver analyzes the session to extract the most useful strategies. It then packages these strategies into structured units, which the researchers call "Gene assets." These are not genetic codes in the biological sense, but rather compact, clear descriptions of a procedure, including exactly when it should be used and what steps it involves. The system also keeps a record of failed attempts, creating a list of things to avoid, ensuring that the agent does not repeat past errors.
The true power of Evolver lies in how it uses this accumulated knowledge. When a new task arrives, the system does not just let the agent wander blindly. Instead, it searches its library of Gene assets to find relevant strategies that match the current situation. It then feeds these instructions into the agent's workflow, guiding its decisions before it even begins to act. This process happens without altering the agent's fundamental programming. The agent remains the same, but it is now equipped with a curated set of historical wisdom that helps it navigate the problem more efficiently. The researchers tested this approach using two different types of advanced coding agents and four distinct benchmarks that measure performance in areas like software repair and terminal operations. They compared the agents' performance when working alone against their performance when assisted by Evolver's library of past experiences.
The results suggest that this method of turning exploration into evolution is highly effective. In a set of 320 paired tests, the agents using Evolver successfully completed tasks more often than those working alone, with the success rate rising from 83.13% to 87.50%. This improvement was consistent across different types of tasks and different underlying agent models. Perhaps even more striking was the gain in efficiency. The agents guided by Evolver used significantly fewer computational resources to reach their goals. On average, the amount of data processed, measured in tokens, dropped by 24.24%. This reduction was not just a side effect of failing less often; even in the cases where both the guided and unguided agents succeeded, the guided version still used 17.57% fewer resources. This indicates that the historical experience helped the agents find shorter, more direct paths to the solution, avoiding unnecessary detours and redundant calculations.
The researchers illustrated these gains with specific examples. In one task involving counting data tokens, an agent assisted by Evolver completed the job using 86.5% fewer resources and made far fewer attempts to call upon the AI model compared to its unassisted counterpart. In another instance involving a complex software bug, the guided agent solved the problem with 58.2% less data usage. These cases demonstrate that the system does not merely help agents succeed where they would otherwise fail; it also helps them succeed with greater elegance and speed. The study concludes that by treating past exploration as a persistent, reusable asset, it is possible to create agents that continuously improve their capabilities. This approach offers a practical path forward, allowing artificial intelligence systems to learn from their history and evolve their skills without the heavy cost of retraining their core models, effectively turning every completed task into a stepping stone for the next.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.