Harness RSI
A method that confines recursive self-improvement to the agent's software harness layer: model parameters are frozen, and only prompts, memory mechanisms, workflows, and executable components undergo feedback-driven continual revision. Revisions persist across runs, yielding greater controllability and rollback capability.
Timeline (last 90 days)
The paper "Self-Evolving Agent Harness" proposes achieving self-evolution by revising the agent's software harness while keeping the model weights unchanged.
The core idea of Harness RSI is not to modify model parameters, but to let the software harness continuously revise itself in a feedback-driven loop, with adopted revisions persisting across runs.
The paper characterizes the search process of the Harness from the perspective of evolutionary optimization and analyzes its reliability using continual PAC learning.
All Facts (3)
The paper "Self-Evolving Agent Harness" proposes achieving self-evolution by revising the agent's software harness while keeping the model weights unchanged.
50%UnverifiedThe core idea of Harness RSI is not to modify model parameters, but to let the software harness continuously revise itself in a feedback-driven loop, with adopted revisions persisting across runs.
50%UnverifiedThe paper characterizes the search process of the Harness from the perspective of evolutionary optimization and analyzes its reliability using continual PAC learning.
50%