New Self-Supervised Method Enables AI Agents to Improve Without External Validation Data
Researchers introduced Retrospective Harness Optimization (RHO), a self-supervised technique that allows AI agents to improve their skills and tools using only past task attempts, without requiring labeled validation data. The method works by having agents analyze their own previous attempts, generate improvements, and select the best updates through self-evaluation. Testing across software engineering and knowledge work tasks showed significant improvements, including raising performance on SWE-Bench Pro from 59% to 78% in a single optimization round.
Researchers at arXiv have published a new approach called Retrospective Harness Optimization (RHO) that addresses a practical challenge in deploying AI agents: the difficulty of obtaining labeled validation data to improve agent performance. RHO is a self-supervised method that optimizes an agent's harness—its collection of skills, tools, and workflows—using only historical trajectories from past task attempts. The system works by selecting a diverse set of challenging tasks from previous attempts, re-solving them in parallel, and having the agent evaluate its own rollouts using self-validation and self-consistency checks before generating and selecting candidate improvements. The researchers evaluated RHO across three domains including software engineering, technical work, and knowledge work, demonstrating that a single optimization round improved performance on SWE-Bench Pro from 59% to 78% without external grading. The analysis shows that RHO effectively targets and corrects the agent's prior failure modes, resulting in altered behavior patterns and sustained higher accuracy over extended task sequences.
What's missing
The paper does not discuss potential limitations of self-preference mechanisms, such as whether agents might reinforce their own biases or errors, or how the method performs when agents have systematic blind spots that self-evaluation cannot detect. Additionally, generalization to domains beyond software engineering and knowledge work, computational costs of the optimization process, and comparison with other self-improvement approaches are not addressed in the abstract.
What different sources said
- arXiv cs.AICenter
Evolving Agents in the Dark: Retrospective Harness Optimization via Self-Preference
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.