AI agents get better by anticipating errors in their own software changes

edited by Lisa Lock, reviewed by Robert Egan

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources Like Athena guiding Odysseus, the IDM anticipates task-level outcomes to guide an optional revision before evaluation. Only the resulting child’s task agent is evaluated. Final forecasts and observed outcomes provide evidence for improving the IDM across generations.

Credit: arXiv (2026). DOI: 10.48550/arxiv.2610.13991 Some artificial intelligence agents can modify the programs and instructions governing their own work, with the goal of improving performance. These changes, however, can introduce errors, while testing unsuccessful modifications consumes resources.

Researchers at the University of Virginia, Microsoft Research and Tsinghua University recently developed forecasting software that helps AI agents anticipate the consequences of proposed edits before testing the modified software.

The approach, introduced in an article posted to the arXiv preprint server, was evaluated on programming tasks, predicting whether scientific papers would be accepted for publication, designing rewards for simulated robots and grading advanced mathematical solutions. Initial results suggest that using forecasts to guide software changes helped agents improve more efficiently.

The researchers' approach builds on HyperAgents, a framework combining two AI components or agents. The first performs tasks, while the second modifies the code and instructions used by both. Earlier versions and their results are preserved in an archive to inform subsequent changes.

The team introduced a new component called the Improvement Dynamics Model, which examines a proposed modification, the agent being modified and earlier results. It also predicts outcomes for individual tasks, distinguishing between incorrect answers and attempts that never finish.

Distinguishing between incorrect answers and unfinished attempts is valuable, as the two can require different interventions. An incorrect answer may require better reasoning, while an unfinished attempt could indicate that the software needs fixing.

If the forecast, adjusted for previous prediction bias, indicates worse performance than the current agent's measured performance, the modifying component receives the forecast and one additional revision opportunity.

The resulting agent is tested regardless of the final forecast. Changes affect its software and instructions, while the underlying language model weights remain fixed.

In one example outlined by the researchers, an edit placed instructions outside quotation marks, breaking the program. The modifying agent recognized the error but exhausted its tool calls before repairing it.

The forecaster predicted unfinished tasks, prompting another revision. After correcting the quotation marks and checking the syntax, the revised agent completed 11 tasks.

The researchers conducted five runs with 40 improvement rounds each for every experimental condition. They found that their forecasting approach developed stronger coding agents earlier than HyperAgents.

On 166 programming tasks kept separate from improvement and selection, selected agents achieved a median success rate of about 30%, compared with 22% for HyperAgents and 16% for their shared starting agent. Success required producing a software patch that passed task-specific tests.

The researchers also compared the approaches with matching budgets for language model tokens, the units of text processed or generated, and found that the forecasting approach retained a performance advantage. This comparison accounted for tokens used in modification, forecasting, forecast updates and task evaluation, but did not measure total computing cost.

The team's forecasting software can also improve. Every five rounds, the system proposes updates and checks their predictions against subsequently recorded outcomes before deciding whether to adopt them.

The team also tested whether the experience it gained could be transferred to new tasks by applying a forecaster developed jointly on paper-acceptance prediction and reward design for simulated robots to mathematical grading. During these tests, the researchers started with fresh task-solving and modifying programs.

Across 100 unseen grading tasks, selected agents exactly matched expert grades with a median accuracy of 55%. Agents using a fresh forecaster achieved 51%, while those relying on HyperAgents reached 47%.

These benchmark findings do not establish that AI agents can replace human reviewers, graders or programmers. Instead, they suggest that forecasting can help agents identify helpful software changes, and that experience predicting these changes could assist improvement on a different task.

Written for you by our author Ingrid Fadelli, edited by Lisa Lock, and fact-checked and reviewed by Robert Egan —this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You'll get an ad-free account as a thank-you.

Zixuan Xie et al., Predict Before You Modify: Learning Improvement Dynamics for Self-Evolving Agents, arXiv (2026). DOI: 10.48550/arxiv.2610.13991 Journal information: arXiv

Freelance journalist with BSc Psychology and MA International Journalism. Covers AI, robotics, neuroscience, and astrophysics since 2018. Full profile →

BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile →

Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →