On 11 November 2026, researchers submitted a paper describing Verify by Doing (VbD), a method for improving an agent's execution and evaluation procedures within a task when external feedback is unavailable. The paper says the method keeps the model's weights fixed while changing the procedures around it.[1]
The authors describe two components: an Actor that executes tasks and an Evaluator that uses tools to inspect artifacts and the state of the environment. The Evaluator's findings provide an internal signal for selecting outputs and guiding changes to both components. Proposed Evaluators are compared on common candidate outputs, while proposed Actor updates require support from their own fresh outputs.[1]
The abstract reports the highest overall scores among the compared methods without external feedback on AutomationBench, τ³ Banking and selected GDPval tasks. It claims gains of 3.3, 4.1 and 2.3 percentage points, respectively, over the strongest baselines on the main backbone. The authors also report ablations supporting active verification and Evaluator evolution, and say improved verification selected successful outputs from fixed candidate pools without generating new ones. These are the authors' reported results, not independent findings.[1]
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
52f7e2436f25b89f3e48d97d22aec3ac5acb7e02a5d933f2c192eb5ae2722969