TL;DR
Verify by Doing (VbD) is a novel framework that enables LLM agents to self-improve at test time without any external feedback or ground truth. By co-evolving their execution ("Actor") and verification ("Evaluator") procedures through active tool-based inspection, agents can refine their internal success signals and recover successful outputs that were previously overlooked.
What Problem Does This Paper Address?
Agent self-improvement typically relies on feedback—knowing whether a task was successful. While benchmarks provide this via ground-truth labels, real-world tasks often lack external "oracles." In these scenarios, agents must infer success from their own observations.
However, if an agent’s internal evaluation process is flawed or incomplete, it creates a "blind spot." Simply generating more attempts won't help if the agent can't recognize a correct answer when it sees one. Existing methods often focus on improving the agent's actions (the "Actor") while keeping the evaluation process fixed, or they rely on held-out data to improve the judge. VbD tackles the challenge of within-task adaptation, where both execution and verification must evolve together from scratch, using only the evidence available in the environment.
Methods
VbD introduces a co-evolutionary loop that updates two distinct "harnesses"—collections of instructions, skills, and rubrics—for the Actor and the Evaluator.
- Active Verification: Unlike passive critiquing, the Evaluator in VbD uses tools to actively probe the environment state (e.g., checking database entries or reading specific spreadsheet cells) to gather evidence of task completion.
- Co-Evolution Loop: In each round, the agent generates new candidates using both current and proposed Actor harnesses. Simultaneously, a proposed Evaluator is tested against the current one.
- Comparative Selection: A fixed "Reviewer" compares the reports from the current and proposed Evaluators on the same set of candidates. An Evaluator update is only adopted if it provides better-supported findings and clearer evidence.
- Separate Retention: VbD makes three distinct decisions: which output to select as the final answer, whether to adopt the new Evaluator, and whether to adopt the new Actor. This decoupling prevents "lucky" successes from misguiding the improvement of the agent's core procedures.
Datasets
The authors evaluated VbD across three diverse benchmarks requiring different types of evidence:
- AutomationBench: 600 tasks across six workflow domains (e.g., Finance, HR, Sales) where success is measured by the final application state.
- $\tau^3$ Banking: 97 tasks involving dialogue and actions that must satisfy complex banking policies.
- GDPval: 133 tasks focused on professional deliverables like spreadsheets and documents, scored by a high-reasoning model (GPT-5.6-sol) using official rubrics.
Key Insights and Ideas
- Verification is a Target for Improvement: The "aha moment" is that the internal signal used for self-improvement shouldn't be static. By learning what to inspect and how to judge, the agent becomes a better self-teacher.
- Shared Evidence Discovery: A discovery made during verification (e.g., finding a specific API call that reveals a hidden error) can be directly translated into a new skill for the Actor to use in future attempts.
- Recovering Success from the Past: Improved verification often allows the agent to identify a successful output that was already generated in an earlier round but was initially rejected by a weaker evaluator.
- Robustness to Regression: VbD is exceptionally good at rejecting "regressions"—proposed updates that actually make the agent perform worse. As shown in Figure 4(a), VbD achieved a True Negative Rate (TNR) of 89.3%, significantly higher than prior co-evolution methods like RQGM.
Key Results
VbD consistently achieved the highest scores among all methods operating without external feedback across two different model backbones (DeepSeek-V4-Flash and Qwen3.8-Flash-Next).
- AutomationBench: VbD reached a 37.0% pass rate, outperforming the strongest baseline (RQGM) by 3.3 percentage points.
- $\tau^3$ Banking: It achieved a 49.5% pass rate, a 14.4 point gain over the base model and 4.1 points over the best baseline.
- GDPval: VbD led in all nine sectors, achieving 82.4% overall rubric quality.
- Efficiency: VbD's second-round performance often exceeded the final-round performance of other methods, sometimes at a lower cumulative token cost.
Limitations and Future Work
- Shared Errors: If the Actor, Evaluator, and Reviewer all share the same fundamental misunderstanding of a task requirement, VbD cannot easily correct it.
- Environment Constraints: The method requires environments that can be reset and inspected. Irreversible actions or delayed effects pose a challenge for active verification.
- Inference Cost: The iterative nature of co-evolving two harnesses incurs significant token costs compared to single-shot execution.
- Future Directions: The authors suggest exploring cross-task transfer, where lessons learned in one task (like a new verification skill) are carried over to future, different tasks. They also note that the current study was limited to three rounds of evolution; longer-term dynamics remain an open question.
Summary updated