Verify by Doing: Co-Evolving Agent Execution and Evaluation without External Feedback
Organizations: Ant International · Zhejiang University
Abstract
Agent self-improvement requires feedback about task success, yet external feedback may be unavailable at test time. Agents must then infer success from observable evidence, and incomplete verification can misguide subsequent updates. We introduce Verify by Doing (VbD), a method that co-evolves Actor and Evaluator harnesses within a single task while keeping model weights fixed. The Evaluator actively inspects artifacts and environment state with tools; its findings ground an internal signal for output selection and guide updates to both execution and verification procedures. Proposed Evaluators undergo comparative review on common candidates, while Actor updates require support from their own fresh outputs. Across AutomationBench, Banking, and selected GDPval tasks, VbD achieves the highest overall scores among the compared methods operating without external feedback, with gains of 3.3, 4.1, and 2.3 percentage points over the strongest baselines on the main backbone. Ablations support the contributions of active verification and Evaluator evolution, and selection from fixed candidate pools shows that improved verification can recover successful outputs without generating new ones.
Figures & tables
| Method | Finance | Sales | Marketing | Operations | Support | HR | Overall |
|---|---|---|---|---|---|---|---|
| Base model | |||||||
| BoN | |||||||
| Self-preference | |||||||
| HGM | |||||||
| RHO | |||||||
| RHI |
| Method | Pass rate (%) | |
|---|---|---|
| Base model | ||
| BoN | ||
| HGM | ||
| RHO | ||
| RHI | ||
| RQGM |
| Method | Prof | Govt | Manuf | RealEst | Fin | Info | Whsle | Health | Retail | Overall |
|---|---|---|---|---|---|---|---|---|---|---|
| Base model | ||||||||||
| BoN | ||||||||||
| HGM | ||||||||||
| RHO | ||||||||||
| RHI | ||||||||||
| RQGM |
Appendix figures & tables55 assets
Supplementary material from the paper’s appendix.
Appendix
| Parameter | Value | Scope |
|---|---|---|
| Backbone assignment | One model per run | Fixed weights; all roles within a task. |
| Context window | 1M tokens | Formatted input plus reserved output. |
| Tool calling | Native | All compared methods. |
| Controller temperature | 0 | Evaluator, editor, and comparative review. |
| Editable components | Instructions, skills, rubrics; persistent check plans | Both harnesses; persistent check plans are Actor state. |
| Evolution rounds | 3 | Initialize in Round 1; test updates in Rounds 2–3. |
| Parameter | DeepSeek-V4-Flash-0731 | Qwen3.8-Flash-Next |
|---|---|---|
| Backbone parameters | 284B | 125B |
| Activated parameters per token | 13B | 6B |
| Reasoning effort (highest level) | max | xhigh |
| Method | Updated components | Adaptation / selection procedure |
| Base model | No updates | One Actor attempt. |
| BoN ( ) | No updates | Select among six candidates from one round. |
| Self-preference | No updates | Agentic Evaluation and output retention under VbD’s schedule. |
| HGM | Actor | Posterior-guided Actor policy search with a fixed Evaluator. |
| RHO | Actor | Diagnose prior executions of the current task; guide updates by self-preference with a fixed Evaluator. |
| RHI | Actor instructions | Revise instructions from task-local harness and pairwise preference history with a fixed Evaluator. |
| Method | Adapted components | Feedback | Scope |
|---|---|---|---|
| Actor harness optimization | |||
| GEPA [ Agrawal et al., 2026 ] | Prompts and instructions | Validation metric + reflection | Across a task set |
| HGM [ Wang et al., 2026b ] | Agent code | Benchmark utility | Descendant search |
| RHO [ Pan et al., 2026 ] | Tools, skills, instructions | Diagnosis + self-preference | Future tasks (offline) |
| RHI [ Lee et al., 2026 ] | Agent loop instructions | Pairwise preference history | Within a task |
| TTHE [ Nie et al., 2026 ] | Executable harness | Traces + proxy signals | Across batches |
| Method | Finance | Sales | Marketing | Operations | Support | HR | Overall |
|---|---|---|---|---|---|---|---|
| Base model | |||||||
| BoN | |||||||
| HGM | |||||||
| RHO | |||||||
| RHI | |||||||
| RQGM |
| Method | Pass rate (%) | |
|---|---|---|
| Base model | ||
| BoN | ||
| HGM | ||
| RHO | ||
| RHI | ||
| RQGM |
| Method | Prof | Govt | Manuf | RealEst | Fin | Info | Whsle | Health | Retail | Overall |
|---|---|---|---|---|---|---|---|---|---|---|
| Base model | ||||||||||
| BoN | ||||||||||
| HGM | ||||||||||
| RHO | ||||||||||
| RHI | ||||||||||
| RQGM |
| Method | Proposed | Accepted | TP | FP | TN |
|---|---|---|---|---|---|
| HGM | 1,200 | 227 | 158 | 56 | 108 |
| RHO | 1,432 | 698 | 302 | 121 | 89 |
| RHI | 1,723 | 1,412 | 198 | 240 | 97 |
| RQGM | 1,200 | 231 | 164 | 49 | 121 |
| VbD | 1,104 | 207 | 156 | 45 | 374 |
| Round | Round 1 | Round 2 | Round 3 |
|---|---|---|---|
| Actor update adopted | No | Yes | No |
| Evaluator update adopted | – | Yes | No |
| Selected output | Round 1 | Round 2 | Round 2 |
| Offline outcome | Fail | Pass | Pass |
| Round | Round 1 | Round 2 | Round 3 |
|---|---|---|---|
| Actor update adopted | No | No | No |
| Evaluator update adopted | – | No | Yes |
| Selected output | Round 1 | Round 1 | Round 2 |
| Offline outcome | Fail | Fail | Pass |
| Metric | Round 1 | Round 2 | Round 3 |
|---|---|---|---|
| Tool calls per attempt | 63.8 | 51.8 | 54.6 |
| Attempts with Verify (%) | 20.4 | 24.0 | 25.3 |
| Attempts loading skills (%) | 0.0 | 22.2 | 23.1 |
| Action | Recorded operation |
|---|---|
| Search | Discover APIs or search for relevant records. |
| Read | Retrieve a specified resource or state. |
| Edit | Create, update, or delete records or artifacts. |
| Execute | Send, publish, submit, or transform data. |
| Verify | Read the state after an earlier action. |
| Skill | Explicitly load a skill. |
LLM Agent Verification
LLM agent verification checks whether agents’ actions, evidence, policies, and completion claims satisfy explicit requirements.
See all 166 papers