Autonomous Evaluation and Refinement of Digital Agents

We show that domain-general automatic evaluators can significantly improve the performance of agents for web navigation and device control. We experiment with multiple evaluation models that trade off between inference cost, modularity of design, and accuracy. We validate the performance of these models in several popular benchmarks for digital agents, finding between 74.4 and 92.9% agreement with oracle evaluation metrics. Finally, we use these evaluators to improve the performance of existing agents via fine-tuning and inference-time guidance. Without any additional supervision, we improve state-of-the-art performance by 29% on the popular benchmark WebArena, and achieve around 75% relative improvement in device control settings.

Paper

Similar papers

Reviewer tJw67/10 · confidence 4/52024-05-07

Summary

This paper tackles an important problem of agentic research: how to automate the evaluation of domain-general agents? Current agent benchmarks rely heavily on hard-coded rules for evaluation, which is expensive to construct and hard to scale. This paper proposes a multimodal approach to automate this process. They first train a VLM-based Captioner to generate a textual description of the agent's observation (i.e., the screenshot). Then the LLM-based Reasoner will take this description along with the actions, and the original user instruction to judge whether the execution is successful or not. Given such an automatic evaluator, they further show that it can be used to improve the performance of existing agents by providing inference-time guidance for self-refinement.

Rating

7

Confidence

4

Ethics flag

1

Reasons to accept

- This paper is well-motivated and clearly presented. - The problem being tackled is important and impactful. - The proposed approach, although quite simple, is effective on two evaluated benchmarks.

Reasons to reject

- The key to Captioner's success is the collection of 1,272 manually curated caption annotations, whose collection is slow and expensive. Is there any way to automate this process such that it can be scaled up? - The proposed approach, despite being effective on WebArena and AitW, might not work well in domains where screenshots cannot capture the change of state. For instance, if an agent executes "mv file_a foler_b" in the terminal. The action will be executed completely in the backend and nothing will be shown unless both folders are explicitly open and displayed on the screen.

Questions to authors

What would you perceive as the biggest challenge in automating agent evaluation?

Reviewer zFha7/10 · confidence 4/52024-05-09

Summary

This paper designs two lines of automatic evaluation for digital agents: (1) Using VLMs to caption the trajectory and then prompting the LLM to reason if the trajectory is correct. (2) Using an end-to-end VLM to evaluate the trajectory (GPT-4V in the paper). These methods show a high correlation with ground truth evaluation metrics, and the paper demonstrates that the agents can be improved by using reflective prompting or filtered BC.

Rating

7

Confidence

4

Ethics flag

1

Reasons to accept

- The paper is well-written and easy to follow. - The framework is interesting and holds promise for digital agent development. The paper has the potential to provide insights for the community. - The experiments involving filtered BC are interesting and address the data scarcity issue in some agent learning settings well. - The paper curates data and tasks in the iOS domain, which may benefit future studies on digital agents.

Reasons to reject

- Using LLMs to evaluate the trajectory of agent policies and improving the agent seems common in existing literature. For example, [1] also uses LLM evaluators to conduct filtered BC, and [2] uses reflection. - The evaluation of filtered BC in the iOS domain is not very persuasive because Table 2 only shows the comparison of the number of successful tasks for different methods. It is unclear if solving 14 out of 52 tasks represents a real improvement or just randomness. [1] Du et al., Vision-Language Models as Success Detectors. [2] Wang et al., JARVIS-1: Open-world Multi-task Agents with Memory-Augmented Multimodal Language Models.

Questions to authors

I think the methods in this paper are promising for digital agent development. But the weaknesses on originality and evaluation may limit the contribution of this paper. I would like to raise my assessment for the paper in the followup phase if the authors could well address the issues and make a rounded discussion.

Reviewer saqr6/10 · confidence 4/52024-05-13

Summary

This paper empirically studies model's automatic evaluation of trajectories on digital agents, such as agents operating on the web and Android. The authors focus on two methods to construct the evaluator, one is an end-to-end approach that directly uses a pretrained VLM, another is a modular approach that first transcribes the screenshot to captions with a VLM, and then feeds all text to a LM. Both methods demonstrate high accuracies against oracle evaluator and human judge. The authors then utilize such automatic evaluators to improve the policy agent, through reflexion and filtered behavioral cloning respectively. In both settings, the evaluation is able to improve the tasks.

Rating

6

Confidence

4

Ethics flag

1

Reasons to accept

1. Empirically studying whether current models are able to evaluate digital agents’ trajectory success answers a previously unknown question and is useful practically as well. 2. The authors successfully utilize the model evaluators to improve the performance on two difficult tasks, with different approaches. 3. This paper studies both end-to-end and modular approaches, across both proprietary and open-source models, which can provide practical guidance for different readers.

Reasons to reject

As an empirical paper, I expect the experiments to be conducted in a more comprehensive manner with more details, and in general the empirical results are relatively sparse. Some examples: (1) while the authors talk about both trajectory-level and step-level evaluators, step-level evaluation is only applied for the iOS agent refinement, and there is no study how accurate such step-level evaluation is. From Table 2, I guess step-level evaluation may not be very accurate given small improvement over self-training; (2) end-to-end methods only evaluate GPT-4V, while I would like to see the performance of more open VLMs; (3) In Table 2, filtered BC only finishes 3 more tasks than self-training in a test of 52 tasks. I am not sure how significant this result is.

Questions to authors

1. The high accuracy in Table 1 is nice, but I am feeling this is because the special properties of the two environments – for example, the final screenshot is typically enough to classify whether it satisfies the goal. I wonder whether the authors think such high accuracies are general, and whether there are environments where the final state is difficult to classify, I would like to see more discussions on this aspect as potential limitations of the paper. 2. Webarena by design is a text-only task originally, I see the authors are using GPT4-V on that, are the authors experimenting with VisualWebArena or am i missing sth? ``` After rebuttal, I bump up my score to 6 given the clarifications and added results to increase the significance of the previous results. ```

Reviewer zFha2024-06-03

The authors, Thank you for your responses. My concerns are addressed and I would like to raise the score. It would be better if the discussion and the new results for the rebuttal phase can be included in the revised paper, and the empirical experiments can be more organized. Reviewer zFha

Authorsrebuttal2024-06-04

Dear Reviewer zFha, Thanks again for your helpful feedback! We will definitely incorporate all the discussions and new results into the paper, and improve the organization of the experiments.

Reviewer saqr2024-06-04

Thank you for the clarifications! The added results on AitW is nice and significant, I hereby increase my score.

Program Chairsdecision2024-07-10

Decision

Accept

© 2026 NYSGPT2525 LLC