Multi-Agent LLM Committees for Autonomous Software Beta Testing

Manual software beta testing is costly and time-consuming, while single-agent large language model (LLM) approaches suffer from hallucinations and inconsistent behavior. We propose a multi-agent committee framework in which diverse vision-enabled LLMs collaborate through a three-round voting protocol to reach consensus on testing actions. The framework combines model diversity, persona-driven behavioral variation, and visual user interface understanding to systematically explore web applications. Across 84 experimental runs with 9 testing personas and 4 scenarios, multi-agent committees achieve an 89.5 percent overall task success rate. Configurations with 2 to 4 agents reach 91.7 to 100 percent success, compared to 78.0 percent for single-agent baselines, yielding improvements of 13.7 to 22.0 percentage points. At the action level, the system attains a 93.1 percent success rate with a median per-action latency of 0.71 seconds, enabling real-time and continuous integration testing. Vision-enabled agents successfully identify user interface elements, with navigation and reporting achieving 100 percent success and form filling achieving 99.2 percent success. We evaluate the framework on WebShop and OWASP benchmarks, achieving 74.7 percent success on WebShop compared to a 50.1 percent published GPT-3 baseline, and 82.0 percent success on OWASP Juice Shop security testing with coverage of 8 of the 10 OWASP Top 10 vulnerability categories. Across 20 injected regressions, the committee achieves an F1 score of 0.91 for bug detection, compared to 0.78 for single-agent baselines. The open-source implementation enables reproducible research and practical deployment of LLM-based software testing in CI/CD pipelines.

Exploring Embedding Priors in Prompt-Tuning for Improved Interpretability and Control

Prompt-Tuning is an efficient method for adapting pre-trained language models to new tasks with minimal computational overhead by modifying prompt embeddings. In this work, we investigate how crucial the phenomenon of embedding collapse, frequently observed in Prompt-Tuning, is for the final performance of the model. To address this question, we designed embedding priors and compared them with posteriors of the converged Soft and Deep Prompt-Tuning methods. Our findings suggest that priors strongly affect the position of the tuned embeddings, and models can effectively work with embeddings from different parts of activation spaces, including completely new regions. As the final Prompt-Tuning capabilities are limited, we hypothesize that controllable Prompt-Tuning posteriors may serve as a good starting point for tasks such as chain-of-thought (COT) distillation. Our experiments also show that generated trajectories are not localized in the activation space of the models. However, there are distinct clusters of activations for distant tasks (e.g., NLP and arithmetic), while activations between NLP tasks (e.g., Question-Answering and MLM) lie in the same cluster. These observations raise questions about the importance of a single activation cluster for the generalization abilities of large language models.