Bloom: Open-Source AI Behavior Evaluation Framework

Bloom: A New Era for Automated AI Model Evaluation

Evaluating the behavior of frontier AI models is a critical but challenging task for developers and researchers. Traditional methods are often slow to build and can quickly become outdated. To address this, Anthropic has released Bloom, an open-source agentic framework designed to automate and scale the creation of behavioral evaluations for AI models.

Bloom empowers builders to quickly quantify specific behaviors, like bias or sycophancy, by automatically generating a wide range of test scenarios. This approach significantly speeds up the evaluation process, allowing teams to move from concept to results in days rather than weeks.

How Bloom Works: An Automated Four-Stage Pipeline

Bloom operates through a clever, four-stage pipeline that uses AI agents to manage the entire evaluation process, reflecting recent advancements in AI agent development. A developer simply provides a description of the behavior they want to test and a configuration file. From there, Bloom takes over.

• 1. Understanding: The first agent analyzes the behavior description and any example transcripts to build a deep contextual understanding of what needs to be measured.
• 2. Ideation: Next, an ideation agent generates diverse and creative evaluation scenarios designed to provoke the target behavior. Each scenario defines the situation, user persona, and system prompt.
• 3. Rollout: The scenarios are then executed in parallel. An agent simulates user and tool interactions to create a realistic environment for the model being tested.
• 4. Judgment: Finally, a judge model scores each interaction for the presence of the specified behavior, and a meta-judge provides a high-level analysis of the entire test suite.

This dynamic process produces fresh scenarios on each run, preventing the staleness of fixed evaluation sets while ensuring reproducibility through a "seed" configuration file.

Key Features for Developers

Bloom is built with the needs of AI builders in mind, offering a powerful set of features to streamline safety and alignment research.

• High Configurability: Developers can customize every stage of the pipeline, from the models used as agents to the length and modality of interactions.
• Dynamic Scenarios: Unlike static benchmarks, Bloom generates new test cases on the fly, providing a more robust and realistic measure of model behavior.
• Proven Reliability: Bloom's automated judgments show a strong correlation with human evaluations (a Spearman correlation of 0.86 with Claude Opus 4.1 as the judge).
• Developer-Friendly Integrations: The framework integrates with Weights & Biases for large-scale experiments and exports transcripts compatible with the industry-standard Inspect format.

Impact and Practical Applications

For developers, Bloom's biggest advantage is its efficiency. It removes the need to spend significant time on evaluation pipeline engineering, freeing up resources to focus on interpreting results and improving models. As the AI development landscape matures, open-source solutions like Bloom are becoming essential counterparts to the integrated toolsets offered by major platforms, reflected in updates like Azure's December AI & API offerings for developers.

In a case study on "self-preferential bias," Bloom not only replicated the findings from a manual evaluation but also uncovered deeper insights. It revealed that increasing a model's reasoning effort could reduce bias by prompting the model to recognize a conflict of interest. This demonstrates Bloom's power not just to validate known issues but to discover new behavioral nuances.

By providing a scalable and reliable framework for testing complex behaviors, Bloom is an essential tool for any team focused on building safe and aligned AI systems. It is now available for the community to use for applications like testing jailbreak vulnerabilities, measuring evaluation awareness, and generating sabotage traces.

Discover more cutting-edge AI apps and apps on Appse, your go-to directory for the latest AI innovations.

Source: Anthropic Research: Bloom