AI Research

CRAFT: Clustering Rubrics to Diagnose Weak LLM Capa... | AI Research

Key Takeaways

CRAFT: Clustering Rubrics to Diagnose Weak LLM Capabilities and Generate Targeted Fine-Tuning Data introduces a new method for improving large language model...
Evaluations should do more than measure a models current performance.
They should tell us what to fix for the next model iteration and provide a way to generate targeted post training data.
Most evaluation pipelines identify weak examples, topics, or categories, but they leave the underlying capability failure implicit: they say where a model fails, not why.
We introduce CRAFT, a method that converts any rubric based evaluation dataset into a model specific diagnosis of weak capabilities.

Paper AbstractExpand

Evaluations should do more than measure a models current performance. They should tell us what to fix for the next model iteration and provide a way to generate targeted post training data. Most evaluation pipelines identify weak examples, topics, or categories, but they leave the underlying capability failure implicit: they say where a model fails, not why. We introduce CRAFT, a method that converts any rubric based evaluation dataset into a model specific diagnosis of weak capabilities. CRAFT treats each grading criterion as a capability probe: it extracts a capability description from every prompt rubric pair, clusters these descriptions into a hierarchical capability tree, scores the target model at every node, and selects low performing nodes dynamically across tree levels, at the granularity where each failure is clearest. The selected weak capabilities then direct the generation of targeted supervised finetuning data. Holding the data generation, finetuning, and evaluation setup fixed, we compare CRAFT against prompt level EvalTree clustering and untargeted random generation on four open source models, two professional domains (finance and legal), and 13 held out benchmarks disjoint from the diagnostic data. CRAFT achieves the strongest finance domain average for all four models under repeated temperature decoding; on legal domain, it is strongest for three of four models and remains within the decoding variance bands of the best baseline on the fourth. Diagnosing weaknesses at the level of rubric criteria, rather than prompts or categories, thus yields both a sharper picture of what a model cannot do and measurably better models after finetuning on that diagnosis.

CRAFT: Clustering Rubrics to Diagnose Weak LLM Capabilities and Generate Targeted Fine-Tuning Data introduces a new method for improving large language models by moving beyond simple performance scores. While most evaluation pipelines tell developers where a model is failing—such as a low score on a math or legal test—they rarely explain why. CRAFT addresses this by using rubric-based evaluation data to create a detailed, model-specific map of a model's weaknesses. By identifying exactly which skills a model lacks, developers can generate highly targeted training data to fix those specific issues in future iterations.

Turning Rubrics into Capability Maps

The core of the CRAFT approach is treating rubric criteria as "capability probes." Instead of looking at a prompt as a single unit, CRAFT breaks down every prompt-rubric pair into a specific skill description. These descriptions are then organized into a hierarchical tree, where broad categories (like "Corporate Finance") branch out into increasingly specific skills (like "Capex budgeting"). Because the tree is built dynamically, it can capture weaknesses at the exact level of detail where they occur, whether that is a broad conceptual failure or a narrow, technical error.

Diagnosing and Targeting Weaknesses

Once the capability tree is built, CRAFT evaluates the target model against every node in the hierarchy. This creates an empirical "heat map" of the model's performance. The system then automatically selects the nodes where the model is underperforming. By focusing on these specific weak nodes, the system can generate synthetic training data that directly addresses the identified gaps. This ensures that the fine-tuning process is not just adding more data, but is specifically tailored to the model's unique shortcomings.

Proven Performance Gains

The researchers tested CRAFT against two other methods: a prompt-level clustering approach and a random data selection baseline. Across four different open-source models and two professional domains—finance and legal—CRAFT consistently outperformed the alternatives. In the finance domain, CRAFT achieved the strongest average performance for all four models. In the legal domain, it was the top performer for three models and performed within the expected range of the best baseline for the fourth. These results demonstrate that diagnosing weaknesses at the level of rubric criteria provides a much sharper and more effective guide for model improvement than traditional, broader evaluation methods.

Comments (0)

No comments yet

Be the first to share your thoughts!