Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2512.20638 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866916067246342144 |
|---|---|
| author | Bohacek, Maty Scherrer, Nino Dufour, Nicholas Leung, Thomas Bregler, Christoph Chan, Stephanie C. Y. |
| author_facet | Bohacek, Maty Scherrer, Nino Dufour, Nicholas Leung, Thomas Bregler, Christoph Chan, Stephanie C. Y. |
| contents | The evaluation of large language models relies heavily on standardized benchmarks. These benchmarks provide useful aggregated metrics, but can obscure (i) particular sub-areas where the models are weak ("model gaps") and (ii) imbalanced coverage in the benchmarks themselves ("benchmark gaps"). To automatically uncover both types of gaps, we propose a simple new method using concept activations from sparse autoencoders, to identify fine-grained gaps on a per-concept basis. The method also benefits from grounding evaluation in the model's internal representations, as well as easy comparison across benchmarks. We applied the method to five popular open-source models and more than a dozen benchmarks, as illustrative examples. As validation of the approach, we found that our automatic, unsupervised method was able to recover model gaps that have been previously documented in the literature (e.g. relating to sycophancy), in addition to identifying novel model gaps. We were also able to automatically uncover benchmark gaps: core concepts that should fall within the scope of a given benchmark. Our "competency gaps" method can be used to complement existing benchmarks, by providing a concept-level decomposition of model behavior, and by helping benchmark developers iterate upon benchmark design. Code is available at https://competency-gaps.github.io. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2512_20638 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Uncovering Competency Gaps in Large Language Models and Their Benchmarks Bohacek, Maty Scherrer, Nino Dufour, Nicholas Leung, Thomas Bregler, Christoph Chan, Stephanie C. Y. Computation and Language Artificial Intelligence Machine Learning The evaluation of large language models relies heavily on standardized benchmarks. These benchmarks provide useful aggregated metrics, but can obscure (i) particular sub-areas where the models are weak ("model gaps") and (ii) imbalanced coverage in the benchmarks themselves ("benchmark gaps"). To automatically uncover both types of gaps, we propose a simple new method using concept activations from sparse autoencoders, to identify fine-grained gaps on a per-concept basis. The method also benefits from grounding evaluation in the model's internal representations, as well as easy comparison across benchmarks. We applied the method to five popular open-source models and more than a dozen benchmarks, as illustrative examples. As validation of the approach, we found that our automatic, unsupervised method was able to recover model gaps that have been previously documented in the literature (e.g. relating to sycophancy), in addition to identifying novel model gaps. We were also able to automatically uncover benchmark gaps: core concepts that should fall within the scope of a given benchmark. Our "competency gaps" method can be used to complement existing benchmarks, by providing a concept-level decomposition of model behavior, and by helping benchmark developers iterate upon benchmark design. Code is available at https://competency-gaps.github.io. |
| title | Uncovering Competency Gaps in Large Language Models and Their Benchmarks |
| topic | Computation and Language Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2512.20638 |