Elamur Research⌕
☰
home / papers / 2026 / Interpreto: An Explainability Library for Transformers

Interpreto: An Explainability Library for Transformers

Paper · 2026 · Antonin Poché, Thomas Mullor, Gabriele Sarti et al.

Summary

Explaining how Transformer language models reach their outputs matters for debugging, detecting bias and safety work. Existing tools are fragmented, though. Some handle only attributions, others cover only one stage of concept analysis, and many support a single task or lack documentation and metrics. The authors present Interpreto, an MIT-licensed Python library for HuggingFace models, from BERT-style encoders to LLMs. It offers one interface for both classification and text generation, split into two modules.

The attributions module scores how much input tokens, words or sentences contribute to a prediction. It uses perturbation- and gradient-based methods and adds faithfulness metrics and visualization. The concepts module covers the full workflow. It splits a model via nnsight to gather activations, learns concepts (largely through overcomplete), labels them (for example with an LLM), and estimates their effect on outputs.

Correctness is checked with unit tests, sanity checks on hand-built models, and tests across many architectures. A demo gallery with runnable snippets accompanies the release. The library's value lies in making recent interpretability research easier to apply and benchmark from a single package.

Key points

  • Attributions module: provides 10 methods (e.g., LIME, KernelSHAP, Occlusion) and 2 evaluation metrics, with granularity control. It supports SequenceClassification and CausalLM models.
  • Concepts module: offers 15 concept-learning options (e.g., Semi-NMF, SAEs), 3 interpretation methods, concept-importance estimation and 7 metrics (e.g., ConSim, Stability). It also supports MaskedLM and Seq2SeqLM.
  • Testing and requirements: tested on more than 15 architectures. Requires Python 3.10–3.13, torch ≥ 2.0, transformers ≥ 4.22 and nnsight ≥ 0.5.1. Installed via uv pip install interpreto.
  • Demo gallery: covers 6 models. The classifiers are DistilBERT/IMDB, BERT/emotion and RoBERTa/AG-News. The generators are GPT-2, Qwen3-0.6B and Llama 3.1 8B.
  • Example cost: learning 20 Semi-NMF concepts on Qwen3-0.6B from 100 AG-News samples, then labeling them with GPT-4.1-nano, takes under 3 minutes on an RTX 3080.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🔬Institut de Recherche en Informatique de ToulouseToulouse · France35 works · 1,872 citations🔬CentraleSupélecGif-sur-Yvette · France10 works · 1,330 citations🔬Mathématiques et Informatique pour la Complexité et les SystèmesGif-sur-Yvette · France3 works · 262 citations🔬IRT Saint ExupéryToulouse · France2 works · 34 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Interpretability · 130 works

Cite

Cite as Antonin Poché, Thomas Mullor, Gabriele Sarti, Frédéric Boisnard, Corentin Friedrich, Charlotte Claye, Francois Hoofd, Raphael Bernas, Nicholas Asher, Celine Hudelot and Fanny Jourdan. 2026. Interpreto: An Explainability Library for Transformers. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-demo.1

@inproceedings{poche2026interpreto,
  title = {{Interpreto: An Explainability Library for Transformers}},
  author = {Antonin Poché, Thomas Mullor, Gabriele Sarti, Frédéric Boisnard, Corentin Friedrich, Charlotte Claye, Francois Hoofd, Raphael Bernas, Nicholas Asher, Celine Hudelot and Fanny Jourdan},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-demo.1},
  url = {https://doi.org/10.18653/v1/2026.acl-demo.1}
}

References

3 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Guillaume Alain and Yoshua Bengio. 2016. Understanding intermediate layers using linear classifier probes. In Proceedings of the International Conference on Learning Representations (ICLR). find
  2. Outlook (later). We plan to extend library coverage to ViT (Vision Transformer) to later aim for multi-modal transformer support. Out of scope. We do not plan to include circuit-level MI methods, data attribution, or feature visualization; the library will remain centered on HuggingFace model workfl… — The Falcon Series of Open Language Models
  3. Ethical statement Giuseppe Attanasio, Eliana Pastor, Chiara Di Bonaventura, and Debora Nozza. 2023. ferret: a framework for benchmarking explainers on transformers. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL). find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/poche-2026-interpreto-explainability-library-transformers/