The leading workshop on analysing and interpreting neural networks for NLP Co-located with EMNLP 2026 in Budapest, Hungary
EMNLP 2026
Joined May 2023
- Tweets183
- Following18
- Followers860
- Likes80
Time for our closing session! 🏆 The Best Paper Award (Main Track) goes to “Does Out-of-Sight Equal Out-of-Mind in CoT Monitorability?” by Pedro Lobato Ferreira, Wilker Aziz, and Ivan Titov. Congratulations!
The Best Paper Award (Reproducibility Track) goes to “Probing Without Labels: Contrast-only Reproductions of Interpretability Results” by Stefan F. Schouten and Peter Bloem, with a $500 prize from our generous sponsor NDIF @ndif_team . Congratulations! 🏆
That’s a wrap on BlackboxNLP 2026! Thank you to all our speakers, authors, reviewers, and attendees. See you next year at #EMNLP2027 in Cancún!
Our last oral: “Does Out-of-Sight Equal Out-of-Mind in CoT Monitorability?” by Pedro Lobato Ferreira, Wilker Aziz, and Ivan Titov, on how well monitors reading a model’s chain of thought can flag misalignment, reward hacking, and misleading context reliance.
Takeaways: for this problem, latent reasoning turns out to be as monitorable as explicit reasoning, though the latent monitors needed training. Verbalizing latents, e.g. with the logit lens or auxiliary decoders, looks like a promising way forward.
From our reproducibility track: “Probing Without Labels: Contrast-only Reproductions of Interpretability Results” by Stefan F. Schouten and Peter Bloem revisits interpretability findings using contrast pairs instead of labeled data.
Our second oral session kicks off with “Probing for Knowledge Attribution in Large Language Models” by Ivo Brink, Alexander Boer, and Dennis Ulmer. A simple linear probe can tell whether an answer comes from the context or the model’s own memory
Time for our final keynote! Michael Hahn @mhahn29 is presenting “From Transformers to Programs and Automata: The Theory of Opening the Black Box,” exploring when we can recover faithful representations of what Transformers compute.
Next, Michael presented D-RASP, a method for decompiling Transformers trained on algorithmic tasks into simple programs. The result is the most direct evidence so far that Transformers often implement interpretable RASP programs.
Starting off, Michael showed that a model can only hide so much computation before it starts showing up in its chain of thought, with bounds set by properties of the model itself
Starting soon in room P3! Come join us
After lunch, at 14:00, @mhahn29 will talk about turning Transformer computations into explicit programs and automata, and what thinking traces do (and don’t) reveal about how models reach their answers.
What a turnout for Poster Session 1, with packed aisles across all our tracks! Missed it? Poster Session 2 is at 16:00, see you there!
Our third oral: "A circuit for predicting hierarchical structure in-context in LLMs” by Tankred Saanum, Can Demircan, Samuel J. Gershman, and Eric Schulz. They find adaptive induction heads that learn in-context what to attend to, guided by heads that infer latent contexts.
Our second oral: “How Much Sonority Structure Do Phonotactic Language Models Learn from Distribution?” by Jia Can Xu and Mario Giulianelli asks whether models trained on sound sequences pick up sonority, the rise and fall of sounds within syllables, just from data.
Our first oral presentation is starting! “Data Attribution of Emergent Misalignment with Persona Features” by Clemens Vetter, David Kaczér, Lucie Flek, and Florian Mai traces emergent misalignment back to the training data behind it
