Benjamin Heinzerling

Natural Language Understanding Team Deputy Director @RIKEN_AIP, Associate Professor @TohokuNLP, based in Sendai, Japan. PhD @UniHeidelberg @HITStudies

[Benjamin Heinzerling @benbenhh]

We've released a new version of our pretained byte pair embeddings in 275 languages. Now with pip install, automatic download of embeddings and sentencepiece models, convenient subword segmentation, and tons of pretty UMAP visualizations. https://t.co/wNbg5nYKhx

[Joe Stacey @_joestacey_]

Excited to share my first postdoc paper with @SheffieldNLP ! 🤩

In this work we argue that supervised uncertainty quantification (UQ) needs better evaluation

Want to know more? Here's a little summary 🧵

[Joe Stacey @_joestacey_]

This work has been really fun to work on! Massive thank you my fantastic collaborators @OrgadHadas @inuikentaro @benbenhh and @NafiseSadat

You can find the paper here:
https://t.co/RusU6LD50C

[Masaki @MskSakata]

#NLP2026
明日 3/11（水）11:15–12:45 @ B会場（B6-16）でポスター発表します！

Japan ⊂ Eastern Asia ⊂ Asia ... のような概念の階層構造が，⾔語モデルの中間層にどのように・どの程度埋め込まれているかについて明らかにしました。

https://t.co/wtb8d6dA5l

[R. Maruyama @rmaruy]

Benjamin Heinzerling @benbenhh さんの『人工知能』のコラム、10名の博士課程学生が一斉に雇用されるGraduiertenkolleg（Research Training Group）というドイツの仕組みを紹介していて、とても面白い。https://t.co/BHVh9WCdV6

[AACL 2026 @aaclmeeting]

The deadline for AACL-IJCNLP tutorial proposals is on August 14th AoE. We are interested in introductory as well as cutting edge half/full day tutorials. Introductory tutorials can be non-CL/non-NLP but interesting to the community.

https://t.co/JFmdfndkVW

[Joe Stacey @_joestacey_]

Super excited to have a cool post-doc lined up for after my PhD, working on Uncertainty Quantification with @NafiseSadat at @SheffieldNLP, and @benbenhh and @inuikentaro from RIKEN, Japan.

Could not be happier 🙂

[Benjamin Heinzerling @benbenhh]

@_joestacey_ @NafiseSadat @SheffieldNLP @inuikentaro Really excited and can't wait to start working with you!

[sho_yokoi @sho_yokoi]

「言語モデルの内部機序：解析と解釈」のスライドを公開しました
https://t.co/yf3Fjf8kZg

言語処理学会年次大会 #NLP2025 のチュートリアル資料です。w/ @benbenhh さん @goro_koba さん
皆で言語モデルの中を凝視していきましょう🔬

[sho_yokoi @sho_yokoi]

3月10日(月) に言語モデルの解釈性のチュートリアルをします🔥
@tohoku_nlp の @benbenhh さん @goro_koba さんと一緒に、モデルの内部機序（内部表象・計算メカニズム）について既存のどんな資料よりも詳しく分かりやすく解説します @ #NLP2025 長崎/オンライン。お楽しみに〜！

[Sheffield NLP @SheffieldNLP]

🚀 Reminder! 🚀

The deadline for the Postdoc in Uncertainty Quantification for Foundation Models at @sheffieldNLP with @NafiseSadat @benbenhh @inuikentaro is March 2nd—just a few days left! 🗓️
📆 Initial contract: 12 months (with a possibility of an additional 2-year extension)

[Nafise Sadat Moosavi @NafiseSadat]

🚀 Postdoc Opportunity! 🚀

With @benbenhh & in collaboration with @inuikentaro, we’re hiring a Postdoc at @sheffieldNLP for Uncertainty Quantification in Foundation Models, with a chance to spend time at RIKEN, Japan!

📅 Deadline: 2 March
🔗 More info: https://t.co/dbc4DznLhd

[NLP2026 UTSUNOMIYA @anlpmeeting]

#NLP2025 では、3/10（月）午後 に、4件のチュートリアルを行います！
言語モデルの内部機序：解析と解釈（仮）
Benjamin Heinzerling  先生（理化学研究所／東北大学）
横井 祥 先生（東北大学／理化学研究所）
小林 悟郎 先生（東北大学／理化学研究所）
https://t.co/1rZMwe6bJM

[Yufang Hou @yufanghou]

📢 Academic Job Alert: Deadline Approaching Fast – Only 2 Days Left! (Apply by November 30th)

I'm currently seeking 2 postdoctoral researchers and 3 PhD students to join my team at IT:U-Interdisciplinary Transformation University Austria (@itu_austria)!

[Benjamin Heinzerling @benbenhh]

preprint: https://t.co/5e77T0cY7s

joint work with @inuikentaro at @tohoku_nlp

[Benjamin Heinzerling @benbenhh]

LLMs give the right answer to questions like "When was Karl Popper born?" but do they have the right internal representation? We find evidence that they do: There are directions in activation space that have a causal, interpretable effect on the expression of numeric properties

[Benjamin Heinzerling @benbenhh]

By activation patching along these directions we can make the LM express earlier/smaller and larger/later values across all properties and LMs we looked at. This suggests that monotonic representation of numeric properties consistently emerges during pretraining

[Benjamin Heinzerling @benbenhh]

Dimension reduction via PLS reveals the monotonic structure of LM representations: There are clearly visible directions in activation space that correlate with numeric entity attributes such as birthyear, death year, and geo-coordinates

[Benjamin Heinzerling @benbenhh]

@tbmihaylov @fedelopez77 Congrats, Fede!