Lightnews — Scholar-powered news

Nils Feldhus @nfel.bsky.social · 6d

🙏 Many thanks to the institutions that supported this research:
@tuberlin.bsky.social
@bifold.berlin

Looking forward to presenting this in 🇨🇳 Suzhou early November!

3

Nils Feldhus @nfel.bsky.social · 6d

Our synthesis reveals a growing demand for more rigorous, causal evaluation. By outlining the state of the art and identifying key challenges, this survey provides a roadmap for future research toward making models more transparent.

This survey has been accepted at @blackboxnlp.bsky.social at EMNLP

Concept description evaluation techniques categorized by metric, study, and the underlying quality being measured. Metrics are grouped into conceptual families: predictive simulation, input-based evaluation, output-based evaluation, semantic similarity, and human judgment.

1 5

Nils Feldhus @nfel.bsky.social · 6d

We consider concept descriptions in open-vocabulary settings, the evolving landscape of automated and human metrics for evaluating them, and the datasets that underpin this research.

This is a companion paper to our PRISM paper that was accepted at NeurIPS last week: bsky.app/profile/lkop...

Concept description techniques categorized by component/abstraction (Neurons, SAEs, Circuits, Attention Heads), description source, and target dataset.

1 3

Nils Feldhus @nfel.bsky.social · 6d

🔍 Are you curious about uncovering the underlying mechanisms and identifying the roles of model components (neurons, …) and abstractions (SAEs, …)?

We provide the first survey of concept description generation and evaluation methods.

Joint effort w/ @lkopf.bsky.social

📄 arxiv.org/abs/2510.01048

Overview of descriptions for model components (neurons, attention heads) and model abstractions (SAE features, circuits).

1 3 17

Reposted by Nils Feldhus

Laura Kopf @lkopf.bsky.social · 19d

Happy to share that our PRISM paper has been accepted at #NeurIPS2025 🎉

In this work, we introduce a multi-concept feature description framework that can identify and score polysemantic features.

📄 Paper: arxiv.org/abs/2506.15538

#NeurIPS #MechInterp #XAI

1 3 25

Nils Feldhus @nfel.bsky.social · Aug 25

The submission deadline of the inaugural Young Researchers workshop at INLG 2025 has been extended by 5 days.
We're excited to receive your 2p position papers showcasing your NLG-related research until August 31, 2025! @siggen.bsky.social

ynlg-workshop.github.io

bsky.app/profile/nfel...

1

Reposted by Nils Feldhus

Nils Feldhus @nfel.bsky.social · Aug 12

Excited to announce the first-ever Workshop for Young Researchers in Natural Language Generation (YNLG), supported by @siggen.bsky.social, taking place on October 29, 2025 in Hanoi, Vietnam, co-located with INLG 2025.
Call for Submissions is out now!

ynlg-workshop.github.io

1 8 9

Nils Feldhus @nfel.bsky.social · Aug 12

We welcome position papers on ongoing research in Natural Language Generation.

Let’s build a vibrant, collaborative community of the next generation of NLG researchers!

Stoked to organize it together with @patuchen.bsky.social, Adarsa, Rudali, Michela, and Alyssa! 🤩

3

Nils Feldhus @nfel.bsky.social · Aug 12

📍 Venue: Vietnam Institute for Advanced Study in Mathematics, Hanoi, Vietnam — during INLG 2025.

💡 If you'll be in the region for INLG or EMNLP 2025, this is a great opportunity to connect and share your work!

📅 Submission deadline: August 26, 2025

1

Nils Feldhus @nfel.bsky.social · Aug 12

YNLG aims to foster a welcoming community of early-career researchers in NLG. This is your chance to:
📌 Present your work-in-progress research
📌 Receive constructive feedback from senior researchers and peers
📌 Discuss current trends and future directions in NLG

1 2

Nils Feldhus @nfel.bsky.social · Aug 12

Excited to announce the first-ever Workshop for Young Researchers in Natural Language Generation (YNLG), supported by @siggen.bsky.social, taking place on October 29, 2025 in Hanoi, Vietnam, co-located with INLG 2025.
Call for Submissions is out now!

ynlg-workshop.github.io

1 8 9

Nils Feldhus @nfel.bsky.social · Jul 28

Thoroughly enjoying the range of topics at the first #ACL2025NLP poster session!

Our FitCF poster presentation on counterfactual example generation at #ACL2025 has been moved to Tuesday, July 29, at 16:00-17:30.

bsky.app/profile/nfel...

Nils Feldhus @nfel.bsky.social · Jul 26

Qianli Wang ( @qiaw99.bsky.social ) et al.:
"FitCF: A Framework for Automatic Feature Importance-guided Counterfactual Example Generation"
📄 ACL Anthology: aclanthology.org/2025.finding...
🏟️ ACL Findings, July 28 @ Hall 4/5
Poster presentation: 18:00-19:30 (pres. by Qianli and myself)

FitCF: A Framework for Automatic Feature Importance-guided Counterfactual Example Generation

Qianli Wang, Nils Feldhus, Simon Ostermann, Luis Felipe Villa-Arenas, Sebastian Möller, Vera Schmitt. Findings of the Association for Computational Linguistics: ACL 2025. 2025.

aclanthology.org

2

Nils Feldhus @nfel.bsky.social · Jul 26

We introduce the TableEval benchmark and investigate the effectiveness and robustness of text-based and multimodal LLMs on table understanding through a cross-domain & cross-modality evaluation.

Joint work by DFKI SLT incl. Fabio Barth, Raia Abu Ahmad, @malteos.bsky.social @pjox.bsky.social

1 3

Nils Feldhus @nfel.bsky.social · Jul 26

Ekaterina Borisova et al.:
"Table Understanding and (Multimodal) LLMs: A Cross-Domain Case Study on Scientific vs. Non-Scientific Data"
📄 ACL Anthology: aclanthology.org/2025.trl-1.10/
📊 July 31, TRL Workshop @ Room 2.15
Oral presentation: 09:40-09:55 (pres. by Ekaterina)
Poster session: 16:35-17:15

Table Understanding and (Multimodal) LLMs: A Cross-Domain Case Study on Scientific vs. Non-Scientific Data

Ekaterina Borisova, Fabio Barth, Nils Feldhus, Raia Abu Ahmad, Malte Ostendorff, Pedro Ortiz Suarez, Georg Rehm, Sebastian Möller. Proceedings of the 4th Table Representation Learning Workshop. 2025.

aclanthology.org

1 1

Nils Feldhus @nfel.bsky.social · Jul 26

Our contribution to the FEVER shared task: Our EFC framework stays competitive against this year's baseline while significantly reducing the average runtime per claim achieved through semantic filtering strategies for veracity prediction.

Joint work by the XplaiNLP group incl. @jingyng.bsky.social

1 1

Nils Feldhus @nfel.bsky.social · Jul 26

Max Upravitelev et al.:
"Exploring Semantic Filtering Heuristics For Efficient Claim Verification"
📄 ACL Anthology: aclanthology.org/2025.fever-1...
🤒 FEVER Workshop, July 31 @ Room 2.31
Oral presentation: 10:10-10:30 (presented by Max)

Exploring Semantic Filtering Heuristics For Efficient Claim Verification

Max Upravitelev, Premtim Sahitaj, Arthur Hilbert, Veronika Solopova, Jing Yang, Nils Feldhus, Tatiana Anikina, Simon Ostermann, Vera Schmitt. Proceedings of the Eighth Fact Extraction and VERification...

aclanthology.org

1 1

Nils Feldhus @nfel.bsky.social · Jul 26

We investigate how rationale generation is affected by readability level control, and find that explanations are adaptable, but the observed distinction between readability levels does not fully match the desired complexity.

Joint work with @hakimov.bsky.social.

1 1 1

Nils Feldhus @nfel.bsky.social · Jul 26

Yi-Sheng Hsu et al.:
"Free-text Rationale Generation under Readability Level Control"
📄 arXiv (Camera-ready): arxiv.org/abs/2407.01384
💎 GEM^2 Workshop July 31 @ Hall C
Poster session: 11:30-12:30 and 14:00-15:00 (presented by myself)

Free-text Rationale Generation under Readability Level Control

Free-text rationales justify model decisions in natural language and thus become likable and accessible among approaches to explanation across many tasks. However, their effectiveness can be hindered ...

arxiv.org

1

Nils Feldhus @nfel.bsky.social · Jul 26

Using saliency scores, label flip verification and few-shot prompting, our FitCF method outperforms three state-of-the-art baselines on counterfactual example generation.

Joint work with @simost.bsky.social, Luis Felipe Villa-Arenas, Sebastian Möller, Vera Schmitt.

Code: github.com/qiaw99/FitCF

1 1

Nils Feldhus @nfel.bsky.social · Jul 26

Qianli Wang ( @qiaw99.bsky.social ) et al.:
"FitCF: A Framework for Automatic Feature Importance-guided Counterfactual Example Generation"
📄 ACL Anthology: aclanthology.org/2025.finding...
🏟️ ACL Findings, July 28 @ Hall 4/5
Poster presentation: 18:00-19:30 (pres. by Qianli and myself)

FitCF: A Framework for Automatic Feature Importance-guided Counterfactual Example Generation

Qianli Wang, Nils Feldhus, Simon Ostermann, Luis Felipe Villa-Arenas, Sebastian Möller, Vera Schmitt. Findings of the Association for Computational Linguistics: ACL 2025. 2025.

aclanthology.org

1

Nils Feldhus @nfel.bsky.social · Jul 26

🚆On my way to Vienna! #ACL2025NLP #ACL2025
Together with my amazing colleagues from TU Berlin, DFKI, Saarland & Potsdam, I will present 4 papers on counterfactuals (Findings), free-text rationales (GEM), fact checking (FEVER oral), table understanding (TRL oral).
Excited to meet old and new friends!

1 7

Nils Feldhus @nfel.bsky.social · Jul 23

The Alternative Annotator Test (Calderon et al., 2025):
aclanthology.org/2025.acl-lon...

Promises and Pitfalls of LLM Annotations in Dataset Labeling (Horych et al., 2025): aclanthology.org/2025.finding...

The Promises and Pitfalls of LLM Annotations in Dataset Labeling: a Case Study on Media Bias Detection

Tomáš Horych, Christoph Mandl, Terry Ruas, Andre Greiner-Petter, Bela Gipp, Akiko Aizawa, Timo Spinde. Findings of the Association for Computational Linguistics: NAACL 2025. 2025.

aclanthology.org

4

Nils Feldhus @nfel.bsky.social · Jul 23

Liking this direction a lot!
Thanks for the pointers! I only knew the first one so far.
These are some of my favs:

Co-DETECT (Xiong et al., 2025): arxiv.org/abs/2507.05010

LLMs as Span Annotators (Kasner et al., 2025):
arxiv.org/abs/2504.08697

Co-DETECT: Collaborative Discovery of Edge Cases in Text Classification

We introduce Co-DETECT (Collaborative Discovery of Edge cases in TExt ClassificaTion), a novel mixed-initiative annotation framework that integrates human expertise with automatic annotation guided by...

arxiv.org

1 5

Nils Feldhus @nfel.bsky.social · Jul 3

We respectfully disagree and would like to clarify your concerns about the novelty of our work.

4

Reposted by Nils Feldhus

Mahdi Dhaini @mahdidh.bsky.social · Jun 26

Very happy to be at #FAccT2025 in Athens, where I presented our work "Gender Bias in Explainability: Investigating Performance Disparity in Post-hoc Methods"

📄Paper: dl.acm.org/doi/10.1145/...

At #FAccT2025? Let's connect if you're interested in improving the usability of explainability methods!

1 1 9