Papers by Bernhard Schölkopf
On the Emergence and Test-Time Use of Structural Information in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | a controlled environment is required to study how language models learn structural information from observational data. |
| Approach: | They propose a natural language dataset based on linguistic structural transformations to study how language models learn abstract structures and utilize the learnt structural information at test-time. |
| Outcome: | The proposed model can generate new knowledge outside the training corpus in a controlled environment. |
Competition of Mechanisms: Tracing How Language Models Handle Facts and Counterfactuals (2024.acl-long)
Copied to clipboard
| Challenge: | Existing interpretability research focused on analyzing a single mechanism . et al., 2023) focused on finding how models copy or recall factual knowledge . |
| Approach: | They propose a competition of mechanisms that focuses on the interplay of multiple mechanisms instead of individual mechanisms . they uncover how and where the competition of mechanism happens within LLMs using logit inspection and attention modification methods. |
| Outcome: | The proposed model is based on two interpretability methods, logit inspection and attention modification. |
Orthogonal Finetuning Made Scalable (2025.emnlp-main)
Copied to clipboard
| Challenge: | a recent shift in foundation models has slowed the adoption of finetuning methods . however, its high runtime and memory demands limit its scalability . |
| Approach: | They propose an input-centric reformulation that uses matrix-vector multiplications instead of cubic multiplication . they extend OFTv2 to support finetuning quantized foundation models and show it outperforms QLoRA . |
| Outcome: | The proposed model outperforms the popular QLoRA in training stability, efficiency, and memory usage. |
Moûsai: Efficient Text-to-Music Diffusion Models (2024.acl-long)
Copied to clipboard
| Challenge: | Recent years have seen the rapid development of large generative models for text; however, little research has explored the connection between text and another “language” of communication – music. |
| Approach: | They develop a text-to-music generation model that can generate multiple minutes of high-quality stereo music at 48kHz from textual descriptions. |
| Outcome: | The proposed model can generate multiple minutes of high-quality stereo music at 48kHz from textual descriptions. |
CausalCite: A Causal Formulation of Paper Citations (2024.findings-acl)
Copied to clipboard
Ishan Agrawal, Zhijing Jin, Ehsan Mokhtarian, Siyuan Guo, Yuen Chen, Mrinmaya Sachan, Bernhard Schölkopf
| Challenge: | citation counts are often criticized for failing to accurately reflect the true impact of a paper. |
| Approach: | They propose a method to measure the impact of a paper on follow-up papers by comparing similar papers by cosine similarity. |
| Outcome: | The proposed method is based on a new causal inference method, TextMatch. |
PaperMentor: A Human-Centered Multi-Agent Writing Tutor for AI Research Papers in Overleaf (2026.acl-demo)
Copied to clipboard
Jiarui Liu, Terry Jingchen Zhang, Ryan Faulkner, Xuanqiang Angelo Huang, Vilém Zouhar, Dominik Glandorf, Isabel Dahlgren, Rishit Dagli, Yuen Chen, Felix Leeb, Van Q. Truong, Punya Syon Pandey, Yves Bicker, Suvajit Majumder, Wenyuan Jiang, Zeju Qiu, Sankalan Pal Chowdhury, Mrinmaya Sachan, Bernhard Schölkopf, Mona T. Diab, Zhijing Jin
| Challenge: | Emerging AI-powered writing assistants focus on grammar fixes or simulating peer review with final scores, yet they fall short of providing concrete, actionable suggestions that help students improve their papers during drafting. |
| Approach: | They propose a human-centered writing assistant system that delivers actionable suggestions as Overleaf-native inline comments while leaving the actual writing entirely to human authors. |
| Outcome: | The proposed system outperforms a baseline with the skill library and provides actionable suggestions while leaving the actual writing to human authors. |
Are Language Models Consequentialist or Deontological Moral Reasoners? (2025.emnlp-main)
Copied to clipboard
Keenan Samway, Max Kleiman-Weiner, David Guzman Piedrahita, Rada Mihalcea, Bernhard Schölkopf, Zhijing Jin
| Challenge: | Recent studies have focused on the moral judgments in large language models rather than their underlying moral reasoning process. |
| Approach: | They propose a taxonomy of moral rationales to classify reasoning traces according to consequentialism and deontology . they use trolley problems to analyze moral reasoning tracing in large language models . |
| Outcome: | The proposed taxonomy of moral rationales sheds light on consequentialism and deontology . it systematically classifies reasoning traces according to two main ethical theories . |
DARS: Dynamic Action Re-Sampling to Enhance Coding Agent Performance by Adaptive Tree Traversal (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to developing LLM-powered coding agents struggle with sub-optimal decision-making. |
| Approach: | They propose a novel inference time compute scaling approach that recovers from sub-optimal decisions by branching out a trajectory at certain key decision points by taking an alternative action given the history of the trajectory and execution feedback of the previous attempt. |
| Outcome: | The proposed approach achieves a pass@1 rate of 47% on the SWE-Bench Lite benchmark, outperforming state-of-the-art (SOTA) open-source frameworks. |
Improving Large Language Model Safety with Contrastive Representation Learning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing defenses against large language models (LLMs) are limited by their ability to generate responses to diverse inputs. |
| Approach: | They propose a model defense framework that finetunes a large-scale model using a triplet-based loss combined with adversarial hard negative mining to encourage separation between benign and harmful representations. |
| Outcome: | The proposed model defense outperforms previous representation engineering-based defenses while improving robustness against input-level and embedding-space attacks. |
Test of Time: Rethinking Temporal Signal of Benchmark Contamination (2026.acl-long)
Copied to clipboard
Terry Jingchen Zhang, Gopal Dev, Ning Wang, Max Obreiter, Punya Syon Pandey, Keenan Samway, Wenyuan Jiang, Yinya Huang, Bernhard Schölkopf, Mrinmaya Sachan, Zhijing Jin
| Challenge: | Existing work on benchmarks containing publicly available information has been interpreted as a temporal signal for benchmark contamination. |
| Approach: | They show that LLM-transformed questions can produce remarkably different temporal patterns compared to fill-in-the-blank questions directly retrieved from the very same documents. |
| Outcome: | The proposed model can produce different temporal patterns compared to fill-in-the-blank questions retrieved from the same documents. |
How Robust Are Router-LLMs? Analysis of the Fragility of LLM Routing Capabilities (2026.eacl-long)
Copied to clipboard
| Challenge: | Large language model (LLM) routing has emerged as a promising solution to balancing computational costs and performance. |
| Approach: | They propose a framework that categorizes router performance across a broad spectrum of query types . large language models have revolutionized natural language processing . |
| Outcome: | The proposed framework categorizes router performance across a broad spectrum of query types . it integrates privacy and safety assessments to reveal hidden risks . |
The Odyssey of Commonsense Causality: From Foundational Benchmarks to Cutting-Edge Reasoning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Despite its significance, a systematic exploration of commonsense causality is lacking. |
| Approach: | They focus on taxonomies, benchmarks, acquisition methods, qualitative reasoning, and quantitative measurements in commonsense causality. |
| Outcome: | The proposed method synthesizes insights from over 200 representative articles and provides a practical guide for beginners. |
When Do Language Models Endorse Limitations on Human Rights Principles? (2026.findings-eacl)
Copied to clipboard
Keenan Samway, Miu Nicole Takagi, Rada Mihalcea, Bernhard Schölkopf, Ilias Chalkidis, Daniel Hershcovich, Zhijing Jin
| Challenge: | a recent study evaluated how large language models navigate trade-offs involving the Universal Declaration of Human Rights. |
| Approach: | They evaluate how large language models navigate trade-offs involving the Universal Declaration of Human Rights (UDHR) they use 1,152 synthetically generated scenarios across 24 rights articles and eight languages . |
| Outcome: | The proposed models accept limiting economic, social, and cultural rights more often than political and civil rights, the authors show . their models show significant cross-linguistic variation with elevated endorsement rates of rights-limiting actions in Chinese and Hindi compared to English or Romanian . |
A diverse Multilingual News Headlines Dataset from around the World (2024.naacl-short)
Copied to clipboard
| Challenge: | Babel Briefings features 4.7 million news headlines from August 2020 to November 2021, across 30 languages and 54 locations worldwide with English translations of all articles included. |
| Approach: | They propose to use a weighted similarity metric to group articles into clusters about the same event and visualize the event signatures of the event. |
| Outcome: | The proposed dataset features 4.7 million headlines from August 2020 to November 2021 across 30 languages and 54 locations worldwide with English translations of all articles included. |
Beyond Good Intentions: Reporting the Research Landscape of NLP for Social Good (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in natural language processing (NLP) have created a vast number of applications that are aimed at social good applications. |
| Approach: | They propose a dataset with three tasks that can help identify NLP4SG papers and characterize the NLP landscape by: (1) identifying the papers that address a social problem, (2) mapping them to the corresponding UN Sustainable Development Goals, and (3) identifying their methods. |
| Outcome: | The proposed dataset can help identify NLP4SG papers and characterize the NLP landscape by: (1) identifying the papers that address a social problem, (2) mapping them to the corresponding UN Sustainable Development Goals (SDGs), and (3) identifying their methods. |
Implicit Personalization in Language Models: A Systematic Study (2024.findings-emnlp)
Copied to clipboard
Zhijing Jin, Nils Heil, Jiarui Liu, Shehzaad Dhuliawala, Yahang Qi, Bernhard Schölkopf, Rada Mihalcea, Mrinmaya Sachan
| Challenge: | Existing studies have focused on the implicit personalization problem, but no unified framework exists to study it. |
| Approach: | They propose a mathematical formulation and a moral reasoning framework to study the phenomenon of Implicit Personalization (IP) they propose 'direct intervention' to estimate causal effect of mediator variable that cannot be directly intervened upon. |
| Outcome: | The proposed method estimates the causal effect of a mediator variable that cannot be directly intervened upon. |
Analyzing the Role of Semantic Representations in the Era of Large Language Models (2024.naacl-long)
Copied to clipboard
Zhijing Jin, Yuen Chen, Fernando Gonzalez Adauto, Jiarui Liu, Jiayi Zhang, Julian Michael, Bernhard Schölkopf, Mona Diab
| Challenge: | Existing studies show the benefits of semantic representations in NLP tasks . Existing work using AMR is concerned with trainable models . |
| Approach: | They propose an AMR-driven chain-of-thought prompting method that uses AMR . they propose to use it to predict which input examples AMR may help or hurt on . |
| Outcome: | The proposed method hurts performance more than it helps on five different tasks. |
Original or Translated? A Causal Analysis of the Impact of Translationese on Machine Translation Performance (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing work on translationese neglects important factors and conclusions are mostly correlational but not causal. |
| Approach: | They use a dataset where MT training data are also labeled with human translation directions to examine the impact of translationese on machine translation evaluation. |
| Outcome: | The proposed model learns in the same direction as human translation directions. |
Do LLMs Think Fast and Slow? A Causal Study on Sentiment Analysis (2024.findings-emnlp)
Copied to clipboard
Zhiheng Lyu, Zhijing Jin, Fernando Gonzalez Adauto, Rada Mihalcea, Bernhard Schölkopf, Mrinmaya Sachan
| Challenge: | Sentiment analysis aims to identify the sentiment expressed in a piece of text, often in the form of a review. |
| Approach: | They propose a causal discovery task that distinguishes whether a review "primes" the sentiment and a traditional prediction task to model the sentiment using the review as input. |
| Outcome: | The proposed model improves by 32.13 F1 points on a zero-shot five-class SA. |