Austria

Deterministically controlling the target generation language of large multilingual language models (LLMs) remains a fundamental challenge, particularly in zero-shot settings where neither explicit language prompts nor fine-tuning are available. In this work, we investigate whether sparse autoencoder (SAE) features, previously shown to correlate with interpretable model behaviors, can be leveraged to steer the generated language of LLMs during inference. Leveraging pretrained SAEs on the residual streams of Gemma-2B and Gemma-9B, we identify features whose activations differ most significantly between English and four target languages: Chinese, Japanese, Spanish, and French. By modifying just a single SAE feature at one transformer layer, we achieve controlled language shifts with up to 90% success, as measured by FastText language classification, while preserving semantic fidelity according to LaBSE (Language-Agnostic BERT Sentence Embedding) similarity. Our analysis reveals that language steering is most effective in mid-to-late transformer layers and is amplified by specific attention heads disproportionately associated with language-sensitive SAE features. These results demonstrate the promise of sparse feature steering as a lightweight and interpretable mechanism for controllable multilingual generation.

ACL 2025

Causal Language Control in Multilingual Transformers via Sparse Feature Steering

steering

translation

interpretability

multilingual

poster

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Variation in human perspectives has drawn increasing attention in natural language processing. The widespread human annotation disagreement challenges the traditional paradigm of a single "ground truth" and raises concerns about the limitations of conventional label aggregation methods and the uniform models built upon them, which often discard minority opinions and obscure valuable individual perspectives. This thesis proposal investigates three core dimensions of perspective-oriented research: (1) annotation formats that better capture the nuance and uncertainty of individual judgments; (2) modeling approaches that leverage socio-demographic features to improve prediction for underrepresented or minority viewpoints; and (3) personalized generation that tailor outputs to individual users’ preferences and communicative styles. Through this work, we aim to advance methods that more faithfully reflect the diversity of human interpretation, enhancing both inclusiveness and fairness in language technologies.

Proposal: From One-Fit-All to Perspective Aware Modeling

We investigate feature universality in Gemma-2 language models (Gemma-2-2B & Gemma-2-9B), asking whether models with a fourfold difference in scale still converge on comparable internal concepts. Using the sparse autoencoder (SAE) dictionary learning pipeline, we used pretrained SAEs on each model’s residual-stream activations, aligned the resulting monosemantic features via activation correlation, and compared the matched feature spaces with metrics such as SVCCA and RSA. Middle layers yield the strongest overlap, indicating that this is where both models most similarly represent concepts, while early and late layers show much less similarity. Preliminary experiments extending the analysis from single tokens to multi-token subspaces show that semantically similar subspaces tend to interact similarly with LLMs. These results offer further evidence that large language models carve the world into broadly similar, interpretable features despite size differences, reinforcing universality as a foundation for cross-model interpretability.

Semantic Convergence: Investigating Shared Representations Across Scaled LLMs

Large language models often suffer from language confusion, a phenomenon where responses are partially or entirely generated in unintended languages. This can critically impact user experience in low-resource settings. We hypothesize that conventional supervised fine-tuning exacerbates this issue because the softmax objective focuses probability mass only on the single correct token but does not explicitly penalize cross-lingual mixing. 
Interestingly, by observing loss trajectories during the pretraining phase, we observe that models fail to learn to distinguish between monolingual and language-confused text. Additionally, we find that ORPO, which adds penalties for unwanted output styles to standard SFT, effectively suppresses language-confused generations even at high decoding temperatures without degrading overall model performance. Our findings suggest that incorporating appropriate penalty terms can mitigate language confusion in low-resource settings with limited data.

Controlling Language Confusion in Multilingual LLMs

We introduce a modified sequence tagging architecture, proposed in (Omelianchuk et al., 2020), for the Grammatical Error Correction of the Russian language. We propose language-specific operation set and preprocessing algorithm as well as a classification scheme which makes distinct predictions for insertions and other operations. The best versions of our models outperform previous approaches and set new SOTA on the two Russian GEC benchmarks -- RU-Lang8 and GERA, while achieve competitive performance on RULEC-GEC.

Grammatical Error Correction via Sequence Tagging for Russian

Text corpora in non-English clinical contexts are sparse, where synthetic data generation with Large Language Models (LLMs) appears as a promising strategy to overcome this data gap. In order to test the quality of LLMs in generating synthetic data, we applied them to our novel German Medical Interview Questions Corpus (GerMedIQ), consisting of 4,524 unique question-response pairs in German. We augmented our corpus by asking a cohort of models to produce suitable responses to the same questions. Structural and semantic evaluations of the synthetic responses revealed that while augmented responses may meet the grammatical requirements, most models were not able to produce semantically comparable responses to humans. Also, an LLM-as-a-judge experiment showcased that human responses were consistently rated more appropriate than synthetic ones. We find that data augmentation with LLMs in non-English and clinical domain contexts has to be performed carefully.

GerMedIQ: A Resource for Simulated and Synthesized Anamnesis Interview Responses in German

This paper examines differences between stream-of-consciousness narratives written by humans and those generated by large language models (LLMs) to assess narrative coherence and personality expression. We generated texts by prompting LLMs (Llama-3.1-8B \& DeepSeek-R1-Distill-Llama-8B) with the first half of essays either revealing the personality characteristics (Big Five) or hiding them. Our analysis revealed consistently low similarity between LLM-generated continuations and original human texts, as measured by cosine similarity, perplexity, and BLEU scores. Including explicit personality traits significantly enhanced Llama-3.1-8B's performance, particularly in BLEU scores. Further analysis of personality expression showed varying alignment patterns between LLMs and human texts. Further analysis of personality expression found higher openness for all models. Additionally, Llama-3.1-8B models exhibited higher Extraversion but low Agreeableness, while DeepSeek-R1-Distill-Llama-8B's showed dramatic personality shifts during its thinking process, particularly when provided with personality traits in the prompt.

Unstructured Minds, Predictable Machines: A Comparative Study of Narrative Cohesion in Human and LLM Stream-of-Consciousness Writing

This study investigates the use of Large Language Models (LLMs) for political stance detection in informal online discourse, where language is often sarcastic, ambiguous, and context-dependent. We explore whether providing contextual information, specifically user profile summaries derived from historical posts, can improve classification accuracy. Using a real-world political forum dataset, we generate structured profiles that summarize users' ideological leaning, recurring topics, and linguistic patterns. We evaluate seven state-of-the-art LLMs across baseline and context-enriched setups through a comprehensive cross-model evaluation. Our findings show that contextual prompts significantly boost accuracy, with improvements ranging from +17.5\% to +38.5\%, achieving up to 74\% accuracy that surpasses previous approaches. We also analyze how profile size and post selection strategies affect performance, showing that strategically chosen political content yields better results than larger, randomly selected contexts. These findings underscore the value of incorporating user-level context to enhance LLM performance in nuanced political classification tasks.

Exploiting contextual information to improve stance detection in informal political discourse with LLMs

Homelessness is a persistent issue, impacting millions worldwide, and over 770,000 people experienced homelessness in the U.S. in 2024. Social stigmatization is a significant barrier to alleviation, shifting public perception, and influencing policy. Online discourse on platforms such as Reddit shape public opinion. To address this, the project leverages natural language processing and large language models (LLMs) to mitigate bias against people experiencing homelessness (PEH) in online spaces. The goal is to promote awareness, reduce harmful biases, inform policy, and improve the fairness of generative AI. We gather Reddit data for 10 U.S. cities, then perform zero-shot classification, and finally, mitigation using Llama 3.2 Instruct and Qwen 2.5 7B Instruct models. The initial results highlighted the differing classifications between models and indicated that many mitigated outputs remained biased. This suggests the need for potential model refinement for the mitigation of text related to PEH.

Local LLM Zero-Shot Analysis of Homelessness Discourse on Reddit

Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP), but their success remains largely confined to high-resource, general-purpose domains. In contrast, applying LLMs to low-resource domains poses significant challenges due to limited training data, domain drift, and strict terminology constraints. This survey provides an overview of the current landscape in domain-specific, low-resource QA with LLMs. We begin by analyzing the coverage and representativeness of specialized-domain QA datasets against large-scale reference datasets what we refer to as ParentQA. Building on this analysis, we survey data-centric strategies to enhance input diversity, including data augmentation techniques. We further discuss evaluation metrics for specialized tasks and consider ethical concerns. By mapping current methodologies and outlining open research questions, this survey aims to guide future efforts in adapting LLMs for robust and responsible use in resource-constrained, domain-specific environments.

QA Analysis in Medical and Legal Domains: A Survey of Data Augmentation in Low-Resource Settings

With Large Language Model (LLM)-based applications becoming more common due to strong performance across many tasks, prompt optimization has emerged as a way to extract better solutions from frozen, often commercial LLMs that are not specifically adapted to a task. LLM-assisted prompt optimization methods provide a promising alternative to manual/human prompt engineering, where LLM “reasoning” can be used to make them optimizing agents. However, the cost of using LLMs for prompt optimization via commercial APIs remains high, especially for heuristic methods like evolutionary algorithms (EAs), which need many iterations to converge, and thus, tokens, API calls, and rate-limited network overhead. We propose GenDLN, an efficient genetic algorithm-based prompt pair optimization framework that leverages commercial API free tiers. Our approach allows teams with limited resources (NGOs, non-profits, academics…) to efficiently use commercial LLMs for EA-based prompt optimization. We conduct experiments on CLAUDETTE for legal terms of service classification and MRPC for paraphrase detection, performing in line with selected prompt optimization baselines, at no cost. Our code is available in <omitted>.

Downloads

Next from ACL 2025

Proposal: From One-Fit-All to Perspective Aware Modeling

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2025

Proposal: From One-Fit-All to Perspective Aware Modeling

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads