United States

Large language models (LLMs) are increasingly used as evaluative tools across languages, yet bias research remains overwhelmingly Anglocentric, with most studies conducted in English using Latin-script names. It remains unclear whether bias patterns generalize across linguistic contexts. We investigate this question and introduce the stereotype perceptual map, a framework for analyzing how ethnic groups are positioned along evaluative dimensions. Using 900,000 model responses over 45,000 names spanning 9 ethnicities, we evaluate model behavior across prompt languages (English, Chinese, Thai), writing scripts (Latin, Chinese, Thai), evaluative domains (competence, warmth), and models (GPT, DeepSeek). We find that ethnic bias hierarchies are jointly shaped by local linguistic context and model origin and differ substantially between Western-centric and Sinocentric models. DeepSeek exhibits highly stable rankings across conditions in math competence judgments, consistently placing Chinese at the top, followed by Russian, and White, Hispanic, and Black names at the bottom. GPT, by contrast, shows strong script-dependent reordering: Latin-script conditions form one stable cluster, while native-script conditions form another, with substantially lower cross-cluster correlations. We term this script-gated bias: transliterating the same names into a non-Latin script can activate a different evaluative frame and produce rankings that are sometimes inversely correlated with Latin-script results. Warmth evaluations are less stable than competence across both models. Our findings demonstrate that multilingual bias cannot be characterized through single-language, single-script audits. For multilingual users, code-switching between languages can toggle between different bias regimes. Fairness evaluations for multilingual LLMs must therefore account for deployment language, writing system, and model origin to capture the full range of potentially harmful bias these systems exhibit.

ACL 2026

Through the Looking Glass of Multilingual AI: Contrasting Language- and Name Script-Dependent Ethnic Hierarchies in GPT and DeepSeek

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields 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.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

poster

We propose a comprehensive research agenda to detect, measure, and mitigate racial bias in Natural Language Processing (NLP) systems deployed in criminal justice contexts. Our preliminary work demonstrates that racial descriptors systematically alter embedding similarity scores and retrieval rankings across six models, with bias being race-specific and models showing rank displacements of 1.82 to 7.44 positions, on average. This empirically indicates that even small shifts in similarity scores can displace relevant records outside top-10 results, leading to systematic under-retrieval of records involving certain demographic groups. Building on these findings, this thesis proposes four research questions: (1) developing and evaluating debiasing techniques including counterfactual data augmentation, adversarial training, and fairness-constrained fine-tuning; (2) validating synthetic findings on authentic law enforcement data through IRB-approved partnerships; (3) investigating intersectional bias patterns across race, gender, and age; and (4) we extend beyond embedding-level analysis to examine how bias propagates across modern multi-stage retrieval pipelines from embeddings to cross-encoders to LLMs. Expected contributions include empirical comparisons of debiasing methods, bias benchmarks for criminal justice NLP, deployment guidelines for fairness-aware retrieval systems, and the first comprehensive analysis of multi-stage bias propagation in retrieval pipelines.

Thesis Proposal: Auditing and Mitigating Demographic Bias in Multi-Stage Retrieval Systems for Criminal Justice Applications

Scenario-based text generation has broad applications across education and creative writing, but remains underexplored in controllable text generation. We introduce the Contextual Diversity Measure (CDM), a metric that quantifies semantic diversity for scenario generation given an abstract semantic role labeling constraints, and validate it through controlled experiments. Statistical analysis across four embedding models demonstrates that CDM successfully distinguishes between high-diversity and low-diversity text pairs, with all tests achieving significance at $p < 0.05$ on both subsets of the dataset, with effect sizes from small-to-medium (Cohen's $d$ from 0.292 to 0.508) and medium-to-large ($d$: 0.677--1.195). Baseline comparisons indicate that CDM achieves excellent discrimination accuracy (100\% and 91.9\% for the two subsets), with discriminative power up to $5.5\times$ greater than the best~baseline.

Contextual Diversity Measure (CDM) for Controllable Story Generation in Large Language Models

The development of fact-checking systems for verifying the factuality of text generated by large language models (LLMs) has been advancing. In the verdict prediction step of such systems, the system determines whether claims in the generated text are supported by retrieved evidence, formulated as a natural language inference (NLI) task. This study extends the label set for verdict prediction to capture claim-evidence relationships that humans would commonly interpret as supported or refuted, even in the absence of strict logical entailment or contradiction. It also constructs a Japanese dataset comprising 28,147 instances from two sources based on this extended label set. We analyze the causes of annotation disagreement and find that ambiguity in the boundary of acceptable inference, interpretive characteristics of negative cases, and incomplete information in the evidence affect annotation variability. Using this dataset, we evaluate the performance of prompt-based verdict prediction methods and show that prompts that explicitly elicit chain-of-thought reasoning improve F1 by 4 percentage points compared to baseline.

Constructing a Japanese Verdict Prediction Dataset for Fact-Checking of LLM-Generated Texts

We disentangle multilingual sentence embeddings into language-dependent and language-agnostic components, leveraging the latter to improve cross-lingual similarity estimation. Previous studies focused on encoder-based approaches that use only the input sentence; in contrast, this study examines the effectiveness of disentanglement methods across a broader range of sentence embeddings, including decoder-based approaches and those that utilize prompts. Experimental results demonstrate that embedding disentanglement is effective for a wide variety of sentence embeddings.

Disentangling Meaning and Language Components in Diverse Multilingual Sentence Embeddings

Vision-Language Models (VLMs) excel at multimodal tasks, but they remain vulnerable to hallucinations that are factually incorrect or ungrounded in the input image. Recent work suggests that hallucination detection using internal representations is more efficient and accurate than approaches that rely solely on model outputs. However, existing internal-representation-based methods typically rely on a single representation or detector, limiting their ability to capture diverse hallucination signals. In this paper, we propose EnsemHalDet, an ensemble-based hallucination detection framework that leverages multiple internal representations of VLMs, including attention outputs and hidden states. EnsemHalDet trains independent detectors for each representation and combines them through ensemble learning. Experimental results across multiple VQA datasets and VLMs show that EnsemHalDet consistently outperforms prior methods and single-detector models in terms of AUC. These results demonstrate that ensembling diverse internal signals significantly improves robustness in multimodal hallucination detection.

EnsemHalDet: Robust VLM Hallucination Detection via Ensemble of Internal State Detectors

Lean proofs are built as sequences of tactic-induced state transitions, yet learned models often represent proof steps primarily through tactic strings or raw proof-state text. Building on Delta Tokens, which encode a proof step by the local edit it induces between successive proof states, we train an encoder-only Transformer to learn contextualized representations of Lean proof steps from state changes. We then use these step representations to study complete proofs as trajectories in a learned latent space. We first show that the Delta-based Transformer yields better held-out next-tactic retrieval than a matched surface-syntax control, supporting the representational choice used in the trajectory analysis. We then analyze proof trajectories using path length, endpoint span, directness, curvature, and torsion. Across the LeanWorkbook slice used here, longer proofs become increasingly indirect within a relatively bounded latent span: path length grows sharply with proof length while endpoint span changes little, mean step size decreases, curvature rises modestly, and torsion falls. Qualitative case studies show that these geometric patterns align with recognizable proof organizations, including immediate closure, aligned accumulation, scaffolded enrichment, bookkeeping-heavy restructuring, and repeated local contradiction work. The dataset is small and heavily skewed toward short proofs, so the claims are necessarily limited. Within those limits, the results suggest that learned state-change representations recover nontrivial structure in how proofs unfold and provide a promising basis for future trajectory-aware theorem proving.

Proofs as Trajectories: Learning Lean Proof-Step Representations from State Changes

Large language models (LLMs) exhibit substantial cross-lingual variation in mathematical reasoning performance, but it remains unclear whether these differences reflect language-specific parameters or a shared mechanism that manifests differently by language. We present a cross-lingual mechanistic analysis of mathematical reasoning in LLMs, enabling us to localize and compare model parameters that support mathematical reasoning across languages. We find that mathematical reasoning is supported by a partially shared set of parameters, with the strongest overlap concentrated in intermediate model layers. We further observe that English consistently produces the largest set of math-relevant parameters, whereas lower-resource languages reveal smaller sets of relevant parameters. These results indicate that math-related reasoning in multilingual LLMs is neither fully language-invariant nor fully language-specific, but instead relies on a shared parameter core with systematic cross-lingual differences.

LLM Parameters for Math Across Languages: Shared or Separate?

Large language models show strong capabilities in natural language generation and have been applied to translate complex structured data into human-readable insights. While these models excel at surface-level fluency, they remain unreliable as they produce factually inaccurate outputs and struggle with consistent logical inference beyond surface-level patterns. Moreover, they often lack a clear sense of relevance and produce shallow or uninformative insights. This thesis proposal argues that a key source of these limitations is the fact that tasks are often underspecified, requiring models to make implicit assumptions about missing context. We will investigate how such underspecification leads to unintentional assumptions and how these affect faithfulness and evaluation. We will examine how models can identify missing premises and surface multiple plausible interpretations to improve factuality of their inferences and make their evaluation more rigorous. We will also study how underlying assumptions influence the perceived interestingness of insights. Finally, we will explore how to improve reasoning to enable deeper inferences, focusing on code generation and qualitative reasoning. By shifting focus from surface-level generation to assumption-aware deeper inferences, this work aims to improve reliability, interpretability, and user controlability in NLG.

Thesis Proposal: Intentional Inference for Insight Generation

Large language models and LLM-based agents are widely deployed as personal chat assistants, enterprise copilots, and autonomous workflow agents. Across all these applications, memory (the ability to retain, access, and reason over information accumulated over long contexts and multiple interactions) plays a crucial role in determining the reliability of any agent. We introduce RECON (Reasoning over Extended Contexts with Obfuscated Narratives), a benchmark for evaluating compositional temporal reasoning over long contexts. RECON uses investigation case file (100k tokens) and tests agents on \textbf{six memory-intensive tasks}: reconstructing multi-hop evidence chains, propagating cascading invalidations, resolving source conflicts, counterfactual reasoning, satisfying temporal constraints, and temporal fact retrieval. Recent memory benchmarks evaluate whether agents can retrieve scattered facts or detect if a fact has changed; RECON evaluates what happens after the change, whether agents can trace which downstream conclusions are affected, which survive through independent support, and how alternative timelines would have unfolded. We evaluate systems across three architectural families comprising long-context LLMs, retrieval-augmented generation (RAG) agents, and commercial memory agents and find that existing memory mechanisms are not yet sufficient for reliable reasoning over long, evolving narratives. Our benchmark and code are available at \url{https://anonymous.4open.science/r/recon-benchmark/README.md}.

RECON: Benchmarking Agent Memory for Compositional Reasoning over Long Contexts

To accelerate scientific knowledge acquisition, LLMs are increasingly used to synthesize multiple papers into structured tables by inferring schemas and values. While value generation within a fixed schema can often be reduced to extractive question answering, the schema generation problem, determining which dimensions to compare a set of documents, lacks a formal mapping to standard NLP tasks. In this work, we formulate schema generation as a reinforcement learning problem and investigate whether these dimensions can be induced without access to gold-standard schemas. We design a multi-faceted reward framework capturing schema coverage, non-redundancy, relevance, and format, and train a small language model on a literature review dataset. Our approach yields consistent improvements over the untuned base model across intrinsic, reference-based, and LLM-judge metrics, and remains competitive with supervised fine-tuned models at 5$\times$ the parameter count on structural and diversity dimensions.\footnote{Code will be released upon publication.}

Downloads

Next from ACL 2026

Thesis Proposal: Auditing and Mitigating Demographic Bias in Multi-Stage Retrieval Systems for Criminal Justice Applications

Similar lecture

Through the Looking Glass of Multilingual AI: Contrasting Language- and Name Script-Dependent Ethnic Hierarchies in GPT and DeepSeek

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 2026

Thesis Proposal: Auditing and Mitigating Demographic Bias in Multi-Stage Retrieval Systems for Criminal Justice Applications

Similar lecture

Through the Looking Glass of Multilingual AI: Contrasting Language- and Name Script-Dependent Ethnic Hierarchies in GPT and DeepSeek

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads