Large Language Models (LLMs) can produce convincing geometric arguments, yet their outputs are not reliable enough to be treated as proofs without independent verification. In parallel, symbolic geometry tools (e.g. automated theorem provers in dynamic geometry systems) offer strong rigor guarantees, but require formalized inputs and can struggle with problem formalization, auxiliary construction, and proof presentation. This survey synthesizes work at the intersection of these lines: hybrid LLM–symbolic systems for geometry that (i) translate natural language and diagrams into formal constraints, (ii) search for solution plans and proof steps using learned or heuristic methods, and (iii) verify the resulting steps using symbolic provers or proof assistants. We propose a taxonomy organized around (a) the role of the LLM in the pipeline (parser, strategist, prover, critic), (b) the target proof artifact (answer-only, informal proof, semi-formal step trace, or kernel-checked formal proof), and (c) the verification backend (numeric testing, algebraic provers, synthetic provers, and proof-assistant kernels). We review representative systems in NLP and AI (e.g. GeoS, Inter-GPS, FormalGeo, AlphaGeometry, AutoGPS, and recent heuristic-only deductive solvers), and connect them to broader neurosymbolic paradigms for faithful reasoning (e.g. SatLM, LINC, and autoformalization). Finally, we outline evaluation protocols emphasizing step-level soundness and robustness, and we discuss open problems in multimodal formalization, handling of non-degeneracy conditions, human-readable certified proofs, and reproducibility.

Recent years have produced a small but rapidly growing set of results where Large Language Models (LLMs) - usually embedded in a search-and-verification loop - advance the state of the art on problems previously regarded as "open" in the pragmatic sense of lacking a best-known construction, bound, or proof certificate. This paper surveys that emerging line of work with a Big Picture emphasis: what makes these successes possible, what should count as "solved", and what design patterns generalize? We (i) propose an evidence ladder for interpreting "LLM solved an open problem" claims, (ii) map mathematical subfields by difficulty dimensions that matter for LLM-based discovery, (iii) curate a timeline of key breakthroughs leading to verifiable discovery systems, and (iv) synthesize the techniques and frameworks - tool use, retrieval, search, and verification - that repeatedly appear in successful case studies. We give particular attention to formal-methods backends common in security and verification contexts, including Linear Temporal Logic (LTL) and Satisfiability Modulo Theories (SMT) solvers, as scalable middle-layer verifiers between lightweight tests and proof assistants. We close with an evaluation and reproducibility checklist aimed at making the next wave of claims easier to trust, reproduce, and build upon, while separating peer-reviewed or certificate-backed results from fast-moving community reports that are useful signals but not yet stable evidence.

Current hallucination detection systems operate under a flawed assumption: that model outputs deviating from factual grounding are uniformly problematic regardless of task context, modality, or cultural setting. Through analysis of computational humor as a motivating case study, we demonstrate that identical model behaviors require radically different evaluations depending on context. We propose reframing hallucination detection as task-output alignment assessment, introducing a three-dimensional framework spanning factual grounding requirements, novelty requirements, and risk tolerance.

When Large Language Models (LLMs) combined with prompt-based approaches as human simulacra emerged, they promised revolutionary shortcuts. Models trained on vast internet corpora may replicate human behavior and communication through text-based alignment. The initial optimism of the NLP community positioned LLMs as universal human proxies capable of replacing participants in surveys, generating authentic social media content, and simulating diverse cultural perspectives. We systematically dismantle this "myth of universal generalization" and document a shift toward methodological rigor. Our research reveals fundamental limitations: LLMs exhibit inhuman response patterns in psychometric assessments and produce detectable synthetic content. We analyze the difference between superficial linguistic fluency and genuine human-like representation, and reframe the current paradigm from asking "can LLMs replace humans?" to "under what validated conditions might LLMs serve as useful research components in social sciences?" Our work shows how interconnected research efforts challenge foundational assumptions and establishes best practices for deploying LLMs as human simulacra.

A socio-technical gap exists between how NLP systems are developed and evaluated and how people use them in practice. To help close this gap, I propose a direction for scientific progress in NLP centered on advancing trustworthy AI-mediated communication between humans, using cross-lingual and cross-cultural interaction as a stress test for this goal – settings where common ground is hard-won, miscommunication can go unnoticed, and human users often lack the means to independently evaluate AI outputs. I outline a research agenda emphasizing two complementary requirements spanning both sides of the interaction. On the model side, I study how multilingual systems access and use knowledge across languages, and when they systematically privilege sources in certain languages. On the user side, I design decision-support mechanisms and evaluate how they shape user’s reliance on imperfect outputs. Taken together, these results motivate future work for aligning multilingual NLP with real communicative practice, with the goal of building AI systems that more reliably serve diverse communities. This paper summarizes and draws heavily on my PhD thesis proposal.

Transformers have dominated sequence processing tasks for the past seven years—most notably language modeling. However, the inherent quadratic complexity of their attention mechanism remains a significant bottleneck as context length increases. We review and distill the recent efforts to overcome this bottleneck, including advances in (sub-quadratic) attention variants, recurrent neural networks, state space models, and hybrid architectures. We critically analyze approaches regarding compute and memory complexity, benchmark results, and fundamental limitations to assess whether the dominance of pure-attention transformers may soon be challenged, which we consider possible, particularly in domain-specific and edge-device applications.

Cross-lingual transfer has become a central paradigm for extending natural language processing (NLP) technologies to low-resource languages. By leveraging supervision from high-resource languages, multilingual language models can achieve strong task performance with little or no labeled target-language data. However, it remains unclear to what extent cross-lingual transfer can substitute for language-specific efforts. In this paper, we synthesize prior research findings and data collection results on Luxembourgish, which, despite its typological proximity to high-resource languages and its presence in a multilingual context, remains insufficiently represented in modern NLP technologies. Across findings, we observe a fundamental interdependence between cross-lingual transfer and language-specific efforts. Cross-lingual transfer can substantially improve target-language performance, but its success depends critically on the availability of sufficiently high-quality, task-aligned target-language data. At the same time, such resources, particularly in low-resource settings, are typically too limited in scale to drive strong performance on their own. Instead, such resources reach their full potential only when leveraged within a cross-lingual framework. We therefore argue that cross-lingual transfer and language-specific efforts should not be viewed as competing alternatives. Instead, they function as complementary components of a sustainable low-resource NLP pipeline. Based on these insights, we provide practical guidelines for integrating and balancing cross-lingual transfer with language-specific development in sustainable low-resource NLP pipelines.

This paper reflects on twenty years of building NLP resources and research infrastructure for Arabic, a language spoken by hundreds of millions yet historically underserved relative to languages such as English or Chinese. The first decade focused on foundational linguistic infrastructure; the second shifted toward computational social science, social media analysis, and socially oriented applications. Rather than cataloguing outputs, the paper examines what the experience of building them revealed. Three counterintuitive lessons emerge: building datasets is as much a social process as a technical one; communities formed around shared tasks often matter more than the tasks themselves; and moving from language resources to computational social science exposes challenges that traditional NLP training does not address. We discuss three failures: a depression detection corpus that never reached clinical practice, a period of spreading across too many shared tasks without sufficient depth, and a long-standing assumption that Modern Standard Arabic infrastructure would transfer cleanly to dialectal tasks. These experiences suggest that the hardest problems in developing NLP for underserved communities are not linguistic but social, institutional, and epistemic, and require competencies the field rarely teaches.

This work aims to shine a spotlight on the topic of metalanguage. We first define metalanguage, link it to NLP and LLMs, and then discuss our two labs’ metalanguage-centered efforts. Finally, we discuss four dimensions of metalanguage and metalinguistic tasks, offering a list of understudied future research directions.

Language models have changed from unreliable text generators to highly-capable large models with trillions of parameters. Capability increases come hand-in-hand with increases in scale, making understanding the internal representations of models more challenging. Since millions of users increasing rely on language models to interact with external tools or make decisions in medium or high-stakes scenarios, we need to establish control over model behavior and know when to trust model outputs. In this paper, we discuss our contributions on harnessing the latent spaces by proposing steering vectors for control and developing latent space-based model calibrators for trust. Together, our contributions help demystify the latent spaces of language models and offer new insights into how to harness model internals to build more trustworthy language technology.

Language models are increasingly used to quantify cultural phenomena, but what makes such measurement distinctively cultural? This paper argues that NLP work on culture is a material-discursive practice: the apparatus—model, data, annotation, evaluation—participates in constituting the cultural reality it measures, rather than passively recording it. Drawing on Karen Barad’s concept of the agential cut—the contingent boundary between phenomenon and instrument—I show that the apparatus’s substantive design choices draw such boundaries, and that the boundary is entangled from the start because language models have already internalized much of the cultural material they measure. I illustrate this through three case studies on television and film dialogue and two examinations of the apparatus itself: erasure of character names as cultural markers, and attunement to historically distant Restoration drama. This big picture analysis proposes a research program that is theory-driven, empirically rigorous, and culturally contingent, treating each agential cut as a conscious commitment.

Memorisation in deep learning is undergoing a paradigm shift; it is increasingly recognised as a mechanism that can support, rather than hinder, generalisation. This is particularly relevant in NLP, where language combines compositional, generalisable structure with non-compositional expressions such as idioms, requiring memorisation from models and humans alike. My PhD work investigated memorisation in transformer models in generic terms, and through the lens of (non-)compositionality, from both data and model-internal perspectives. I analysed which training examples require memorisation, whether memorisation supports generalisation, and where memorisation occurs within model layers. I also studied how transformers process non-compositional idiom translations and how they balance compositional generalisation with non-compositional memorisation. Based on my findings, I stress that memorisation is an inherent part of learning natural language, can be beneficial, and is partially predictable. Yet it is not cleanly separable from generalisation, both at the level of data and of model parameters. Here, I summarise those findings and reflect on my PhD work.