Papers

This study explored how in-context learning (ICL) in large language models can be disrupted by data poisoning attacks in the setting of public health sentiment analysis. Using tweets of Human Metapneumovirus (HMPV), small adversarial perturbations such as synonym replacement, negation insertion, and randomized perturbation were introduced into the support examples. Even these minor manipulations caused major disruptions, with sentiment labels flipping in up to 67% of cases. To address this, a Spectral Signature Defense was applied, which filtered out poisoned examples while keeping the data's meaning and sentiment intact. After defense, ICL accuracy remained steady at around 46.7%, and logistic regression validation reached 100% accuracy, showing that the defense successfully preserved the dataset's integrity. Overall, the findings extend prior theoretical studies of ICL poisoning to a practical, high-stakes setting in public health discourse analysis, highlighting both the risks and potential defenses for robust LLM deployment. This study also highlights the fragility of ICL under attack and the value of spectral defenses in making AI systems more reliable for health-related social media monitoring.

Recently proposed evaluation benchmarks aim to characterize the effective context length and the forgetting tendencies of large language models (LLMs). However, these benchmarks often rely on simplistic 'needle in a haystack' retrieval or continuation tasks that may not accurately reflect the performance of these models in information-dense scenarios. Thus, rather than simple next token prediction, we argue for evaluating these models on more complex reasoning tasks that requires them to induce structured relational knowledge from the text - such as graphs from potentially noisy natural language content. While the input text can be viewed as generated in terms of a graph, its structure is not made explicit and connections must be induced from distributed textual cues, separated by long contexts and interspersed with irrelevant information. Our findings reveal that LLMs begin to exhibit memory drift and contextual forgetting at much shorter effective lengths when tasked with this form of relational reasoning, compared to what existing benchmarks suggest. With these findings, we offer recommendations for the optimal use of popular LLMs for complex reasoning tasks. We further show that even models specialized for reasoning, such as OpenAI o1, remain vulnerable to early memory drift in these settings. These results point to significant limitations in the models' ability to abstract structured knowledge from unstructured input and highlight the need for architectural adaptations to improve long-range reasoning.

Large language models (LLMs) are increasingly adept at solving complex problems, such as mathematical reasoning and automatic evaluation. However, performance often degrades when prompts intertwine task instructions with rigid formatting requirements. This entanglement creates competing goals for the model, hindering its reasoning capabilities. To address this, we introduce Deco-G, a decoding framework that explicitly decouples format adherence from problem solving. Deco-G delegates format adherence to a separate Format Estimation Module (FEM), which performs probabilistic lookahead to estimate future format compliance rate and reweighs token probabilities, allowing the LLM to focus solely on task resolution. To make this approach both practical and efficient, we introduce three key innovations: instruction-aware distillation, a flexible trie-building algorithm, and HMM state pruning. Experiments across mathematical reasoning, event argument extraction, and LLM-as-a-judge demonstrate that Deco-G constantly gains over prompting or structured generation baselines, with guaranteed format compliance. We release our code at https://github.com/haikangdeng/deco-g.

An important aspect subtending language understanding and production is the ability to independently encode positional and symbolic information of the words within a sentence. In Transformers, positional information is typically encoded using Positional Encodings (PEs). One such popular PE, namely Rotary PE (RoPE), has been widely used due to its empirical success. Recently, it has been argued that part of RoPE's success emerges from its ability to encode robust positional and semantic information using large and small frequencies, respectively. In this work, we perform a deeper dive into the positional versus symbolic dichotomy of attention heads behavior, both at the theoretical and empirical level. We provide general definitions of what it means for a head to behave positionally or symbolically, prove that these are two mutually exclusive behaviors and develop a metric to quantify them. We apply our framework to analyze Transformer-based LLMs using RoPE and find that all heads exhibit a strong correspondence between behavior and frequency use. Finally, we introduce canonical tasks designed to be either purely positional or symbolic, and demonstrate that the Transformer performance causally relates to the ability of attention heads to leverage the appropriate frequencies. In particular, we show that we can control the Transformer performance by controlling which frequencies the attention heads can access. Altogether, our work provides a detailed understanding of RoPE, and how its properties relate to model behavior.

This article presents a comparative study of large language models (LLMs) in the task of sentiment analysis of financial market news. This work aims to analyze the performance difference of these models in this important natural language processing task within the context of finance. LLM models are compared with classical approaches, allowing for the quantification of the benefits of each tested model or approach. Results show that large language models outperform classical models in the vast majority of cases.

Biomedical Named Entity Recognition (BioNER), task6 in BioASQ (A challenge in large-scale biomedical semantic indexing and question answering), is crucial for extracting information from scientific literature but faces hurdles such as distinguishing between similar entity types like genes and chemicals. This study evaluates the GLiNER-BioMed model on a BioASQ dataset and introduces a targeted dictionary-based post-processing strategy to address common misclassifications. While this post-processing approach demonstrated notable improvement on our development set, increasing the micro F1-score from a baseline of 0.79 to 0.83, this enhancement did not generalize to the blind test set, where the post-processed model achieved a micro F1-score of 0.77 compared to the baselines 0.79. We also discuss insights gained from exploring alternative methodologies, including Conditional Random Fields. This work highlights the potential of dictionary-based refinement for pre-trained BioNER models but underscores the critical challenge of overfitting to development data and the necessity of ensuring robust generalization for real-world applicability.

As Large Language Models (LLMs) gain wider adoption, ensuring their reliable handling of Personally Identifiable Information (PII) across diverse regulatory contexts has become essential. This work introduces a scalable multilingual data curation framework designed for high-quality PII annotation across 13 underrepresented locales, covering approximately 336 locale-specific PII types. Our phased, human-in-the-loop annotation methodology combines linguistic expertise with rigorous quality assurance, leading to substantial improvements in recall and false positive rates from pilot, training, and production phases. By leveraging inter-annotator agreement metrics and root-cause analysis, the framework systematically uncovers and resolves annotation inconsistencies, resulting in high-fidelity datasets suitable for supervised LLM fine-tuning. Beyond reporting empirical gains, we highlight common annotator challenges in multilingual PII labeling and demonstrate how iterative, analytics-driven pipelines can enhance both annotation quality and downstream model reliability.

The 2025 Multimodal Models for Low-Resource Contexts and Social Impact (MMLoSo) Language Challenge addresses one of India's most pressing linguistic gaps: the lack of resources for its diverse low-resource languages (LRLs). In this study, we investigate whether enforcing cross-lingual similarity in specific internal layers of a decoder-only multilingual large language model (LLM) can improve translation quality from LRL to high-resource language (HRL). Specifically, we combine Centered Kernel Alignment (CKA), a similarity metric that encourages representations of different languages to align, with REPINA, a regularization method that constrains parameter updates to remain close to the pretrained model, into a joint method we call TRepLiNa. In this research project, we experiment with zero-shot, few-shot, and fine-tuning settings using Aya-23 8B with QLoRA across MMLoSo shared task language pairs (Mundari, Santali, Bhili) with Hindi/English pivots. Our results show that aligning mid-level layers using TRepLiNa (CKA+REPINA) is a low-cost, practical approach to improving LRL translation, especially in data-scarce settings.

Finding an optimal word representation algorithm is particularly important in terms of domain specific data, as the same word can have different meanings and hence, different representations depending on the domain and context. While Generative AI and transformer architecture does a great job at generating contextualized embeddings for any given work, they are quite time and compute extensive, especially if we were to pre-train such a model from scratch. In this work, we focus on the scientific domain and finding the optimal word representation algorithm along with the tokenization method that could be used to represent words in the scientific domain. The goal of this research is two fold: 1) finding the optimal word representation and tokenization methods that can be used in downstream scientific domain NLP tasks, and 2) building a comprehensive evaluation suite that could be used to evaluate various word representation and tokenization algorithms (even as new ones are introduced) in the scientific domain. To this end, we build an evaluation suite consisting of several downstream tasks and relevant datasets for each task. Furthermore, we use the constructed evaluation suite to test various word representation and tokenization algorithms.

Referring Expression Comprehension and Segmentation are critical tasks for assessing the integration of language understanding and image comprehension, serving as benchmarks for Multimodal Large Language Models (MLLMs) capabilities. To address these challenges, we propose a new strategy, CoT Referring, which enhances model reasoning across modalities through a structured, chain-of-thought training data structure. Our approach systematically parses textual structures to a sequential referring step, where in each step it identifies relationships and ensures consistent reference alignment, thereby improving accuracy in complex query scenarios. We restructure the training data to enforce a new output form, providing new annotations for existing datasets and compiling an evaluation benchmark from existing resources. This benchmark is designed explicitly for complex referring cases. We also integrate detection and segmentation capabilities into a unified MLLM framework, training it with a novel adaptive weighted loss to optimize performance. Experimental results on our curated benchmark and RefCOCO/+/g demonstrate the effectiveness of our approach, with a notable increase of 2.5%+ over baseline models.

Open-ended survey responses provide valuable insights in marketing research, but low-quality responses not only burden researchers with manual filtering but also risk leading to misleading conclusions, underscoring the need for effective evaluation. Existing automatic evaluation methods target LLM-generated text and inadequately assess human-written responses with their distinct characteristics. To address such characteristics, we propose a two-stage evaluation framework specifically designed for human survey responses. First, gibberish filtering removes nonsensical responses. Then, three dimensions-effort, relevance, and completeness-are evaluated using LLM capabilities, grounded in empirical analysis of real-world survey data. Validation on English and Korean datasets shows that our framework not only outperforms existing metrics but also demonstrates high practical applicability for real-world applications such as response quality prediction and response rejection, showing strong correlations with expert assessment.

Industrial question-answering (QA) systems require higher safety and reliability than general-purpose dialogue models, as errors in high-risk scenarios such as equipment fault diagnosis can have severe consequences. Although multi-agent large language models enhance reasoning depth, they suffer from uncontrolled iterations and unverifiable outputs, and conventional distillation methods struggle to transfer collaborative reasoning capabilities to lightweight, deployable student models. To address these challenges, we propose Knowledge Graph-guided Multi-Agent System Distillation (KG-MASD). Our approach formulates distillation as a Markov Decision Process and incorporates a knowledge graph as a verifiable structured prior to enrich state representation and ensure convergence. By integrating collaborative reasoning with knowledge grounding, KG-MASD generates high-confidence instruction-tuning data and jointly distills reasoning depth and verifiability into compact student models suitable for edge deployment. Experiments on an industrial QA dataset show that KG-MASD improves accuracy by 2.4 per cent to 20.1 per cent over baselines and significantly enhances reliability, enabling trustworthy AI deployment in safety-critical industrial scenarios. Code and data are available at https://github.com/erwinmsmith/KG-MAD/.

We present OpenStaxQA, an evaluation benchmark specific to college-level educational applications based on 43 open-source college textbooks in English, Spanish, and Polish, available under a permissive Creative Commons license. We finetune and evaluate large language models (LLMs) with approximately 7 billion parameters on this dataset using quantized low rank adapters (QLoRa). Additionally we also perform a zero-shot evaluation on the AI2 reasoning challenge dev dataset in order to check if OpenStaxQA can lead to an improved performance on other tasks. We also discuss broader impacts relevant to datasets such as OpenStaxQA.

This work presents our participation in the EvalLLM 2025 challenge on biomedical Named Entity Recognition (NER) and health event extraction in French (few-shot setting). For NER, we propose three approaches combining large language models (LLMs), annotation guidelines, synthetic data, and post-processing: (1) in-context learning (ICL) with GPT-4.1, incorporating automatic selection of 10 examples and a summary of the annotation guidelines into the prompt, (2) the universal NER system GLiNER, fine-tuned on a synthetic corpus and then verified by an LLM in post-processing, and (3) the open LLM LLaMA-3.1-8B-Instruct, fine-tuned on the same synthetic corpus. Event extraction uses the same ICL strategy with GPT-4.1, reusing the guideline summary in the prompt. Results show GPT-4.1 leads with a macro-F1 of 61.53% for NER and 15.02% for event extraction, highlighting the importance of well-crafted prompting to maximize performance in very low-resource scenarios.

With the increasing adoption of Large Language Models (LLMs), more customization is needed to ensure privacy-preserving and safe generation. We address this objective from two critical aspects: unlearning of sensitive information and robustness to jail-breaking attacks. We investigate various constrained optimization formulations that address both aspects in a unified manner, by finding the smallest possible interventions on LLM weights that either make a given vocabulary set unreachable or embed the LLM with robustness to tailored attacks by shifting part of the weights to a safer region. Beyond unifying two key properties, this approach contrasts with previous work in that it doesn't require an oracle classifier that is typically not available or represents a computational overhead. Surprisingly, we find that the simplest point-wise constraint-based intervention we propose leads to better performance than max-min interventions, while having a lower computational cost. Comparison against state-of-the-art defense methods demonstrates superior performance of the proposed approach.

The Transformer architecture has become the de facto standard for Large Language Models (LLMs), demonstrating remarkable capabilities in language understanding and generation. However, its application in conversational AI is fundamentally constrained by its stateless nature and the quadratic computational complexity ($O(L^2)$) with respect to sequence length $L$. Current models emulate memory by reprocessing an ever-expanding conversation history with each turn, leading to prohibitive costs and latency in long dialogues. This paper introduces the Reactive Transformer (RxT), a novel architecture designed to overcome these limitations by shifting from a data-driven to an event-driven paradigm. RxT processes each conversational turn as a discrete event in real-time, maintaining context in an integrated, fixed-size Short-Term Memory (STM) system. The architecture features a distinct operational cycle where a generator-decoder produces a response based on the current query and the previous memory state, after which a memory-encoder and a dedicated Memory Attention network asynchronously update the STM with a representation of the complete interaction. This design fundamentally alters the scaling dynamics, reducing the total user-facing cost of a conversation from quadratic ($O(N^2 \cdot T)$) to linear ($O(N \cdot T)$) with respect to the number of interactions $N$. By decoupling response generation from memory updates, RxT achieves low latency, enabling truly real-time, stateful, and economically viable long-form conversations. We validated our architecture with a series of proof-of-concept experiments on synthetic data, demonstrating superior performance and constant-time inference latency compared to a baseline stateless model of comparable size.

Although large language models (LLMs) are increasingly implicated in interpersonal and societal decision-making, their ability to navigate explicit conflicts between legitimately different cultural value systems remains largely unexamined. Existing benchmarks predominantly target cultural knowledge (CulturalBench), value prediction (WorldValuesBench), or single-axis bias diagnostics (CDEval); none evaluate how LLMs adjudicate when multiple culturally grounded values directly clash. We address this gap with CCD-Bench, a benchmark that assesses LLM decision-making under cross-cultural value conflict. CCD-Bench comprises 2,182 open-ended dilemmas spanning seven domains, each paired with ten anonymized response options corresponding to the ten GLOBE cultural clusters. These dilemmas are presented using a stratified Latin square to mitigate ordering effects. We evaluate 17 non-reasoning LLMs. Models disproportionately prefer Nordic Europe (mean 20.2 percent) and Germanic Europe (12.4 percent), while options for Eastern Europe and the Middle East and North Africa are underrepresented (5.6 to 5.8 percent). Although 87.9 percent of rationales reference multiple GLOBE dimensions, this pluralism is superficial: models recombine Future Orientation and Performance Orientation, and rarely ground choices in Assertiveness or Gender Egalitarianism (both under 3 percent). Ordering effects are negligible (Cramer's V less than 0.10), and symmetrized KL divergence shows clustering by developer lineage rather than geography. These patterns suggest that current alignment pipelines promote a consensus-oriented worldview that underserves scenarios demanding power negotiation, rights-based reasoning, or gender-aware analysis. CCD-Bench shifts evaluation beyond isolated bias detection toward pluralistic decision making and highlights the need for alignment strategies that substantively engage diverse worldviews.

Generative large language models (LLMs) are now used extensively for text classification in computational social science (CSS). In this work, focus on the steps before and after LLM prompting – conceptualization of concepts to be classified and using LLM predictions in downstream statistical inference – which we argue have been overlooked in much of LLM-era CSS. We claim LLMs can tempt analysts to skip the conceptualization step, creating conceptualization errors that bias downstream estimates. Using simulations, we show that this conceptualization-induced bias cannot be corrected for solely by increasing LLM accuracy or post-hoc bias correction methods. We conclude by reminding CSS analysts that conceptualization is still a first-order concern in the LLM-era and provide concrete advice on how to pursue low-cost, unbiased, low-variance downstream estimates.

Multi-turn medical question answering (QA) aims to model realistic clinical diagnosis, where a doctor gathers patient information across multiple turns of conversation. Existing multi-turn medical conversation systems have shown promising progress, but they often rely on accumulated conversation histories as memory, leaving clinical evidence fragmented across turns. We propose GraphMed-LT, a patient-specific graph memory approach with latent clinical thought refinement for multi-turn medical conversations. GraphMed-LT extracts patient-specific clinical triplets from patient responses, retrieves relevant knowledge triplets, and organises them into an incrementally updated graph memory. The graph memory is projected into graph-conditioned evidence tokens and refined inside a trainable doctor agent through hidden-state feedback, enabling the agent to update its internal clinical context before asking follow-up questions or producing the final answer. Experiments on three multi-turn medical QA benchmarks show that GraphMed-LT consistently outperforms existing multi-turn medical conversation baselines across multiple LLM backbones, achieving up to a 6.3 percentage-point absolute improvement over the strongest baseline. Further analyses show that GraphMed-LT asks more answerable follow-up questions and provides consistent gains across medical specialties.

Instruction-tuning plays a vital role in enhancing the task-solving abilities of large language models (LLMs), improving their usability in generating helpful responses on various tasks. However, previous work has demonstrated that they are sensitive to minor variations in instruction phrasing. In this paper, we explore whether introducing perturbations in instruction-tuning data can enhance LLMs' resistance against noisy instructions. We focus on how instruction-tuning with perturbations, such as removing stop words or shuffling words, affects LLMs' performance on the original and perturbed versions of widely-used benchmarks (MMLU, BBH, GSM8K). We further assess learning dynamics and potential shifts in model behavior. Surprisingly, our results suggest that instruction-tuning on perturbed instructions can, in some cases, improve downstream performance. These findings highlight the importance of including perturbed instructions in instruction-tuning, which can make LLMs more resilient to noisy user inputs.

Language models can be sampled multiple times to access the distribution underlying their responses, but existing methods cannot efficiently synthesize rich epistemic signals across different long-form responses. We introduce Consensus Graphs (ConGrs), a flexible DAG-based data structure that represents shared information, as well as semantic variation in a set of sampled LM responses to the same prompt. We construct ConGrs using a light-weight lexical sequence alignment algorithm from bioinformatics, supplemented by the targeted usage of a secondary LM judge. Further, we design task-dependent decoding methods to synthesize a single, final response from our ConGr data structure. Our experiments show that synthesizing responses from ConGrs improves factual precision on two biography generation tasks by up to 31% over an average response and reduces reliance on LM judges by more than 80% compared to other methods. We also use ConGrs for three refusal-based tasks requiring abstention on unanswerable queries and find that abstention rate is increased by up to 56%. We apply our approach to the MATH and AIME reasoning tasks and find an improvement over self-verification and majority vote baselines by up to 6 points of accuracy. We show that ConGrs provide a flexible method for capturing variation in LM responses and using the epistemic signals provided by response variation to synthesize more effective responses.

In specialized domains, humans often compare new problems against similar examples, highlight nuances, and draw conclusions instead of analyzing information in isolation. When applying reasoning in specialized contexts with LLMs on top of a RAG, the pipeline can capture contextually relevant information, but it is not designed to retrieve comparable cases or related problems. While RAG is effective at extracting factual information, its outputs in specialized reasoning tasks often remain generic, reflecting broad facts rather than context-specific insights. In finance, it results in generic risks that are true for the majority of companies. To address this limitation, we propose a peer-aware comparative inference layer on top of RAG. Our contrastive approach outperforms baseline RAG in text generation metrics such as ROUGE and BERTScore in comparison with human-generated equity research and risk.

Time series reasoning is crucial to decision-making in diverse domains, including finance, energy, and scientific discovery. While existing time series foundation models (TSFMs) can capture low-level dynamic patterns and provide accurate forecasting, further analysis usually requires additional background knowledge and sophisticated reasoning, which are lacking in most TSFMs but can be achieved through Large Language Models (LLMs). On the other hand, without expensive post-training, LLMs often struggle with the numerical understanding of time series data. Although it is intuitive to integrate the two types of models, developing effective training recipes that align the two modalities for reasoning tasks is still an open challenge. To this end, we propose TS-Reasoner that aligns the latent representations of TSFMs with the textual inputs of LLMs for downstream understanding/reasoning tasks. Specifically, we propose a simple yet effective method to curate diverse, synthetic pairs of time series and textual captions for alignment training. We then develop a two-stage training recipe that applies instruction fine-tuning after the alignment pretraining. Unlike existing works that train an LLM to take time series as inputs, we leverage a pretrained TSFM and freeze it during training. Experiments on several benchmarks demonstrate that TS-Reasoner not only outperforms a wide range of open-source LLMs, Vision-Language Models (VLMs), and Time Series LLMs of comparable scale, but also does so with remarkable data efficiency, e.g., using less than half the training data.

As military organisations consider integrating large language models (LLMs) into command and control (C2) systems for planning and decision support, understanding their behavioural tendencies is critical. This study develops a benchmarking framework for evaluating aspects of legal and moral risk in targeting behaviour by comparing LLMs acting as agents in multi-turn simulated conflict. We introduce four metrics grounded in International Humanitarian Law (IHL) and military doctrine: Civilian Target Rate (CTR) and Dual-use Target Rate (DTR) assess compliance with legal targeting principles, while Mean and Max Simulated Non-combatant Casualty Value (SNCV) quantify tolerance for civilian harm. We evaluate three frontier models, GPT-4o, Gemini-2.5, and LLaMA-3.1, through 90 multi-agent, multi-turn crisis simulations across three geographic regions. Our findings reveal that off-the-shelf LLMs exhibit concerning and unpredictable targeting behaviour in simulated conflict environments. All models violated the IHL principle of distinction by targeting civilian objects, with breach rates ranging from 16.7% to 66.7%. Harm tolerance escalated through crisis simulations with MeanSNCV increasing from 16.5 in early turns to 27.7 in late turns. Significant inter-model variation emerged: LLaMA-3.1 selected an average of 3.47 civilian strikes per simulation with MeanSNCV of 28.4, while Gemini-2.5 selected 0.90 civilian strikes with MeanSNCV of 17.6. These differences indicate that model selection for deployment constitutes a choice about acceptable legal and moral risk profiles in military operations. This work seeks to provide a proof-of-concept of potential behavioural risks that could emerge from the use of LLMs in Decision Support Systems (AI DSS) as well as a reproducible benchmarking framework with interpretable metrics for standardising pre-deployment testing.

We introduce ALHD, the first large-scale comprehensive Arabic dataset explicitly designed to distinguish between human- and LLM-generated texts. ALHD spans three genres (news, social media, reviews), covering both MSA and dialectal Arabic, and contains over 400K balanced samples generated by three leading LLMs and originated from multiple human sources, which enables studying generalizability in Arabic LLM-genearted text detection. We provide rigorous preprocessing, rich annotations, and standardized balanced splits to support reproducibility. In addition, we present, analyze and discuss benchmark experiments using our new dataset, in turn identifying gaps and proposing future research directions. Benchmarking across traditional classifiers, BERT-based models, and LLMs (zero-shot and few-shot) demonstrates that fine-tuned BERT models achieve competitive performance, outperforming LLM-based models. Results are however not always consistent, as we observe challenges when generalizing across genres; indeed, models struggle to generalize when they need to deal with unseen patterns in cross-genre settings, and these challenges are particularly prominent when dealing with news articles, where LLM-generated texts resemble human texts in style, which opens up avenues for future research. ALHD establishes a foundation for research related to Arabic LLM-detection and mitigating risks of misinformation, academic dishonesty, and cyber threats.

We introduce the SEER (Span-based Emotion Evidence Retrieval) Benchmark to test Large Language Models' (LLMs) ability to identify the specific spans of text that express emotion. Unlike traditional emotion recognition tasks that assign a single label to an entire sentence, SEER targets the underexplored task of emotion evidence detection: pinpointing which exact phrases convey emotion. This span-level approach is crucial for applications like empathetic dialogue and clinical support, which need to know how emotion is expressed, not just what the emotion is. SEER includes two tasks: identifying emotion evidence within a single sentence, and identifying evidence across a short passage of five consecutive sentences. It contains new annotations for both emotion and emotion evidence on 1200 real-world sentences. We evaluate 14 open-source LLMs and find that, while some models approach average human performance on single-sentence inputs, their accuracy degrades in longer passages. Our error analysis reveals key failure modes, including overreliance on emotion keywords and false positives in neutral text.

In this paper, we design a signalling game-based emergent communication environment to generate state-of-the-art emergent languages in terms of similarity to human language. This is done with hyperparameter optimization, using XferBench as the objective function. XferBench quantifies the statistical similarity of emergent language to human language by measuring its suitability for deep transfer learning to human language. Additionally, we demonstrate the predictive power of entropy on the transfer learning performance of emergent language as well as corroborate previous results on the entropy-minimization properties of emergent communication systems. Finally, we report generalizations regarding what hyperparameters produce more realistic emergent languages, that is, ones which transfer better to human language.

We present Omni-Embed-Nemotron, a unified multimodal retrieval embedding model developed to handle the increasing complexity of real-world information needs. While Retrieval-Augmented Generation (RAG) has significantly advanced language models by incorporating external knowledge, existing text-based retrievers rely on clean, structured input and struggle with the visually and semantically rich content found in real-world documents such as PDFs, slides, or videos. Recent work such as ColPali has shown that preserving document layout using image-based representations can improve retrieval quality. Building on this, and inspired by the capabilities of recent multimodal models such as Qwen2.5-Omni, we extend retrieval beyond text and images to also support audio and video modalities. Omni-Embed-Nemotron enables both cross-modal (e.g., text - video) and joint-modal (e.g., text - video+audio) retrieval using a single model. We describe the architecture, training setup, and evaluation results of Omni-Embed-Nemotron, and demonstrate its effectiveness in text, image, and video retrieval.

We introduce CSAR, an algorithm for inducing morphemes from emergent language corpora of parallel utterances and meanings. It is a greedy algorithm that (1) weights morphemes based on mutual information between forms and meanings, (2) selects the highest-weighted pair, (3) removes it from the corpus, and (4) repeats the process to induce further morphemes (i.e., Count, Select, Ablate, Repeat). The effectiveness of CSAR is first validated on procedurally generated datasets and compared against baselines for related tasks. Second, we validate CSAR's performance on human language data to show that the algorithm makes reasonable predictions in adjacent domains. Finally, we analyze a handful of emergent languages, quantifying linguistic characteristics like degree of synonymy and polysemy.

Kernel change-point detection (KCPD) has become a widely used tool for identifying structural changes in complex data. While existing theory establishes consistency under independence assumptions, real-world sequential data such as text exhibits strong dependencies. We establish new guarantees for KCPD under $m$-dependent data: specifically, we prove consistency in the number of detected change points and weak consistency in their locations under mild additional assumptions. We perform an LLM-based simulation that generates synthetic $m$-dependent text to validate the asymptotics. To complement these results, we present the first comprehensive empirical study of KCPD for text segmentation with modern embeddings. Across diverse text datasets, KCPD with text embeddings outperforms baselines in standard text segmentation metrics. We demonstrate through a case study on Taylor Swift's tweets that KCPD not only provides strong theoretical and simulated reliability but also practical effectiveness for text segmentation tasks.