Backchannels and fillers are important linguistic expressions in dialogue, but are under-represented in modern transformer-based language models (LMs). Our work studies the representation of them in language models using three fine-tuning strategies. The models are trained on three dialogue corpora in English and Japanese, where backchannels and fillers are preserved and annotated, to investigate how fine-tuning can help LMs learn their representations. We first apply clustering analysis to the learnt representation of backchannels and fillers, and have found increased silhouette scores in representations from fine-tuned models, which suggests that fine-tuning enables LMs to distinguish the nuanced semantic variation in different backchannel and filler use. We also use natural language generation (NLG) metrics to confirm that the utterances generated by fine-tuned language models resemble human-produced utterances more closely. Our findings suggest the potentials of transforming general LMs into conversational LMs that are more capable of producing human-like languages adequately.

We present Federation of Agents (FoA), a distributed orchestration framework that transforms static multi-agent coordination into dynamic, capability-driven collaboration. FoA introduces Versioned Capability Vectors (VCVs): machine-readable profiles that make agent capabilities searchable through semantic embeddings, enabling agents to advertise their capabilities, cost, and limitations. Our aarchitecturecombines three key innovations: (1) semantic routing that matches tasks to agents over sharded HNSW indices while enforcing operational constraints through cost-biased optimization, (2) dynamic task decomposition where compatible agents collaboratively break down complex tasks into DAGs of subtasks through consensus-based merging, and (3) smart clustering that groups agents working on similar subtasks into collaborative channels for k-round refinement before synthesis. Built on top of MQTT,s publish-subscribe semantics for scalable message passing, FoA achieves sub-linear complexity through hierarchical capability matching and efficient index maintenance. Evaluation on HealthBench shows 13x improvements over single-model baselines, with clustering-enhanced laboration particularly effective for complex reasoning tasks requiring multiple perspectives. The system scales horizontally while maintaining consistent performance, demonstrating that semantic orchestration with structured collaboration can unlock the collective intelligence of heterogeneous federations of AI agents.

Multilingual Large Language Models (LLMs) are increasingly used worldwide, making it essential to ensure they are free from gender bias to prevent representational harm. While prior studies have examined such biases in high-resource languages, low-resource languages remain understudied. In this paper, we propose a template-based probing methodology, validated against real-world data, to uncover gender stereotypes in LLMs. As part of this framework, we introduce the Domain-Specific Gender Skew Index (DS-GSI), a metric that quantifies deviations from gender parity. We evaluate four prominent models, GPT-4o mini, DeepSeek R1, Gemini 2.0 Flash, and Qwen QwQ 32B, across four semantic domains, focusing on Persian, a low-resource language with distinct linguistic features. Our results show that all models exhibit gender stereotypes, with greater disparities in Persian than in English across all domains. Among these, sports reflect the most rigid gender biases. This study underscores the need for inclusive NLP practices and provides a framework for assessing bias in other low-resource languages.

Recent benchmarks for medical Large Vision-Language Models (LVLMs) emphasize leaderboard accuracy, overlooking reliability and safety. We study sycophancy -- models' tendency to uncritically echo user-provided information -- in high-stakes clinical settings. We introduce EchoBench, a benchmark to systematically evaluate sycophancy in medical LVLMs. It contains 2,122 images across 18 departments and 20 modalities with 90 prompts that simulate biased inputs from patients, medical students, and physicians. We evaluate medical-specific, open-source, and proprietary LVLMs. All exhibit substantial sycophancy; the best proprietary model (Claude 3.7 Sonnet) still shows 45.98% sycophancy, and GPT-4.1 reaches 59.15%. Many medical-specific models exceed 95% sycophancy despite only moderate accuracy. Fine-grained analyses by bias type, department, perceptual granularity, and modality identify factors that increase susceptibility. We further show that higher data quality/diversity and stronger domain knowledge reduce sycophancy without harming unbiased accuracy. EchoBench also serves as a testbed for mitigation: simple prompt-level interventions (negative prompting, one-shot, few-shot) produce consistent reductions and motivate training- and decoding-time strategies. Our findings highlight the need for robust evaluation beyond accuracy and provide actionable guidance toward safer, more trustworthy medical LVLMs.

In recent years, deep neural network-based large language models (LLMs) have demonstrated remarkable performance (Comanici et al., 2025; OpenAI et al., 2024; DeepSeek-AI et al., 2025). The development of these models has largely followed what Richard Sutton termed "the bitter lesson"-that the most effective approach to improving AI performance has historically been to leverage greater computational resources, larger models, and more data, rather than incorporating human knowledge or specialized architectures (Sutton, 2019). This trend has been formalized through scaling laws, which quantify the relationship between model performance and factors such as model size, dataset size, and computational budget through power law relationships (Kaplan et al., 2020). While these scaling laws provide valuable quantitative predictions for model performance, they also reveal a concerning limitation: the power law relationship suggests that achieving further significant improvements may require prohibitively large increases in model and data size, making continued scaling increasingly impractical and resource-intensive. One promising approach to address these limitations is to develop a deeper understanding of the underlying mechanisms that drive transformer models' success in natural language processing (NLP). The No Free Lunch theorem establishes that no single algorithm can perform optimally across all problem domains, highlighting the fundamental importance of understanding both the characteristics of the data and the properties of the algorithms that process it (Wolpert & Macready, 1997).

Recent papers show LLMs achieve near-random accuracy in causal relation classification, raising questions about whether such failures arise from limited pretraining exposure or deeper representational gaps. We investigate this under uncertainty-based evaluation, testing whether pretraining exposure to causal examples improves causal understanding >18K PubMed sentences -- half from The Pile corpus, half post-2024 -- across seven models (Pythia-1.4B/7B/12B, GPT-J-6B, Dolly-7B/12B, Qwen-7B). We analyze model behavior through: (i) causal classification, where the model identifies causal relationships in text, and (ii) verbatim memorization probing, where we assess whether the model prefers previously seen causal statements over their paraphrases. Models perform four-way classification (direct/conditional/correlational/no-relationship) and select between originals and their generated paraphrases. Results show almost identical accuracy on seen/unseen sentences (p > 0.05), no memorization bias (24.8% original selection), and output distribution over the possible options is almost flat, with entropic values near the maximum (1.35/1.39), confirming random guessing. Instruction-tuned models show severe miscalibration (Qwen: > 95% confidence, 32.8% accuracy, ECE=0.49). Conditional relations induce highest entropy (+11% vs. direct). These findings suggest that failures in causal understanding arise from the lack of structured causal representation, rather than insufficient exposure to causal examples during pretraining.

Phonemization, the conversion of text into phonemes, is a key step in text-to-speech. Traditional approaches use rule-based transformations and lexicon lookups, while more advanced methods apply preprocessing techniques or neural networks for improved accuracy on out-of-domain vocabulary. However, all systems struggle with names, loanwords, abbreviations, and homographs. This work presents OLaPh (Optimal Language Phonemizer), a framework that combines large lexica, multiple NLP techniques, and compound resolution with a probabilistic scoring function. Evaluations in German and English show improved accuracy over previous approaches, including on a challenging dataset. To further address unresolved cases, we train a large language model on OLaPh-generated data, which achieves even stronger generalization and performance. Together, the framework and LLM improve phonemization consistency and provide a freely available resource for future research.

ABSTRACT In many IoT applications, the central interest lies not in individual sensor signals but in their differences, yet interpreting such differences requires expert knowledge. We propose DiffNator, a framework for structured explanations of differences between two time series. Using the Time-series Observations of Real-world IoT (TORI) dataset, we generate paired sequences and train a model that combine a time-series encoder with a frozen LLM to output JSON-formatted explanations. Experimental results show that DiffNator generates accurate difference explanations and substantially outperforms both a visual question answering (VQA) baseline and a retrieval method using a pre-trained time-series encoder. Index T erms-- Time series analysis, natural language generation, difference explanation, industrial IoT, large language model 1. INTRODUCTION The spread of the Internet of Things (IoT) has enabled large-scale collection of sensor data from industrial machinery.

Large language model-based artificial conversational agents (like ChatGPT) give answers to all kinds of questions, and often enough these answers are correct. Just on the basis of that capacity alone, we may attribute knowledge to them. But do these models use this knowledge as a basis for their own conversational behaviour? I argue this is not the case, and I will refer to this failure as a `disconnect'. I further argue this disconnect is fundamental in the sense that with more data and more training of the LLM on which a conversational chatbot is based, it will not disappear. The reason is, as I will claim, that the core technique used to train LLMs does not allow for the establishment of the connection we are after. The disconnect reflects a fundamental limitation on the capacities of LLMs, and explains the source of hallucinations. I will furthermore consider the ethical version of the disconnect (ethical conversational knowledge not being aligned with ethical conversational behaviour), since in this domain researchers have come up with several additional techniques to influence a chatbot's behaviour. I will discuss how these techniques do nothing to solve the disconnect and can make it worse.

Recent advances in generalist robot manipulation leverage pre-trained Vision-Language Models (VLMs) and large-scale robot demonstrations to tackle diverse tasks in a zero-shot manner. A key challenge remains: scaling high-quality, action-labeled robot demonstration data, which existing methods rely on for robustness and generalization. To address this, we propose a method that benefits from videos without action labels - featuring humans and/or robots in action - enhancing open-vocabulary performance and enabling data-efficient learning of new tasks. Our method extracts dense, dynamic 3D point clouds at the hand or gripper location and uses a proposed 3D dynamics predictor for self-supervision. This predictor is then tuned to an action predictor using a smaller labeled dataset for action alignment. We show that our method not only learns from unlabeled human and robot demonstrations - improving downstream generalist robot policies - but also enables robots to learn new tasks without action labels (i.e., out-of-action generalization) in both real-world and simulated settings.