The head of Google’s DeepMind says artificial intelligence could usher in an era of ‘incredible productivity’ and ‘radical abundance’. But who will it benefit? And why does he wish the tech giants had moved more slowly?

Recent advances in natural language processing (NLP) have opened up greater opportunities to enable fine-tuned large language models (LLMs) to behave as more powerful interactive agents through improved instruction-following ability. However, understanding how this impacts confidence calibration for reliable model output has not been researched in full. In this work, we examine various open-sourced LLMs, identifying significant calibration degradation after instruction tuning in each. Seeking a practical solution, we look towards label smoothing, which has been shown as an effective method to regularize for overconfident predictions but has yet to be widely adopted in the supervised fine-tuning (SFT) of LLMs. We first provide insight as to why label smoothing is sufficient to maintain calibration throughout the SFT process. However, settings remain where the effectiveness of smoothing is severely diminished, in particular the case of large vocabulary LLMs (LV-LLMs). We posit the cause to stem from the ability to become over-confident, which has a direct relationship with the hidden size and vocabulary size, and justify this theoretically and experimentally. Finally, we address an outstanding issue regarding the memory footprint of the cross-entropy loss computation in the label smoothed loss setting, designing a customized kernel to dramatically reduce memory consumption without sacrificing speed or performance in comparison to existing solutions for non-smoothed losses.

Stochasticity in language model fine-tuning, often caused by the small batch sizes typically used in this regime, can destabilize training by introducing large oscillations in generation quality. A popular approach to mitigating this instability is to take an Exponential moving average (EMA) of weights throughout training. While EMA reduces stochasticity, thereby smoothing training, the introduction of bias from old iterates often creates a lag in optimization relative to vanilla training. In this work, we propose the Bias-Corrected Exponential Moving Average (BEMA), a simple and practical augmentation of EMA that retains variance-reduction benefits while eliminating bias. BEMA is motivated by a simple theoretical model wherein we demonstrate provable acceleration of BEMA over both a standard EMA and vanilla training. Through an extensive suite of experiments on Language Models, we show that BEMA leads to significantly improved convergence rates and final performance over both EMA and vanilla training in a variety of standard LM benchmarks, making BEMA a practical and theoretically motivated intervention for more stable and efficient fine-tuning.

Existing AIG (AI-generated) text detectors struggle in real-world settings despite succeeding in internal testing, suggesting that they may not be robust enough. We rigorously examine the machine-learning procedure to build these detectors to address this. Most current AIG text detection datasets focus on zero-shot generations, but little work has been done on few-shot or one-shot generations, where LLMs are given human texts as an example. In response, we introduce the Diverse Adversarial Corpus of Texts Yielded from Language models (DACTYL), a challenging AIG text detection dataset focusing on one-shot/few-shot generations. We also include texts from domain-specific continued-pre-trained (CPT) language models, where we fully train all parameters using a memory-efficient optimization approach. Many existing AIG text detectors struggle significantly on our dataset, indicating a potential vulnerability to one-shot/few-shot and CPT-generated texts. We also train our own classifiers using two approaches: standard binary cross-entropy (BCE) optimization and a more recent approach, deep X-risk optimization (DXO). While BCE-trained classifiers marginally outperform DXO classifiers on the DACTYL test set, the latter excels on out-of-distribution (OOD) texts. In our mock deployment scenario in student essay detection with an OOD student essay dataset, the best DXO classifier outscored the best BCE-trained classifier by 50.56 macro-F1 score points at the lowest false positive rates for both. Our results indicate that DXO classifiers generalize better without overfitting to the test set. Our experiments highlight several areas of improvement for AIG text detectors.

Large language models (LLMs) are increasingly deployed in sensitive contexts where fairness and inclusivity are critical. Pronoun usage, especially concerning gender-neutral and neopronouns, remains a key challenge for responsible AI. Prior work, such as the MISGENDERED benchmark, revealed significant limitations in earlier LLMs' handling of inclusive pronouns, but was constrained to outdated models and limited evaluations. In this study, we introduce MISGENDERED+, an extended and updated benchmark for evaluating LLMs' pronoun fidelity. We benchmark five representative LLMs, GPT-4o, Claude 4, DeepSeek-V3, Qwen Turbo, and Qwen2.5, across zero-shot, few-shot, and gender identity inference. Our results show notable improvements compared with previous studies, especially in binary and gender-neutral pronoun accuracy. However, accuracy on neopronouns and reverse inference tasks remains inconsistent, underscoring persistent gaps in identity-sensitive reasoning. We discuss implications, model-specific observations, and avenues for future inclusive AI research.

This paper presents our system for SemEval-2025 Task 8: DataBench, Question-Answering over Tabular Data. The primary objective of this task is to perform question answering on given tabular datasets from diverse domains under two subtasks: DataBench QA (Subtask I) and DataBench Lite QA (Subtask II). To tackle both subtasks, we developed a zero-shot solution with a particular emphasis on leveraging Large Language Model (LLM)-based code generation. Specifically, we propose a Python code generation framework utilizing state-of-the-art open-source LLMs to generate executable Pandas code via optimized prompting strategies. Our experiments reveal that different LLMs exhibit varying levels of effectiveness in Python code generation. Additionally, results show that Python code generation achieves superior performance in tabular question answering compared to alternative approaches. Although our ranking among zero-shot systems is unknown at the time of this paper's submission, our system achieved eighth place in Subtask I and sixth place in Subtask~II among the 30 systems that outperformed the baseline in the open-source models category.

Hate speech detection on Chinese social networks presents distinct challenges, particularly due to the widespread use of cloaking techniques designed to evade conventional text-based detection systems. Although large language models (LLMs) have recently improved hate speech detection capabilities, the majority of existing work has concentrated on English datasets, with limited attention given to multimodal strategies in the Chinese context. In this study, we propose MMBERT, a novel BERT -based multimodal framework that integrates textual, speech, and visual modalities through a Mixture-of-Experts (MoE) architecture. To address the instability associated with directly integrating MoE into BERT - based models, we develop a progressive three-stage training paradigm. MMBERT incorporates modality-specific experts, a shared self-attention mechanism, and a router-based expert allocation strategy to enhance robustness against adversarial perturbations. Empirical results in several Chinese hate speech datasets show that MMBERT significantly surpasses fine-tuned BERT -based encoder models, fine-tuned LLMs, and LLMs utilizing in-context learning approaches.

Large language models (LLMs) are trained on large corpora, yet it is unclear whether they can reason about the information present within their training data. We design experiments to study out-of-context abduction in LLMs, the ability to infer the most plausible explanations for observations using relevant facts present in training data. We train treatment LLMs on names and behavior descriptions of fictitious chatbots, but not on examples of dialogue with the chatbots. We find that OpenAI's GPT 4o LLM can correctly infer at least one chatbot's name after observing example responses characteristic of that chatbot. We also find that previously training GPT 4o on descriptions of a chatbot's behavior allows it to display behaviors more characteristic of the chatbot when iteratively trained to display such behaviors. Our results have implications for situational awareness in LLMs and, therefore, for AI safety.

Style change detection - identifying the points in a document where writing style shifts - remains one of the most important and challenging problems in computational authorship analysis. At PAN 2025, the shared task challenges participants to detect style switches at the most fine-grained level: individual sentences. The task spans three datasets, each designed with controlled and increasing thematic variety within documents. We propose to address this problem by modeling the content of each problem instance - that is, a series of sentences - as a whole, using a Sequential Sentence Pair Classifier (SSPC). The architecture leverages a pre-trained language model (PLM) to obtain representations of individual sentences, which are then fed into a bidirectional LSTM (BiLSTM) to contextualize them within the document. The BiLSTM-produced vectors of adjacent sentences are concatenated and passed to a multi-layer perceptron for prediction per adjacency. Building on the work of previous PAN participants classical text segmentation, the approach is relatively conservative and lightweight. Nevertheless, it proves effective in leveraging contextual information and addressing what is arguably the most challenging aspect of this year's shared task: the notorious problem of "stylistically shallow", short sentences that are prevalent in the proposed benchmark data. Evaluated on the official PAN-2025 test datasets, the model achieves strong macro-F1 scores of 0.923, 0.828, and 0.724 on the EASY, MEDIUM, and HARD data, respectively, outperforming not only the official random baselines but also a much more challenging one: claude-3.7-sonnet's zero-shot performance.

As large language models (LLMs) become increasingly embedded in our daily lives, evaluating their quality and reliability across diverse contexts has become essential. While comprehensive benchmarks exist for assessing LLM performance in English, there remains a significant gap in evaluation resources for other languages. Moreover, because most LLMs are trained primarily on data rooted in European and American cultures, they often lack familiarity with non-Western cultural contexts. To address this limitation, our study focuses on the Persian language and Iranian culture. We introduce 19 new evaluation datasets specifically designed to assess LLMs on topics such as Iranian law, Persian grammar, Persian idioms, and university entrance exams. Using these datasets, we benchmarked 41 prominent LLMs, aiming to bridge the existing cultural and linguistic evaluation gap in the field.