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.

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.

Artificial intelligence - driven adaptive learning systems are reshaping education through data - driven adaptation of learning experiences. Yet many of these systems lack transparency, offering limited insight into how decisions are made. Most explainable AI (XAI) techniques focus on technical outputs but neglect user roles and comprehension. This paper proposes a hybrid framework that integrates traditional XAI techniques with generative AI models and u ser personalisation to generate multimodal, personalised explanations tailored to user needs. We redefine explainability as a dynamic communication process tailored to user roles and learning goals. We outline the framework ' s design, key XAI limitations in education, and research directions on accuracy, fairness, and personalisation. Our aim is to move towards explainable AI that enhances transparency while supporting user - centred experiences.

Code retrieval aims to provide users with desired code snippets based on users' natural language queries. With the development of deep learning technologies, adopting pre-trained models for this task has become mainstream. Considering the retrieval efficiency, most of the previous approaches adopt a dual-encoder for this task, which encodes the description and code snippet into representation vectors, respectively. However, the model structure of the dual-encoder tends to limit the model's performance, since it lacks the interaction between the code snippet and description at the bottom layer of the model during training. To improve the model's effectiveness while preserving its efficiency, we propose a framework, which adopts Self-AdaPtive Model Distillation for Efficient CodE Retrieval, named SPENCER. SPENCER first adopts the dual-encoder to narrow the search space and then adopts the cross-encoder to improve accuracy. To improve the efficiency of SPENCER, we propose a novel model distillation technique, which can greatly reduce the inference time of the dual-encoder while maintaining the overall performance. We also propose a teaching assistant selection strategy for our model distillation, which can adaptively select the suitable teaching assistant models for different pre-trained models during the model distillation to ensure the model performance. Extensive experiments demonstrate that the combination of dual-encoder and cross-encoder improves overall performance compared to solely dual-encoder-based models for code retrieval. Besides, our model distillation technique retains over 98% of the overall performance while reducing the inference time of the dual-encoder by 70%.

Large Language Models (LLMs) have shown remarkable abilities in structured reasoning and symbolic tasks, with coding emerging as a particular area of strength. This success has sparked growing interest in applying LLMs to mathematics, both in informal problem-solving and formal theorem proving. However, progress in formal mathematics has proven to be significantly more difficult, despite surface-level similarities between programming and proof construction. This discrepancy raises important questions about how LLMs ``reason'', how they are supervised, and whether they internally track a notion of computational or deductive state. In this article, we address the state-of-the-art of the discipline, focusing on recent models and benchmarks, and explore three central issues at the intersection of machine learning and mathematical cognition: (i) the trade-offs between formal and informal mathematics as training domains; (ii) the deeper reasons why proof generation remains more brittle than code synthesis; (iii) and the question of whether LLMs represent, or merely mimic, a notion of evolving logical state. Our goal is not to draw hard boundaries, but to identify where the current limits lie, and how they might be extended.

Prompt tuning (PT), as an emerging resource-efficient fine-tuning paradigm, has showcased remarkable effectiveness in improving the task-specific transferability of vision-language models. This paper delves into a previously overlooked information asymmetry issue in PT, where the visual modality mostly conveys more context than the object-oriented textual modality. Correspondingly, coarsely aligning these two modalities could result in the biased attention, driving the model to merely focus on the context area. To address this, we propose DAPT, an effective PT framework based on an intuitive decouple-before-align concept. First, we propose to explicitly decouple the visual modality into the foreground and background representation via exploiting coarse-and-fine visual segmenting cues, and then both of these decoupled patterns are aligned with the original foreground texts and the hand-crafted background classes, thereby symmetrically strengthening the modal alignment. To further enhance the visual concentration, we propose a visual pull-push regularization tailored for the foreground-background patterns, directing the original visual representation towards unbiased attention on the region-of-interest object. We demonstrate the power of architecture-free DAPT through few-shot learning, base-to-novel generalization, and data-efficient learning, all of which yield superior performance across prevailing benchmarks. Our code will be released at https://github.com/Ferenas/DAPT.

Evaluating the performance of visual language models (VLMs) in graphic reasoning tasks has become an important research topic. However, VLMs still show obvious deficiencies in simulating human-level graphic reasoning capabilities, especially in complex graphic reasoning and abstract problem solving, which are less studied and existing studies only focus on simple graphics. To evaluate the performance of VLMs in complex graphic reasoning, we propose ReasonBench, the first evaluation benchmark focused on structured graphic reasoning tasks, which includes 1,613 questions from real-world intelligence tests. ReasonBench covers reasoning dimensions related to location, attribute, quantity, and multi-element tasks, providing a comprehensive evaluation of the performance of VLMs in spatial, relational, and abstract reasoning capabilities. We benchmark 11 mainstream VLMs (including closed-source and open-source models) and reveal significant limitations of current models. Based on these findings, we propose a dual optimization strategy: Diagrammatic Reasoning Chain (DiaCoT) enhances the interpretability of reasoning by decomposing layers, and ReasonTune enhances the task adaptability of model reasoning through training, all of which improves VLM performance by 33.5\%. All experimental data and code are in the repository: https://huggingface.co/datasets/cistine/ReasonBench.

We present an online tutoring system that learns to provide effective feedback to students after they answer questions incorrectly. Using data from one million students, the system learns which assistance action (e.g., one of multiple hints) to provide for each question to optimize student learning. Employing the multi-armed bandit (MAB) framework and offline policy evaluation, we assess 43,000 assistance actions, and identify trade-offs between assistance policies optimized for different student outcomes (e.g., response correctness, session completion). We design an algorithm that for each question decides on a suitable policy training objective to enhance students' immediate second attempt success and overall practice session performance. We evaluate the resulting MAB policies in 166,000 practice sessions, verifying significant improvements in student outcomes. While MAB policies optimize feedback for the overall student population, we further investigate whether contextual bandit (CB) policies can enhance outcomes by personalizing feedback based on individual student features (e.g., ability estimates, response times). Using causal inference, we examine (i) how effects of assistance actions vary across students and (ii) whether CB policies, which leverage such effect heterogeneity, outperform MAB policies. While our analysis reveals that some actions for some questions exhibit effect heterogeneity, effect sizes may often be too small for CB policies to provide significant improvements beyond what well-optimized MAB policies that deliver the same action to all students already achieve. We discuss insights gained from deploying data-driven systems at scale and implications for future refinements. Today, the teaching policies optimized by our system support thousands of students daily.