Comprehensive evaluations of language models (LM) during both development and deployment phases are necessary because these models possess numerous capabilities (e.g., mathematical reasoning, legal support, or medical diagnostic) as well as safety risks (e.g., racial bias, toxicity, or misinformation). The average score across a wide range of benchmarks provides a signal that helps guide the use of these LMs in practice. Currently, holistic evaluations are costly due to the large volume of benchmark questions, making frequent evaluations impractical. A popular attempt to lower the cost is to compute the average score on a subset of the benchmark. This approach, unfortunately, often renders an unreliable measure of LM performance because the average score is often confounded with the difficulty of the questions in the benchmark subset. Item response theory (IRT) was designed to address this challenge, providing a reliable measurement by careful controlling for question difficulty. Unfortunately, question difficulty is expensive to estimate. Facing this challenge, we train a model that predicts question difficulty from its content, enabling a reliable measurement at a fraction of the cost. In addition, we leverage this difficulty predictor to further improve the evaluation efficiency through training a question generator given a difficulty level. This question generator is essential in adaptive testing, where, instead of using a random subset of the benchmark questions, informative questions are adaptively chosen based on the current estimation of LLM performance. Experiments on 22 common natural language benchmarks and 172 LMs show that this approach is more reliable and efficient compared to current common practice.

Efficient deployment of deep neural networks on resource-constrained devices demands advanced compression techniques that preserve accuracy and interoperability. This paper proposes a machine learning framework that augments Knowledge Distillation (KD) with Integrated Gradients (IG), an attribution method, to optimise the compression of convolutional neural networks. We introduce a novel data augmentation strategy where IG maps, precomputed from a teacher model, are overlaid onto training images to guide a compact student model toward critical feature representations. This approach leverages the teacher's decision-making insights, enhancing the student's ability to replicate complex patterns with reduced parameters. Experiments on CIFAR-10 demonstrate the efficacy of our method: a student model, compressed 4.1-fold from the MobileNet-V2 teacher, achieves 92.5% classification accuracy, surpassing the baseline student's 91.4% and traditional KD approaches, while reducing inference latency from 140 ms to 13 ms--a tenfold speedup. We perform hyperparameter optimisation for efficient learning. Comprehensive ablation studies dissect the contributions of KD and IG, revealing synergistic effects that boost both performance and model explainability. Our method's emphasis on feature-level guidance via IG distinguishes it from conventional KD, offering a data-driven solution for mining transferable knowledge in neural architectures. This work contributes to machine learning by providing a scalable, interpretable compression technique, ideal for edge computing applications where efficiency and transparency are paramount.

We study Multimodal Large Language Models (MLLMs) with in-context learning for food preparation task planning. In this context, we identify two key challenges: cross-modal distraction and geometric feasibility. Cross-modal distraction occurs when the inclusion of visual input degrades the reasoning performance of a MLLM. Geometric feasibility refers to the ability of MLLMs to ensure that the selected skills are physically executable in the environment. To address these issues, we adapt Chain of Thought (CoT) with Self-Consistency to mitigate reasoning loss from cross-modal distractions and use affordance predictor as skill preconditions to guide MLLM on geometric feasibility. We construct a dataset to evaluate the ability of MLLMs on quantity estimation, reachability analysis, relative positioning and collision avoidance. We conducted a detailed evaluation to identify issues among different baselines and analyze the reasons for improvement, providing insights into each approach. Our method reaches a success rate of 76.7% on the entire dataset, showing a substantial improvement over the CoT baseline at 36.7%.

Large language models (LLMs) have demonstrated remarkable capabilities in various natural language processing tasks. However, achieving strong performance in specialized domains like mathematical reasoning and non-English languages often requires extensive training on massive datasets. This paper investigates a contrasting approach: strategic fine-tuning on a small, high-quality, bilingual (English-French) dataset to enhance both the reasoning capabilities and French language proficiency of a large language model. Rather than relying on scale, we explore the hypothesis that targeted data curation and optimized training can achieve competitive, or even superior, performance. We demonstrate, through targeted supervised fine-tuning (SFT) on only 2,000 carefully selected samples, significant improvements in mathematical reasoning. Specifically, Pensez 7B exhibits an increase in accuracy of the base model up to 20% on the AIME25 and a 12% increase on a French MATH level 5 benchmark. These results challenge the prevailing assumption that massive datasets are aprerequisite for strong reasoning performance in LLMs, highlighting the potential of strategic data curation and optimized fine-tuning for enhancing both specialized skills and multilingual capabilities. Our findings have implications for the efficient development of high-performing, multilingual LLMs, especially in resource-constrained scenarios.

Large language models (LLMs) exhibit strong reasoning abilities, often attributed to few-shot or zero-shot chain-of-thought (CoT) prompting. While effective, these methods require labor-intensive prompt engineering, raising the question of whether reasoning can be induced without reliance on explicit prompts. In this work, we unlock the reasoning capabilities of LLMs without explicit prompting. Inspired by zero-shot CoT and CoT-decoding, we propose a novel decoding strategy that systematically nudges LLMs to continue reasoning, thereby preventing immature reasoning processes. Specifically, we monitor the model's generation and inject a designated phrase whenever it is likely to conclude its response prematurely, before completing the reasoning process. Our experimental evaluations on diverse reasoning benchmarks demonstrate that our proposed strategy substantially improves LLM reasoning capabilities, highlighting the potential of decoding-based interventions as an alternative to traditional prompting techniques.

The use of large language models (LLMs) to simulate human behavior has gained significant attention, particularly through personas that approximate individual characteristics. Persona-based simulations hold promise for transforming disciplines that rely on population-level feedback, including social science, economic analysis, marketing research, and business operations. Traditional methods to collect realistic persona data face significant challenges. They are prohibitively expensive and logistically challenging due to privacy constraints, and often fail to capture multi-dimensional attributes, particularly subjective qualities. Consequently, synthetic persona generation with LLMs offers a scalable, cost-effective alternative. However, current approaches rely on ad hoc and heuristic generation techniques that do not guarantee methodological rigor or simulation precision, resulting in systematic biases in downstream tasks. Through extensive large-scale experiments including presidential election forecasts and general opinion surveys of the U.S. population, we reveal that these biases can lead to significant deviations from real-world outcomes. Our findings underscore the need to develop a rigorous science of persona generation and outline the methodological innovations, organizational and institutional support, and empirical foundations required to enhance the reliability and scalability of LLM-driven persona simulations. To support further research and development in this area, we have open-sourced approximately one million generated personas, available for public access and analysis at https://huggingface.co/datasets/Tianyi-Lab/Personas.

Generating images with embedded text is crucial for the automatic production of visual and multimodal documents, such as educational materials and advertisements. However, existing diffusion-based text-to-image models often struggle to accurately embed text within images, facing challenges in spelling accuracy, contextual relevance, and visual coherence. Evaluating the ability of such models to embed text within a generated image is complicated due to the lack of comprehensive benchmarks. In this work, we introduce TextInVision, a large-scale, text and prompt complexity driven benchmark designed to evaluate the ability of diffusion models to effectively integrate visual text into images. We crafted a diverse set of prompts and texts that consider various attributes and text characteristics. Additionally, we prepared an image dataset to test Variational Autoencoder (VAE) models across different character representations, highlighting that VAE architectures can also pose challenges in text generation within diffusion frameworks. Through extensive analysis of multiple models, we identify common errors and highlight issues such as spelling inaccuracies and contextual mismatches. By pinpointing the failure points across different prompts and texts, our research lays the foundation for future advancements in AI-generated multimodal content.

Large language models (LLMs) have revolutionized code generation, automating programming with remarkable efficiency. However, these advancements challenge programming skills, ethics, and assessment integrity, making the detection of LLM-generated code essential for maintaining accountability and standards. While, there has been some research on this problem, it generally lacks domain coverage and robustness, and only covers a small number of programming languages. To this end, we propose a framework capable of distinguishing between human- and LLM-written code across multiple programming languages, code generators, and domains. We use a large-scale dataset from renowned platforms and LLM-based code generators, alongside applying rigorous data quality checks, feature engineering, and comparative analysis using evaluation of traditional machine learning models, pre-trained language models (PLMs), and LLMs for code detection. We perform an evaluation on out-of-domain scenarios, such as detecting the authorship and hybrid authorship of generated code and generalizing to unseen models, domains, and programming languages. Moreover, our extensive experiments show that our framework effectively distinguishes human- from LLM-written code and sets a new benchmark for this task.

Transfer Learning (TL) is a powerful tool that enables robots to transfer learned policies across different environments, tasks, or embodiments. To further facilitate this process, efforts have been made to combine it with Learning from Demonstrations (LfD) for more flexible and efficient policy transfer. However, these approaches are almost exclusively limited to offline demonstrations collected before policy transfer starts, which may suffer from the intrinsic issue of covariance shift brought by LfD and harm the performance of policy transfer. Meanwhile, extensive work in the learning-from-scratch setting has shown that online demonstrations can effectively alleviate covariance shift and lead to better policy performance with improved sample efficiency. This work combines these insights to introduce online demonstrations into a policy transfer setting. We present Policy Transfer with Online Demonstrations, an active LfD algorithm for policy transfer that can optimize the timing and content of queries for online episodic expert demonstrations under a limited demonstration budget. We evaluate our method in eight robotic scenarios, involving policy transfer across diverse environment characteristics, task objectives, and robotic embodiments, with the aim to transfer a trained policy from a source task to a related but different target task. The results show that our method significantly outperforms all baselines in terms of average success rate and sample efficiency, compared to two canonical LfD methods with offline demonstrations and one active LfD method with online demonstrations. Additionally, we conduct preliminary sim-to-real tests of the transferred policy on three transfer scenarios in the real-world environment, demonstrating the policy effectiveness on a real robot manipulator.

Lifelong Reinforcement Learning (LRL) holds significant potential for addressing sequential tasks, but it still faces considerable challenges. A key difficulty lies in effectively preventing catastrophic forgetting and facilitating knowledge transfer while maintaining reliable decision-making performance across subsequent tasks in dynamic environments. To tackle this, we propose a novel framework, SDW (Similarity-Driven Weighting Framework), which leverages large-language-model-generated dynamic functions to precisely control the training process. The core of SDW lies in two functions pre-generated by large models: the task similarity function and the weight computation function. The task similarity function extracts multidimensional features from task descriptions to quantify the similarities and differences between tasks in terms of states, actions, and rewards. The weight computation function dynamically generates critical training parameters based on the similarity information, including the proportion of old task data stored in the Replay Buffer and the strategy consistency weight in the loss function, enabling an adaptive balance between learning new tasks and transferring knowledge from previous tasks. By generating function code offline prior to training, rather than relying on large-model inference during the training process, the SDW framework reduces computational overhead while maintaining efficiency in sequential task scenarios. Experimental results on Atari and MiniHack sequential tasks demonstrate that SDW significantly outperforms existing lifelong reinforcement learning methods.