A central question in multilingual language modeling is whether large language models (LLMs) develop a universal concept representation, disentangled from specific languages. In this paper, we address this question by analyzing latent representations (latents) during a word translation task in transformer-based LLMs. We strategically extract latents from a source translation prompt and insert them into the forward pass on a target translation prompt. By doing so, we find that the output language is encoded in the latent at an earlier layer than the concept to be translated. Building on this insight, we conduct two key experiments. First, we demonstrate that we can change the concept without changing the language and vice versa through activation patching alone. Second, we show that patching with the mean over latents across different languages does not impair and instead improves the models' performance in translating the concept. Our results provide evidence for the existence of language-agnostic concept representations within the investigated models.

Grading programming assignments is crucial for guiding students to improve their programming skills and coding styles. This study presents an automated grading framework, CodEv, which leverages Large Language Models (LLMs) to provide consistent and constructive feedback. We incorporate Chain of Thought (CoT) prompting techniques to enhance the reasoning capabilities of LLMs and ensure that the grading is aligned with human evaluation. Our framework also integrates LLM ensembles to improve the accuracy and consistency of scores, along with agreement tests to deliver reliable feedback and code review comments. The results demonstrate that the framework can yield grading results comparable to human evaluators, by using smaller LLMs. Evaluation and consistency tests of the LLMs further validate our approach, confirming the reliability of the generated scores and feedback.

This paper explores ideas and provides a potential roadmap for the development and evaluation of physics-specific large-scale AI models, which we call Large Physics Models (LPMs). These models, based on foundation models such as Large Language Models (LLMs) - trained on broad data - are tailored to address the demands of physics research. LPMs can function independently or as part of an integrated framework. This framework can incorporate specialized tools, including symbolic reasoning modules for mathematical manipulations, frameworks to analyse specific experimental and simulated data, and mechanisms for synthesizing theories and scientific literature. We begin by examining whether the physics community should actively develop and refine dedicated models, rather than relying solely on commercial LLMs. We then outline how LPMs can be realized through interdisciplinary collaboration among experts in physics, computer science, and philosophy of science. To integrate these models effectively, we identify three key pillars: Development, Evaluation, and Philosophical Reflection. Development focuses on constructing models capable of processing physics texts, mathematical formulations, and diverse physical data. Evaluation assesses accuracy and reliability by testing and benchmarking. Finally, Philosophical Reflection encompasses the analysis of broader implications of LLMs in physics, including their potential to generate new scientific understanding and what novel collaboration dynamics might arise in research. Inspired by the organizational structure of experimental collaborations in particle physics, we propose a similarly interdisciplinary and collaborative approach to building and refining Large Physics Models. This roadmap provides specific objectives, defines pathways to achieve them, and identifies challenges that must be addressed to realise physics-specific large scale AI models.

This paper introduce LongViTU, a large-scale (~121k QA pairs, ~900h videos), automatically generated dataset for long-form video understanding. We developed a systematic approach that organizes videos into a hierarchical tree structure and incorporates self-revision mechanisms to ensure high-quality QA pairs. Each QA pair in LongViTU features: 1) long-term context (average certificate length of 4.6 minutes); 2) rich knowledge and condensed reasoning (commonsense, causality, planning, etc.); and 3) explicit timestamp labels for relevant events. LongViTU also serves as a benchmark for instruction following in long-form and streaming video understanding. We evaluate the open-source state-of-the-art long video understanding model, LongVU, and the commercial model, Gemini-1.5-Pro, on our benchmark. They achieve GPT-4 scores of 49.9 and 52.3, respectively, underscoring the substantial challenge posed by our benchmark. Further supervised fine-tuning (SFT) on LongVU led to performance improvements of 12.0% on our benchmark, 2.2% on the in-distribution (ID) benchmark EgoSchema, 1.0%, 2.2% and 1.2% on the out-of-distribution (OOD) benchmarks VideoMME (Long), WorldQA and OpenEQA, respectively. These outcomes demonstrate LongViTU's high data quality and robust OOD generalizability.

In continual learning, where task data arrives in a sequence, fine-tuning on later tasks will often lead to performance degradation on earlier tasks. This is especially pronounced when these tasks come from diverse domains. In this setting, how can we mitigate catastrophic forgetting of earlier tasks and retain what the model has learned with minimal computational expenses? Inspired by other merging methods, and L2-regression, we propose Sequential Fine-tuning with Averaging (SFA), a method that merges currently training models with earlier checkpoints during the course of training. SOTA approaches typically maintain a data buffer of past tasks or impose a penalty at each gradient step. In contrast, our method achieves comparable results without the need to store past data, or multiple copies of parameters for each gradient step. Furthermore, our method outperforms common merging techniques such as Task Arithmetic, TIES Merging, and WiSE-FT, as well as other penalty methods like L2 and Elastic Weight Consolidation. In turn, our method offers insight into the benefits of merging partially-trained models during training across both image and language domains.

In a paper published in 1951, Shannon, having just published the foundational papers of information theory, proposed a "guessing game" of next word prediction to estimate the entropy of English (Shannon, 1951). Nearly 70 years later, training a high-capacity transformer network (Vaswani et al., 2017) on this task, provided the generative pre-training backbone for Large Language Models (Radford et al., 2018; Devlin et al., 2019; Radford et al., 2019; Brown et al., 2020). Less well known is the fact that in 1954, Fred Attneave (Attneave, 1954) proposed an analog of Shannon's task for images. To quote "We may divide the picture into arbitrarily small elements which we "transmit" to a subject (S) in a cumulative sequence, having them guess at the color of each successive element until they are correct. This method of analysis resembles the scanning process used in television and facsimile systems and accomplishes the like purpose of transforming two spatial dimensions into a single sequence in time".

Structured image understanding, such as interpreting tables and charts, requires strategically refocusing across various structures and texts within an image, forming a reasoning sequence to arrive at the final answer. However, current multimodal large language models (LLMs) lack this multihop selective attention capability. In this work, we introduce ReFocus, a simple yet effective framework that equips multimodal LLMs with the ability to generate "visual thoughts" by performing visual editing on the input image through code, shifting and refining their visual focuses. Specifically, ReFocus enables multimodal LLMs to generate Python codes to call tools and modify the input image, sequentially drawing boxes, highlighting sections, and masking out areas, thereby enhancing the visual reasoning process. We experiment upon a wide range of structured image understanding tasks involving tables and charts. ReFocus largely improves performance on all tasks over GPT-4o without visual editing, yielding an average gain of 11.0% on table tasks and 6.8% on chart tasks. We present an in-depth analysis of the effects of different visual edits, and reasons why ReFocus can improve the performance without introducing additional information. Further, we collect a 14k training set using ReFocus, and prove that such visual chain-of-thought with intermediate information offers a better supervision than standard VQA data, reaching a 8.0% average gain over the same model trained with QA pairs and 2.6% over CoT.

The success of social media platforms has facilitated the emergence of various forms of online abuse within digital communities. This abuse manifests in multiple ways, including hate speech, cyberbullying, emotional abuse, grooming, and sexting. In this paper, we present a comprehensive analysis of the different forms of abuse prevalent in social media, with a particular focus on how emerging technologies, such as Language Models (LMs) and Large Language Models (LLMs), are reshaping both the detection and generation of abusive content within these networks. We delve into the mechanisms through which social media abuse is perpetuated, exploring the psychological and social impact. Additionally, we examine the dual role of advanced language models-highlighting their potential to enhance automated detection systems for abusive behavior while also acknowledging their capacity to generate harmful content. This paper aims to contribute to the ongoing discourse on online safety and ethics, offering insights into the evolving landscape of cyberabuse and the technological innovations that both mitigate and exacerbate it.

Studying public sentiment during crises is crucial for understanding how opinions and sentiments shift, resulting in polarized societies. We study Weibo, the most popular microblogging site in China, using posts made during the outbreak of the COVID-19 crisis. The study period includes the pre-COVID-19 stage, the outbreak stage, and the early stage of epidemic prevention. We use Llama 3 8B, a Large Language Model, to analyze users' sentiments on the platform by classifying them into positive, negative, sarcastic, and neutral categories. Analyzing sentiment shifts on Weibo provides insights into how social events and government actions influence public opinion. This study contributes to understanding the dynamics of social sentiments during health crises, fulfilling a gap in sentiment analysis for Chinese platforms. By examining these dynamics, we aim to offer valuable perspectives on digital communication's role in shaping society's responses during unprecedented global challenges.

Existing benchmarks for evaluating long-context language models (LCLMs) primarily focus on long-context recall, requiring models to produce short responses based on a few critical snippets while processing thousands of irrelevant tokens. We introduce LongProc (Long Procedural Generation), a new benchmark that requires both the integration of highly dispersed information and long-form generation. LongProc consists of six diverse procedural generation tasks, such as extracting structured information from HTML pages into a TSV format and executing complex search procedures to create travel plans. These tasks challenge LCLMs by testing their ability to follow detailed procedural instructions, synthesize and reason over dispersed information, and generate structured, long-form outputs (up to 8K tokens). Furthermore, as these tasks adhere to deterministic procedures and yield structured outputs, they enable reliable rule-based evaluation. We evaluate 17 LCLMs on LongProc across three difficulty levels, with maximum numbers of output tokens set at 500, 2K, and 8K. Notably, while all tested models claim a context window size above 32K tokens, open-weight models typically falter on 2K-token tasks, and closed-source models like GPT-4o show significant degradation on 8K-token tasks. Further analysis reveals that LCLMs struggle to maintain long-range coherence in long-form generations. These findings highlight critical limitations in current LCLMs and suggest substantial room for improvement. Data and code available at: https://princeton-pli.github.io/LongProc