Deep networks are frequently tuned to novel tasks and continue learning from ongoing data streams. Such sequential training requires consolidation of new and past information, a challenge predominantly addressed by retaining the most important data points - formally known as coresets. Traditionally, these coresets consist of entire samples, such as images or sentences. However, recent transformer architectures operate on tokens, leading to the famous assertion that an image is worth 16x16 words. Intuitively, not all of these tokens are equally informative or memorable. Going beyond coresets, we thus propose to construct a deeper-level data summary on the level of tokens. Our respectively named core tokensets both select the most informative data points and leverage feature attribution to store only their most relevant features. We demonstrate that core tokensets yield significant performance retention in incremental image classification, open-ended visual question answering, and continual image captioning with significantly reduced memory. In fact, we empirically find that a core tokenset of 1% of the data performs comparably to at least a twice as large and up to 10 times larger coreset. Deep learning models are rarely deployed in static environments and instead need to be continuously adjusted to their ever-changing surroundings. As distribution shifts occur over time, we require models to retain their performance on previous settings while accommodating new examples (Hadsell et al., 2020; Kudithipudi et al., 2022; Mundt et al., 2023). Naively, re-training a model from scratch for every change in the environment quickly becomes unfeasible for complex architectures with billions of parameters (Touvron et al., 2023; Brown et al., 2020). Consequently, research has focused on identifying small, representative subsets of the training data (Lopez-Paz & Ranzato, 2017), which can later be replayed continuously to avoid (catastrophic) forgetting (McCloskey & Cohen, 1989) of previously acquired information (Graham et al., 2021; Hassani et al., 2021; Xu et al., 2021).

This paper proposes a novel data-driven control strategy for maintaining connectivity in networked multi-robot systems. Existing approaches often rely on a pre-determined communication model specifying whether pairwise robots can communicate given their relative distance to guide the connectivity-aware control design, which may not capture real-world communication conditions. To relax that assumption, we present the concept of Data-driven Connectivity Barrier Certificates, which utilize Control Barrier Functions (CBF) and Gaussian Processes (GP) to characterize the admissible control space for pairwise robots based on communication performance observed online. This allows robots to maintain a satisfying level of pairwise communication quality (measured by the received signal strength) while in motion. Then we propose a Data-driven Connectivity Maintenance (DCM) algorithm that combines (1) online learning of the communication signal strength and (2) a bi-level optimization-based control framework for the robot team to enforce global connectivity of the realistic multi-robot communication graph and minimally deviate from their task-related motions. We provide theoretical proofs to justify the properties of our algorithm and demonstrate its effectiveness through simulations with up to 20 robots.

Embodied AI represents a paradigm in AI research where artificial agents are situated within and interact with physical or virtual environments. Despite the recent progress in Embodied AI, it is still very challenging to learn the generalizable manipulation skills that can handle large deformation and topological changes on soft-body objects, such as clay, water, and soil. In this work, we proposed an effective policy, namely GP2E behavior cloning policy, which can guide the agent to learn the generalizable manipulation skills from soft-body tasks, including pouring, filling, hanging, excavating, pinching, and writing. Concretely, we build our policy from three insights:(1) Extracting intricate semantic features from point cloud data and seamlessly integrating them into the robot's end-effector frame; (2) Capturing long-distance interactions in long-horizon tasks through the incorporation of our guided self-attention module; (3) Mitigating overfitting concerns and facilitating model convergence to higher accuracy levels via the introduction of our two-stage fine-tuning strategy. Through extensive experiments, we demonstrate the effectiveness of our approach by achieving the 1st prize in the soft-body track of the ManiSkill2 Challenge at the CVPR 2023 4th Embodied AI workshop. Our findings highlight the potential of our method to improve the generalization abilities of Embodied AI models and pave the way for their practical applications in real-world scenarios.

In the rapidly evolving field of deep learning, specialized models have driven significant advancements in tasks such as computer vision and natural language processing. However, this specialization leads to a fragmented ecosystem where models lack the adaptability for broader applications. To overcome this, we introduce AutoFusion, an innovative framework that fuses distinct model parameters(with the same architecture) for multi-task learning without pre-trained checkpoints. Using an unsupervised, end-to-end approach, AutoFusion dynamically permutes model parameters at each layer, optimizing the combination through a loss-minimization process that does not require labeled data. We validate AutoFusion's effectiveness through experiments on commonly used benchmark datasets, demonstrating superior performance over established methods like Weight Interpolation, Git Re-Basin, and ZipIt. Our framework offers a scalable and flexible solution for model integration, positioning it as a powerful tool for future research and practical applications.

The rapid advancement of large language models (LLMs) has highlighted the need for robust evaluation frameworks that assess their core capabilities, such as reasoning, knowledge, and commonsense, leading to the inception of certain widely-used benchmark suites such as the H6 benchmark. However, these benchmark suites are primarily built for the English language, and there exists a lack thereof for under-represented languages, in terms of LLM development, such as Thai. On the other hand, developing LLMs for Thai should also include enhancing the cultural understanding as well as core capabilities. To address these dual challenge in Thai LLM research, we propose two key benchmarks: Thai-H6 and Thai Cultural and Linguistic Intelligence Benchmark (ThaiCLI). Through a thorough evaluation of various LLMs with multi-lingual capabilities, we provide a comprehensive analysis of the proposed benchmarks and how they contribute to Thai LLM development. Furthermore, we will make both the datasets and evaluation code publicly available to encourage further research and development for Thai LLMs.

As the field of Multimodal Large Language Models (MLLMs) continues to evolve, their potential to revolutionize artificial intelligence is particularly promising, especially in addressing mathematical reasoning tasks. Current mathematical benchmarks predominantly focus on evaluating MLLMs' problem-solving ability, yet there is a crucial gap in addressing more complex scenarios such as error detection, for enhancing reasoning capability in complicated settings. To fill this gap, we formally formulate the new task: multimodal error detection, and introduce ErrorRadar, the first benchmark designed to assess MLLMs' capabilities in such a task. ErrorRadar evaluates two sub-tasks: error step identification and error categorization, providing a comprehensive framework for evaluating MLLMs' complex mathematical reasoning ability. It consists of 2,500 high-quality multimodal K-12 mathematical problems, collected from real-world student interactions in an educational organization, with rigorous annotation and rich metadata such as problem type and error category. Through extensive experiments, we evaluated both open-source and closed-source representative MLLMs, benchmarking their performance against educational expert evaluators. Results indicate significant challenges still remain, as GPT-4o with best performance is still around 10% behind human evaluation. The dataset will be available upon acceptance.

This paper introduces an innovative approach to analyzing and improving multi-hop reasoning in AI systems by drawing inspiration from Hamiltonian mechanics. We propose a novel framework that maps reasoning chains in embedding spaces to Hamiltonian systems, allowing us to leverage powerful analytical tools from classical physics. Our method defines a Hamiltonian function that balances the progression of reasoning (kinetic energy) against the relevance to the question at hand (potential energy). Using this framework, we analyze a large dataset of reasoning chains from a multi-hop question-answering task, revealing intriguing patterns that distinguish valid from invalid reasoning. We show that valid reasoning chains have lower Hamiltonian energy and move in ways that make the best trade-off between getting more information and answering the right question. Furthermore, we demonstrate the application of this framework to steer the creation of more efficient reasoning algorithms within AI systems. Our results not only provide new insights into the nature of valid reasoning but also open up exciting possibilities for physics-inspired approaches to understanding and improving artificial intelligence.

While recent advancements in large language model (LLM) alignment have enabled the effective identification of malicious objectives involving scene nesting and keyword rewriting, our study reveals that these methods remain inadequate at detecting malicious objectives expressed through context within nested harmless objectives. This study identifies a previously overlooked vulnerability, which we term Attack via Implicit Reference (AIR). AIR decomposes a malicious objective into permissible objectives and links them through implicit references within the context. This method employs multiple related harmless objectives to generate malicious content without triggering refusal responses, thereby effectively bypassing existing detection techniques.Our experiments demonstrate AIR's effectiveness across state-of-the-art LLMs, achieving an attack success rate (ASR) exceeding 90% on most models, including GPT-4o, Claude-3.5-Sonnet, and Qwen-2-72B. Notably, we observe an inverse scaling phenomenon, where larger models are more vulnerable to this attack method. These findings underscore the urgent need for defense mechanisms capable of understanding and preventing contextual attacks. Furthermore, we introduce a cross-model attack strategy that leverages less secure models to generate malicious contexts, thereby further increasing the ASR when targeting other models.Our code and jailbreak artifacts can be found at https://github.com/Lucas-TY/llm_Implicit_reference.

Ensuring faithfulness to context in large language models (LLMs) and retrieval-augmented generation (RAG) systems is crucial for reliable deployment in real-world applications, as incorrect or unsupported information can erode user trust. Despite advancements on standard benchmarks, faithfulness hallucination-where models generate responses misaligned with the provided context-remains a significant challenge. In this work, we introduce FaithEval, a novel and comprehensive benchmark tailored to evaluate the faithfulness of LLMs in contextual scenarios across three diverse tasks: unanswerable, inconsistent, and counterfactual contexts. These tasks simulate real-world challenges where retrieval mechanisms may surface incomplete, contradictory, or fabricated information. FaithEval comprises 4.9K high-quality problems in total, validated through a rigorous four-stage context construction and validation framework, employing both LLM-based auto-evaluation and human validation. Our extensive study across a wide range of open-source and proprietary models reveals that even state-of-the-art models often struggle to remain faithful to the given context, and that larger models do not necessarily exhibit improved faithfulness.Project is available at: \url{https://github.com/SalesforceAIResearch/FaithEval}.

Creating specialized large language models requires vast amounts of clean, special purpose data for training and fine-tuning. With only a handful of existing large-scale, domain-specific datasets, creation of new datasets is required in most applications. This requires the development of new application-specific filtering of web-scale data. Filtering with a high-performance, general-purpose LLM such as GPT-4o can be highly effective, but this is extremely expensive at web-scale. This paper proposes SIEVE, a lightweight alternative that matches GPT-4o accuracy at a fraction of the cost. SIEVE can perform up to 500 filtering operations for the cost of one GPT-4o filtering call. The key to SIEVE is a seamless integration of GPT-4o and lightweight T5 models, using active learning to fine-tune T5 in the background with a small number of calls to GPT-4o. Once trained, it performs as well as GPT-4o at a tiny fraction of the cost. We experimentally validate SIEVE on the OpenWebText dataset, using five highly customized filter tasks targeting high quality and domain-specific content. Our results demonstrate the effectiveness and efficiency of our method in curating large, high-quality datasets for language model training at a substantially lower cost (1%) than existing techniques. To further validate SIEVE, experiments show that SIEVE and GPT-4o achieve similar accuracy, with human evaluators preferring SIEVE's filtering results to those of GPT-4o.