Large Language Models (LLMs) often encounter conflicts between their learned, internal (parametric knowledge, PK) and external knowledge provided during inference (contextual knowledge, CK). Understanding how LLMs models prioritize one knowledge source over the other remains a challenge. In this paper, we propose a novel probing framework to explore the mechanisms governing the selection between PK and CK in LLMs. Using controlled prompts designed to contradict the model's PK, we demonstrate that specific model activations are indicative of the knowledge source employed. We evaluate this framework on various LLMs of different sizes and demonstrate that mid-layer activations, particularly those related to relations in the input, are crucial in predicting knowledge source selection, paving the way for more reliable models capable of handling knowledge conflicts effectively.

The recent development of vision language models (VLMs) has led to significant advances in visual-language integration through visual instruction tuning, and they have rapidly evolved in the field of remote sensing image understanding, demonstrating their powerful capabilities. However, existing RSVLMs mainly focus on image-level or frame-level understanding, making it difficult to achieve fine-grained pixel-level visual-language alignment. Additionally, the lack of mask-based instructional data limits their further development. In this paper, we propose a mask-text instruction tuning method called Aquila-plus, which extends the capabilities of RSVLMs to achieve pixel-level visual understanding by incorporating fine-grained mask regions into language instructions. To achieve this, we first meticulously constructed a mask region-text dataset containing 100K samples, and then designed a visual-language model by injecting pixel-level representations into a large language model (LLM). Specifically, Aquila-plus uses a convolutional CLIP as the visual encoder and employs a mask-aware visual extractor to extract precise visual mask features from high-resolution inputs. Experimental results demonstrate that Aquila-plus outperforms existing methods in various region understanding tasks, showcasing its novel capabilities in pixel-level instruction tuning.

Continual learning (CL) is crucial for language models to dynamically adapt to the evolving real-world demands. To mitigate the catastrophic forgetting problem in CL, data replay has been proven a simple and effective strategy, and the subsequent data-replay-based distillation can further enhance the performance. However, existing methods fail to fully exploit the knowledge embedded in models from previous tasks, resulting in the need for a relatively large number of replay samples to achieve good results. In this work, we first explore and emphasize the importance of attention weights in knowledge retention, and then propose a SElective attEntion-guided Knowledge Retention method (SEEKR) for data-efficient replay-based continual learning of large language models (LLMs). Specifically, SEEKR performs attention distillation on the selected attention heads for finer-grained knowledge retention, where the proposed forgettability-based and task-sensitivity-based measures are used to identify the most valuable attention heads. Experimental results on two continual learning benchmarks for LLMs demonstrate the superiority of SEEKR over the existing methods on both performance and efficiency. Explicitly, SEEKR achieves comparable or even better performance with only 1/10 of the replayed data used by other methods, and reduces the proportion of replayed data to 1%.

Teasers are an effective tool for promoting content in entertainment, commercial and educational fields. However, creating an effective teaser for long videos is challenging for it requires long-range multimodal modeling on the input videos, while necessitating maintaining audiovisual alignments, managing scene changes and preserving factual accuracy for the output teasers. Due to the lack of a publicly-available dataset, progress along this research direction has been hindered. In this work, we present DocumentaryNet, a collection of 1,269 documentaries paired with their teasers, featuring multimodal data streams of video, speech, music, sound effects and narrations. With DocumentaryNet, we propose a new two-stage system for generating teasers from long documentaries. The proposed TeaserGen system first generates the teaser narration from the transcribed narration of the documentary using a pretrained large language model, and then selects the most relevant visual content to accompany the generated narration through language-vision models. For narration-video matching, we explore two approaches: a pretraining-based model using pretrained contrastive language-vision models and a deep sequential model that learns the mapping between the narrations and visuals. Our experimental results show that the pretraining-based approach is more effective at identifying relevant visual content than directly trained deep autoregressive models.

Although rapid advancements in Large Language Models (LLMs) are facilitating the integration of artificial intelligence-based applications and services in healthcare, limited research has focused on the systematic evaluation of medical notes for guideline adherence. This paper introduces GuidelineGuard, an agentic framework powered by LLMs that autonomously analyzes medical notes, such as hospital discharge and office visit notes, to ensure compliance with established healthcare guidelines. By identifying deviations from recommended practices and providing evidence-based suggestions, GuidelineGuard helps clinicians adhere to the latest standards from organizations like the WHO and CDC. This framework offers a novel approach to improving documentation quality and reducing clinical errors.

We show that variational learning can significantly improve the accuracy and calibration of Low-Rank Adaptation (LoRA) without a substantial increase in the cost. We replace AdamW by the Improved Variational Online Newton (IVON) algorithm to finetune large language models. For Llama-2 with 7 billion parameters, IVON improves the accuracy over AdamW by 2.8% and expected calibration error by 4.6%. The accuracy is also better than the other Bayesian alternatives, yet the cost is lower and the implementation is easier. Our work provides additional evidence for the effectiveness of IVON for large language models.

Natural Language Processing (NLP) operations, such as semantic sentiment analysis and text synthesis, may often impair users' privacy and demand significant on device computational resources. Centralized learning (CL) on the edge offers an alternative energy-efficient approach, yet requires the collection of raw information, which affects the user's privacy. While Federated learning (FL) preserves privacy, it requires high computational energy on board tiny user devices. We introduce split learning (SL) as an energy-efficient alternative, privacy-preserving tiny machine learning (TinyML) scheme and compare it to FL and CL in the presence of Rayleigh fading and additive noise. Our results show that SL reduces processing power and CO2 emissions while maintaining high accuracy, whereas FL offers a balanced compromise between efficiency and privacy. Hence, this study provides insights into deploying energy-efficient, privacy-preserving NLP models on edge devices.

Large language models (LLMs) are often equipped with multi-sample decoding strategies. An LLM implicitly defines an arithmetic code book, facilitating efficient and embarrassingly parallelizable \textbf{arithmetic sampling} to produce multiple samples using quasi-random codes. Traditional text generation methods, such as beam search and sampling-based techniques, have notable limitations: they lack parallelizability or diversity of sampled sequences. This study explores the potential of arithmetic sampling, contrasting it with ancestral sampling across two decoding tasks that employ multi-sample inference: chain-of-thought reasoning with self-consistency and machine translation with minimum Bayes risk decoding. Our results demonstrate that arithmetic sampling produces more diverse samples, significantly improving reasoning and translation performance as the sample size increases. We observe a $\mathbf{3\text{-}5\%}$ point increase in accuracy on the GSM8K dataset and a $\mathbf{0.45\text{-}0.89\%}$ point increment in COMET score for WMT19 tasks using arithmetic sampling without any significant computational overhead.

We assess the difficulty of gender resolution in literary-style dialogue settings and the influence of gender stereotypes. Instances of the test suite contain spoken dialogue interleaved with external meta-context about the characters and the manner of speaking. We find that character and manner stereotypes outside of the dialogue significantly impact the gender agreement of referents within the dialogue.

We introduce Visual Premise Proving (VPP), a novel task tailored to refine the process of chart question answering by deconstructing it into a series of logical premises. Each of these premises represents an essential step in comprehending a chart's content and deriving logical conclusions, thereby providing a granular look at a model's reasoning abilities. This approach represents a departure from conventional accuracy-based evaluation methods, emphasizing the model's ability to sequentially validate each premise and ideally mimic human analytical processes. A model adept at reasoning is expected to demonstrate proficiency in both data retrieval and the structural understanding of charts, suggesting a synergy between these competencies. However, in our zero-shot study using the sophisticated MATCHA model on a scientific chart question answering dataset, an intriguing pattern emerged. The model showcased superior performance in chart reasoning (27\%) over chart structure (19\%) and data retrieval (14\%). This performance gap suggests that models might more readily generalize reasoning capabilities across datasets, benefiting from consistent mathematical and linguistic semantics, even when challenged by changes in the visual domain that complicate structure comprehension and data retrieval. Furthermore, the efficacy of using accuracy of binary QA for evaluating chart reasoning comes into question if models can deduce correct answers without parsing chart data or structure. VPP highlights the importance of integrating reasoning with visual comprehension to enhance model performance in chart analysis, pushing for a balanced approach in evaluating visual data interpretation capabilities.