The long-tail recommendation is a challenging task for traditional recommender systems, due to data sparsity and data imbalance issues. The recent development of large language models (LLMs) has shown their abilities in complex reasoning, which can help to deduce users' preferences based on very few previous interactions. However, since most LLM-based systems rely on items' semantic meaning as the sole evidence for reasoning, the collaborative information of user-item interactions is neglected, which can cause the LLM's reasoning to be misaligned with task-specific collaborative information of the dataset. To further align LLMs' reasoning to task-specific user-item interaction knowledge, we introduce collaborative retrieval-augmented LLMs, CoRAL, which directly incorporate collaborative evidence into the prompts. Based on the retrieved user-item interactions, the LLM can analyze shared and distinct preferences among users, and summarize the patterns indicating which types of users would be attracted by certain items. The retrieved collaborative evidence prompts the LLM to align its reasoning with the user-item interaction patterns in the dataset. However, since the capacity of the input prompt is limited, finding the minimally-sufficient collaborative information for recommendation tasks can be challenging. We propose to find the optimal interaction set through a sequential decision-making process and develop a retrieval policy learned through a reinforcement learning (RL) framework, CoRAL. Our experimental results show that CoRAL can significantly improve LLMs' reasoning abilities on specific recommendation tasks. Our analysis also reveals that CoRAL can more efficiently explore collaborative information through reinforcement learning.

This micro-paper describes a trick to speed up inference of transformers with RoPE (such as LLaMA, Mistral, PaLM, and Gemma). For these models, a large portion of the first transformer layer can be precomputed, which results in slightly lower latency and lower cost-per-token. Because this trick optimizes only one layer, the relative savings depend on the total number of layers. For example, the maximum savings for a model with only 4 layers (such as Whisper tiny) is limited to 25%, while a 32-layer model is limited to 3% savings. See https://github.com/OpenMachine-ai/transformer-tricks for code and more transformer tricks.

Large language model (LLM) has achieved promising performance in multilingual machine translation tasks through zero/few-shot prompts or prompt-tuning. However, due to the mixture of multilingual data during the pre-training of LLM, the LLM-based translation models face the off-target issue in both prompt-based methods, including a series of phenomena, namely instruction misunderstanding, translation with wrong language and over-generation. For this issue, this paper introduces an Auto-Constriction Turning mechanism for Multilingual Neural Machine Translation (ACT-MNMT), which is a novel supervised fine-tuning mechanism and orthogonal to the traditional prompt-based methods. In this method, ACT-MNMT automatically constructs a constrained template in the target side by adding trigger tokens ahead of the ground truth. Furthermore, trigger tokens can be arranged and combined freely to represent different task semantics, and they can be iteratively updated to maximize the label likelihood. Experiments are performed on WMT test sets with multiple metrics, and the experimental results demonstrate that ACT-MNMT achieves substantially improved performance across multiple translation directions and reduce the off-target phenomena in the translation.

The recent huge advance of Large Language Models (LLMs) is mainly driven by the increase in the number of parameters. This has led to substantial memory capacity requirements, necessitating the use of dozens of GPUs just to meet the capacity. One popular solution to this is storage-offloaded training, which uses host memory and storage as an extended memory hierarchy. However, this obviously comes at the cost of storage bandwidth bottleneck because storage devices have orders of magnitude lower bandwidth compared to that of GPU device memories. Our work, Smart-Infinity, addresses the storage bandwidth bottleneck of storage-offloaded LLM training using near-storage processing devices on a real system. The main component of Smart-Infinity is SmartUpdate, which performs parameter updates on custom near-storage accelerators. We identify that moving parameter updates to the storage side removes most of the storage traffic. In addition, we propose an efficient data transfer handler structure to address the system integration issues for Smart-Infinity. The handler allows overlapping data transfers with fixed memory consumption by reusing the device buffer. Lastly, we propose accelerator-assisted gradient compression/decompression to enhance the scalability of Smart-Infinity. When scaling to multiple near-storage processing devices, the write traffic on the shared channel becomes the bottleneck. To alleviate this, we compress the gradients on the GPU and decompress them on the accelerators. It provides further acceleration from reduced traffic. As a result, Smart-Infinity achieves a significant speedup compared to the baseline. Notably, Smart-Infinity is a ready-to-use approach that is fully integrated into PyTorch on a real system. We will open-source Smart-Infinity to facilitate its use.

The academic intelligence of large language models (LLMs) has made remarkable progress in recent times, but their social intelligence performance remains unclear. Inspired by established human social intelligence frameworks, particularly Daniel Goleman's social intelligence theory, we have developed a standardized social intelligence test based on real-world social scenarios to comprehensively assess the social intelligence of LLMs, termed as the Situational Evaluation of Social Intelligence (SESI). We conducted an extensive evaluation with 13 recent popular and state-of-art LLM agents on SESI. The results indicate the social intelligence of LLMs still has significant room for improvement, with superficially friendliness as a primary reason for errors. Moreover, there exists a relatively low correlation between the social intelligence and academic intelligence exhibited by LLMs, suggesting that social intelligence is distinct from academic intelligence for LLMs. Additionally, while it is observed that LLMs can't ``understand'' what social intelligence is, their social intelligence, similar to that of humans, is influenced by social factors.

We present DeepSeek-VL, an open-source Vision-Language (VL) Model designed for real-world vision and language understanding applications. Our approach is structured around three key dimensions: Data Construction: We strive to ensure our data is diverse, scalable and extensively covers real-world scenarios including web screenshots, PDFs, OCR, charts, and knowledge-based content (expert knowledge, textbooks), aiming for a comprehensive representation of practical contexts. Further, we create a use case taxonomy from real user scenarios and construct an instruction-tuning dataset accordingly. The fine-tuning with this dataset substantially improves the model's user experience in practical applications. Model Architecture: Considering efficiency and the demands of most real-world scenarios, DeepSeek-VL incorporates a hybrid vision encoder that efficiently processes high-resolution images (1024 x 1024) within a fixed token budget, while maintaining a relatively low computational overhead. This design choice ensures the model's ability to capture critical semantic and detailed information across various visual tasks. Training Strategy: We posit that a proficient Vision-Language Model should, foremost, possess strong language abilities. To ensure the preservation of LLM capabilities during pretraining, we investigate an effective VL pretraining strategy by integrating LLM training from the beginning and carefully managing the competitive dynamics observed between vision and language modalities. Starting with a focus on text, we gradually adjust the ratio to facilitate a balanced integration of both modalities.

With appropriate data selection and training techniques, Large Language Models (LLMs) have demonstrated exceptional success in various medical examinations and multiple-choice questions. However, the application of LLMs in medical dialogue generation-a task more closely aligned with actual medical practice-has been less explored. This gap is attributed to the insufficient medical knowledge of LLMs, which leads to inaccuracies and hallucinated information in the generated medical responses. In this work, we introduce the Medical dialogue with Knowledge enhancement and clinical Pathway encoding (MedKP) framework, which integrates an external knowledge enhancement module through a medical knowledge graph and an internal clinical pathway encoding via medical entities and physician actions. Evaluated with comprehensive metrics, our experiments on two large-scale, real-world online medical consultation datasets (MedDG and KaMed) demonstrate that MedKP surpasses multiple baselines and mitigates the incidence of hallucinations, achieving a new state-of-the-art. Extensive ablation studies further reveal the effectiveness of each component of MedKP. This enhancement advances the development of reliable, automated medical consultation responses using LLMs, thereby broadening the potential accessibility of precise and real-time medical assistance.

We present the Thought Graph as a novel framework to support complex reasoning and use gene set analysis as an example to uncover semantic relationships between biological processes. Our framework stands out for its ability to provide a deeper understanding of gene sets, significantly surpassing GSEA by 40.28% and LLM baselines by 5.38% based on cosine similarity to human annotations. Our analysis further provides insights into future directions of biological processes naming, and implications for bioinformatics and precision medicine.

Legal judgment prediction suffers from the problem of long case documents exceeding tens of thousands of words, in general, and having a non-uniform structure. Predicting judgments from such documents becomes a challenging task, more so on documents with no structural annotation. We explore the classification of these large legal documents and their lack of structural information with a deep-learning-based hierarchical framework which we call MESc; "Multi-stage Encoder-based Supervised with-clustering"; for judgment prediction. Specifically, we divide a document into parts to extract their embeddings from the last four layers of a custom fine-tuned Large Language Model, and try to approximate their structure through unsupervised clustering. Which we use in another set of transformer encoder layers to learn the inter-chunk representations. We analyze the adaptability of Large Language Models (LLMs) with multi-billion parameters (GPT-Neo, and GPT-J) with the hierarchical framework of MESc and compare them with their standalone performance on legal texts. We also study their intra-domain(legal) transfer learning capability and the impact of combining embeddings from their last layers in MESc. We test these methods and their effectiveness with extensive experiments and ablation studies on legal documents from India, the European Union, and the United States with the ILDC dataset and a subset of the LexGLUE dataset. Our approach achieves a minimum total performance gain of approximately 2 points over previous state-of-the-art methods.

Recent advances in instruction-tuned Large Vision-Language Models (LVLMs) have imbued the models with the ability to generate high-level, image-grounded explanations with ease. While such capability is largely attributed to the rich world knowledge contained within the Large Language Models (LLMs), our work reveals their shortcomings in fine-grained visual categorization (FGVC) across six different benchmark settings. Most recent state-of-the-art LVLMs like LLaVa-1.5, InstructBLIP and GPT-4V not only severely deteriorate in terms of classification performance, e.g., average drop of 65.58 in EM for Stanford Dogs for LLaVA-1.5, but also struggle to generate an accurate explanation with detailed attributes based on the concept that appears within an input image despite their capability to generate holistic image-level descriptions. In-depth analyses show that instruction-tuned LVLMs exhibit modality gap, showing discrepancy when given textual and visual inputs that correspond to the same concept, preventing the image modality from leveraging the rich parametric knowledge within the LLMs. In an effort to further the community's endeavor in this direction, we propose a multiple granularity attribute-centric evaluation benchmark, Finer, which aims to establish a ground to evaluate LVLMs' fine-grained visual comprehension ability and provide significantly improved explainability.