APE: Faster and Longer Context-Augmented Generation via Adaptive Parallel Encoding

Context-augmented generation (CAG) techniques, including RAG and ICL, require the efficient combination of multiple contexts to generate responses to user queries. Directly inputting these contexts as a sequence introduces a considerable computational burden by re-encoding the combined selection of contexts for every request. To address this, we explore the promising potential of parallel encoding to independently pre-compute and cache each context's KV states. This approach enables the direct loading of cached states during inference while accommodating more contexts through position reuse across contexts. However, due to misalignments in attention distribution, directly applying parallel encoding results in a significant performance drop. To enable effective and efficient CAG, we propose Adaptive Parallel Encoding ($\textbf{APE}$), which brings shared prefix, attention temperature, and scaling factor to align the distribution of parallel encoding with sequential encoding. Results on RAG and ICL tasks demonstrate that APE can preserve 98% and 93% sequential encoding performance using the same inputs while outperforming parallel encoding by 3.6% and 7.9%, respectively. It also scales to many-shot CAG, effectively encoding hundreds of contexts in parallel. Efficiency evaluation shows that APE can achieve an end-to-end 4.5$\times$ speedup by reducing 28$\times$ prefilling time for a 128K-length context.

DeepSeek-V3 Technical Report

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks. The model checkpoints are available at https://github.com/deepseek-ai/DeepSeek-V3.

Graph Structure Learning for Spatial-Temporal Imputation: Adapting to Node and Feature Scales

Spatial-temporal data collected across different geographic locations often suffer from missing values, posing challenges to data analysis. Existing methods primarily leverage fixed spatial graphs to impute missing values, which implicitly assume that the spatial relationship is roughly the same for all features across different locations. However, they may overlook the different spatial relationships of diverse features recorded by sensors in different locations. To address this, we introduce the multi-scale Graph Structure Learning framework for spatial-temporal Imputation (GSLI) that dynamically adapts to the heterogeneous spatial correlations. Our framework encompasses node-scale graph structure learning to cater to the distinct global spatial correlations of different features, and feature-scale graph structure learning to unveil common spatial correlation across features within all stations. Integrated with prominence modeling, our framework emphasizes nodes and features with greater significance in the imputation process. Furthermore, GSLI incorporates cross-feature and cross-temporal representation learning to capture spatial-temporal dependencies. Evaluated on six real incomplete spatial-temporal datasets, GSLI showcases the improvement in data imputation.

S$^{2}$FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity

Current PEFT methods for LLMs can achieve either high quality, efficient training, or scalable serving, but not all three simultaneously. To address this limitation, we investigate sparse fine-tuning and observe a remarkable improvement in generalization ability. Utilizing this key insight, we propose a family of Structured Sparse Fine-Tuning (S$^{2}$FT) methods for LLMs, which concurrently achieve state-of-the-art fine-tuning performance, training efficiency, and inference scalability. S$^{2}$FT accomplishes this by "selecting sparsely and computing densely". It selects a few heads and channels in the MHA and FFN modules for each Transformer block, respectively. Next, it co-permutes weight matrices on both sides of the coupled structures in LLMs to connect the selected components in each layer into a dense submatrix. Finally, S$^{2}$FT performs in-place gradient updates on all submatrices. Through theoretical analysis and empirical results, our method prevents overfitting and forgetting, delivers SOTA performance on both commonsense and arithmetic reasoning with 4.6% and 1.3% average improvements compared to LoRA, and surpasses full FT by 11.5% when generalizing to various domains after instruction tuning. Using our partial backpropagation algorithm, S$^{2}$FT saves training memory up to 3$\times$ and improves latency by 1.5-2.7$\times$ compared to full FT, while delivering an average 10% improvement over LoRA on both metrics. We further demonstrate that the weight updates in S$^{2}$FT can be decoupled into adapters, enabling effective fusion, fast switch, and efficient parallelism for serving multiple fine-tuned models.

A privacy-preserving distributed credible evidence fusion algorithm for collective decision-making

The theory of evidence reasoning has been applied to collective decision-making in recent years. However, existing distributed evidence fusion methods lead to participants' preference leakage and fusion failures as they directly exchange raw evidence and do not assess evidence credibility like centralized credible evidence fusion (CCEF) does. To do so, a privacy-preserving distributed credible evidence fusion method with three-level consensus (PCEF) is proposed in this paper. In evidence difference measure (EDM) neighbor consensus, an evidence-free equivalent expression of EDM among neighbored agents is derived with the shared dot product protocol for pignistic probability and the identical judgment of two events with maximal subjective probabilities, so that evidence privacy is guaranteed due to such irreversible evidence transformation. In EDM network consensus, the non-neighbored EDMs are inferred and neighbored EDMs reach uniformity via interaction between linear average consensus (LAC) and low-rank matrix completion with rank adaptation to guarantee EDM consensus convergence and no solution of inferring raw evidence in numerical iteration style. In fusion network consensus, a privacy-preserving LAC with a self-cancelling differential privacy term is proposed, where each agent adds its randomness to the sharing content and step-by-step cancels such randomness in consensus iterations. Besides, the sufficient condition of the convergence to the CCEF is explored, and it is proven that raw evidence is impossibly inferred in such an iterative consensus. The simulations show that PCEF is close to CCEF both in credibility and fusion results and obtains higher decision accuracy with less time-comsuming than existing methods.

Y-Mol: A Multiscale Biomedical Knowledge-Guided Large Language Model for Drug Development

Large Language Models (LLMs) have recently demonstrated remarkable performance in general tasks across various fields. However, their effectiveness within specific domains such as drug development remains challenges. To solve these challenges, we introduce \textbf{Y-Mol}, forming a well-established LLM paradigm for the flow of drug development. Y-Mol is a multiscale biomedical knowledge-guided LLM designed to accomplish tasks across lead compound discovery, pre-clinic, and clinic prediction. By integrating millions of multiscale biomedical knowledge and using LLaMA2 as the base LLM, Y-Mol augments the reasoning capability in the biomedical domain by learning from a corpus of publications, knowledge graphs, and expert-designed synthetic data. The capability is further enriched with three types of drug-oriented instructions: description-based prompts from processed publications, semantic-based prompts for extracting associations from knowledge graphs, and template-based prompts for understanding expert knowledge from biomedical tools. Besides, Y-Mol offers a set of LLM paradigms that can autonomously execute the downstream tasks across the entire process of drug development, including virtual screening, drug design, pharmacological properties prediction, and drug-related interaction prediction. Our extensive evaluations of various biomedical sources demonstrate that Y-Mol significantly outperforms general-purpose LLMs in discovering lead compounds, predicting molecular properties, and identifying drug interaction events.

Small Tunes Transformer: Exploring Macro & Micro-Level Hierarchies for Skeleton-Conditioned Melody Generation

Recently, symbolic music generation has become a focus of numerous deep learning research. Structure as an important part of music, contributes to improving the quality of music, and an increasing number of works start to study the hierarchical structure. In this study, we delve into the multi-level structures within music from macro-level and micro-level hierarchies. At the macro-level hierarchy, we conduct phrase segmentation algorithm to explore how phrases influence the overall development of music, and at the micro-level hierarchy, we design skeleton notes extraction strategy to explore how skeleton notes within each phrase guide the melody generation. Furthermore, we propose a novel Phrase-level Cross-Attention mechanism to capture the intrinsic relationship between macro-level hierarchy and micro-level hierarchy. Moreover, in response to the current lack of research on Chinese-style music, we construct our Small Tunes Dataset: a substantial collection of MIDI files comprising 10088 Small Tunes, a category of traditional Chinese Folk Songs. This dataset serves as the focus of our study. We generate Small Tunes songs utilizing the extracted skeleton notes as conditions, and experiment results indicate that our proposed model, Small Tunes Transformer, outperforms other state-of-the-art models. Besides, we design three novel objective evaluation metrics to evaluate music from both rhythm and melody dimensions.

Unveiling Language Skills under Circuits

The exploration of language skills in language models (LMs) has always been one of the central goals in mechanistic interpretability. However, existing circuit analyses often fall short in representing the full functional scope of these models, primarily due to the exclusion of Feed-Forward layers. Additionally, isolating the effect of a single language skill from a text, which inherently involves multiple entangled skills, poses a significant challenge. To address these gaps, we introduce a novel concept, Memory Circuit, a minimum unit that fully and independently manipulates the memory-reading functionality of a language model, and disentangle the transformer model precisely into a circuit graph which is an ensemble of paths connecting different memory circuits. Based on this disentanglement, we identify salient circuit paths, named as skill paths, responsible for three crucial language skills, i.e., the Previous Token Skill, Induction Skill and In-Context Learning (ICL) Skill, leveraging causal effect estimation through interventions and counterfactuals. Our experiments on various datasets confirm the correspondence between our identified skill paths and language skills, and validate three longstanding hypotheses: 1) Language skills are identifiable through circuit dissection; 2) Simple language skills reside in shallow layers, whereas complex language skills are found in deeper layers; 3) Complex language skills are formed on top of simpler language skills. Our codes are available at: https://github.com/Zodiark-ch/Language-Skill-of-LLMs.

FlatFusion: Delving into Details of Sparse Transformer-based Camera-LiDAR Fusion for Autonomous Driving

The integration of data from diverse sensor modalities (e.g., camera and LiDAR) constitutes a prevalent methodology within the ambit of autonomous driving scenarios. Recent advancements in efficient point cloud transformers have underscored the efficacy of integrating information in sparse formats. When it comes to fusion, since image patches are dense in pixel space with ambiguous depth, it necessitates additional design considerations for effective fusion. In this paper, we conduct a comprehensive exploration of design choices for Transformer-based sparse cameraLiDAR fusion. This investigation encompasses strategies for image-to-3D and LiDAR-to-2D mapping, attention neighbor grouping, single modal tokenizer, and micro-structure of Transformer. By amalgamating the most effective principles uncovered through our investigation, we introduce FlatFusion, a carefully designed framework for sparse camera-LiDAR fusion. Notably, FlatFusion significantly outperforms state-of-the-art sparse Transformer-based methods, including UniTR, CMT, and SparseFusion, achieving 73.7 NDS on the nuScenes validation set with 10.1 FPS with PyTorch.

Rethinking Affect Analysis: A Protocol for Ensuring Fairness and Consistency

Evaluating affect analysis methods presents challenges due to inconsistencies in database partitioning and evaluation protocols, leading to unfair and biased results. Previous studies claim continuous performance improvements, but our findings challenge such assertions. Using these insights, we propose a unified protocol for database partitioning that ensures fairness and comparability. We provide detailed demographic annotations (in terms of race, gender and age), evaluation metrics, and a common framework for expression recognition, action unit detection and valence-arousal estimation. We also rerun the methods with the new protocol and introduce a new leaderboards to encourage future research in affect recognition with a fairer comparison. Our annotations, code, and pre-trained models are available on \hyperlink{https://github.com/dkollias/Fair-Consistent-Affect-Analysis}{Github}.