Direct Preference Optimization (DPO) is a widely used offline preference optimization algorithm that reparameterizes reward functions in reinforcement learning from human feedback (RLHF) to enhance simplicity and training stability. In this work, we propose SimPO, a simpler yet more effective approach. The effectiveness of SimPO is attributed to a key design: using the average log probability of a sequence as the implicit reward. This reward formulation better aligns with model generation and eliminates the need for a reference model, making it more compute and memory efficient. Additionally, we introduce a target reward margin to the Bradley-Terry objective to encourage a larger margin between the winning and losing responses, further enhancing the algorithm's performance. We compare SimPO to DPO and its latest variants across various state-of-the-art training setups, including both base and instruction-tuned models like Mistral and Llama3. We evaluated on extensive instruction-following benchmarks, including AlpacaEval 2, MT-Bench, and the recent challenging Arena-Hard benchmark. Our results demonstrate that SimPO consistently and significantly outperforms existing approaches without substantially increasing response length. Specifically, SimPO outperforms DPO by up to 6.4 points on AlpacaEval 2 and by up to 7.5 points on Arena-Hard. Our top-performing model, built on Llama3-8B-Instruct, achieves a remarkable 53.7 length-controlled win rate on AlpacaEval 2 -- surpassing Claude 3 Opus on the leaderboard, and a 36.5 win rate on Arena-Hard -- making it the strongest 8B open-source model.
相關內容
The field of emotion recognition of conversation (ERC) has been focusing on separating sentence feature encoding and context modeling, lacking exploration in generative paradigms based on unified designs. In this study, we propose a novel approach, InstructERC, to reformulate the ERC task from a discriminative framework to a generative framework based on Large Language Models (LLMs). InstructERC makes three significant contributions: (1) it introduces a simple yet effective retrieval template module, which helps the model explicitly integrate multi-granularity dialogue supervision information. (2) We introduce two additional emotion alignment tasks, namely speaker identification and emotion prediction tasks, to implicitly model the dialogue role relationships and future emotional tendencies in conversations. (3) Pioneeringly, we unify emotion labels across benchmarks through the feeling wheel to fit real application scenarios. InstructERC still perform impressively on this unified dataset. Our LLM-based plugin framework significantly outperforms all previous models and achieves comprehensive SOTA on three commonly used ERC datasets. Extensive analysis of parameter-efficient and data-scaling experiments provides empirical guidance for applying it in practical scenarios.
RoarGraph: A Projected Bipartite Graph for Efficient Cross-Modal Approximate Nearest Neighbor Search
Approximate Nearest Neighbor Search (ANNS) is a fundamental and critical component in many applications, including recommendation systems and large language model-based applications. With the advancement of multimodal neural models, which transform data from different modalities into a shared high-dimensional space as feature vectors, cross-modal ANNS aims to use the data vector from one modality (e.g., texts) as the query to retrieve the most similar items from another (e.g., images or videos). However, there is an inherent distribution gap between embeddings from different modalities, and cross-modal queries become Out-of-Distribution (OOD) to the base data. Consequently, state-of-the-art ANNS approaches suffer poor performance for OOD workloads. In this paper, we quantitatively analyze the properties of the OOD workloads to gain an understanding of their ANNS efficiency. Unlike single-modal workloads, we reveal OOD queries spatially deviate from base data, and the k-nearest neighbors of an OOD query are distant from each other in the embedding space. The property breaks the assumptions of existing ANNS approaches and mismatches their design for efficient search. With insights from the OOD workloads, we propose pRojected bipartite Graph (RoarGraph), an efficient ANNS graph index built under the guidance of query distribution. Extensive experiments show that RoarGraph significantly outperforms state-of-the-art approaches on modern cross-modal datasets, achieving up to 3.56x faster search speed at a 90% recall rate for OOD queries.
Graph Neural Networks (GNNs) have demonstrated effectiveness in collaborative filtering tasks due to their ability to extract powerful structural features. However, combining the graph features extracted from user-item interactions and auxiliary features extracted from user genres and item properties remains a challenge. Currently available fusion methods face two major issues: 1) simple methods such as concatenation and summation are generic, but not accurate in capturing feature relationships; 2) task-specific methods like attention mechanisms and meta paths may not be suitable for general feature fusion. To address these challenges, we present GraphTransfer, a simple but universal feature fusion framework for GNN-based collaborative filtering. Our method accurately fuses different types of features by first extracting graph features from the user-item interaction graph and auxiliary features from users and items using GCN. The proposed cross fusion module then effectively bridges the semantic gaps between the interaction scores of different features. Theoretical analysis and experiments on public datasets show that GraphTransfer outperforms other feature fusion methods in CF tasks. Additionally, we demonstrate the universality of our framework via empirical studies in three other scenarios, showing that GraphTransfer leads to significant improvements in the performance of CF algorithms.
Deep & Cross Network and its derivative models have become an important paradigm for click-through rate (CTR) prediction due to their effective balance between computational cost and performance. However, these models face four major limitations: (1) the performance of existing explicit feature interaction methods is often weaker than that of implicit deep neural network (DNN), undermining their necessity; (2) many models fail to adaptively filter noise while increasing the order of feature interactions; (3) the fusion methods of most models cannot provide suitable supervision signals for their different sub-networks; (4) while most models claim to capture high-order feature interactions, they often do so implicitly and non-interpretably through DNN, which limits the trustworthiness of the model's predictions. To address the identified limitations, this paper proposes the next generation deep cross network: Deep Cross Network v3 (DCNv3), along with its two sub-networks: Linear Cross Network (LCN) and Exponential Cross Network (ECN) for CTR prediction. DCNv3 ensures interpretability in feature interaction modeling while linearly and exponentially increasing the order of feature interactions to achieve genuine Deep Crossing rather than just Deep & Cross. Additionally, we employ a Self-Mask operation to filter noise and reduce the number of parameters in the Cross Network by half. In the fusion layer, we use a simple yet effective multi-loss trade-off and calculation method, called Tri-BCE, to provide appropriate supervision signals. Comprehensive experiments on six datasets demonstrate the effectiveness, efficiency, and interpretability of DCNv3. The code, running logs, and detailed hyperparameter configurations are available at: //github.com/salmon1802/DCNv3.
Given a list L of elements and a property that L exhibits, ddmin is a well-known test input minimization algorithm designed to automatically eliminate irrelevant elements from L. This algorithm is extensively adopted in test input minimization and software debloating. Recently, ProbDD, an advanced variant of ddmin, has been proposed and achieved state-of-the-art performance. Employing Bayesian optimization, ProbDD predicts the likelihood of each element in L being essential, and statistically decides which elements and how many should be removed each time. Despite its impressive results, the theoretical probabilistic model of ProbDD is complex, and the specific factors driving its superior performance have not been investigated. In this paper, we conduct the first in-depth theoretical analysis of ProbDD, clarifying trends in probability and subset size changes while simplifying the probability model. Complementing this analysis, we perform empirical experiments, including success rate analysis, ablation studies, and analysis on trade-offs and limitations, to better understand and demystify this state-of-the-art algorithm. Our success rate analysis shows how ProbDD addresses bottlenecks of ddmin by skipping inefficient queries that attempt to delete complements of subsets and previously tried subsets. The ablation study reveals that randomness in ProbDD has no significant impact on efficiency. Based on these findings, we propose CDD, a simplified version of ProbDD, reducing complexity in both theory and implementation. Besides, the performance of CDD validates our key findings. Comprehensive evaluations across 76 benchmarks in test input minimization and software debloating show that CDD can achieve the same performance as ProbDD despite its simplification. These insights provide valuable guidance for future research and applications of test input minimization algorithms.
Generalized Category Discovery (GCD) is a challenging task in which, given a partially labelled dataset, models must categorize all unlabelled instances, regardless of whether they come from labelled categories or from new ones. In this paper, we challenge a remaining assumption in this task: that all images share the same domain. Specifically, we introduce a new task and method to handle GCD when the unlabelled data also contains images from different domains to the labelled set. Our proposed `HiLo' networks extract High-level semantic and Low-level domain features, before minimizing the mutual information between the representations. Our intuition is that the clusterings based on domain information and semantic information should be independent. We further extend our method with a specialized domain augmentation tailored for the GCD task, as well as a curriculum learning approach. Finally, we construct a benchmark from corrupted fine-grained datasets as well as a large-scale evaluation on DomainNet with real-world domain shifts, reimplementing a number of GCD baselines in this setting. We demonstrate that HiLo outperforms SoTA category discovery models by a large margin on all evaluations.
Reinforcement Learning from Human Feedback (RLHF) is a popular method for aligning Language Models (LM) with human values and preferences. RLHF requires a large number of preference pairs as training data, which are often used in both the Supervised Fine-Tuning and Reward Model training and therefore publicly available datasets are commonly used. In this work, we study to what extent a malicious actor can manipulate the LMs generations by poisoning the preferences, i.e., injecting poisonous preference pairs into these datasets and the RLHF training process. We propose strategies to build poisonous preference pairs and test their performance by poisoning two widely used preference datasets. Our results show that preference poisoning is highly effective: injecting a small amount of poisonous data (1-5\% of the original dataset), we can effectively manipulate the LM to generate a target entity in a target sentiment (positive or negative). The findings from our experiments also shed light on strategies to defend against the preference poisoning attack.
Humans are capable of continuously manipulating a wide variety of deformable objects into complex shapes. This is made possible by our intuitive understanding of material properties and mechanics of the object, for reasoning about object states even when visual perception is occluded. These capabilities allow us to perform diverse tasks ranging from cooking with dough to expressing ourselves with pottery-making. However, developing robotic systems to robustly perform similar tasks remains challenging, as current methods struggle to effectively model volumetric deformable objects and reason about the complex behavior they typically exhibit. To study the robotic systems and algorithms capable of deforming volumetric objects, we introduce a novel robotics task of continuously deforming clay on a pottery wheel. We propose a pipeline for perception and pottery skill-learning, called RoPotter, wherein we demonstrate that structural priors specific to the task of pottery-making can be exploited to simplify the pottery skill-learning process. Namely, we can project the cross-section of the clay to a plane to represent the state of the clay, reducing dimensionality. We also demonstrate a mesh-based method of occluded clay state recovery, toward robotic agents capable of continuously deforming clay. Our experiments show that by using the reduced representation with structural priors based on the deformation behaviors of the clay, RoPotter can perform the long-horizon pottery task with 44.4% lower final shape error compared to the state-of-the-art baselines.
The remarkable success of Large Language Models (LLMs) across diverse tasks has driven the research community to extend their capabilities to molecular applications. However, most molecular LLMs employ adapter-based architectures that do not treat molecule and text modalities equally and lack a supervision signal for the molecule modality. To address these issues, we introduce UniMoT, a Unified Molecule-Text LLM adopting a tokenizer-based architecture that expands the vocabulary of LLM with molecule tokens. Specifically, we introduce a Vector Quantization-driven tokenizer that incorporates a Q-Former to bridge the modality gap between molecule and text. This tokenizer transforms molecules into sequences of molecule tokens with causal dependency, encapsulating high-level molecular and textual information. Equipped with this tokenizer, UniMoT can unify molecule and text modalities under a shared token representation and an autoregressive training paradigm, enabling it to interpret molecules as a foreign language and generate them as text. Following a four-stage training scheme, UniMoT emerges as a multi-modal generalist capable of performing both molecule-to-text and text-to-molecule tasks. Extensive experiments demonstrate that UniMoT achieves state-of-the-art performance across a wide range of molecule comprehension and generation tasks.
Deep learning has shown great potential for modeling the physical dynamics of complex particle systems such as fluids (in Lagrangian descriptions). Existing approaches, however, require the supervision of consecutive particle properties, including positions and velocities. In this paper, we consider a partially observable scenario known as fluid dynamics grounding, that is, inferring the state transitions and interactions within the fluid particle systems from sequential visual observations of the fluid surface. We propose a differentiable two-stage network named NeuroFluid. Our approach consists of (i) a particle-driven neural renderer, which involves fluid physical properties into the volume rendering function, and (ii) a particle transition model optimized to reduce the differences between the rendered and the observed images. NeuroFluid provides the first solution to unsupervised learning of particle-based fluid dynamics by training these two models jointly. It is shown to reasonably estimate the underlying physics of fluids with different initial shapes, viscosity, and densities. It is a potential alternative approach to understanding complex fluid mechanics, such as turbulence, that are difficult to model using traditional methods of mathematical physics.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191