Reward Models (RMs) play a crucial role in aligning LLMs with human preferences, enhancing their performance by ranking outputs during inference or iterative training. However, the degree to which an RM generalizes to new tasks is often not known a priori (e.g. some RMs may excel at scoring creative writing vs. math reasoning). Therefore, using only one fixed RM while training LLMs can be suboptimal. Moreover, optimizing LLMs with multiple RMs simultaneously can be prohibitively computationally-intensive and challenging due to conflicting signals from different RMs, potentially degrading performance. To address these challenges, we introduce LASeR (Learning to Adaptively Select Rewards), which iteratively trains LLMs using multiple RMs, selecting and utilizing the most well-suited RM for each instance to rank outputs and generate preference data, framed as a multi-armed bandit problem. Our results on commonsense and math reasoning tasks demonstrate that LASeR can boost iterative LLM optimization by optimizing for multiple RMs, improving the absolute average accuracy of Llama-3-8B over three datasets by 2.67% over training with ensemble RM scores while also showing superior training efficiency (e.g., a 2x speedup). Moreover, on WildChat, a benchmark of instruction-following prompts, we find that using Llama-3-8B LASeR leads to a 71.45% AlpacaEval win rate over sequentially optimizing multiple RMs. Extending to long-context generation tasks, we find that on Llama-3-8B, LASeR achieves an average improvement of 2.64 F1 and 2.42 F1 on single- and multi-document QA over random RM selection when used with best-of-n sampling. LASeR is robust to noisy rewards and generalizes to multiple settings. Finally, LASeR's RM selection changes depending on the underlying task or instance and we verify the presence of conflicting preferences from multiple RMs that can be mitigated using LASeR.
相關內容
Instruction tuning constitutes a prevalent technique for tailoring Large Vision Language Models (LVLMs) to meet individual task requirements. To date, most of the existing approaches are confined to single-task adaptation, whereas the requirements in real-world scenarios are inherently varied and continually evolving. Thus an ideal LVLM should sustain continual instruction tuning in the face of stream-task distributions (i.e., different domains, emerging capabilities, and new datasets) while minimizing the forgetting of previously acquired knowledge. To achieve this, we propose a new benchmark for COntinuAl inStruction Tuning on LVLMs (COAST), which encompasses the aforementioned domain-incremental, capability-incremental, and dataset-incremental configurations. In terms of methodology, we propose Continual LLaVA, a rehearsal-free method tailored for continual instruction tuning in LVLMs. To circumvent the additional overhead associated with experience replay, we freeze LVLMs and construct the dual increment embeddings for each input instruction to facilitate parameter-efficient tuning. Specifically, the increment embeddings can be decomposed into two principal components: 1) intrinsic increment embeddings to encode task-specific characteristics. To achieve this, we set up a low-rank pool containing candidate embeddings, from which we select the relevant ones based on their similarity with the user instructions; 2) contextual increment embeddings to investigate the inter-dependencies across tasks. In this regard, the low-rank embeddings chosen in the previous tasks are aggregated via learnable weighted sum to provide complementary hints. Extensive experiments indicate that the proposed Continual LLaVA outperforms previous methods by significantly reducing the forgetting during the continual instruction tuning process.
Recent trends in Generative AI have emerged towards fine-tuning foundational large language models (LLMs) to create domain-specific LLMs for automation and chatbot-like applications. Specialized applications for analytics-heavy domains such as Financial report generation require specific writing styles that comprise compound and creative sentences with minimized hallucinations. In this work, we explore the self-corrective auto-regressive qualities of LLMs to learn creativity in writing styles with minimal prompting. We propose a novel two-stage fine-tuning (FT) strategy wherein in the first stage public domain financial reports are used to train for writing styles while allowing the LLM to hallucinate. In the second stage the examples of hallucinations are manually corrected and further used to fine-tune the LLM. The finally trained LLM learns to generate specific financial report sections using minimal instructions and tabular data inputs while ensuring low fine-tuning costs. Our proposed two-stage fine-tuning boosts the accuracy of financial questions answering by two-folds while reducing hallucinations by over 50%. Also, the fine-tuned model has lower perplexity, improved ROUGE, TER and BLEU scores, higher creativity and knowledge density with lower uncertainty and cross entropy than base LLMs. Thus, the proposed framework can be generalized to train creativity in LLMs by first allowing them to hallucinate.
We propose a single-channel Deep Cascade Fusion of Diarization and Separation (DCF-DS) framework for back-end speech recognition, combining neural speaker diarization (NSD) and speech separation (SS). First, we sequentially integrate the NSD and SS modules within a joint training framework, enabling the separation module to leverage speaker time boundaries from the diarization module effectively. Then, to complement DCF-DS training, we introduce a window-level decoding scheme that allows the DCF-DS framework to handle the sparse data convergence instability (SDCI) problem. We also explore using an NSD system trained on real datasets to provide more accurate speaker boundaries during decoding. Additionally, we incorporate an optional multi-input multi-output speech enhancement module (MIMO-SE) within the DCF-DS framework, which offers further performance gains. Finally, we enhance diarization results by re-clustering DCF-DS outputs, improving ASR accuracy. By incorporating the DCF-DS method, we achieved first place in the realistic single-channel track of the CHiME-8 NOTSOFAR-1 challenge. We also perform the evaluation on the open LibriCSS dataset, achieving a new state-of-the-art performance on single-channel speech recognition.
Classification tasks are typically handled using Machine Learning (ML) models, which lack a balance between accuracy and interpretability. This paper introduces a new approach for classification tasks using Large Language Models (LLMs) in an explainable method. Unlike ML models, which rely heavily on data cleaning and feature engineering, this method streamlines the process using LLMs. This paper proposes a method called "Language Model Learning (LML)" powered by a new method called "Data-Augmented Prediction (DAP)." The classification is performed by LLMs using a method similar to that used by humans who manually explore and understand the data to decide classifications. In the process of LML, a dataset is summarized and evaluated to determine the features leading to each label the most. In the DAP process, the system uses the data summary and a row of the testing dataset to automatically generate a query to retrieve relevant rows from the dataset for context-aware classification. LML and DAP unlock new possibilities in areas that require explainable and context-aware decisions by ensuring satisfactory accuracy even with complex data. The system scored an accuracy above 90% in some test cases, confirming the effectiveness and potential of the system to outperform ML models in various scenarios. The source code is available at //github.com/Pro-GenAI/LML-DAP
Head pose estimation (HPE) requires a sophisticated understanding of 3D spatial relationships to generate precise yaw, pitch, and roll angles. Previous HPE models, primarily CNN-based, rely on cropped close-up human head images as inputs and often lack robustness in real-world scenario. Vision Language Models (VLMs) can analyze entire images while focusing on specific objects through their attention mechanisms. In this paper, we propose a novel framework to improve the HPE accuracy by leveraging the object detection grounding capability of a VLM, referred to as CogVLM. We empirically find that directly LoRA fine-tuning of this VLM for the HPE task fails to achieve desirable HPE accuracy, while some model merging methods can improve accuracy but frequently produce blended invalid response formats, struggling to handle both object detection and HPE tasks simultaneously. To integrate HPE capability into CogVLM effectively, we develop a novel LoRA layer-based model merging method. This merging approach applies a high cosine similarity threshold and a winner-takes-all layer selection strategy, aligning attention to the HPE task while preserving original object detection knowledge. It successfully resolves issues with blended invalid response formats and improves accuracy. Results show that our HPE-CogVLM achieves a 31.5\% reduction in Mean Absolute Error over the current state-of-the-art CNN model, 6DRepNet, in cross-dataset evaluation. Furthermore, HPE-CogVLM outperforms both directly LoRA fine-tuned and task arithmetic-based merged VLMs across all HPE metrics.
Deep neural networks (DNNs) typically employ an end-to-end (E2E) training paradigm which presents several challenges, including high GPU memory consumption, inefficiency, and difficulties in model parallelization during training. Recent research has sought to address these issues, with one promising approach being local learning. This method involves partitioning the backbone network into gradient-isolated modules and manually designing auxiliary networks to train these local modules. Existing methods often neglect the interaction of information between local modules, leading to myopic issues and a performance gap compared to E2E training. To address these limitations, we propose the Multilaminar Leap Augmented Auxiliary Network (MLAAN). Specifically, MLAAN comprises Multilaminar Local Modules (MLM) and Leap Augmented Modules (LAM). MLM captures both local and global features through independent and cascaded auxiliary networks, alleviating performance issues caused by insufficient global features. However, overly simplistic auxiliary networks can impede MLM's ability to capture global information. To address this, we further design LAM, an enhanced auxiliary network that uses the Exponential Moving Average (EMA) method to facilitate information exchange between local modules, thereby mitigating the shortsightedness resulting from inadequate interaction. The synergy between MLM and LAM has demonstrated excellent performance. Our experiments on the CIFAR-10, STL-10, SVHN, and ImageNet datasets show that MLAAN can be seamlessly integrated into existing local learning frameworks, significantly enhancing their performance and even surpassing end-to-end (E2E) training methods, while also reducing GPU memory consumption.
SaSR-Net: Source-Aware Semantic Representation Network for Enhancing Audio-Visual Question Answering
Audio-Visual Question Answering (AVQA) is a challenging task that involves answering questions based on both auditory and visual information in videos. A significant challenge is interpreting complex multi-modal scenes, which include both visual objects and sound sources, and connecting them to the given question. In this paper, we introduce the Source-aware Semantic Representation Network (SaSR-Net), a novel model designed for AVQA. SaSR-Net utilizes source-wise learnable tokens to efficiently capture and align audio-visual elements with the corresponding question. It streamlines the fusion of audio and visual information using spatial and temporal attention mechanisms to identify answers in multi-modal scenes. Extensive experiments on the Music-AVQA and AVQA-Yang datasets show that SaSR-Net outperforms state-of-the-art AVQA methods.
The rapid advancement of Large Language Models (LLMs) has led to their increased integration into mobile devices for personalized assistance, which enables LLMs to call external API functions to enhance their performance. However, challenges such as data scarcity, ineffective question formatting, and catastrophic forgetting hinder the development of on-device LLM agents. To tackle these issues, we propose Alopex, a framework that enables precise on-device function calls using the Fox LLM. Alopex introduces a logic-based method for generating high-quality training data and a novel ``description-question-output'' format for fine-tuning, reducing risks of function information leakage. Additionally, a data mixing strategy is used to mitigate catastrophic forgetting, combining function call data with textbook datasets to enhance performance in various tasks. Experimental results show that Alopex improves function call accuracy and significantly reduces catastrophic forgetting, providing a robust solution for integrating function call capabilities into LLMs without manual intervention.
Despite existing 3D cloth simulators producing realistic results, they predominantly operate on discrete surface representations (e.g. points and meshes) with a fixed spatial resolution, which often leads to large memory consumption and resolution-dependent simulations. Moreover, back-propagating gradients through the existing solvers is difficult, and they cannot be easily integrated into modern neural architectures. In response, this paper re-thinks physically plausible cloth simulation: We propose NeuralClothSim, i.e., a new quasistatic cloth simulator using thin shells, in which surface deformation is encoded in neural network weights in the form of a neural field. Our memory-efficient solver operates on a new continuous coordinate-based surface representation called neural deformation fields (NDFs); it supervises NDF equilibria with the laws of the non-linear Kirchhoff-Love shell theory with a non-linear anisotropic material model. NDFs are adaptive: They 1) allocate their capacity to the deformation details and 2) allow surface state queries at arbitrary spatial resolutions without re-training. We show how to train NeuralClothSim while imposing hard boundary conditions and demonstrate multiple applications, such as material interpolation and simulation editing. The experimental results highlight the effectiveness of our continuous neural formulation. See our project page: //4dqv.mpi-inf.mpg.de/NeuralClothSim/.
State-of-the-art Convolutional Neural Network (CNN) benefits a lot from multi-task learning (MTL), which learns multiple related tasks simultaneously to obtain shared or mutually related representations for different tasks. The most widely-used MTL CNN structure is based on an empirical or heuristic split on a specific layer (e.g., the last convolutional layer) to minimize different task-specific losses. However, this heuristic sharing/splitting strategy may be harmful to the final performance of one or multiple tasks. In this paper, we propose a novel CNN structure for MTL, which enables automatic feature fusing at every layer. Specifically, we first concatenate features from different tasks according to their channel dimension, and then formulate the feature fusing problem as discriminative dimensionality reduction. We show that this discriminative dimensionality reduction can be done by 1x1 Convolution, Batch Normalization, and Weight Decay in one CNN, which we refer to as Neural Discriminative Dimensionality Reduction (NDDR). We perform ablation analysis in details for different configurations in training the network. The experiments carried out on different network structures and different task sets demonstrate the promising performance and desirable generalizability of our proposed method.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191