As a scientific researcher, you need to search and browse a large amount of academic literature every day to obtain the latest scientific and technological progress and research results. However, traditional retrieval and reading methods can no longer meet the needs of researchers.
ChatPaper, a document knowledge tool that integrates retrieval, reading, and knowledge question-and-answer. Help you quickly improve the efficiency of searching and reading papers, obtain the latest research trends in the field, and make scientific research work more easily.
Combined with the cutting-edge dynamic subscription function, select arXiv's popular new papers of the day to form a summary of papers, so that everyone can understand cutting-edge trends more quickly.
If you want to have an in-depth dialogue on a certain paper, you can directly copy the link of the paper to your browser or go directly to the ChatPaper page: https://www.aminer.cn/chat/g/
List of Featured New Papers for July 26, 2023:
1. Contrastive Example-Based Control paper details page
https://www.aminer.cn/pub/64a29620d68f896efa28f818/
The paper discusses the challenges in reinforcement learning that practical problems rarely fit the model of a Markov decision process (MDP), interaction with the environment is often expensive, and specifying reward functions is challenging. To address these challenges, previous studies proposed data-driven approaches that fully learn from samples of transfer dynamics and examples of high-reward states. These methods typically learn a reward function from high-reward states, label transitions with this reward function, and apply offline reinforcement learning algorithms to these transitions. While these methods can achieve good results on many tasks, they can be complex, often requiring regularization and time-difference updates. This paper proposes an offline, example-based control method that learns an implicit model of multi-step transfers instead of a reward function. We show that this implicit model can represent Q-values for example-based control problems. Across a range of state-based and image-based offline control tasks, our method outperforms baseline methods using learned reward functions; additional experiments demonstrate improved robustness and scalability to dataset size.
2. LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition paper details page
https://www.aminer.cn/pub/64c09a963fda6d7f06e3e219/
The paper introduces a framework called LoraHub, which aims to achieve adjustable adaptability to unseen tasks by combining LoRA modules trained on different tasks. The paper states that with LoraHub, multiple LoRA modules can be fluently combined without human expertise, just by getting a few examples from new tasks. This combination requires neither additional model parameters nor gradients. Experimental results show that LoraHub can effectively simulate performance in context learning with a small number of examples, without requiring contextual examples next to each inference input. An important contribution of this research is the establishment of a LoRA community where users can share their trained LoRA modules to facilitate their application to new tasks. This resource is expected to expand the application and advancement of general intelligence and LLMs in production.
3. ARB: Advanced Reasoning Benchmark for Large Language Models paper details page
https://www.aminer.cn/pub/64c09a9c3fda6d7f06e3e9dd/
Despite the impressive performance of large language models on various quantitative reasoning and knowledge benchmarks, many benchmarks gradually lose their usefulness as language models score higher and higher, although they have not yet reached expert level. To address this, the researchers introduced a new benchmark called the ARB, which contains advanced reasoning questions across domains including mathematics, physics, biology, chemistry and law. They evaluated the performance of state-of-the-art models such as GPT-4 and Claude on ARB and found that current models scored well below 50% on more challenging tasks. To improve automatic and assisted evaluation capabilities, they introduce a scoring criterion-based evaluation method that allows GPT-4 to score its own intermediate reasoning steps. In addition, they performed human evaluation on a subset of ARB's symbols and found some degree of agreement between the annotators' and GPT-4's scores.
4. Predicting Code Coverage without Execution paper details page
https://www.aminer.cn/pub/64c09a9c3fda6d7f06e3e898/
For the problem of calculating code coverage, the paper points out that the resources required to calculate code coverage are large, and the context of the entire program is required to calculate the coverage of code fragments. To reduce the cost of computing code coverage, the authors propose to use machine learning to predict code coverage, requiring only the context of the source code. The authors propose a new evaluation task called "Code Coverage Prediction for Large Language Models (LLMs)", which aims to evaluate the ability of LLMs to understand code execution. The authors create a dataset called COVERAGEEVAL by performing tests and collecting code coverage information, and report the performance of four state-of-the-art LLMs for code-related tasks. Finally, the authors also demonstrate that code coverage is valuable as an indicator and pre-training data source for the overall performance of LLMs on software engineering tasks.
5. Decision-Focused Learning: Foundations, State of the Art, Benchmark and Future Opportunities paper details page
https://www.aminer.cn/pub/64c09a9c3fda6d7f06e3e956/
The paper mainly discusses Decision-focused learning (DFL), an emerging machine learning paradigm. DFL aims to train models to optimize decisions, integrating prediction and optimization in one end-to-end system. This paradigm promises to transform decision making in many real-world applications facing uncertainty, where the estimation of unknown parameters in these decision models often presents an important obstacle. This paper provides a comprehensive review of DFL and an in-depth analysis of various techniques used to integrate machine learning and optimization models, proposes a taxonomy of DFL methods classified according to their characteristics, and conducts extensive empirical evaluation of these methods. Benchmark datasets and tasks suitable for DFL. Finally, the study provides valuable insights into current and potential future directions of DFL research.
6.Group Activity Recognition in Computer Vision: A Comprehensive Review, Challenges, and Future Perspectives paper details page
https://www.aminer.cn/pub/64c09a9c3fda6d7f06e3e93c/
This paper mainly discusses the problem of group activity recognition in computer vision. Group activity recognition can effectively simulate hierarchical relationships in scenes by identifying group relationships, and accurately extract discriminative spatio-temporal features from groups, which has broad application prospects. The paper first reviews the relevant literature and different methods for group activity recognition, including traditional methods and state-of-the-art methods based on spatial structure, descriptors, non-deep learning, hierarchical recurrent neural networks, relational models, and attention mechanisms. Next, the paper introduces the relational network and relational architecture of each module. The paper then explores methods for group activity recognition and compares their performance to state-of-the-art techniques. The paper summarizes existing challenges and provides a comprehensive guide for newcomers to understand group activity recognition. Finally, the paper also reviews new directions and possibilities for group activity recognition.
7.FacTool: Factuality Detection in Generative AI – A Tool Augmented Framework for Multi-Task and Multi-Domain Scenarios 论文详情页
https://www.aminer.cn/pub/64c09a9c3fda6d7f06e3e92d/
The paper points out the advantages of generative pre-trained models in synthesizing high-quality text, but also presents the challenge of identifying factual errors in the generated text. Specifically, the paper points out the following issues: (1) As generative models handle more and more diverse tasks, the risk of containing factual errors is also increasing. (2) The generated texts tend to be long and lack the fine-grainedness of well-defined facts for individual facts. (3) Lack of clear evidence during the fact-checking process. With the above challenges in mind, this paper proposes FacTool, a task- and domain-independent framework (e.g. ChatGPT) for detecting factual errors in text generated by large language models. The paper demonstrates the effectiveness of the proposed method by conducting experiments on four different tasks including knowledge-based question answering, code generation, mathematical reasoning, and scientific literature review.
8.Analyzing Chain-of-Thought Prompting in Large Language Models via Gradient-based Feature Attributions 论文详情页
https://www.aminer.cn/pub/64c09a9c3fda6d7f06e3e869/
The paper points out that while generative pre-training models produce high-quality text, they also bring the challenge of identifying factual errors in the generated text. Specific issues are: (1) As more kinds of tasks are handled by generative models, there is an increased risk of including factual errors. (2) The generated texts tend to be longer and lack well-defined factual granularity. (3) Lack of clear evidence during the fact-checking process. In view of the above issues, this paper proposes a task- and domain-oriented framework, FacTool, for detecting factual errors in text generated by large-scale language models such as ChatGPT. Experiments on four different tasks (knowledge-based question answering, code generation, mathematical reasoning, and scientific literature review) demonstrate the effectiveness of the approach.
9.Strivec: Sparse Tri-Vector Radiance Fields paper details page
https://www.aminer.cn/pub/64c09a963fda6d7f06e3e1eb/
The paper proposes a novel neural representation called Strivec for modeling 3D scenes as radiation fields with sparsely distributed and compact factorized local tensor feature grids. The method utilizes tensor decomposition to model tensor grids. Unlike the recent TensoRF method, which uses global tensors and focuses on vector-matrix decomposition, the Strivec method utilizes a set of local tensors and applies the classical The CANDECOMP/PARAFAC (CP) decomposition decomposes each tensor into three vectors capable of representing the local feature distribution along the spatial axis and compactly encoding the local neural field. At the same time, the authors also apply multi-scale tensor grids to discover geometric and appearance commonalities, and exploit the spatial coherence of multi-scale three-vector decomposition. Finally, properties of the radiation field are regressed by aggregating neural features from multiple local tensors at multiple scales. These three-vector tensors are sparsely distributed around the actual scene surface, discovered by fast coarse reconstruction, exploiting the sparsity of 3D scenes. Experiments demonstrate that our model can achieve better rendering quality with significantly fewer parameters than previous methods, including TensoRF and Instant-NGP.
How to use ChatPaper?
The method of using ChatPaper is very simple. Open the AMiner homepage and enter the ChatPaper page from the navigation bar at the top of the page or the lower right corner.
On the ChatPaper page, you can choose to have a dialogue based on a single document or a dialogue based on the entire library (personal library), and you can choose to upload a local PDF or directly search for documents on AMiner.