Jasper Playing Jazz, Jazzoo
Posts by Collection
portfolio
publications
Meta-DM: Applications of Diffusion Models on Few-Shot Learning
Published in arxiv preprint, 2023
In the field of few-shot learning (FSL), extensive research has focused on improving network structures and training strategies. However, the role of data processing modules has not been fully explored. Therefore, in this paper, we propose Meta-DM, a generalized data processing module for FSL problems based on diffusion models. Meta-DM is a simple yet effective module that can be easily integrated with existing FSL methods, leading to significant performance improvements in both supervised and unsupervised settings. We provide a theoretical analysis of Meta-DM and evaluate its performance on several algorithms. Our experiments show that combining Meta-DM with certain methods achieves state-of-the-art results.
Recommended citation: Hu, W., Jiang, X., Liu, J., Yang, Y., & Tian, H. (2023). Meta-DM: Applications of Diffusion Models on Few-Shot Learning. arXiv preprint arXiv:2305.08092. https://arxiv.org/pdf/2305.08092.pdf
AllTogether: Investigating the Efficacy of Spliced Prompt for Web Navigation using Large Language Models
Published in arxiv preprint, 2023
Large Language Models (LLMs) have emerged as promising agents for web navigation tasks, interpreting objectives and interacting with web pages. However, the efficiency of spliced prompts for such tasks remains underexplored. We introduces AllTogether, a standardized prompt template that enhances task context representation, thereby improving LLMs’ performance in HTML-based web navigation. We evaluate the efficacy of this approach through prompt learning and instruction finetuning based on open-source Llama-2 and API-accessible GPT models. Our results reveal that models like GPT-4 outperform smaller models in web navigation tasks. Additionally, we find that the length of HTML snippet and history trajectory significantly influence performance, and prior step-by-step instructions prove less effective than real-time environmental feedback. Overall, we believe our work provides valuable insights for future research in LLM-driven web agents.
Recommended citation: Liu, Jiarun, Wentao Hu, and Chunhong Zhang. "AllTogether: Investigating the Efficacy of Spliced Prompt for Web Navigation using Large Language Models." arXiv preprint arXiv:2310.18331 (2023). https://arxiv.org/pdf/2310.18331v1.pdf
BM24 at COLIEE 2024: Case and Statute Retrieval by Semantic Similarity
Published in , 2024
Automatic retrieval techniques are playing a more and more role in legal processes. COLIEE, well-known as an international competition interested in novel methods in legal text retrieval, releases Legal Case Retrieval as their task1 and Statute Law Retrieval as their task3. This paper reports team BM24’s methods to solve task1 and task3. We use matched case / statute pairs provided in the training set to fine tune an embedding model which is then used to build a vector store for semantic retrieval. Specifically, for task 1, we summarize the long case text into segments and choose one informative segment to represent the whole case text to preprocess the long case texts.
talks
[KG] A Survey on Knowledge Graphs
Published:
An overview of knowledge graph principles, KGE, knowledge acquisition, and major applications is presented.
[LM] Out of One, Many
Published:
Mainly prompt experiment, no large deal. Working on bias and skewed margin, now a popular field.
[RL] Hierarchical Reinforcement Learning: A Comprehensive Survey
Published:
The mainstream algorithms and developments in hierarchical reinforcement learning are described in detail, from theory to mainstream algorithms.
[LM] Mind’s Eye
Published:
The research proposes a new paradigm for large models to implement reasoning about the physical world Mind’s Eye, by getting information from real simulations of physical problems in the Mujoco physics engine, and inputting auxiliary information along with the problems themselves to enable large models, the improvement in UTOPIA baseline is significant and can be done so that small models plus simulations can outperform large models after the effect improvement is significant.
[LM] Chain-of-Thought
Published:
CoT is a form of discrete cue learning. More specifically, contextual learning of language models adds many textual logical representations of thought in between compared to the previous traditional contextual learning (i.e., a series of texts as input for the large model to complete the output).
[RL] A Survey of Hierarchical Reinforcement Learning
Published:
A paper writing try, also get to summrize the HRL knowledge.
[RL] Option-Critic
Published:
We have read an important paper on option-based hierarchical reinforcement learning, The Option-Critic Architecture, and have critically reviewed the derivation of the equations to understand the SMDP process in option-state augmented space and the corresponding algorithmic framework.
[LM] Challenging BIG-Bench tasks
Published:
CoT was already available when BIG-Bench first came out, but CoT did not perform well on small scale models (emergent effect could not be achieved), so BIG-Bench did not mention using CoT; but after that, PaLM / Davinci-002 / code-davinvi-002 and other larger scale models appeared. So there was a motivation to verify the effect of CoT on the new baseline of BIG-Bench. Sure enough, CoT is indeed better for many tasks.
[LM] The Debate Over Understanding in AI’s Large Language Models
Published:
The article describes what LLM is, how it is trained and works, and how it works, and then shows that LLM is also brittle and can make mistakes when disturbed. The article points out that there are two opposing voices, one arguing that this is the budding of general intelligence, and the other arguing that LLM only learns the form of language rather than its meaning. The article also shows, with some examples, that shortcut learning, a phenomenon often invoked in machine learning, is also present in LLM, i.e., learning systems that rely on spurious correlations in the data.
[LM][AGI] WebGPT: Browser-assisted question-answering with human feedback
Published:
The study fine-tuned the GPT-3 large model to achieve long-form QA by manipulating browser search, i.e., giving long and meaningful complete answers to open-ended questions. The result is that more than half of the generated results are more satisfying than the answers given by humans, with higher accuracy and information validity.
[ML] A Path Towards Autonomous Machine Intelligence
Published:
Different researchers have different perspectives on the ideas proposed by LeCun, but in general this paper is extremely enlightening and leading. The paper proposes an architecture and training paradigm for building autonomous intelligences, combining concepts such as configurable predictive world models, behavior driven by intrinsic motivation, and hierarchical joint embedding architectures trained by self-supervised learning.
[RL] DAC: The Double Actor-Critic Architecture for Learning Options
Published:
The option framework is reformulated as two parallel augmented MDPs. under this new formulation, all policy optimization algorithms are readily available for learning intra-option policy, termination policy, and master option. we apply AC algorithms on each augmented MDP and The DAC architecture is designed. Combined with the PPO algorithm, an empirical study is conducted on challenging robot simulation tasks.
teaching
The Law Large Language Model Project
Lab, LLM Grounding and Task-specific Tuning, 2023
This project mainly relies on the technical route of prompt learning and partial parameter fine-tuning of open-source large language models to realize a large-model intelligent assistant in the legal vertical field for professionals and the public, targeting legal documents, Q&A and legal data. I am responsible for framework design, deployment fine-tuning, evaluation session and interaction design. Check our project’s code, dataset and checkpoints in github.