Even without changing model parameters, large language models have in-context learning skills and can complete the job even with a small number of instances. Due to its task-agnostic nature, one model can be used for many different tasks. In contrast, conventional techniques for task adaptation, including fine-tuning, change model parameters for each task. Even task-independent in-context learning is rarely chosen by practitioners. This is because they always perform worse than task-specific adaptive techniques. Most previous studies attributed this performance difference to LLM’s constrained context window, which can accommodate only a small number of task cases.
However, they show that a gap remains between in-context learning and fine-tuning techniques, even given the same task examples. This finding reveals whether the performance differences are a general constraint of task-independent strategies for adaptation or specific to in-context learning. Can we formulate a specific adaptation strategy that meets the following requirements?
• Task-agnostic: The same model applies universally to different activities.
• Quality: Achieves comparable accuracy to task-specific approaches across several of these tasks.
• Data scalability: Learning efficiency increases as the number of task instances increases. They start by looking at the causes of quality discrepancies.
They divide LLM’s in-context learning power into two components: obtaining effective task representations and performing probabilistic or inference on those representations. Is this gap caused by the lack of information in the representation, or is it caused by LLM’s inability to analyze the representation? We test this notion empirically by evaluating They conclude that LLMs have strong expressive power and that most of the quality disparity is caused by weak reasoning on the part of LLMs.
We also found that fine-tuning enhanced the base model on both axes, but primarily task-specific inference, contributing 72% of the performance gain. Surprisingly, most methods for narrowing the performance gap, such as prompt engineering and active sample selection, target LLM learned representations only. In contrast, their study explored alternative strategies for enhancing LLM reasoning skills. They refine his LLM using artificially created probabilistic reasoning tasks as a first step towards improving his reasoning skills. This method improves the model’s baseline in-context learning performance, but it also requires fine-tuning each LLM individually.
They go a step further and speculate about the potential for developing reasoning skills in a task- and model-independent manner. They demonstrate that a completely agnostic approach can be taken to enhance reasoning skills. In this study, researchers from Stanford University and Cornell University propose Tart to improve the reasoning ability of LLMs using synthetically taught reasoning modules. Only synthetically generated logistic regression problems are used by Tart to train the Transformer-based inference module, regardless of the downstream task or base LLM. Without further training, this inference module can be built using LLM embeddings to enhance its deduction power.
In particular, Tart achieves the required goals.
• Task-neutral: Tart’s inference module needs to be trained once using fictitious data.
• Quality: Outperforms basic LLM overall and uses task-specific fine-tuning techniques to fill in gaps.
• Data Scalable: Handles 10x more instances of in-context learning.
Tart is task, model and domain independent. Using a single inference module trained on synthetic data, they demonstrated that Tart generalizes across three model families across 14 NLP classification tasks, as well as across different domains. They found that Tart’s performance, in terms of quality, was 18.4% for in-context learning, 3.4% for task-specific adapters, and 3.1% for full task-specific fine-tuning across a range of NLP tasks. shows that it is excellent.
In the RAFT benchmark, Tart brought GPT-Neo’s performance to match GPT-3 and Bloom, outperforming the latter by 4%. Tart solves the inconveniently short context duration barrier of in-context learning and is data-scalable. In LLM, each example uses multiple tokens (often hundreds), while Tart’s inference module uses only two tokens per case (one for context and one for label). To do. The benefit from this data scalability can reach 6.8%. Theoretically, Tart’s generalization skill has been shown to depend primarily on the distributional shift between the synthetic data distribution and the natural text embedding distribution, as assessed by the Wasserstein-1 metric. I’m here.
Here’s an overview of their main contributions:
• Investigate why task-specific fine-tuning outperforms in-context learning while accessing the same information using Representational Inference Decomposition.
• Present Tart is a new task-agnostic approach that performs better than task-specific approaches and does not require real data for training.
• Prove that Tart is valid for different model families across NLP tasks. The same inference module applies to the audio and visual domains as well.
please check out paper and Github link.don’t forget to join 24,000+ ML SubReddit, Discord channeland email newsletterShare the latest AI research news, cool AI projects, and more. If you have any questions regarding the article above or missed something, feel free to email me. Asif@marktechpost.com
🚀 Check out 100’s of AI Tools at the AI Tools Club
Aneesh Tickoo is a consulting intern at MarktechPost. He is currently pursuing his Bachelor of Science in Data Science and Artificial Intelligence from the Indian Institute of Technology (IIT), Bhilai. He spends most of his time working on projects aimed at harnessing the power of machine learning. His research interest is image processing and he is passionate about building solutions around it. He loves connecting with people and collaborating on interesting projects.
