?? The End of Lazy LLMs

In this issue:

1. Active Learning for passive LLMs
2. 2 heads, 1 task
3. CRAG yourself before you wreck yourself

Want to market your brand? I’ve been personally using passionfroot for months now and as much as 40% of my partnerships can be accounted to their platform. They make it easy for companies to find fitting creators for their brand and I’ve found their streamlined collaboration process to be more efficient and more enjoyable for both sides.

Become a Sponsor

1. Efficient Exploration for LLMs

Watching: Double TS (paper)

What problem does it solve? Enhancing large language models (LLMs) requires copious amounts of feedback—a costly and time-consuming process. To streamline this, researchers continuously strive to optimize the way such feedback is collected. The crux of the problem lies in efficiently generating queries that elicit the most informative feedback to improve models with the least possible number of queries, a challenge that's critical for refining these models while conserving resources.

How does it solve the problem? Double Thompson sampling, in conjunction with an epistemic neural network—an artificial neural network capable of uncertainty estimation—offers a solution. This technique operates by generating queries based on a model of uncertainty within the learning agent itself. By sampling from this distribution of uncertainty, the agent can craft more informative queries. The combination of an epistemic neural network for uncertainty estimation and a strategic exploration scheme enables efficient collection of high-quality human feedback, thereby reducing the number of needed queries while maintaining or even enhancing the performance of LLMs.

What's next? This research sets the stage for further development of efficient data collection methods and more sophisticated exploration schemes, ensuring faster and cheaper refinements to LLMs. The exciting question now is how the incorporation of such exploration schemes will generalize across different model architectures, tasks, and domains. The ability to enhance model performance with fewer queries opens up possibilities for more intelligent, adaptive models that can self-improve with minimal human intervention.

2. Two Heads Are Better Than One: Integrating Knowledge from Knowledge Graphs and Large Language Models for Entity Alignment

Watching: LLMEA (paper)

What problem does it solve? Entity alignment is crucial for building extensive and interconnected Knowledge Graphs (KGs), which in turn are foundational for tasks in semantic web and artificial intelligence. The challenge in entity alignment lies in identifying correspondences across different KGs, considering the heterogeneous information they harbor—including structural, relational, and attributive data. Currently, embeddings generated to represent entities' multi-faceted information are difficult to match due to their dissimilarity and the lack of effective mechanisms to leverage these embeddings. Moreover, the potential of leveraging the nuanced semantic understanding of Large Language Models (LLMs) for this purpose has remained untapped.

How does it solve the problem? The proposed Large Language Model-enhanced Entity Alignment framework (LLMEA) addresses the shortcomings of previous entity alignment methods by harnessing the semantic knowledge inherent in LLMs. LLMEA operates by: 1) identifying candidate alignments based on embedding similarities and edit distances, and 2) using LLMs' inference abilities to iteratively process multi-choice questions that pinpoint the final aligned entity. This approach marries structural knowledge from KGs with the rich semantic understanding from LLMs, thereby creating a more nuanced and accurate alignment. By interpreting candidate entities through an LLM and using its robust inference capabilities, the framework can dynamically utilize the context and implicit knowledge that LLMs have been trained on.

What’s next? The results from the public datasets where LLMEA outperforms leading baseline models are promising, but it will be interesting to see how this framework performs at a larger scale, or with KGs that have sparse or noisy information. Future developments might include refinements in candidate selection processes or additional integration of LLMs for ongoing knowledge discovery and alignment tasks.

3. Corrective Retrieval Augmented Generation

Watching: CRAG (paper)

What problem does it solve? Large language models (LLMs) sometimes make errors reminiscent of hallucinations where the text generated does not correspond to reality or lacks accuracy, a problem not always solved by the model's inherent knowledge. While retrieval-augmented generation (RAG) strategies help to offer contextually relevant content by pulling from external databases, RAG's effectiveness is compromised if the quality of these external sources is poor. The primary concern this paper addresses is how to enhance the reliability of information generated by LLMs, especially when the RAG approach brings in suboptimal or irrelevant information.

How does it solve the problem? The proposed method, Corrective Retrieval Augmented Generation (CRAG), incorporates a novel retrieval evaluator that gauges the quality and relevance of documents fetched by the RAG system. Based on the confidence score provided by the evaluator, CRAG can adapt its retrieval strategies accordingly—thereby improving the quality of the generation. Furthermore, by incorporating large-scale web searches, CRAG expands beyond static databases to enhance its source material further. It also employs a decompose-then-recompose algorithm to sift through the fetched documents, emphasizing crucial details and discarding irrelevant content. Its "plug-and-play" nature also ensures that CRAG can be used alongside a multitude of existing RAG-based systems.

What's next? Moving forward, it would be fascinating to see how CRAG's introduction to various contexts impacts the adaptation of LLMs in more complex and nuanced environments, such as legal or technical domains where accuracy is paramount. Testing CRAG's effectiveness in these areas could set new standards for information retrieval in LLMs and possibly pave the way for even more advanced systems that combine the judgment and robust reasoning of AI with the ever-expanding knowledge of the internet. Evaluation in real-world scenarios will be crucial to understanding the limits and possibilities of CRAG-enhanced language models.

Papers of the Week:

• F-Eval: Asssessing Fundamental Abilities with Refined Evaluation Methods
• ChartBench: A Benchmark for Complex Visual Reasoning in Charts
• The Power of Noise: Redefining Retrieval for RAG Systems
• LongHealth: A Question Answering Benchmark with Long Clinical Documents
• Learning to Retrieve In-Context Examples for Large Language Models
• StructChart: Perception, Structuring, Reasoning for Visual Chart Understanding
• Unlearnable Algorithms for In-context Learning