Summary of Deniahl: In-context Features Influence Llm Needle-in-a-haystack Abilities, by Hui Dai et al.
DENIAHL: In-Context Features Influence LLM Needle-In-A-Haystack Abilities
by Hui Dai, Dan Pechi, Xinyi Yang, Garvit Banga, Raghav Mantri
First submitted to arxiv on: 28 Nov 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The Needle-in-a-haystack (NIAH) test evaluates language models’ ability to recall specific information from long input contexts. However, this framework lacks a mechanism to analyze the factors contributing to language models’ abilities or limitations in separating needles from their haystacks. To address this gap, we developed the DENIAHL (Data-oriented Evaluation of NIAH for LLM’s) synthetic benchmark, which ablates features beyond typical context length, including data type, size, and patterns. Our study reveals stark differences between GPT-3.5 and LLaMA 2-7B’s performance on DENIAHL, with drops in recall performance when item size is increased and to some extent when data type changes from numbers to letters. These findings have implications for increasingly large context models, demonstrating that factors beyond item-number impact NIAH capabilities. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Language models are getting better at remembering specific information from long texts. But how do they do it? One way to test this is by looking for a “needle” (a specific piece of text) in a “haystack” (a longer piece of text). However, current tests don’t tell us which features of the language model make it good or bad at finding these needles. To fix this, we created a new test called DENIAHL that looks at different types and sizes of texts to see how well language models do. Our results show that two popular language models, GPT-3.5 and LLaMA 2-7B, perform very differently on this test. We also found that when the size or type of text changes, some language models get worse at finding needles. This matters because as language models get bigger, we need to know which features make them better or worse. |
Keywords
» Artificial intelligence » Context length » Gpt » Language model » Llama » Recall
