Insights into LLM Long-Context Failures: When Transformers Know but Don't Tell

Taiming Lu, Muhan Gao, Kuai Yu, Adam Byerly, Daniel Khashabi
6/20/2024
cs.CL

No official code found. Best implementation: TaiMingLu/know-dont-tell (19 stars, last commit 2024-10-14).

Abstract

Large Language Models (LLMs) exhibit positional bias, struggling to utilize information from the middle or end of long contexts. Our study explores LLMs' long-context reasoning by probing their hidden representations. We find that while LLMs encode the position of target information, they often fail to leverage this in generating accurate responses. This reveals a disconnect between information retrieval and utilization, a "know but don't tell" phenomenon. We further analyze the relationship between extraction time and final accuracy, offering insights into the underlying mechanics of transformer models.

View on arXivView PDF

Code Implementations(1)

TaiMingLu/know-dont-tellCommunity100%
191PythonJun 18, 20242 years ago

Similar Papers

Positional Biases Shift as Inputs Approach Context Window Limits

86% match2025

Distance between Relevant Information Pieces Causes Bias in Long-Context LLMs

86% match1 repo142024

LLMs Explain't: A Post-Mortem on Semantic Interpretability in Transformer Models

84% match2026

Context Length Alone Hurts LLM Performance Despite Perfect Retrieval

84% match2025

Why Does the Effective Context Length of LLMs Fall Short?

83% match1 repo822024

Where is the answer? Investigating Positional Bias in Language Model Knowledge Extraction

83% match1 repo12024

Cite this paper

@article{lu2024insights,
  title  = {Insights into LLM Long-Context Failures: When Transformers Know but Don't Tell},
  author = {Taiming Lu and Muhan Gao and Kuai Yu and Adam Byerly and Daniel Khashabi},
  year   = {2024},
  eprint = {2406.14673},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2406.14673v2}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2406.14673.svg)](https://codeofpaper.com/paper/2406.14673)

Discussion