Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

KIWI: A Dataset of Knowledge-Intensive Writing Instructions for Answering Research Questions

Fangyuan Xu, Kyle Lo, Luca Soldaini, Bailey Kuehl, Eunsol Choi, David Wadden

TL;DR

This work constructs KIWI, a dataset of knowledge-intensive writing instructions in the scientific domain, and finds that KIWI will be a valuable resource to measure progress and improve LLMs' instruction-following capabilities for knowledge intensive writing tasks.

Abstract

Large language models (LLMs) adapted to follow user instructions are now widely deployed as conversational agents. In this work, we examine one increasingly common instruction-following task: providing writing assistance to compose a long-form answer. To evaluate the capabilities of current LLMs on this task, we construct KIWI, a dataset of knowledge-intensive writing instructions in the scientific domain. Given a research question, an initial model-generated answer and a set of relevant papers, an expert annotator iteratively issues instructions for the model to revise and improve its answer. We collect 1,260 interaction turns from 234 interaction sessions with three state-of-the-art LLMs. Each turn includes a user instruction, a model response, and a human evaluation of the model response. Through a detailed analysis of the collected responses, we find that all models struggle to incorporate new information into an existing answer, and to perform precise and unambiguous edits. Further, we find that models struggle to judge whether their outputs successfully followed user instructions, with accuracy at least 10 points short of human agreement. Our findings indicate that KIWI will be a valuable resource to measure progress and improve LLMs' instruction-following capabilities for knowledge intensive writing tasks.

KIWI: A Dataset of Knowledge-Intensive Writing Instructions for Answering Research Questions

TL;DR

This work constructs KIWI, a dataset of knowledge-intensive writing instructions in the scientific domain, and finds that KIWI will be a valuable resource to measure progress and improve LLMs' instruction-following capabilities for knowledge intensive writing tasks.

Abstract

Large language models (LLMs) adapted to follow user instructions are now widely deployed as conversational agents. In this work, we examine one increasingly common instruction-following task: providing writing assistance to compose a long-form answer. To evaluate the capabilities of current LLMs on this task, we construct KIWI, a dataset of knowledge-intensive writing instructions in the scientific domain. Given a research question, an initial model-generated answer and a set of relevant papers, an expert annotator iteratively issues instructions for the model to revise and improve its answer. We collect 1,260 interaction turns from 234 interaction sessions with three state-of-the-art LLMs. Each turn includes a user instruction, a model response, and a human evaluation of the model response. Through a detailed analysis of the collected responses, we find that all models struggle to incorporate new information into an existing answer, and to perform precise and unambiguous edits. Further, we find that models struggle to judge whether their outputs successfully followed user instructions, with accuracy at least 10 points short of human agreement. Our findings indicate that KIWI will be a valuable resource to measure progress and improve LLMs' instruction-following capabilities for knowledge intensive writing tasks.
Paper Structure (60 sections, 4 figures, 11 tables)

This paper contains 60 sections, 4 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Interaction Design
  3. Design goals
  4. Interaction process
  5. Generating the initial answer
  6. Issuing instructions
  7. Answer revisions
  8. Measuring answer quality
  9. Handling incorrect instruction-following
  10. Dataset Collection
  11. (Question, Document set) creation
  12. Source article selection
  13. Question creation and filtering
  14. Analysis of collected questions
  15. Interaction data collection
  16. ...and 45 more sections

Figures (4)

  • Figure 1: An example interaction session between a user and our system. Given a question $q$ and a set of documents $\mathbf{D}$, the model first generates an initial answer $a_0$. At each turn $t$, the user issues an instruction $i_{t}$, for which an updated answer $a_{t}$ is generated. The user provides a rating $r_{t}$ and a comment $c_{t}$ for the generated answer, and optionally edits the answer (such as $a_{3}'$ in turn 3) such that the edited answer completely follows the instruction.
  • Figure 2: Top: distribution of annotator ratings. The left columns represent ratings for stylistic instructions (s), the middle columns for info-seeking instructions (i) and the right columns for overall (o) ratings. Bottom: distribution of user-issued instructions across the session.
  • Figure 3: We present the workflow for collecting KIWI. Detailed description for each step is in § \ref{['sec:task_setup']}.
  • Figure 4: A screenshot of our annotation interface.