Evaluating Human-Language Model Interaction

1 Introduction

• HALIE evaluates interactive systems and interaction traces across targets, perspectives, and criteria, extending assessment beyond outputs, third-party views, and quality alone.The framework distinguishes the LM from a system comprising an LM, user interface, and system logic.Source: §1 · p. 2, Figure 1 · p. 2
• The study examines five tasks spanning goal-oriented to open-ended interaction and compares three GPT-3 variants with AI21 Labs’ Jurassic-1 Jumbo.The models were selected to study model size, instruction tuning, and implementation differences.Source: §1 · p. 2
• Across 1,015 English interaction traces, better non-interactive performance did not always lead to better human-LM interaction.Even the worst-performing model in non-interactive settings could perform as well as or better than other models interactively.Source: §1 · p. 3
• The authors propose HALIE to structure the multidimensional space of human-LM interaction evaluation and instantiate it across varied tasks.They describe this range of tasks as demonstrating the framework’s adaptability and versatility.Source: §1 · p. 3
• The authors provide guidelines for applying HALIE to new scenarios and publicly release the data.Source: §1 · p. 3

2 Framework

• HALIE frames interactive evaluation around tasks, systems, interaction traces, and metrics, rather than treating evaluation as a model-completion assessment alone.The framework describes these components as the basis for studying human-LM interaction.Source: §2 · p. 3

2.1 Solving Tasks Interactively

• The five tasks span goal-oriented to open-ended interactions and concretize common LM query types into well-specified study settings.The selected tasks are social dialogue, question answering, crossword puzzles, text summarization, and metaphor generation.Source: §2.1 · pp. 3–4

2.2 Constructing an Interactive System

• Each interactive system combines an LM, a user interface, and system logic that defines states, user actions, and state transitions.When an action queries the LM, the system constructs a prompt, fetches completions, and displays them; the study holds UI design standard and leaves UI benchmarking to future work.Source: §2.2 · pp. 4–5

2.3 Evaluating Human-LM Interaction

• HALIE evaluates the full interaction process, the interacting user's first-person perspective, and quality and preference criteria across task-specific metrics.Preference measures include enjoyment and satisfaction, while metric designs may combine trace-based, output-based, user-survey, and third-party measures.Source: §2.3 · p. 6

2.4 Guidelines

• The framework is intended as a reference for identifying relevant system components and evaluation dimensions when designing an assessment for a new scenario.Its guidance highlights design choices in prompt construction, querying the LM, and presenting completions.Source: §2.4 · p. 6

2.4.1 How to apply the framework to a new task

• Applying HALIE involves defining the task, specifying the system, and selecting metrics across relevant combinations of targets, perspectives, and criteria.The story-writing example uses alternating user and LM sentences, specified interface actions and transitions, and user judgments of output usefulness by story position; pilots are recommended when user interaction is unknown.Source: §2.4.1 · pp. 6–7

2.4.2 When not to apply the framework

• HALIE assumes a human user interacting with an LM toward an evaluation goal; passive output assessment and LM-as-judge settings are outside scope, while multi-user or multi-LM settings may need extensions.The framework is designed for a single user interacting with an LM.Source: §2.4.2 · p. 7

3.1.1 Targets: Output vs. Process

• In QA, models that performed worst non-interactively sometimes matched or surpassed stronger models interactively; in metaphor writing, output-edit effort and process-effort measures favored different models.TextBabbage had the lowest zero-shot accuracy but performed best interactively in Nutrition; on US Foreign Policy, Jumbo beat Davinci interactively despite weaker zero-shot performance. TextDavinci required the fewest edits overall, while Davinci required the least writing time and queries.Source: §3.1.1 · p. 8

3.1.2 Perspectives: Third-party vs. First-person

• In summarization, annotator judgments differed from users’ evaluations, while crossword users sometimes perceived models as more helpful than their accuracy results warranted.Users significantly preferred TextDavinci, but it significantly improved letter accuracy on only one of five puzzles; on another, both instruction-tuned models reduced accuracy, possibly through confident, fluent misinformation.Source: §3.1.2 · p. 8

3.1.3 Criteria: Quality vs. Preference

• Dialogue users rated TextDavinci highest on most quality metrics, yet showed similar inclination to continue with Davinci, whose responses were most specific.The authors hypothesize that specificity may explain this inclination and recommend further study of its relationship to continued interaction and reuse.Source: §3.1.3 · p. 9, Table 2 · p. 10

3.2 Social Dialogue

• Instruction tuning improved most dialogue quality metrics but reduced specificity; even smaller TextBabbage scored higher in humanness than Davinci or Jumbo.The authors suggest general instruction tuning may make responses more aligned with common metrics at the cost of becoming generic.Source: §3.2 · p. 11

3.3 Question Answering

• TextDavinci assistance beat users alone across QA categories, but TextDavinci and Davinci zero-shot systems outperformed their interactive versions on US Foreign Policy.Across questions, TextDavinci-assisted users also had the highest accuracy, least effort, and highest perceived fluency and helpfulness when using AI assistance.Source: Figure 5 · p. 13, Table 3 · p. 13, §3.3 · pp. 14–15
• In QA, models with the worst zero-shot accuracy could outperform stronger models interactively: TextBabbage tied TextDavinci for Nutrition, while Jumbo surpassed Davinci for US Foreign Policy.The authors suggest that interactive use cases may require retrieving relevant information rather than directly answering questions.Source: §3.3 · p. 15
• TextDavinci was most efficient and accurate in QA at 1.78 queries/question and 69% accuracy; Davinci required 1.5x as many queries and achieved 48% accuracy.Despite its smaller size, TextBabbage outperformed Davinci on most metrics, and instruction-tuned models received the most favorable first-person survey evaluations.Source: §3.3 · p. 15
• QA users increasingly included answer choices and used fewer question prompts over time, while becoming less likely to copy exact MMLU question text.Survey responses described users discovering strategies such as including potential answers or phrasing prompts as declarative statements.Source: §3.3 · pp. 15–16

3.4 Crossword Puzzles

• TextDavinci was significantly preferred for crossword helpfulness on all puzzles, but improved accuracy only on SIT and reduced it for both instruction-tuned models on NYT-1.Across clue categories, models—especially TextDavinci—were most useful for factual clues but struggled with wordplay and phrases.Source: Figure 8 · p. 18, Table 4 · p. 18, §3.4 · p. 18
• Crossword users encountered misinformation across models, especially from TextBabbage, and short, ungrammatical prompts were associated with toxic outputs that could cause psychological harm.Toxic outputs were more common for Davinci (12) and Jumbo (8) than TextBabbage (5) and TextDavinci (3); longer prompts might mitigate them, but this requires further study.Source: §3.4 · pp. 18–20
• In crossword solving, users gradually reduced question-style prompts and increased short keyword prompts, with prompt-to-clue edit distance decreasing particularly for TextDavinci.Users also reported learning strategies such as using synonyms and declarative phrasing, and verifying responses by asking in different ways.Source: Figure 9 · p. 19, §3.4 · p. 21
• Crossword users showed varied engagement: some stopped querying after incorrect responses, while others queried later or sustained interaction, often rephrasing prompts; TextDavinci was rated more enjoyable.Some users found figuring out how to use the AI Teammate entertaining in itself.Source: Table 4 · p. 18, §3.4 · pp. 20–21, Figure 10 · p. 20

3.5 Text Summarization

• The summarization task has users edit model-generated summaries for consistency, relevance, and coherence, while previous human-edited summaries serve as in-context examples for later documents.The study randomly selected 964 XSum documents and assembled 20 sessions with ten documents each.Source: §3.5 · p. 21
• TextDavinci was most helpful to users, with the smallest edit distance and revision score, while third parties rated TextBabbage most consistent and relevant; TextBabbage summaries had density 16.11 versus 4.19 for TextDavinci, 4.86 for Davinci, and 4.09 for Jumbo.This discrepancy highlights that commonly used quality measures and what users find helpful when editing summaries can diverge.Source: Table 6 · p. 23, Table 7 · p. 23, §3.5 · p. 24
• Users edited less with TextDavinci over time but more with TextBabbage, and rated TextDavinci as improving most, leading the authors to hypothesize better in-context learning of user preferences.The summarization prompt concatenated prior document and user-edited summary pairs, enabling the study of learning from user examples.Source: §3.5 · pp. 22, 24, Figure 12 · p. 24

3.6 Metaphor Generation

• The metaphor study asked 32 users to write sentences from seed metaphors in 10-minute sessions, assessing user experience and third-party ratings of aptness, specificity, and imageability.Users could request five model suggestions at a time, while third-party evaluators assessed each sentence independently.Source: §3.6 · pp. 25–26
• Davinci required the least time and fewest queries for metaphor writing, had the highest suggestion acceptance rate, and was perceived as easiest to use and most likely to be reused.The authors hypothesize that users could quickly find acceptable sentences with fewer queries; third-party evaluators also rated Davinci sentences most highly.Source: Table 8 · p. 27, §3.6 · p. 27
• TextDavinci was rated most helpful and satisfying for metaphor generation, yet third parties rated its sentences worst; the authors caution that the small samples limit strong conclusions.They suggest satisfaction and reuse may depend on different factors, such as helpfulness and ease, depending on context.Source: Table 9 · p. 27, §3.6 · p. 27

• Across language-generation and LM benchmarks, non-interactive completion evaluation remains the prevailing norm, including consolidated evaluations such as GEM.Source: §4 · p. 28
• Prior dialogue and story-generation studies target interaction traces but often foreground third-party quality over user preference and first-person experience.Source: §4 · p. 28
• Prior interaction-trace work has used traces to assess completion helpfulness, work time, and system language, ideation, and collaboration capabilities.Source: §4 · p. 28
• Interactive-system research spans search, recommendations, voice assistants, robotics, and creative expression; this work focuses on how LMs mediate user interaction with language-generation systems.Source: §4 · p. 28
• Some interaction research treats user feedback as a model-learning signal, whereas this paper focuses on broader user experience with LMs, including but not limited to summarization in-context learning.Source: §4 · pp. 28–29

5.1 General Challenges in Human Evaluation

• Human evaluation is costly, and subjective judgments make study design and workload important constraints.The authors argue that real users are needed at least once to establish ground truth and assess which automatic metrics correlate with user evaluation.Source: §5.1 · p. 29

5.2 General Challenges in Interactive Evaluation

• Interactive studies must account for user heterogeneity and sequential effects, which can confound results unless model order is controlled.Within-subject designs can help with individual differences but require advance model selection; large, diverse samples or one-model-per-user designs offer alternatives.Source: §5.2 · pp. 29–30
• Latency can shape interaction and user perception, and the experiments excluded some models that were too slow.The tested models had similar acceptable latency, and the authors found no significant preference for the faster ones.Source: §5.2 · p. 30
• Benign prompts elicited toxic content, and the blocklist used in the experiments was only a mild safeguard.The authors call for more extensive safeguards in productionized applications.Source: §5.2 · p. 30

5.3 Limitations Specific to Our Experiment Design

• The five selected tasks are case studies, so findings may not cover the many other tasks and domains where language models are used.The authors identify extending evaluation to additional tasks and novel interaction patterns as future work.Source: §5.3 · p. 30
• Standard interfaces and shared prompts and decoding settings supported model comparisons, but different interfaces or system logic could change interaction outcomes.Crowd-worker recruitment may also limit generalization to actual users, and task studies used different within- or between-subject designs.Source: §5.3 · pp. 30–31

5.4 Future Work

• Users adapted their querying strategies over time, motivating study of longer-term interactions.Accommodation differed by task: QA users reformulated more, while crossword users increasingly copied clue text.Source: §5.4 · pp. 31–32

B.1 Social Dialogue

• Open-domain dialogue lacks a standard evaluation approach because many responses can be appropriate and quality judgments are subjective.The discussion notes that automated metrics may not capture the breadth of possible interactions and have shown little correlation with human judgments.Source: §B.1 · p. 46

B.2 Question Answering

• Interactive QA generally outperformed a human or an LM alone in prior work, and this study examines four models with hundreds of users rather than one model and five trained users.The authors also report statistically significant patterns in relative performance across human-LM interactions.Source: §B.2 · p. 46
• Search-query datasets capture realistic prompts but omit interaction quality, user satisfaction, and changes in querying or use of system outputs over time.These evaluations do not account for users’ ability to adapt their query styles or adjust how they trust and use outputs in decisions.Source: §B.2 · pp. 46–47
• Query reformulation is the iterative re-expression of requests to improve system utility, a behavior studied in search logs and automatic reformulation methods.The cited work includes lexical changes and broader strategies such as specialization.Source: §B.2 · p. 47
• The authors describe their study as the first, to their knowledge, of effective querying strategies using real-user data from an LM-based interactive system.They also connect the task to how output factuality, faithfulness, fluency, and tone may relate to user trust and behavior.Source: §B.2 · p. 47

B.3 Crossword Puzzles

• The section frames interactive crossword solving as part of the broader study of human-AI collaborative games.It directs readers to the question-answering discussion for related shared aspects.Source: §B.3 · p. 47
• In cooperative games, success involves engagement and enjoyment as well as task completion, and creativity may trade off with coherence.A cited text-adventure study found creativity of an AI-generated setting was anti-correlated with coherence.Source: §B.3 · p. 47
• Users in cooperative games can overestimate an AI teammate’s knowledge, while greater user agency may expose them to unexpected toxic outputs.The discussion cites an example of inappropriate output in AI Dungeon and notes similar behavior in the authors’ system.Source: §B.3 · pp. 47–48

B.4 Text Summarization

• Generated summaries can be coherent and fluent yet contain information unsupported by their source documents.Prior work has addressed faithfulness through training-data interventions, consistency losses, and candidate-summary discriminators; this work instead studies interaction in in-context learning.Source: §B.4 · p. 48
• Prior interactive summarization work focused on querying for more information in multi-document settings, whereas this work studies user correction of single-document summaries.Source: §B.4 · p. 48
• Rather than fine-tuning the LM, users’ edits to generated summaries are incorporated into prompts for future interactions.Source: §B.4 · p. 48

B.5 Metaphor Generation

• Computational metaphor generation includes replacing literal phrases with figurative ones and searching knowledge graphs for connections between concepts.Prior comparisons of keyword-focused pipelines and few-shot prompting found decomposition improved successful generation, while few-shot-only generation had the highest coherency.Source: §B.5 · p. 48
• Metaphor evaluation often relies on human judgments of the figurative connection, imagery, and fit to the intended meaning.Prior work also evaluates how well a generated connection fits a seed metaphor in aptness and specificity.Source: §B.5 · pp. 48–49
• Collaborative metaphor writing raises questions of diversity and authorship: writers may value varied outputs, while perceived ownership depends on their mental model of the system.Users may experience the system as an overly involved partner or as a computational aid.Source: §B.5 · p. 49

C Crowdsourcing Details

• Participants were U.S.-based Mechanical Turk workers with approval rates above 97% and over 10,000 approved HITs; study review and informed consent were documented.Data-collection pay ranged from $3 to $10 per task, at around $18 or more per hour; third-party evaluation paid $0.35 per summary pair and $0.15 per metaphor.Source: C Crowdsourcing Details · p. 49

C.1 Social Dialogue

• Social-dialogue users discussed assigned scenarios as if they had experienced them, testing common-scenario understanding in one dataset and empathetic communication in the other.Participants were encouraged to converse naturally, ask questions, and introduce topics.Source: §C.1 · pp. 49–50

C.2 Question Answering

• Question-answering users could query an AI assistant, were encouraged to vary their prompts, and were told not to use search engines or switch tabs.Pilot-derived examples demonstrated successful and unsuccessful prompting.Source: §C.2 · p. 50

C.3 Crossword Puzzles

• Crossword users received instructions on interacting with the AI teammate and puzzle interface, including pilot-derived examples of clue prompts.The instructions remained accessible from the crossword interface.Source: §C.3 · p. 51

C.4 Text Summarization

• In summarization, users first rated a machine-generated summary for consistency, coherence, and relevance, then edited it and rated the revision across the same criteria.Consistency meant that all summary information could be inferred from the article; relevance concerned important information without redundancy, and coherence concerned organization and flow.Source: §C.4 · pp. 51–52

C.5 Metaphor Generation

• For metaphor generation, users wrote as many sentences as possible from a seed metaphor with optional AI suggestions, judging them for aptness, imaginability, and specificity.Instructions encouraged users to satisfy most criteria while acknowledging that not every sentence would meet all of them.Source: §C.5 · pp. 52–53

D.1 Methods for statistical analysis

• Main results used group-wise means and standard errors with a modified Tukey-Kramer post-hoc test for unequal group sizes.Calculations used Python scipy and R stats.Source: §D.1 · p. 53
• To check robustness against likely survey-sample bias, the authors supplemented the analysis with ordinary least squares regressions of model types against survey metrics.Reported table values were intercepts plus beta effect estimates; Bonferroni-corrected significance was used, and regression assumptions were checked using residuals and Q-Q plots.Source: §D.1 · p. 54

D.2 Social Dialogue

• The analyzed dialogue metrics were sensibility, specificity, humanness, interestingness, inclination, and reuse.The authors suggest specificity may influence inclination and identify survey-question refinement and mixed-effects analysis as directions for further research.Source: §D.2 · p. 54
• TextDavinci scored highest on sensibleness, humanness, interestingness, and reuse; Davinci led on specificity, while the two tied on inclination.Regression confirmed the reported pattern; the inclination difference between TextDavinci and Davinci was not significant, though Davinci significantly outperformed Jumbo.Source: §D.2 · p. 54

D.3 Question Answering

• The analysis tracked elapsed time, query count, and user correctness; the authors note that block randomization could help control potential assignment-related confounding.The accuracy analysis conditioned on whether a language model was used, and significant accuracy differences among some participant groups suggested sample bias.Source: §D.3 · pp. 54–55, Table 13 · p. 55
• AI assistance was significantly associated with higher accuracy across all models, and TextDavinci users had the highest accuracy, least time, and fewest queries.Accuracy differed significantly between participants assigned to TextDavinci and Davinci, suggesting potential confounding and a limitation of the results.Source: §D.3 · p. 55

D.4 Crossword Puzzle

• The analysis measured fluency, helpfulness, ease, and enjoyment, and tested whether results varied by prompt.No prompts beyond the arbitrarily chosen intercepts were significant, so the regression table was not included.Source: §D.4 · p. 55
• Controlling for error, TextDavinci had significantly better performance metrics than the other models, while Jumbo and Davinci remained difficult to distinguish.The residual and Q-Q plots showed influential outliers that might explain the lack of clarity around the weakest performers.Source: §D.4 · pp. 55–56
• Further comparisons could test other survey samples, regularization, and nonlinear relationships to clarify model differences.The authors note that these approaches might yield clearer comparisons.Source: §D.4 · p. 56

D.5 Text Summarization

• The appendix reports regression tables for automatic performance metrics, user survey responses, and quality annotations from first-person and third-party evaluation.The tables cover automatic metrics, user responses, and annotation-based quality evaluation, respectively.Source: Table 15 · p. 56, Table 16 · p. 56
• TextDavinci was most helpful and improved most with time, while TextBabbage required the most editing and Jumbo had the worst original summary and lowest helpfulness.Regression confirmed these patterns but did not distinguish TextBabbage from Jumbo.Source: §D.5 · p. 56
• Third-party evaluation rated TextBabbage higher than in the main paper, while users rated summaries higher after editing and evidence only slightly favored TextDavinci.The authors suggest wider samples could help assess TextDavinci's superiority.Source: §D.5 · p. 56

D.6 Metaphor Generation

• The appendix presents separate tables for automated metaphor metrics, user-survey responses, and third-party evaluations of metaphor quality.The user-survey table reports no significant differences among models except the reference model.Source: Table 18 · p. 57, Table 19 · p. 57, Table 20 · p. 57
• Only acceptance and edit distance showed significant relationships in the metaphor-generation regression; most user-survey model comparisons did not establish which models performed better or worse.The reported results indicate higher acceptance for Davinci suggestions and shorter edit distance, while other conclusions await further investigation and replication.Source: §D.6 · pp. 56, 58
• The appendix also lists how the full set of metrics maps to the paper's three evaluation dimensions.Source: Table 21 · p. 59