PaperlayerView exact revision on arXiv

Source-linked AI summary

"I Didn't Make the Micro Decisions": Measuring, Inducing, and Exposing Goal-Level AI Contributions in Collaboration

Eunsu Kim, Jessica R. Mindel, Kyungjin Kim, Sherry Tongshuang Wu

arXiv:2605.21363v2cs.CL

TL;DR

Existing attribution methods focus on final artifacts and do not adequately distinguish whether AI executes instructions or shapes goals. CoTrace decomposes goals into verifiable requirements and traces human and AI influence through dialogue; across four tasks, humans accounted for 75–89% of SHAPER mass, while model goal-shaping increased with goal specificity.

  • Problem

    Existing attribution tools focus on final artifacts and do not adequately distinguish whether AI executes instructions or shapes goals, despite the need to assess AI-assisted work and calibrate reliance.

    Source: §1 · p. 1

  • Method

    CoTrace represents goals as desired outcomes with verifiable requirements and measures human and AI contributions throughout the collaboration process.

    Source: §1 · p. 2, §2 · p. 3

  • Results

    Humans account for 75–89% of SHAPER mass across four tasks, while LLM goal-shaping increases with specificity, from parent goals toward individual requirements.

    Source: §3.1 · p. 5

  • Takeaways & Limitations

    Interaction design and prompting can amplify or suppress goal shaping, making them relevant to calibrating AI initiative.

    Source: §7 · p. 10

    CoTrace requires interaction traces, so it is not a detector of AI contribution from final artifacts alone.

    Source: Limitations · p. 11

Abstract

from arXiv · show

As large language models (LLMs) increasingly shape how users form, refine, and extend their goals, attributing contributions in human-AI collaboration becomes critical for users calibrating their own reliance and for evaluators assessing AI-assisted work. Yet existing methods focus on final artifacts, missing the process through which goals themselves are jointly shaped. We introduce a goal-level attribution framework, CoTrace, that decomposes explicit goals into verifiable requirements and traces both direct contributions and indirect influences across dialogue turns. Applying CoTrace to 638 real-world collaboration logs, we find that while models account for only 11-26% of goal-shaping contribution, they contribute substantially more on introducing lower-level concrete requirements, and make various kinds of indirect contributions. Through controlled simulations, we show that interaction design choices significantly affect model goal-shaping behavior. In a user study, exposing participants to goal-level analyses shifts their perceived contributions by nearly 2 points on a 5-point scale, revealing systematic miscalibration in how users understand their own AI-assisted work.

1 Introduction

COTRACE measures human and AI contributions to task goals by tracing requirement creation, refinement, execution, and indirect influence across dialogue. The paper presents it as an evaluation, design, and reflection tool, with studies showing goal-shaping patterns in real-world logs, sensitivity to interaction choices, and shifts in users’ perceptions.

Motivation

  • COTRACE addresses the limits of outcome-focused attribution, which detects AI involvement in final artifacts but leaves goal-shaping initiative invisible to users and evaluators.The authors position process-level measurement as a basis for evaluating model autonomy, bounding AI initiative, and calibrating awareness of AI contributions.Source: §1 · pp. 1–2

COTRACE framework

  • COTRACE decomposes explicit task goals into granular, verifiable requirements and traces both direct contributions and indirect influence across dialogue.Indirect influence includes actions such as clarifying questions, draft artifacts, or exposed errors that lead the other party to formulate a requirement.Source: §1 · p. 2, Figure 1 · p. 2

As an evaluation suite

  • Across four domains, models largely follow user direction at the macro level but play a larger role in shaping specific requirements, especially in technical tasks.After initial turns, many requirements emerge through mutual influence; the analysis identifies 11 recurring patterns of indirect goal shaping.Source: §1 · p. 2

As a design tool

  • Controlled simulations show that interaction design choices and prompting strategies, including underspecification, substantially affect model goal-shaping behavior.The results suggest these choices can serve as design levers for changing how actively models contribute to goals.Source: §1 · p. 2

As a reflection tool

  • In a study with 10 participants, exposing users to goal-level analyses shifted their perceptions of both their own and the AI’s contributions.Some participants reflected on changing their prompting practices, suggesting the tool may support more intentional collaboration.Source: §1 · p. 2

2 COTRACE: Evaluation Framework for Quantifying Agents’ Goal-Level Contributions in Human–AI Collaboration

COTRACE measures human–LLM goal shaping by representing goals as outcomes linked to checkable requirements and tracing direct and indirect influences across dialogue actions. Its automated pipeline extracts and links goals, requirements, and actions, then aggregates influence into speaker-by-role contributions.

Goal representation

  • COTRACE represents each explicit, actionable goal as a desired outcome paired with independently checkable requirements.This outcome–requirement tuple is the framework’s unit of analysis (Qin et al., 2024; Viswanathan et al., 2025).Source: §2 · p. 3
  • Hierarchical parent and child goals make evolving objectives observable, while atomic actions link requirements to specific utterances.Requirements capture granular design decisions; actions are minimal communicative units and serve as the unit for requirement iteration (Kim et al., 2026).Source: §2 · p. 3

Influence model

  • COTRACE distinguishes direct influence, which explicitly introduces or modifies requirements, from indirect influence, where context later motivates them.Indirect influence includes cases where an AI suggestion or clarifying question prompts a human to develop a concrete requirement.Source: §2 · p. 3

Pipeline

  • The four-stage pipeline extracts outcomes and actions, versions requirements, labels action–requirement influence, and aggregates scores into speaker-by-role contributions.Actions are classified as SHAPER, EXECUTOR, or OTHER; requirements track CREATE, REVISE, and DELETE operations, and influence labels are DIRECT CONNECTION, IMPLICIT CONNECTION, or NO CONNECTION.Source: §2 · p. 3

Validation

  • Manual validation achieved over 90% accuracy, and participants rated alignment with their own perceptions above 4 out of 5 on average.The framework was assessed on goal extraction, requirement extraction, and influence labeling through sampled dialogues and participant review of their own analyses.Source: §2 · p. 4

Applications

  • The paper demonstrates COTRACE for measuring collaborative goal shaping, inducing goal-shaping behavior through interaction design, and exposing AI contributions to users.These applications analyze real-world logs, amplify or suppress model initiative, and aim to improve user awareness and reflection on collaboration practices.Source: §2 · p. 4

3 Measuring Collaborative Goal Shaping In the Wild

Applying CoTrace to 638 human–LLM collaboration logs, this section examines who shapes goals, how that shaping unfolds, and how it differs between chat-based and autonomous-agent settings.

3.1 “Who”: Humans set direction, but models shape the details and specificity

  • Humans account for 75–89% of SHAPER mass, while LLMs account for 96–99% of EXECUTOR mass; LLM goal-shaping contributions increase as goals become more specific.Models rarely shape parent outcomes but contribute more to child outcomes and individual requirements. They introduce more implementation-oriented requirements in technical tasks, while human goal-shaping contributions are relatively greater in open-ended tasks (p < .001).Source: Figure 2 · p. 4, §3.1 · p. 5

3.2 “How”: Goal shaping emerges through execution, not just explicit proposals

  • Indirect goal shaping accumulates after users’ initial requirements, reflecting co-construction through four recurring categories: underspecified intent, artifact-triggered elaboration, problem-triggered revision, and interactional steering.For user-created requirements indirectly influenced by assistants, artifact-triggered elaboration accounts for 60%, followed by underspecified intent/preference (13.3%) and interactional steering (10%). For assistant-created requirements indirectly influenced by users, the largest categories are underspecified intent/preference (46.7%) and interactional steering (36.7%).Source: §3.2 · pp. 5–6

3.3 Goal Shaping Across System Settings: Chat-Based Systems vs. Autonomous Agents

  • Chat-based systems contributed more requirements than autonomous agents in Academic Writing (33.11% vs. 5.33%), Data Analysis (47.03% vs. 5.56%), and Planning (37.74% vs. 18.35%; all p < .001).Despite greater autonomy, agents exercise less goal-shaping initiative.Source: §3.3 · pp. 6–7

4 Inducing and Evaluating Goal Shaping at Inference-Time

In COGYM simulations, prompting and communication-before-action increased assistant goal shaping, while greater requirement creation was essentially uncorrelated with normalized output quality. Assistant-created requirements were satisfied more often than user-created ones mainly because many were implemented immediately.

4 Inducing and Evaluating Goal Shaping at Inference-Time

  • The simulations used COGYM writing, travel-planning, and tabular-analysis tasks, comparing agentic and chat settings with Claude 4.5 Sonnet and Gemini 3.1 Pro.COTRACE measured goal shaping, while overall output quality and requirement satisfaction rate measured downstream outcomes.Source: §4 · p. 7

4.1 Interaction Design and Prompting Can Amplify Model Goal-Shaping

  • Assistant contribution to requirement generation rose from 30.65% to 69.64% with underspecification and to 51.47% with interaction steering.Both interventions decreased user-created requirements (p < .05); gains in assistant-originated requirements were consistent across chat and agentic settings, though smaller in the agentic setting.Source: Table 3 · p. 7, §4.1 · p. 7
  • Assistants contributed directly to 42.9% of requirements in CHAT-COGYM versus 24.5% in AGENTIC-COGYM (p < .001).CHAT-COGYM required a message before tool use, creating opportunities to articulate plans; agentic messages typically followed tool use and reported completed actions (70.8% of the time). CHAT-COGYM also produced longer interactions (17.0 vs. 13.7 turns; p < .001).Source: §4.1 · p. 8

4.2 Does Increased Goal-Shaping Improve Collaboration Outcomes?

  • More requirements created per session were essentially uncorrelated with normalized output quality (Pearson = −0.002, Spearman = −0.011).Assistant-created requirements were satisfied more often than user-created ones (75.1% vs. 62.7%), but 35.3% were implemented in the turn they were introduced; excluding those cases, satisfaction rates were nearly the same (61.5%).Source: §4.2 · p. 8

5 Exposing Goal-Level Dynamics to Users

COTRACE exposure shifted users’ perceptions of contributions and helped evaluators inspect LLM use in long interaction logs. Evaluators’ judgments still differed, leaving final evaluation to them.

5.1 Goal-Level Analysis Can Increase User Self-Awareness

  • After tool exposure, perceived own execution contribution fell (∆= −1.8), LLM execution rose (∆= +0.5), own goal-shaping rose (∆= +0.2), and LLM goal-shaping fell (∆= −0.4).Own and LLM goal-shaping perceptions changed by 1.0 and 1.6 points in absolute magnitude, respectively; LLM satisfaction changed by 0.6 points in absolute magnitude (∆= −0.2).Source: §5.1 · p. 9
  • Nine of ten participants said COTRACE revealed previously unnoticed collaboration aspects, including implicit LLM micro-decisions, and prompted reflection on making their own prompts more specific.Source: §5.1 · p. 9
  • Human–human SHAPER share ranged from 7.4% to 100% (σ ≈37.5 pp), versus 36.2%–92.0% (σ ≈17.9 pp) in human–LLM logs; indirect influence was 14.8% versus 25.3%.Source: §5.1 · p. 9

5.2 Goal-Level Analysis Supports Third-Party Evaluation

  • Evaluators revised their LLM-reliance estimates in three of four cases and said COTRACE helped them identify the extent and purposes of use more quickly than raw-log review.They reported that raw-log review was time-consuming or drew attention mainly to the interaction’s initial portion.Source: §5.2 · p. 10
  • Although evaluators identified broadly similar LLM-use patterns, their grading decisions diverged: one assigned Student A a score of 0, while the other retained 100.COTRACE surfaces interaction-level evidence for review while leaving final pedagogical or evaluative judgment to the evaluator.Source: §5.2 · p. 10

6 Related Work

Existing collaboration evaluation and AI-contribution attribution focus on predefined tasks or final artifacts, leaving how participants jointly shape evolving goals and indirectly influence one another underexplained. This work shifts attribution to the evolving goal structure, tracing direct goal shaping and indirect influence.

6 Related Work

  • Collaboration-evaluation frameworks generally assume predefined tasks, requirements, or criteria, while newer studies consider goals progressively specified through interaction (Kim et al., 2026).Source: §6 · p. 10
  • AI-assisted creation attribution traces who wrote or modified final-artifact spans, but remains outcome-oriented and misses how goals and requirements change during interaction (Siddiqui et al., 2025; Liang et al., 2024; Kumarage et al., 2023).Source: §6 · p. 10
  • This work shifts attribution from final artifacts to evolving goal structures, capturing direct goal shaping and indirect influence during human–AI collaboration.Source: §6 · p. 10

7 Conclusion and Implication

COTRACE traces how humans and AI jointly shape goals and requirements, serving as an evaluation, design, and reflection tool across three studies. The findings point to aligning collaboration with task quality, calibrating AI initiative through interaction design, and making goal-level contributions visible for reflection and third-party evaluation.

7 Conclusion and Implication

  • COTRACE moves beyond artifact-level analysis to trace how humans and AI jointly shape goals and requirements throughout collaboration.Across three studies, it served as an evaluation, design, and reflection tool for understanding and improving collaborative dynamics.Source: §7 · p. 10
  • Increasing model goal shaping does not necessarily improve final outcomes, so training and interventions should better align collaboration with task quality.Goal shaping can be amplified or suppressed through interaction design and prompting, making system design important for calibrating AI initiative.Source: §7 · p. 10
  • Goal-level attribution may support collaborators’ self-awareness and reflection, as well as third-party evaluation of work in education and creative fields where authorship, responsibility, and credit matter.COTRACE aims to make goal-level contributions visible in support of more transparent, accountable, and human-centered collaboration.Source: §7 · p. 10

Limitations

CoTrace requires consented interaction logs and provides analytical estimates, not ground-truth responsibility measures; the user-study findings may not generalize beyond its small sample of frequent AI users in a familiar domain.

Limitations

  • CoTrace requires interaction traces, so it cannot detect AI contributions from final artifacts alone and is most appropriate when logs are available with user consent.Potential settings include reflective self-assessment, classrooms, and review workflows.Source: Limitations · p. 11
  • CoTrace contribution scores are analytical estimates, not ground-truth measures of responsibility, quality, or acceptable AI use, and final judgments should remain with humans.A higher AI contribution score does not necessarily indicate better or worse task performance.Source: Limitations · p. 11
  • The small user-study sample of frequent AI users in a familiar domain limits generalizability, and revised judgments may partly reflect anchoring to interface scores.Participants’ qualitative responses also suggest they relied on linked goals, requirements, and rationales.Source: Limitations · p. 11

Ethics Statement … B COTRACE

The related-work sections motivate evaluating evolving goals through explicit, checkable requirements and attributing contributions to goal formation, not only final artifacts. The ethics statement reports IRB approval, participant consent and compensation, and human validation of LLM-as-a-Judge evaluations.

Ethics Statement

  • All human-subject studies received CMU IRB approval, participants gave consent and received compensation, and human verification validated LLM-as-a-Judge accuracy.The passage gives IRB Study Number STUDY2026 00000006 and notes Claude Code and Cursor were used to improve plot clarity based on original versions.Source: Ethics Statement · p. 11

A.1 Background: Goal Definitions

  • Goals are variously framed as desired end states, requested actions, or SMART objectives, reflecting different emphases on intention, action, and evaluation.The passage cites goal-setting theory, an action-centric human–AI interaction perspective (Shelby et al., 2025), and SMART (Doran, 1981; Michigan Department of Health and Human Services, n.d.).Source: §A.1 · p. 15
  • Because goals can be underspecified, evolve, and be only partially satisfied, the approach uses explicit, evolving, checkable success predicates grounded in observable actions and utterances.Prior work emphasizes clarification and tracking as prerequisites for reliable evaluation (Noseworthy et al., 2017; Ghose et al., 2025; Schöttle & Tillmann, 2018).Source: §A.1 · p. 15
  • Building on goal decomposition and responsibility assignment, the paper represents complex instructions as independently verifiable requirements and tracks goals that persist and evolve across turns.The passage relates this framing to van Lamsweerde (2001), InFoBench (Qin et al., 2024), and OnGoal (Coscia et al., 2025).Source: §A.1 · p. 15

A.2 Evaluating Human–AI Collaboration

  • Prior AI-contribution attribution centers on final artifacts or revision histories, leaving goal- and requirement-shaping contributions less visible.The paper shifts attention to how collaborative objectives are formed, including introducing constraints, proposing directions, surfacing latent requirements, and influencing goal formulation.Source: §A.2 · p. 15
  • The framework attributes direct and indirect contributions as goals evolve, addressing limits in prior collaboration evaluations that assume fixed goals.DiscoverLLM (Kim et al., 2026) considers evolving intent but does not explicitly distinguish direct goal shaping from indirect influence or attribute both across participants.Source: §A.2 · pp. 15–16

B.1 Implementation … B.3 Validation

The appendix defines goals through explicit outcomes and verifiable requirements, describes a pipeline for tracing action-level influence, and outlines validation through human review and participant comparisons.

B.1.1 Operational Definitions

  • Goals are explicit, actionable targets represented by intended outcomes and independently verifiable requirements whose changes are tracked as create, revise, or delete operations.Goals are not inferred from latent intentions; each requirement is a binary-evaluable success predicate.Source: §B.1.1 · pp. 16–17

B.1.2 Pipeline Overview

  • The automated four-stage pipeline extracts outcomes and actions, extracts requirements, labels influence, and quantifies contributions; GPT-5.1-mini handles action extraction and GPT-5.2 the remaining judgments.A separate prompt groups outcomes under high-level intentions.Source: §B.1.2 · p. 17, Step 1c: Intention Extraction · p. 19

Step 1a: Action Extraction

  • Actions are extracted as minimal communicative units, assigned a type and SHAPER, EXECUTOR, or OTHER role, and supported with an evidence quote.The prompt requests all actions from every turn.Source: Step 1a: Action Extraction · pp. 18–19

Step 1b: Outcome Extraction

  • Outcomes are purpose-linked deliverables, organized into parent and child outcomes, while every action is assigned to exactly one outcome.The extraction prompt excludes outputs that are merely mentioned and not adopted as goals.Source: Step 1b: Outcome Extraction · p. 19

Step 2: Requirement Extraction

  • Requirements must be binding, directly grounded, nonreplaceable, and pass/fail testable; the extraction prompt supports creating, revising, or deleting them.Advice itself and tentative suggestions are not requirements; only constraints on advice may qualify.Source: Step 2: Requirement Extraction · p. 20

Step 3: Influence Labeling

  • Influence labeling distinguishes direct and implicit connections from implementation, contribution, and no connection, with scores calibrated to relationship type.Direct connections and implementations receive scores from 4–5; implicit connections and contributions receive scores from 1–3.Source: Step 3: Influence Labeling · pp. 20–21

Deliverable Extraction

B.3 Validation

  • Validation combines human assessment of agreement with the framework’s analysis and user-study participants’ comparison of that analysis with their own task thinking.The two complementary validation forms are described in Appendices B.3.1 and B.3.2.Source: §B.3 · p. 22

B.3.1 Manual validation … Disagreement on Indirect Influence Label

Manual validation and participant feedback show that the tool’s analyses were generally supported, while revealing attribution errors and limitations in how users interpret its granular contributions.

B.3.1 Manual validation

  • For each extraction step, one author validated over 100 entities sampled from 37 dialogues, with accuracy reported in Table 4.The supplied passage does not include Table 4’s accuracy values.Source: §B.3.1 · p. 22, Table 4 · p. 22
  • Goal-extraction errors mainly treated requirement-level contributions as outcomes (4/6, 66.7%), with incorrect author attribution and implausible outcomes accounting for one error each.The remaining errors were incorrect author attribution (1/6, 16.7%) and generation of an implausible outcome (1/6, 16.7%).Source: §B.3.1 · p. 22
  • Requirement-extraction errors most often mistook artifact content for requirements (8/13, 61.5%); other errors involved assistant actions, attribution, minor details, or generation time.Each of those four other categories accounted for 1–2 of the 13 errors.Source: §B.3.1 · p. 22
  • Influence-labeling errors primarily misattributed authorship (6/9, 66.7%) and otherwise labeled weak relationships as influence links (3/9, 33.3%).An example misattribution gave the user primary credit when the model derived concrete requirements from the user’s broader goal.Source: §B.3.1 · pp. 22–23

B.3.2 Validation from User Study

  • Participants’ mean agreement exceeded 4 out of 5 for the tool’s goal hierarchy, requirements, and influence relations.Participants reviewed analyses of their own conversations and rated agreement on a Likert scale.Source: §B.3.2 · p. 23, Figure 5 · p. 23

Disagreement on Extracted Requirements

  • One participant found the graph or timeline’s granular contribution percentages confusing and suggested explaining how they were calculated.P7 questioned the distinction between 50% and 55% goal-shaping contribution.Source: Disagreement on Extracted Requirements · p. 23
  • Another participant said the tool did not capture the cognitive work of interpreting assistant outputs and choosing among alternatives.P8 described using suggestions to prompt ideas before selecting options, despite the tool indicating no contribution.Source: Disagreement on Extracted Requirements · p. 23

Disagreement on Indirect Influence Label

  • One participant considered suggestions such as summit or rooftop views too minor to count as meaningful indirect influence.P6 wanted a stronger “wow factor” from the assistant’s contribution.Source: Disagreement on Indirect Influence Label · p. 23

B.4 Sensitivity to LLM-as-a-Judge Errors

A Monte Carlo sensitivity analysis injected errors into judge outputs at rates up to 30%. All four key findings remained intact across the tested rates, with none preserved in fewer than half of trials at 30%.

B.4 Sensitivity to LLM-as-a-Judge Errors

  • The analysis introduced judge-output errors at rates from 0% to 30%, far exceeding the validation-study error rate.At each rate, labeled items independently received one randomly drawn error from seven empirically observed categories, across 300 trials per rate.Source: §B.4 · p. 24
  • All four key findings remained intact across tested error rates, and none fell below 50% preservation at a 30% error rate.Each rate used 300 independently corrupted versions of the full dataset; 100% preservation means a finding held across all 300 trials.Source: §B.4 · p. 24

B.5 Using Multiple LLMs as Judges … D Experimental Details

The section describes how CoTrace compares alternative LLM judges and reports token usage and estimated cost. It notes that requirement-extraction disagreement reflects systematic differences in segmentation and attribution.

B.5 Using Multiple LLMs as Judges

  • The comparison uses GPT-5.2 and GPT-5.1-mini as default judges and compares GPT-5.2 against Claude 4.5 Sonnet and two Llama judges using GPT-5.1-mini-extracted actions.Table 7 reports agreement between GPT-5.2 and alternative judges.Source: §B.5 · p. 24, Table 7 · p. 25
  • For Outcome and Requirement extraction, outputs are greedily matched across judges to maximize action-set overlap before calculating mean Jaccard similarity.Source: §B.5 · p. 24
  • Mean Jaccard similarity penalizes both missed and over-claimed actions and is sensitive to how finely a judge segments dialogue.Source: §B.5 · p. 24
  • Llama-3.1-8B produced approximately 2.7 times as many requirements as GPT-5.2 and more often treated assistant utterances as origins (77% versus 37%).These behaviors account for particularly low agreement between the two judges on Requirement extraction.Source: §B.5 · pp. 24–25
  • Influence-labeling comparisons hold GPT-5.2’s outcomes, requirements, actions, and candidate pairs fixed, isolating the alternative judges’ labeling step.The reported influence-labeling measure is Spearman’s ρ for the user’s Shaper role contribution.Source: §B.5 · p. 25, Table 7 · p. 25

B.6 Cost

  • Table 8 reports average token usage by dialogue length and step; the text also points to estimated costs by step.Source: §B.6 · p. 25, Table 8 · p. 25

D.1 Qualitative Examples … INTERACTIONAL STEERING

Users shape goals explicitly, while assistants often influence them through advice or execution. Assistant-generated requirements more often add low-level implementation details, especially in technical tasks, while requirements overlap more in open-ended domains.

D.1 Qualitative Examples

  • Figure 7 shows users predominantly shaping goals directly, while assistants more often influence them through advice or implicitly during execution.Users commonly Request, Constrain, or Instruct; assistants commonly Suggest, Recommend, Provide, or Describe.Source: §D.1 · p. 25, Figure 7 · p. 26
  • Assistant-generated requirements more often concern low-level implementation details, while user-generated requirements more often express broader goals or intentions.The qualitative comparison uses PCA projections of requirement embeddings and is intended for interpretation, not formal clustering.Source: §D.1 · p. 26, Figure 8 · p. 27
  • In technical tasks, some semantic regions are populated predominantly by assistant-generated requirements, suggesting models introduce requirement types users rarely specify independently.The plots compare programming, data analysis, writing, and planning requirements.Source: §D.1 · p. 26, Figure 8 · p. 27
  • In less technical, more open-ended domains, user- and assistant-generated requirements occupy largely overlapping semantic regions.The passage frames identifying tasks with greater human–AI complementarity across requirement layers as a future direction.Source: §D.1 · p. 27

D.2 Interaction Patterns of Indirect Influence

  • The indirect-influence analysis manually categorized 60 randomly selected requirements—30 user-generated and 30 assistant-generated—and reports subtype proportions in Tables 10–12.Table 10 is captioned as assigning influence types to the 60 sampled requirements.Source: §D.2 · p. 27, Table 10 · p. 28

INTERACTIONAL STEERING

  • Table 9 presents observed action-pair patterns of indirect influence, marking the preceding influencing action separately from the action that creates or specifies a subsequent requirement.The passage directs readers to Table 9 for interaction-subtype definitions.Source: §D.2 · p. 27, Table 9 · p. 28

E User Simulator Implementation … F.1.1 Survey Items

This appendix details the simulated collaboration protocols and user-study procedures, including how participants reviewed goal-level analyses and reflected on human–LLM collaboration.

E.1 CoGym

  • Agentic-CoGym permits optional teammate messages before tool use, whereas Chat-CoGym requires a message before every tool call while holding tasks, tools, environment, and evaluation constant.This contrast isolates the effect of requiring communication before action; messages can share a plan, explain an action, or request feedback.Source: §E.1 · pp. 28–29

E.2 Intervention Setup

  • Both user interventions use soft guidance; steering encourages varied collaboration, while underspecification elicits tentative preferences revealed gradually without fabricated details.The prompts shape interaction style without prescribing strict turn-by-turn behavior.Source: §E.2 · p. 29

Steering intervention prompt

  • The steering prompt offers five actions—delegate, ask an open question, invite extension, introduce a constraint, or acknowledge and redirect—while telling the user not to steer every turn.The simulated user may instead answer, wait, or let the agent continue.Source: Steering intervention prompt · p. 30

Underspecified intervention prompt

  • The underspecified-user prompt asks for short, vague, hedged responses, gradual disclosure, and no volunteered details or fabricated information.The simulated user can ask for suggestions or naturally say they do not know.Source: Underspecified intervention prompt · p. 30

F Human Study; F.1 User Study: Supporting First-Person Reflection

  • The appendix describes a first-person reflection study and a third-party evaluation case study; the first-person study involved five pairs and four stages of collaboration, tool use, survey, and interview.Ten students completed the same travel-planning task with a chatbot and with a human partner, inspected their chatbot logs through the tool, and were interviewed; session order was counterbalanced.Source: F Human Study · p. 30, §F.1 · p. 30

Instruction for Human-LLM Collaboration

  • The human–LLM task asked participants to naturally collaborate with a chatbot for at least four turns on a one-day itinerary for visiting parents, reflecting their preferences and comfort needs.The supplied context specified a hotel starting area, reasonable travel time, a relaxed pace, and limited walking.Source: Instruction for Human-LLM Collaboration · p. 31

Instruction for Human-Human Collaboration

  • The human–human task prohibited LLM use, allowed Google Search, and asked pairs to fill a shared itinerary template for a one-day trip with a third traveler.The provided traveler preferences included scenic walks, a popular local food place, and a relaxed schedule.Source: Instruction for Human-Human… · p. 31

F.1.1 Survey Items

  • Surveys before and after tool use rated perceived human and chatbot contributions to goal shaping and execution, plus chatbot satisfaction; tool evaluations covered goal and requirement agreement, awareness, provenance, and indirect influence.Participants also provided open-ended explanations, post-tool reflections, and comparisons of their human versus chatbot collaborations.Source: §F.1.1 · pp. 31–33

F.1.2 Responses … F.2.1 Survey Items

Participants said COTRACE surfaced hidden and indirect AI contributions, while they described chatbots as faster at execution and human partners as more involved in goal shaping. The survey items compare evaluators’ judgments of LLM use before and after inspecting COTRACE outputs.

F.1.2 Responses

  • Participants said COTRACE revealed AI work and indirect decisions they had not noticed, which they found surprising.Some also said the tool made them more aware of how and when the LLM shaped outcomes.Source: §F.1.2 · pp. 33–34
  • Participants said visualizing goals and requirements clarified how requirements became actionable steps and prompted reflection on their own prompting.One participant said this led them to consider being more specific about preferences.Source: §F.1.2 · p. 34
  • Participants described chatbots as faster and stronger at execution, while human partners contributed more actively to collaborative goal shaping.They associated human goal shaping with back-and-forth dialogue and generating requirements.Source: §F.1.2 · pp. 34–35
  • Participants said human collaboration involved social considerations and partner-dependent quality, while humans could challenge, clarify, or validate plans.Some described considering a partner’s preferences; others noted that partner knowledge or behavior affected usefulness.Source: §F.1.2 · p. 35
  • Participants said they were more direct and demanding with chatbots, but more considerate, open, and mentally engaged with human partners.Some described chatbot use as transactional, whereas human collaboration felt more like a partnership and required them to think harder.Source: §F.1.2 · pp. 35–36

Q4. Are you satisfied with the chatbot? Why or why not?

  • Participants were generally satisfied with the chatbot, citing speed, convenience, constraint-following, adaptation to new requirements, and reduced cognitive work finding options.They said it helped them focus their effort on debating which options best fit their goal.Source: Q4. Are you satisfied with the… · p. 36
  • Some participants found chatbot responses generic, distrusted their currency or accuracy, or felt the chatbot could not provide personal opinions.One participant said information would need to be verified, while another said the chatbot regurgitated facts rather than offering opinions.Source: Q4. Are you satisfied with the… · p. 36

F.2.1 Survey Items

  • The pre- and post-surveys asked evaluators to rate two students’ LLM reliance, describe their uses, consider score adjustments, and reflect on changes and tool helpfulness.The post-survey also asked whether COTRACE revealed information about students’ LLM use that raw logs had not made apparent.Source: §F.2.1 · p. 37

F.2.2 Responses

After using COTRACE, evaluators revised reliance assessments and gained a more detailed view of students’ goal shaping, but their grading decisions remained unchanged. They also found the goal-level summaries and visualizations faster and easier to assess than raw logs.

F.2.2 Responses

  • In three of four evaluator–student cases, evaluators revised LLM-reliance ratings after COTRACE, but neither changed any grading decision.Table 14 summarizes assessments before and after tool use.Source: Table 14 · p. 38, §F.2.2 · p. 38
  • E1 gained detail about Student A’s code generation, execution, and requests for explanations, and Student B’s use of the LLM to improve cohesion and wording.E1 retained the original grades, assigning Student A zero points because the LLM completed the assignment and keeping Student B’s score.Source: §F.2.2 · pp. 38–39
  • E2 concluded both students shaped goals more actively than raw logs suggested, revised both reliance ratings downward, and left both assignment scores unchanged.This enabled a more informed assessment of their reliance on the LLM.Source: §F.2.2 · pp. 38–39
  • Both evaluators found COTRACE faster and easier to use than raw logs for assessing LLM use and identifying who shaped and executed each goal.They cited cumbersome or lengthy logs and the tool’s goal summaries, contribution visualizations, and links to shaping and execution actions.Source: §F.2.2 · pp. 38–39

G ShareChat Data Sampling … Random or tangential topic (excluded from single-topic task set)

This section describes filtering and labeling ShareChat conversation chunks, excluding random or tangential material from the single-topic task set. The examples span exchanges that shift topics or contain unrelated assistance, including statistical explanations, memo drafting, and regional lake questions.

G ShareChat Data Sampling

  • Conversations were filtered to English samples with at least 8 messages, divided into non-overlapping chunks, and random or tangential chunks were excluded before single-topic labels were aggregated by majority vote.The topic definitions came from Chang et al. (2026); conversations were retained only when they contained at least four user–assistant turns.Source: G ShareChat Data Sampling · p. 39

Prompt for Topic Labeling

  • The labeling prompt required a JSON judgment of single-topic versus random/tangential mode and, for coherent chunks, exactly one label from a predefined taxonomy.The schema also requested a one-sentence topic description and a short justification.Source: Prompt for Topic Labeling · p. 39

Computer programming

  • One conversation excerpt moves from Python addition and debugging to a sci-fi discussion of UFOs, antimatter, and quantum physics.The user shifts from asking for Python help to describing a realistic sci-fi movie and asking about a comment concerning UFOs and an airplane.Source: Computer programming · pp. 40–41

Data analysis

  • The data-analysis exchange asks how to explain a binary finding’s significance with respect to degrees of freedom, and the response discusses overfitting, underfitting, multicollinearity, and test power.Source: Data analysis · p. 44
  • The exchange explains that p-values do not give the probability that the null hypothesis is true or the effect size, while model fit, multicollinearity, and test power affect interpretation.Overfitting can produce misleadingly small p-values, while underfitting or limited power can obscure real relationships.Source: Data analysis · p. 45
  • The pirates-and-global-warming example illustrates that correlation and statistical significance do not establish causation, since the relationship may be spurious or driven by omitted factors.The passage says pirates have no scientifically plausible impact on global temperatures.Source: Data analysis · p. 45
  • The encryption analogy distinguishes a meaningful, deterministic key-based transformation from spurious correlation, while treating statistical significance and degrees of freedom as metaphors rather than directly applicable concepts.A large key space makes brute-force guessing impractical, and the correct key enables reliable decryption.Source: Data analysis · p. 46
  • The conversation then shifts to drafting and expanding a breach-of-contract memo, covering contract validity, breach and evidence, damages and mitigation, defenses, costs, alternatives to litigation, and business strategy.The memo weighs expected recovery against litigation costs and strategic effects, and cautions that it is general information rather than legal advice.Source: Data analysis · p. 47
  • The breach-of-contract memo advises assessing contract validity, breach evidence, damages, mitigation, defenses, litigation costs, alternative dispute resolution, and strategic consequences.It concludes that the decision should balance legal and strategic considerations and notes that the memo is general information, not legal advice.Source: Data analysis · pp. 51–52
  • The statistical explanation links degrees of freedom to model predictors and describes how overfitting, underfitting, multicollinearity, limited power, and p-value misinterpretation complicate significance claims.It notes that p-values do not give the probability that the null hypothesis is true or the effect size.Source: Data analysis · p. 54
  • The exchange contrasts the pirates–global-warming example as spurious correlation with encryption’s deterministic key-based transformation and decryption.It says statistical significance in the pirate example does not establish causation, while the correct key enables reliable decryption.Source: Data analysis · pp. 55–56

Planning

  • The lake response names Como, Garda, Maggiore, and Iseo among regional lakes, then says it could not find a specific reactive-silicate concentration for Lake Endine.The response also mentions smaller lakes and recreational uses, but provides no concentration value for Endine.Source: Planning · pp. 56–57
  • The assistant could not find specific lake silicate concentrations, Sensor 14530 observations, or 2016 digestive-disease hospital access averages, and suggested contacting relevant agencies or databases.For sensor readings, it noted that common weather sites generally provide broader conditions rather than individual sensor outputs.Source: Planning · p. 57

Random or tangential topic (excluded from single-topic task set)

  • For the excluded tangential query, the assistant listed notable regional lakes—including Como, Garda, Maggiore, Iseo, and Lugano—but gave no explicit total count.It also mentioned smaller lakes and described the region’s lakes as recreational destinations.Source: Random or tangential topic (excluded… · p. 58

Related papers

Loading 2605.21363v2…