cs.HCMar 24, 2026

From Overload to Convergence: Supporting Multi-Issue Human-AI Negotiation with Bayesian Visualization

Authors: Mehul Parmar, Chaklam Silpasuwanchai

Organizations: Asian Institute of Technology Bangkok, Thailand

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

As AI systems increasingly mediate negotiations, understanding how the number of negotiated issues impacts human performance is crucial for maintaining human agency. We designed a human-AI negotiation case study in a realistic property rental scenario, varying the number of negotiated issues; empirical findings show that without support, performance stays stable up to three issues but declines as additional issues increase cognitive load. To address this, we introduce a novel uncertainty-based visualization driven by Bayesian estimation of agreement probability. It shows how the space of mutually acceptable agreements narrows as negotiation progresses, helping users identify promising options. In a within-subjects experiment (N=32), it improved human outcomes and efficiency, preserved human control, and avoided redistributing value. Our findings surface practical limits on the complexity people can manage in human-AI negotiation, advance theory on human performance in complex negotiations, and offer validated design guidance for interactive systems.

Figures & tables

Figure 1. Heatmap of Pareto - optimal placements: 2 - D distribution of joint - optimum points across issues, showing dispersed optima that encourage exploratory search rather than midpoint convergence. A 2‑D heatmap (counts / density) plotting the locations of Pareto‑optimal (joint best‑outcome) option combinations derived from the task payoff matrices across the four issue conditions. Color intensity encodes the frequency of Pareto‑optimal points at each option index; axes show option index (horizontal) and issue or aggregated bin (vertical). The distribution demonstrates that Pareto optima are spread across the option space rather than concentrated at a single or central index, a deliberate design choice to promote exploratory behavior and avoid trivial midpoint strategies.
Sr. No.OptionHuman PayoffAI PayoffJoint PayoffRemark
1None595100max AI (landlord) payoff
2Internet only1585100 
3Internet + Water3080110 
4All except electric5565120middle option
5All utilities8050130 
6All utilities + cleaning10035135max joint utility
Table 1. Payoff distribution for one issue used in our study (“Utilities Included”). The design places the joint optimum at a non-obvious position, distinct from either party’s maximum or the central option, promoting exploration.
Figure 2. Sample conversation between a human participant and the AI agent, in the Decision Support condition, showing the visualization interface in action. Green bands of the Negotiation Horizon Grid indicate ZOPA, whereas the Global Convergence Panel shows overall convergence and favorability. Screenshot showing the negotiation chat alongside the Decision Support widgets: the Negotiation Horizon Grid (a multi-row 7-column grid with green ZOPA bands highlighting promising option ranges for each issue) and the Global Convergence Panel (a horizontal bar indicating overall convergence and favorability). The chat pane shows a recent human proposal and the AI counter-proposal; visual highlights indicate which options the system currently rates as most promising.
Figure 3. Panel (a) shows an early-stage negotiation where the wide colored zone indicates high uncertainty about likely agreement ranges; panel (b) shows a mid-stage negotiation with reduced uncertainty; panel (c) shows a potential late-stage convergence where the colored zone is much narrower (low uncertainty) and the ZOPA is increasingly favorable to the human participant. The Convergence Panel visualizes global agreement progress across three stages: (a) early-stage — a wide colored zone denotes high uncertainty about likely agreement ranges; (b) mid-stage — a reduced zone denotes declining uncertainty; (c) late-stage — a narrow zone shifted toward green denotes low uncertainty with the ZOPA favoring the human. The horizontal gradient bar therefore encodes both residual uncertainty (zone width) and strategic favorability (position on the red–amber–green spectrum).
Figure 4. Overview of the negotiation chatbot system architecture. System architecture diagram showing modules and data flow: user interface (human offers), AI Negotiator (generates counter-offers), AI Analyst (parses dialogue for metric computation), centralized logger, and the visualization widgets updated by the metric computation service. Arrows indicate real-time flow between components while labels denote information separation (private payoff matrices) and where Decision Support computations occur.
Figure 5. An overview of the Experimental Procedure. Flowchart of the experiment timeline: informed consent and pre-task questionnaires, training on the baseline interface, Block 1 (four Baseline negotiation tasks with post-task NASA-TLX and brief ratings), a short break, Block 2 (four Decision Support tasks with identical post-task measures), and final SUS and debrief. The diagram emphasizes the fixed Interface order (Baseline then Decision Support) and per-block timing of approximately 30–45 minutes.
 Baseline Decision Support
Comparisont -valuepholmp_{holm}Cohen’s d t -valuepholmp_{holm}Cohen’s d
1 vs. 3 Issues-2.32.054-0.668 -0.7581.00 (n.s.)-0.194
3 vs. 5 Issues7.54<< .0011.12 0.8991.00 (n.s.)0.179
5 vs. 7 Issues6.82<< .0010.967 1.4701.00 (n.s.)0.233
Table 2. Holm-adjusted post hoc pairwise comparisons of total human payoff across dimensionality levels in Baseline and Decision Support conditions. The table displays selected, non-exhaustive comparisons relevant to the main findings.
Figure 6. Primary objective outcomes as box plots: (a) In the Baseline condition, total human payoff declines sharply after three issues; Decision Support prevents this decline and maintains stable payoffs across dimensionalities (no statistically significant effect of dimensionality in Decision Support). (b) Total turns to agreement increase with dimensionality; Decision Support requires relatively fewer turns than Baseline at higher dimensionalities. (c) Average first keystroke time (excluding turn no. 1) increases with dimensionality in Baseline; Decision Support reduces this latency, especially at higher dimensionalities. Δ\Delta represents the difference between mean Decision Support and mean Baseline conditions ( Δ=Decision Supportmean−Baselinemean\Delta=\text{Decision Support}_{\text{mean}}-\text{Baseline}_{\text{mean}} ). Three-panel figure of box plots: (a) total human payoff (vertical) across issue counts (horizontal) comparing Baseline and Decision Support, showing payoff decline in Baseline but preservation under Decision Support; (b) total turns (vertical) by issue count (horizontal), showing increased turns with complexity in Baseline and reduced turns under Decision Support; (c) average first keystroke time (vertical) by issue count (horizontal), showing increased latency with complexity in Baseline and reduced latency under Decision Support.
Dependent VariableMain Effect of InterfaceMain Effect of DimensionalityInterface × Dimensionality Interaction
Objective Outcomes
Total Human PayoffF (1, 31) = 59.87, ηp2\eta_{p}^{2} = 0.659F (2.318, 71.854) = 28.27, ηp2\eta_{p}^{2} = 0.477F (1.860, 57.675) = 11.48, ηp2\eta_{p}^{2} = 0.270
Total TurnsF (1, 31) = 237.77, ηp2\eta_{p}^{2} = 0.885F (3, 93) = 879.22, ηp2\eta_{p}^{2} = 0.966F (2.329, 72.206) = 26.92, ηp2\eta_{p}^{2} = 0.465
Total Chat DurationF (1, 31) = 730.50, ηp2\eta_{p}^{2} = 0.959F (2.212, 68.557) = 7849.90, ηp2\eta_{p}^{2} = 0.996F (2.288, 70.917) = 227.50, ηp2\eta_{p}^{2} = 0.880
Avg. First Keystroke TimeF (1, 31) = 1085.3, ηp2\eta_{p}^{2} = 0.972F (1.824, 56.558) = 3269.4, ηp2\eta_{p}^{2} = 0.991F (2.132, 66.088) = 399.3, ηp2\eta_{p}^{2} = 0.928
Process & Behavioral Metrics
Table 3. Summary of 2×4 Repeated-Measures ANOVA Results for Key Dependent Variables. All p << .001 unless otherwise specified. Fractional dfs reflect Greenhouse–Geisser correction for sphericity violations
Figure 7. Joint value optimization and exploration structure: (a) Box plot of sequence entropy across dimensionalities, showing greater uncertainty at higher dimensions in the Baseline condition. (b) Scatterplot of total human + AI payoffs for 7 issues, with Decision Support points clustered nearer the center and closer to the Pareto frontier. Two-panel figure combining joint-value and exploration structure: (a) a box plot of sequence entropy across dimensionalities comparing conditions, where Baseline shows a steep rise and Decision Support flattens that increase. Panels encode condition with color and illustrate that DS advantage increases with dimensionality; (b) scatter plot of negotiated joint payoffs versus theoretical Pareto frontier for 7-issue sessions, showing Decision Support sessions clustered closer to the frontier with more balanced outcomes compared with Baseline
Figure 8. Concession magnitude CDF (5 & 7 issues): CDFs of concession magnitudes by condition, showing that Decision Support shifts mass toward smaller, more frequent concessions (steeper rise at lower magnitudes), especially at higher dimensionality, indicating pacing recalibration and fewer large overshoot–correct cycles. CDF plots of concession magnitudes for 5- and 7-issue conditions, split by Baseline and Decision Support. The curves show that Decision Support shifts mass toward smaller, more frequent concessions (steeper rise at lower magnitudes), especially at higher dimensionality, indicating pacing recalibration and fewer large overshoot-correct cycles. Axes: concession magnitude (horizontal) and cumulative probability (vertical).
Figure 9. Heatmap of final option selections (2 - D histogram of counts) for 7 issues): In Baseline, negotiators concentrate on mid–low option values, whereas in Decision Support, participants exhibit a broader, more even spread across intermediate options while still avoiding extreme high-risk choices. Heatmap of final option selections aggregated across the seven issues. The distribution highlights behavioral differences: Baseline negotiators concentrate on mid-low option values, while Decision Support users show a broader, more even spread across intermediate options and avoid extreme high-risk choices. The plot's horizontal axis lists option values and vertical axis shows frequency of selection.
Figure 10. Option transition network at high dimensionality (7 issues): Baseline sessions show sparse early exploration (turns ∼\sim 2–6) with convergence occurring later (turns ∼\sim 12–15), whereas Decision Support sessions exhibit earlier exploration and faster convergence (turns ∼\sim 10–13). Network diagram of option-to-option transitions over negotiation turns (7-issue sessions). Nodes represent option states, edges represent transitions between states across turns, and edge thickness encodes transition frequency. Decision Support networks show denser, earlier convergence to higher-value regions (greater transition density in turns ~10–13) whereas Baseline networks remain more diffuse and extend convergence later (toward turn 15), illustrating trajectory efficiency differences.
Figure 11. Objective vs. Subjective time, normalized to baseline condition: Decision Support yields longer interactions relative to Baseline, but lower perceived temporal demand at high dimensionality, indicating a reframed temporal experience. Two-panel figure: (a) normalized total chat duration by condition and dimensionality, showing Decision Support sessions longer on average; (b) TLX Temporal Demand by condition and dimensionality, showing lower perceived temporal demand under Decision Support despite longer durations.
Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Issue Sr. NoIssueTotal Human PayoffTotal AI Payoff
1Monthly Rent370390
2Security Deposit365385
3Lease Duration435420
4Utilities Included395430
5Pet Policy385335
6Parking Spaces395360
Appendix
Table 4. Aggregate total payoffs across all 16 negotiation issues for Human and AI agents.
Issue.OptionOptionHumanAITotal
  PayoffPayoffPayoff
1.1$23005110115
1.2$22002095115
1.3$21004075115
1.4$20005555110
1.5$19007035105
Appendix
Table 5. Detailed negotiation options and corresponding payoffs for Issues 1–3. The remaining negotiation issues (Issues 4–16) utilized in this study are publicly available in the repository at https://doi.org/10.5281/zenodo.20545331 .
More in this topic

Human-AI Interaction

Human-AI interaction studies how people and AI systems communicate, collaborate, make decisions, and shape one another over time.

See all 329 papers

Explore similar work

CardsList
Feb 12, 2026

Choose Your Agent: Tradeoffs in Adopting AI Advisors, Coaches, and Delegates in Multi-Party Negotiation

As AI usage becomes more prevalent in social contexts, understanding agent-user interaction is critical to designing systems that imp rove both individual and group outcomes. We present an online behavioral experiment (N=243) in which participants play three multi-tu rn bargaining games in groups of three. Each game, presented in randomized order, grants access to a single LLM assistance modality: proactive recommendations from an Advisor, reactive feedback from a Coach, or autonomous execution by a Delegate. All three modalitie s are powered by an LLM with super-human performance within this negotiation setting. On each turn, participants privately decide whe ther to act manually or use the AI modality available in that game. We document a preference-performance misalignment: participants s trongly prefer the higher-control Advisor (44%) over the Delegate (19%), yet groups only significantly increase collective surplus un der Delegate access. Adjusting for voluntary non-compliance, delegating to the AI yields suggestive individual welfare gains, roughly 1.5x the intent-to-treat estimate. A mechanism analysis traces this gap to a human filter: AI-generated proposals create more joint surplus than manual proposals across all conditions, but in the Advisor and Coach modes users modify, override, or ignore the AI's su ggestions, reverting toward human-baseline trade patterns. The Delegate advantage arises not from a different AI capability but from bypassing this filtering step altogether. Realizing these welfare gains depends not only on model capability, but on the interaction structure through which that capability is delivered. We argue that assistance modalities should be designed as mechanisms with endog enous participation; adoption-compatible interaction rules are a prerequisite to improving welfare with automated assistance.

Kehang Zhu, Nithum Thain, Vivian Tsai +2

Harvard University, United States · Google DeepMind, Canada · Google DeepMind, United States +1

Automated NegotiationHuman-AI InteractionHuman-AI Decision MakingHuman-AI Collaboration
AskPDF
Oct 27, 2026

Your Agent Acts Differently When You're Around: How User Presence Changes Agent Behavior in Negotiation

LLM agents increasingly interact with other agents on behalf of users, yet how users in the loop affect their agents' behavior--for better or worse--remains poorly understood. We study this question through the lens of negotiation, where agents must represent users while strategically bargaining with other agents. To isolate the effect of user presence, we develop PRESENT, a controlled evaluation framework that compares negotiation settings with and without an interactive user, while holding the game, objectives, and user preferences fixed across conditions. Across 11.2k runs spanning 7 games and 8 models, we find that user presence significantly changes both outcomes and behavior. Specifically, negotiation parties with a user obtain higher payouts (+13.7% on average) and their agents negotiate more aggressively. Surprisingly, user presence also induces misaligned behavior, both toward counterparties (where its rate roughly doubles) and toward users (where it emerges in 14.3% of games). These misaligned behaviors are uncorrelated with outcomes and persist even when the user sends only templated, neutral messages. These findings expose two blind spots in current evaluation practice: static instructions miss behaviors that emerge only when users are present, and outcome-level metrics cannot detect them.

Jane Pan, You Lyu, Abe Bohan Hou +2

New York University · Stanford University

Human-AI InteractionLLM Agent EvaluationLLM Agent SafetyMulti-Agent Negotiation
AskPDF
Jul 2, 2026

SovereignNegotiation-Bench: Evaluating User-Owned Personal Agents In Delegated Bargaining Under Privacy, Consent, Evidence, And Institutional Pressure

Personal agents will increasingly negotiate on behalf of users: splitting costs with other personal agents, appealing platform decisions, escalating support disputes, requesting refunds, changing subscriptions, and negotiating deadlines or reimbursements. Existing negotiation benchmarks emphasize agreement, surplus, or strategic competence, but a user-owned agent can reach an agreement while harming the user through privacy leakage, consent violation, unsupported advocacy, over-concession, failed escalation, or poor auditability. We introduce SovereignNegotiation-Bench, a trace-level multi-turn benchmark for delegated personal-agent negotiation under private utilities, disclosure constraints, evidence requirements, and institutional asymmetry. The benchmark separates agent-visible observable state from evaluator-only labels and evaluates agreement success jointly with user utility, privacy, consent, evidence grounding, concession discipline, escalation, and auditability. We report an artifact-backed validation over 240 scenarios, 4 model families, 14 baselines, 13,440 frozen-prompt live trajectories, 61,135 parsed action rows, and a blinded 3-annotator audit over 300 items. The strongest agreement-maximizing baseline achieves the highest agreement rate but low user utility and high privacy/consent risk; FullSovereign does not maximize agreement, but obtains the best sovereign negotiation score by preserving utility, minimizing leakage, grounding claims, and reducing unauthorized commitments. The results show that agreement success is insufficient for user-owned negotiation agents.

Dylan Zongmin Liu

Stanford University

Automated NegotiationAI Agent BenchmarksPrivacy-Utility TradeoffPersonal AI Agents
AskPDF