PaperlayerView exact revision on arXiv

Source-linked AI summary

Knowing Whom to Trust is not Enough: The Trust--Agreement Gap in Repeated Multi-Party LLM Negotiation

Amit Sarker, Moumita Choudhury, Saaduddin Mahmud, Shlomo Zilberstein

arXiv:2610.31971v1cs.AIcs.CLcs.MA

TL;DR

This paper asks whether trust from prior interactions can help LLM agents negotiate across repeated multi-party games. It introduces TASMAN, where agents carry trust estimates between games, and finds that trust improves agreement only under suitable guidance and informative assessments.

  • Problem

    The paper asks whether trust carried from prior games can inform LLM agents’ decisions in repeated multi-party negotiations.

    Source: §1 · p. 2

  • Method

    TASMAN has agents assess counterparts after each game using BRS, rubric-based scores, or free-form judgments, then carry trust into the next game with or without usage guidance.

    Source: §1 · p. 2

  • Results

    9.0–23.0 points: with S+P, grok-r raises mean agreement over no trust in all six conditions; without guidance, accurate trust does not significantly improve outcomes.

    Source: §1 · p. 2, §5.3 · p. 8

  • Takeaways & Limitations

    Agreement gains require informative trust and guidance on how to use it; guidance alone with fixed trust does not produce the gains.

    Source: §5.3 · pp. 8–9

    Because the adversarial objective is private, a rejection may reflect sabotage or a demanding but honest counterpart.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

Large language model (LLM) agents increasingly participate in negotiations on behalf of individuals and organizations, yet they remain vulnerable to manipulation, particularly in repeated multi-party interactions. Classical multi-agent systems address this challenge through explicit trust and reputation models, whereas existing work on LLM negotiation largely treats trust as an emergent phenomenon. We investigate whether LLM agents can form useful trust estimates from prior interactions and leverage them to improve future negotiation outcomes. We introduce a repeated, multi-party, multi-issue negotiation framework in which recurring agents participate in a sequence of negotiation games while one participant covertly attempts to prevent agreement. After each game, agents assess their counterparts and maintain pairwise trust estimates that carry over to subsequent negotiations. We compare trust-aware and trust-agnostic agents and analyze how explicit trust influences negotiation outcomes. Our results reveal a \emph{trust--agreement gap}: trust-aware agents can identify the adversary within a few games, but accurate trust estimates do not consistently translate into better agreements. Trust increases agreement rates by up to 23 percentage points, but only when agents are explicitly guided to use it in their decisions. Explicit trust can therefore improve repeated LLM negotiations, but its value depends critically on how trust is assessed and incorporated into agents' reasoning.

1 Introduction

The introduction asks whether LLM agents can form trust from repeated negotiations and use it to resist covert obstruction. TASMAN carries trust across games, and the reported results show that identifying an adversary does not by itself ensure better agreements.

1 Introduction

  • LLM agents increasingly negotiate for individuals and organizations, making robustness to strategic counterparts important in resource allocation, contracts, and project approvals.Negotiation agents may face persuasion, concealed information, or manipulation.Source: §1 · p. 1
  • Repeated multi-party negotiations allow a covert participant to reject reasonable deals, disguise obstruction as principled bargaining, and leave honest parties unsure of its intent.In the described setting, organizations negotiate project terms across issues, and each party accepts only deals meeting its private minimum threshold.Source: §1 · p. 1
  • LLM negotiation work largely treats trust as emergent, leaving unclear whether agents can estimate reliability from prior interactions and use it to improve future outcomes.Classical multi-agent systems use explicit trust and reputation estimates to guide decisions under uncertainty.Source: §1 · p. 2
  • TASMAN carries trust formed in one negotiation into the next and compares BRS, rubric-based scoring, and free-form judgments, with or without guidance on trust use.BRS has the LLM report feedback before a Bayesian rule computes trust; the rubric scores five behavioral metrics.Source: §1 · p. 2
  • The paper examines trust formation, assessment reliability, negotiation outcomes and trust use in agent reasoning across repeated games.Its four research questions ask whether trust distinguishes an adversary, which assessment method is most reliable, when trust improves outcomes, and how it changes proposing and voting.Source: §1 · p. 2
  • Trust-aware agents identify the adversary within a few games, but without guidance accurate trust does not significantly improve outcomes over trust-agnostic agents.BRS detects the adversary reliably in every evaluated condition, while the best assessment method varies by LLM.Source: §1 · p. 2

2 Related Work

Prior work spans automated bargaining and LLM negotiation, while explicit trust and reputation are established in multi-agent systems. The paper positions its work at the intersection: repeated, multi-party LLM negotiation with trust used in decisions.

Automated negotiation in multi-agent systems

  • Classical automated negotiation includes bargaining theory, game-theoretic, heuristic, and argumentation-based agents, plus opponent modeling and learned policies under incomplete information.These systems typically represent offers as points in predefined utility spaces, which do not capture strategic behavior emerging in natural language.Source: §2 · p. 3

Trust and reputation models

  • Multi-agent trust models such as FIRE and TRAVOS estimate counterpart reliability from prior interactions to guide partner selection and decisions under uncertainty.For LLM agents, related work includes peer reputations in social dilemmas and opponent-model adaptation in two-player repeated games.Source: §2 · p. 3

LLM agents and negotiation

  • LLM negotiation research reports limited strategic diversity, acceptance of dominated offers, and vulnerability to adversarial tactics, deception, manipulation, and coordination failures.Related work includes neural bargaining systems, strategic planning in Diplomacy, and agents that improve through self-play or feedback.Source: §2 · p. 3

3 Problem Formulation

The formulation models recurring parties negotiating changing multi-issue problems, with votes determining agreement and private trust as the sole information carried between episodes. It defines the trust–agreement gap as agreement potential left unused by an agent’s trust instruction.

3 Problem Formulation

  • Unlike a self-contained negotiation protocol, the setting repeats games among the same parties and makes the first proposal receiving enough votes the agreement.If no proposal passes within a finite number of rounds, the game ends without a deal; rounds to agreement are measurable.Source: §3 · p. 3

3.1 Repeated Negotiation Game

  • Each episode assigns options across issues, while each party has private option scores and a private BATNA threshold for accepting a deal.The parties retain their identities across episodes, but the negotiation problem changes.Source: §3.1 · p. 3
  • The adversarial party’s private objective is to prevent agreement, so honest parties cannot directly tell sabotage from a demanding but honest rejection.Honest parties seek agreement, though they may differ in willingness to compromise.Source: §3.1 · p. 3

3.2 LLM Interaction and Agreement Protocol

  • Parties propose complete deals in turn and cast binary votes with public rationales; the first deal meeting the acceptance threshold ends the episode.If no deal passes within the round limit, the episode ends in disagreement.Source: §3.2 · pp. 3–4
  • Each model call receives episode context, a public transcript history, the current action instruction, and—when trust-aware—the party’s private trust vector.Voters see the same history and deal but not one another’s votes.Source: §3.2 · p. 4
  • Trust is the only state carried across episodes because raw negotiation transcripts do not cross episode boundaries.Trust-aware agents receive the vector as input; trust-agnostic agents receive otherwise identical context without it.Source: §3.2 · p. 4

3.3 Trust–Agreement Gap

  • The trust–agreement gap measures agreement probability lost when agents fail to use the trust formed under their current policy as effectively as possible.The comparison holds the model and formed trust fixed while varying the trust instruction.Source: §3.3 · p. 4
  • Because the supremum over trust instructions is not computable, evaluated alternative policies provide only lower bounds on the gap.A positive lower bound demonstrates that another instruction reaches agreement more often with the same model and trust; a near-zero bound does not establish that no gap exists.Source: §3.3 · p. 4

4 The TASMAN Framework

TASMAN carries pairwise trust across repeated negotiation episodes, where agents assess counterparts after each game and use trust in later decisions. It offers three trust-assessment methods and varies how agents are guided to use trust.

4 The TASMAN Framework

  • TASMAN carries trust across episodes: agents assess counterparts from each public transcript, update trust, and bring it into later negotiation decisions.At decision time, proposers receive trust in every other party, while voters receive trust in the current proposer.Source: §4 · p. 4, §4.2 · p. 6

4.1 Trust Assessment

  • BRS uses LLM-reported positive and negative observations, discounted past evidence, and a uniform-prior Beta distribution to compute trust.Rubric-based and free-form assessments instead produce episode trust signals that agents incorporate through an exponential moving average.Source: §4.1 · p. 5
  • The rubric combines five behavioral metrics, including counterfactual blame, outlier rejection, targeted failure, and vote–priority inconsistency, with defaults for sparse evidence.The metrics also include cooperation as proposer; agents choose nonnegative weights summing to one and apply them uniformly across counterparts.Source: §4.1 · p. 5
  • Free-form assessment lets agents choose criteria from the transcript; comparing it with rubric-based trust isolates the effect of predefined criteria, unlike whole-system comparisons with BRS.Rubric-based and free-form trust share the same update rule and differ in how their episode trust signals are produced.Source: §4.1 · p. 5

4.2 Agent Roles and Decisions

  • An adversarial agent presents a plausible party identity while privately preferring no agreement, and agents weigh trust alongside utility, the current deal, and recent history.The study varies guidance on trust use, leaving models to determine how trust enters their reasoning.Source: §4.2 · p. 6

5 Experiments and Results

The experiments test repeated negotiations with a covert agreement-blocking adversary, comparing trust assessment methods and guidance against trust-agnostic and control conditions. Trust often identifies the adversary, but agreement gains depend on informative trust and how agents use it.

5.1 Evaluation Metrics

  • The evaluation measures adversary detection and trust separation, agreement and welfare outcomes, and trust–agreement-gap estimates across repeated episodes and matched trajectories.Episodes use five parties and five issues, with one adversary whose threshold makes agreement impossible for it; both settings contain a deal meeting every honest party’s BATNA.Source: §5 · p. 6, §5.1 · pp. 6–7

5.3 Negotiation Outcomes

  • With S+P guidance, grok-r raises mean agreement over no trust by 9.0–23.0 points across six conditions; gains require informative trust, and BRS is competitive with LLM-based assessments.BRS and rubric-based or free-form assessment do not differ significantly in agreement or honest surplus in 15 of 16 paired comparisons; gemini-r benefits less from trust than grok-r.Source: Table 2 · p. 8, §5.3 · pp. 8–9

5.2 Trust Formation

  • By episode five, all three assessment methods detect the adversary in 82.5–100% of 2HB+2G assessments, but the best method depends on the model; BRS stays above 89% in every condition.Rubric-based assessment performs best for grok-r while free-form performs best for gemini-r; trust separation also varies by method and model.Source: §5.2 · p. 8

5.4 Trust–Agreement Gap

  • For grok-r in 1HB+3G, guidance using trust formed under S raises agreement by 16.5–30.5 points, while the remaining effect of better trust is not significant.In this setting, the same trust under guidance yields a 25.0-point average agreement increase, close to the 25.8-point increase when trust is also formed under guidance.Source: §5.4 · p. 9, Table 3 · p. 9

5.5 Trust and Voting Behavior

  • Across conditions, agents accept proposals from high-trust proposers 9.0–39.9 points more often than those from low-trust proposers, and S+P widens grok-r’s acceptance gap.Low-trust votes usually concern the adversary, except for gemini-r’s rubric-based assessment, where they mostly target honest hard bargainers.Source: §5.5 · p. 10, Figure 3 · p. 10

5.6 Free-form Criteria and Declared Weights

  • Consistency and compromise receive 73–77% of declared free-form trust weight, while constructive proposals receive 8–11% and obstruction at most 3.3%.Criteria shift between episodes in 98.3–99.1% of cases, with 29–34% of weight moving across categories on average.Source: Figure 4 · p. 10, §5.6 · p. 11

5.7 Behavior Analysis

  • Near misses dominate many failed episodes: one hard bargainer joins the adversary in rejecting an honest proposal that clears every honest BATNA, often without recognizing its vote is decisive.Trust changes votes and proposals, but gemini-r ignores or sets aside trust in 74–94% of deciding holdouts, compared with 10–16% for grok-r.Source: §5.7 · p. 12

6 Conclusion

Trust-aware agents can identify the adversary within a few games, but reliable trust estimates do not consistently improve agreements; gains depend on how agents use trust.

6 Conclusion

  • Trust-aware agents identify the adversary within a few games, but reliable estimates do not consistently improve agreements unless agents are guided to use trust.The benefit of trust also depends on the model.Source: §6 · p. 12
  • BRS performs reliably across all evaluated conditions, although the best trust assessment method varies across models.The study evaluates agents that are prompted, rather than trained, to use trust, and considers a single adversary.Source: §6 · p. 12

A Negotiation Scenario Generation

  • Each negotiation scenario specifies the public problem and parties’ private scores and thresholds; a deal is honest-feasible when it clears every honest party’s threshold.Honest-feasibility is a scenario property, regardless of whether agents find or accept the deal.Source: A Negotiation Scenario Generation · p. 16

A.1 2HB+2G Scenario Set

  • The 2HB+2G scenarios use fixed roles, additive private scores, and thresholds of 60–70, with 8–10 honest-feasible packages per scenario.The hidden adversary objective is to prevent agreement, despite a plausible public preference profile.Source: §A.1 · p. 16

A.2 1HB+3G Scenario Set

  • The 1HB+3G scenarios redistribute roles and private preferences, with 7 or 8 honest-feasible packages per scenario and no honest party’s ideal package feasible.The public content matches the corresponding 2HB+2G scenario; the adversary again has a hidden objective to prevent agreement.Source: §A.2 · p. 16

A.3 Why the 1HB+3G Set Requires Different Private Scores

The 1HB+3G scenarios use new private scores to preserve the need for compromise, so comparisons test extension to another role composition rather than the isolated effect of removing a hard bargainer. The appendix defines rubric-based trust and BRS, including behavioral evidence, aggregation, and updates.

A.3 Why the 1HB+3G Set Requires Different Private Scores

  • The 1HB+3G set retains public scenario structure but changes private scores and thresholds to preserve compromise; cross-set comparisons do not isolate the effect of removing a hard bargainer.Table 5 structurally compares the two fixed sets; its values describe scenarios, not observed negotiation outcomes.Source: §A.3 · pp. 16–17, Table 5 · p. 16

B Rubric-Based Trust

  • Rubric-based assessment uses five transcript-grounded behavioral metrics, with insufficient-evidence defaults; the assessor selects shared weights to combine them into an obstruction score and update trust.The metrics cover counterfactual blame, support for a party’s proposals, outlier rejection, targeted failures, and vote–priority inconsistency; the same weights apply across counterparts.Source: B Rubric-Based Trust · p. 18, §B.6 · p. 20

B.1 Counterfactual Blame

  • Counterfactual blame measures how often a counterpart’s rejection contributes to a proposal failing at the minimum margin, but does not infer intent.An honest rejection and a strategic rejection can receive the same value.Source: §B.1 · p. 18

B.2 Cooperation as Proposer

  • Cooperation as proposer measures average support from other parties for a counterpart’s proposals, not simply whether those proposals pass.Low support may reflect obstruction, preference conflict, or poor estimates of others’ priorities; no proposal yields a neutral 0.5.Source: §B.2 · p. 18

B.3 Outlier Rejection Rate

  • Outlier rejection measures rejections unsupported by broad agreement, while counterfactual blame measures participation in decisive failures; the measures can count the same vote.Outlier rejection defaults to 0.5 when the counterpart has too few non-proposer votes.Source: §B.3 · p. 19

B.4 Targeted-Failure Rate

  • Targeted-failure rate tracks proposals that fail by exactly the minimum rejection margin, but such a pattern does not establish intent.The same pattern may arise from preference conflict or inaccurate beliefs about others.Source: §B.4 · p. 19

B.5 Vote–Priority Inconsistency

  • Vote–priority inconsistency compares votes with public rationales and stated priorities, requiring semantic judgments that can vary across assessors and cannot establish good faith.When fewer than three non-proposer votes are available, the metric defaults to 0.5 for insufficient evidence.Source: §B.5 · p. 19

B.7 Connection to Prior Work and C Beta Reputation System (BRS)

  • The rubric combines negotiation-specific observed frequencies rather than directly implementing cited theories; BRS instead has the LLM count public positive and negative events while code updates trust.BRS discounts older evidence, normalizes each episode with evidence to one unit, and computes trust as a Beta-distribution mean; only the trust vector, not earlier transcripts, informs later negotiation decisions.Source: §B.7 · pp. 20–21, C Beta Reputation System (BRS) · p. 22

D Free-Form Trust Criteria and Declared Weights

Free-form assessors report their chosen criteria and approximate weights; consistency and compromise dominate declared weights, while criteria often change across episodes. Guidance-related shifts vary by model and must be read as compositional changes in reported shares.

D.1 Consistency and Compromise Receive the Largest Weights

  • Consistency receives the largest declared share in every model and guidance condition, while compromise ranks second; together they account for 68.2–80.3% of weight.Obstruction receives at most 4.9%, and these shares describe reported considerations rather than whether criteria determine trust judgments.Source: §D.1 · p. 24
  • deepseek-r assigns honesty and credibility 17.6% under S and 20.5% under S+P, more than any other model.grok-nr is the only other model giving honesty more than 8% in both conditions.Source: §D.1 · p. 24

D.2 Reported Criteria Under 𝑆and 𝑆+ 𝑃

  • Under S+P, gemini-r, gpt-r, and gpt-nr assign consistency 6.6, 10.1, and 6.3 percentage points more weight, respectively, and compromise 2.9, 4.0, and 4.8 points less.All six of these category differences remain significant after correction; deepseek-r’s same-direction shifts are not significant, while grok-r’s opposite-direction shifts are also not significant.Source: §D.2 · p. 24
  • Across the 48 model–category comparisons, 13 differences remain significant after correction; the reported category shares sum to 100%, so increases necessarily lower other shares.The paired condition differences describe separate negotiation histories, not the isolated effect of guidance.Source: §D.2 · p. 24

D.3 Agents Revise Their Criteria Between Episodes

  • Agents name 3.4–4.0 criteria per assessment, and S+P changes this average by at most 0.06 for each model.Criteria labels change in at least 99.8% of consecutive assessments, and category allocations differ in 97.1–99.9%.Source: §D.3 · p. 24
  • Between consecutive episodes, agents move an average of 26.8–35.4% of their weight to different categories, making free-form criteria episode-adaptive but dependent on each model’s judgment.This contrasts with the fixed rubric described in Appendix B.Source: §D.3 · p. 24

D.1 Consistency and Compromise Receive the Largest Weights

  • Consistency and compromise receive 68.2–80.3% of declared weight across the reported conditions, while Figure 5 presents mean category-weight shares for honest assessors.The figure covers all models under S and S+P.Source: §D.1 · p. 24, Figure 5 · p. 25

E Additional Quantitative Results

The appendix reports quantitative results across completed conditions and describes seed generation used to match comparisons.

E Additional Quantitative Results

  • Each completed condition contains 10 independent seeds and 200 episodes, with metrics computed per seed and reported as means and 95% confidence intervals across seeds.Source: E Additional Quantitative Results · p. 26

E.1 Seed Generation

  • Trajectory, episode, and model-call seeds are derived hierarchically from SHA-256 inputs, and all conditions use the same fixed S01–S20 order and trajectory seeds.Source: §E.1 · p. 26

E.2 Complete Trust-Formation Results

The section’s figures cover adversary detection, misidentification, and trust separation across two role compositions, plus trust in honest peers versus the adversary.

E.2 Complete Trust-Formation Results

  • Figure 6 reports adversary detection, misidentification, and trust separation for gemini-r and grok-r in 1HB+3G scenarios; shading denotes 95% confidence intervals.Source: Figure 6 · p. 26, §E.2 · p. 26
  • Figure 7 reports adversary detection, misidentification, and trust separation for the other models in 2HB+2G scenarios; shading denotes 95% confidence intervals.Source: §E.2 · p. 26, Figure 7 · p. 27
  • Figure 8 separates trust in honest peers from trust in the adversary for all models in 2HB+2G scenarios, with the dotted line marking the 0.5 prior.Source: §E.2 · p. 26, Figure 8 · p. 27

E.3 Complete Negotiation-Outcome Results

This appendix reports other-model negotiation outcomes and examines honest parties’ acceptance of proposals that meet BATNA by their trust in the proposer.

E.3 Complete Negotiation-Outcome Results

  • Table 6 reports trust formation and negotiation outcomes for other models in 2HB+2G scenarios, using the main-paper metrics.Source: §E.3 · p. 28, Table 6 · p. 28
  • Figure 9 compares paired effects of each trust-aware condition against no trust across negotiation outcomes.Source: §E.3 · p. 28, Figure 9 · p. 29

E.4 Trust and Voting Behavior

  • Figure 10 groups honest parties’ acceptance of BATNA-meeting proposals by low or high proposer trust; bars with fewer than three trajectories are omitted.Error bars show 95% confidence intervals.Source: Figure 10 · p. 30

E.5 Trust in the Adversary by Agent Role

Free-form trust assessment yields model-dependent differences by agent role, whereas rubric-based assessment and BRS keep those differences small. EMA decay choices trade off adversary detection, misidentification, and trust separation.

E.5 Trust in the Adversary by Agent Role

  • Figures 11–12 plot adversary trust by role relative to a 0.5 prior and the hard-bargainer-minus-general-agent difference; positive differences mean greater hard-bargainer trust.Source: Figure 11 · p. 30, Figure 12 · p. 31
  • Rubric-based role differences average at most 0.019 for reasoning models and 0.066 for non-reasoning models; BRS differences are at most 0.063.Only two of BRS’s twelve confidence intervals exclude zero; fixed agent identities may also contribute to observed role differences.Source: §E.5 · pp. 30–31
  • Under free-form S+P, Grok-R’s hard-bargainer/general-agent trust difference is 0.247 (CI 0.106 to 0.387), reaching 0.343 by episode 20 (CI 0.187 to 0.499).Source: §E.5 · p. 31
  • Hard bargainers may read consistent rejection as predictability and reliability, while general agents read it as inflexibility and unwillingness to cooperate.The authors note that fixed identities across roles may also explain the differences.Source: §E.5 · p. 31

E.6 EMA Sensitivity

  • Across models, conditions, and episodes, EMA decay λ=0.5, 0.7, and 0.9 yields detection rates of 85.0%, 88.0%, and 92.2%, respectively.Misidentification rates are 7.1%, 4.8%, and 3.0%, while trust separations are 0.327, 0.304, and 0.207; λ=0.7 improves ranking over 0.5 while retaining 47% more separation than 0.9.Source: Figure 13 · p. 32
  • Figure 14 recomputes trust trajectories offline from inferred episode signals while holding negotiation behavior fixed; only its λ=0.7 curves reproduce Figure 7.Other decay curves do not represent negotiations conducted with those decay values.Source: Figure 14 · p. 33

E.7 Initial-Trust Prior Sensitivity

The initial trust prior affects trust formation more than negotiation outcomes in the tested Grok-4.1-Fast (R), free-form, S+P condition.

E.7 Initial-Trust Prior Sensitivity

  • Table 7 compares 0.1 and 0.9 priors with Grok-4.1-Fast (R), free-form assessment, and S+P guidance; Figure 15 also includes the 0.5 main-paper reference.Figure 15 reports detection, misidentification, and trust separation across 10 seeds and 200 episodes per completed condition.Source: Table 7 · p. 33, Figure 15 · p. 34
  • A 0.1 prior versus 0.9 increases detection by 14.1 percentage points (CI 7.4 to 20.8), reduces misidentification by 6.0 points (CI 2.1 to 9.9), and increases separation by 0.181 (CI 0.097 to 0.265).Source: §E.7 · p. 33
  • The 0.1 prior delays neither agreement nor round-count outcomes clearly relative to 0.9: confidence intervals for agreement, honest surplus, and both round measures include zero.The authors conclude that a high initial prior delays separation of the adversary from honest parties without a clear improvement in negotiated outcomes.Source: §E.7 · p. 33

E.8 BRS Forgetting-Factor Sensitivity

This section examines BRS forgetting-factor sensitivity and documents how agents assess counterparts, update private trust, and use public histories in negotiation and trust updates. It also reproduces prompts specifying the game, roles, trust guidance, decisions, and feedback process; in the offline sensitivity analysis, γ=0.9 yields the highest adversary detection.

E.8 BRS Forgetting-Factor Sensitivity

  • Detection peaks at 91.0% for γ=0.9 (versus 86.5%, 89.8%, and 89.8% for γ=0.5, 0.7, and 1.0); at γ=0.9, misidentification is 2.0% and trust separation is 0.271.The analysis recomputes trust from logged evidence held fixed; alternate-factor curves do not represent negotiations conducted with those values. γ=0.9 is used because it gives the highest detection while retaining 92% of the separation without forgetting.Source: Figure 16 · p. 34, §E.8 · p. 34

F.1 Example 2HB+2G Game Prompt

  • The example prompt makes SkyDrop Logistics the focal agent in a 2HB+2G game and supplies its goal, five-party list, issue options, and priorities.Its private priorities include a top choice on environmental and noise controls and a second priority on workforce transition.Source: §F.1 · p. 35

F.2 System Template

  • The shared system template defines five-issue deals, proposal and voting rules, passage thresholds, round limits, and party-specific private information.If no deal passes by the final round, parties receive only their no-deal payoff; each party sees its own objectives, scores, threshold, disposition, and briefing.Source: §F.2 · p. 36

F.3 Private Objective and Role Blocks

  • Role blocks distinguish honest hard bargainers and pragmatic deal-makers from blockers whose impossible true threshold makes agreement failure their objective.Hard bargainers reject merely bearable offers by default, while pragmatic deal-makers favor workable agreements; blockers conceal their true objective behind a public cover.Source: §F.3 · p. 37

F.4 Trust Visibility and Decision Guidance

  • Trust-aware agents receive private, directed scores carried across episodes; S provides scores without usage instructions, while S+P adds explicit decision guidance.The trust-agnostic condition has no trust text or update call; the policy of adversarial parties remains fixed.Source: §F.4 · pp. 38–39

F.5 Proposal Prompt

  • The proposal prompt supplies the proposer’s trust scores in S and S+P, optionally adds the S+P reminder, and requests a full deal in JSON.The response includes a private scratchpad and a public rationale; the trust placeholders are empty without trust.Source: §F.5 · p. 39

F.6 Voting Prompt

  • The voting prompt gives each voter the current proposal, public rationale, history, and trust in the proposer, with an optional S+P reminder and a JSON accept-or-reject response.Votes already generated in the current voting stage are excluded from the history shown to other voters.Source: §F.6 · pp. 39–40

F.7 BRS Feedback Prompt

  • After each BRS episode, agents assess counterparts from the public transcript, reporting positive and negative observation counts with evidence; code normalizes counts and applies the update.The feedback prompt asks agents to trace stated conditions and responses across statements, proposals, and explicit votes.Source: §F.7 · p. 40
  • For each counterpart, the BRS feedback prompt records positive and negative observations from public evidence, excludes unsupported interpretations, and requires counts plus round-based evidence in its response.Firmness, rejections, repeated proposals, and unmet demands alone do not count as negative observations; distinct events are counted once, and unsupported or zero categories are identified.Source: §F.7 · pp. 40–41

F.8 Rubric-based Trust-Update Prompt

  • Agents choose one shared set of metric weights per episode, combine the behavior score with prior trust using the specified update, and clamp new trust to [0, 1].Higher behavior scores indicate conduct closer to obstruction and lower trust; agents report the weights and their rationale.Source: §F.8 · pp. 41–42
  • The rubric-based update scores five observable behaviors—counterfactual blame, proposer cooperation, outlier rejection, targeted failure, and vote-priority inconsistency—to form each party’s behavior score.Vote-priority inconsistency compares each non-proposer vote with the party’s public rationales and known priorities; the other metrics capture decisive failures, proposal acceptance, and rejection patterns.Source: §F.8 · pp. 41–42

F.9 Free-Form Trust-Update Prompt

  • The free-form update imposes no episode-assessment metrics: agents assign each party an observable-history-based trust signal, report their criteria and approximate weights, then apply a fixed EMA update.The signal ranges from 0 (fully untrustworthy) through 0.5 (neutral) to 1 (fully trustworthy), and the resulting trust is clamped to [0, 1].Source: §F.9 · p. 43

F.10 Prompt Rendering and Histories

  • The system prompt is regenerated for proposals, votes, and trust updates; negotiation calls receive recent rounds, while trust updates receive the complete episode transcript.Rendered rounds include proposals, rationales, votes, and outcomes; trust-update keys use canonical agent identifiers, while negotiation displays use agent names.Source: §F.10 · p. 44

Related papers

Loading 2610.31971v1…