acceptodds
Under review as a conference paper at ICLR 2027

On the Behavior of LLMs as Mediators in Bilateral Trade

Abstract

Large Language Model (LLM) agents increasingly act on behalf of people in commerce, including as mediators that help a buyer and a seller reach a deal. Evaluating such mediators is challenging since (i) their pricing policy is latent: we observe the prices they propose, not the rule that produces them, and (ii) outcomes such as deal rates depend on how the counterparties respond, so they cannot separate a good policy from a favorable environment. To address these challenges, we study LLM mediators in multi-round bilateral trade with private valuations, pair them with persona-parameterized deterministic agents whose responses are known by construction, and recover each mediator's revealed policy with a counterfactual audit that patches one input of the negotiation history at a time. Across 504 market configurations and three frontier model families, we find three consistent behaviors: buyer-facing proposals concede toward the buyer while seller-facing proposals stay near compromise prices, later proposals are driven mainly by the mediator's own earlier proposals rather than by the parties' offers, and the mediator filters offers by trajectory, tracking concessions that continue the bargaining path while ignoring offers that reverse it. We show that this filtering removes the gain from retracting a concession but leaves an incentive to stall, and we derive the feasibility horizon, surplus decomposition, and mechanism-design ceilings of the environment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLMs, Agent Negotiations, Model Behavior

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_u6zROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies LLMs acting as mediators in bilateral trade rather than as negotiators for one party. Its main methodological contribution is a counterfactual policy audit that perturbs components of the negotiation history and measures how the mediator's proposed price responds, with the goal of recovering a local revealed policy. The paper combines this analysis with a controlled environment of deterministic counterparties, theoretical characterizations of feasible trade and welfare benchmarks, and experiments across several frontier LLM families.

The main empirical conclusion is that LLM mediators exhibit side-dependent behavior, substantial dependence on their own previous proposals, and asymmetric responses to offers depending on whether they continue or reverse the previous bargaining trajectory. The paper further argues that these behaviors induce economically meaningful incentives, including robustness to retracted concessions but incentives for parties to stall.

Critical strengths

A meaningful evaluation perspective beyond endpoint metrics. Treating the mediator's revealed behavioral policy and the strategic incentives it induces as the central objects of study is both conceptually well motivated and potentially useful beyond bilateral bargaining. This is the strongest aspect of the work.

Good integration of controlled experimentation and economic analysis. Using counterparties with known policies makes it possible to disentangle mediator behavior from opponent behavior, while the theoretical benchmarks give the empirical outcomes interpretable reference points. The cross-model replication also provides some evidence that the main qualitative observations are not idiosyncratic to one LLM family.

Critical weaknesses

The evidence supporting the central “revealed policy” claims is too local relative to the breadth of the conclusions. The audit relies on one-variable-at-a-time perturbations around only a small number of base histories. Importantly, the paper itself notes that the audited histories share a particular negotiation pattern. For a highly nonlinear, history-dependent LLM policy, this leaves open whether self-anchoring and trajectory filtering are genuinely stable policy properties or behaviors specific to a narrow region of the state space. This matters directly because policy identification, rather than benchmark performance, is the paper's main contribution.

The economic conclusions remain strongly tied to a constructed environment. The deterministic agents use designed concession and acceptance rules, while the alternative LLM-counterparty environment behaves very differently and does not support the same welfare analysis. The current experiments therefore convincingly characterize behavior within the testbed, but provide weaker evidence that the identified incentives characterize realistic mediation settings. This substantially limits the scope of the broader conclusions.

Questions for authors

How stable are the reported self-anchoring and trajectory-filtering effects across a substantially larger and randomly or systematically stratified collection of negotiation histories, especially histories with different acceptance/rejection patterns? If these effects weaken materially outside the current audit histories, my assessment of the main contribution would change.

Because the audit changes one input at a time, have the authors evaluated whether the resulting local policy approximation predicts mediator outputs on joint perturbations of multiple history variables or held-out negotiation states? Such an experiment would help distinguish identification of an underlying policy from a collection of local marginal response curves.

Reviewer_ebHQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how LLM agents behave as mediators in bilateral negoiations between buyers and sellers with private valuations. The authors introduce a controlled negotiation environment with deterministic counterparties and a counterfactual auditing method that perturbs individual elements of negotiation histories to recover the mediator's implicit pricing policy.

Across Gemini 3.5 Flash, GPT-5.6, and Claude Sonnet 5, the authors identify three recurring behaviors: asymmetric pricing toward buyers and sellers, increasing reliance on the mediator's own previous proposals, and selective responsiveness to offers that follow existing concession trajectoies. They also analyze the resulting economic incentives, compare outcomes against heuristic mediators and mechanism design bounds, and evaluate robustness with LLM counterparties.

Critical strengths
  1. Counterfactual policy identification is a useful contribution. The auditing approach provides direct estimates of how individual negotiation variables affect pricing decisions. This is more informative than comparing deal rates and could generalize to other LLM agents making numerical decisions.

  2. The economic analysis provides useful explanations for observed behavior. The connection between trajectory filtering, self-anchoring, and bargaining incentives is interesting. The authors also replicate the main qualitative patterns across three model families and demonstrate that objective prompting can have counterintuitive effects on negotiation outcomes.

Critical weaknesses
  1. The ecnomic comparison is heavily influenced by the simulated environment and relatively weak baselines. Deterministic agents follow hand-designed concession schedules and restrictive first-round acceptance thresholds. Furthermore, all heuristic mediators initially forward the seller's asking price, and none reaches agreement in round one. Since early agreement is the primary source of Gemini's advantage, this design gives the LLM an important structural advantage. Comparisons against adaptive, deadline aware mediators would be necessary to establish whether the observed improvements reflect stronger mediation policies.

  2. The policy and incentive conclusions are insufficiently validated beyond a small set of histories. The detailed counterfactual analysis uses only three base histories per decision point, all involving previous seller acceptance followed by buyer rejection. Consequently, the measured self anchoring may partly reflect rational use of accetance information. The theoretical argument that trajectory filtering incentivizes stalling also assumes counterparties optimize the next proposed price rather than expected negotiation utility. No experiments demonstrate that strategically adapting agents actually exploit this behavior.

Questions for authors
  1. Given that none of the heuristic mediators closes agreements in round one, how much of Gemini's advantage would remain against adaptive or deadline aware mediators that explicitly optimize for early agreement?

  2. Does the strong self anchoring effect persist when the audit uses negotiation histories without prior seller acceptance, or when counterparties respond strategically to the mediator's observed policy?

Reviewer_uuBdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates LLM mediators in multi-round bilateral trade with private valuations. It also analyzes feasibility and surplus benchmarks and compares economic outcomes across mediator objectives. The main contribution is an audit of observable mediation behavior beyond aggregate agreement rates.

Critical strengths
  • The study is useful and interesting.
  • The paper provides informative evidence about model and environment dependence, also provide some theoretical insights.
Critical weaknesses
  • The validity and consistency of the inferred mediation policy are not sufficiently established. The counterfactual audit measures how numerical proposals respond to edited inputs, but the stronger interpretation of these responses as a coherent reasoning mechanism needs additional validation. Appendix G explicitly requests explanations of calculations, reference values, and strategies, yet the paper does not systematically assess whether these explanations reproduce the proposed prices or agree with the measured input sensitivities. Such agreement would be a useful consistency check, although it would not by itself establish the faithfulness of the explanations.
  • More fundamentally, the central round-3 audit uses only three histories per decision point, all containing a seller acceptance followed by buyer rejection. Editing a price while retaining its acceptance or rejection annotation also changes the information available about willingness to trade.
  • The three-family comparison is a strength, but the evidence does not establish that all reported behaviors are equally robust across models and configurations. Appendix A states that the round-5 and temperature-zero analyses, among other diagnostics, are Gemini-only. The models are also evaluated with different provider-specific low-reasoning configurations, without a sensitivity analysis establishing how the findings depend on reasoning budget.
Questions for authors
  • If I understand correctly, the paper also does not sufficiently validate whether the assembled local policy predicts responses to combined input changes or subsequent negotiation rounds. This matters because the conclusions about strategic incentives extend beyond the single-step price responses directly measured by the audit. How do the authors address it?
  • How sensitive are the main findings to the deterministic agents’ first-round acceptance thresholds and persona-derived parameters?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.