pith. sign in

arxiv: 2605.16575 · v1 · pith:YGEPOTPTnew · submitted 2026-05-15 · 💻 cs.AI

Counterparty Modeling is Not Strategy: The Limits of LLM Negotiators

Romain Cosentino , Sarath Shekkizhar , Adam Earle , Silvio Savarese This is my paper

Pith reviewed 2026-05-20 18:09 UTC · model grok-4.3

classification 💻 cs.AI
keywords LLM negotiationpreference modelingstrategic bargainingmulti-attribute negotiationcounterparty preferencesbargaining outcomesconcession strategiesLLM agents
0
0 comments X

The pith

LLM negotiators accurately model their counterpart's preferences but do not use that knowledge to make strategic offers that improve their outcomes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether large language model agents can negotiate effectively in a multi-attribute bargaining setting once they know what the other party values. It establishes that agents quickly and accurately capture the counterparty's preferences in their reasoning, yet this information does not lead to better final agreements for the informed side. Agents respond to what they believe the other values but fail to link those responses to gains on their own high-value attributes, so outcomes remain shaped by initial offers rather than actual utility differences. The work shows that even forcing agents to state explicit concession-for-reciprocity trades improves the appearance of strategy in individual turns without raising the efficiency of the agreements reached.

Core claim

In a controlled multi-attribute bargaining environment, LLM agents model a counterparty's preferences accurately and early in their reasoning traces, but this does not reliably improve outcomes for the informed side. Turn-level analyses reveal that agents often respond to what they believe the counterparty values without consistently pairing those moves with gains on their own high-value attributes. Sellers tend to be more accommodating overall, and in asymmetric-information conditions the informed side frequently makes weakly compensated concessions. Because agents do not leverage the underlying utility structure, final agreements are heavily dictated by surface-level opening anchors rather

What carries the argument

The consistent pairing of concessions with gains on own high-value attributes, which agents fail to perform reliably even after accurate preference modeling.

If this is right

  • Informed agents achieve no reliable advantage over uninformed agents despite accurate modeling of the counterparty.
  • Sellers make more accommodating offers overall regardless of information condition.
  • In asymmetric settings the informed side tends to make concessions that are not offset by gains on its own priorities.
  • Final agreements remain correlated with initial offers rather than with the parties' true utility weights.
  • Requiring explicit statements of concession-for-reciprocity improves turn-level strategy appearance but does not raise final agreement efficiency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the same gap between modeling and strategic use appears in repeated or reputation-based negotiations, current LLM agents may require separate utility-tracking modules to function autonomously.
  • Training or prompting regimes that reward explicit utility maximization at each turn could close the observed performance gap.
  • The dominance of opening anchors suggests that negotiation benchmarks should include controls that randomize or remove initial offers to isolate the contribution of preference information.

Load-bearing premise

That strategic bargaining is defined by consistently trading concessions for gains on high-value attributes and that surface-level opening anchors dominate outcomes when this pairing is absent.

What would settle it

A controlled run in which agents that explicitly track and optimize their own cumulative utility across turns reach agreements with measurably higher own-utility scores than agents that receive the same preference information but do not perform this tracking.

Figures

Figures reproduced from arXiv: 2605.16575 by Adam Earle, Romain Cosentino, Sarath Shekkizhar, Silvio Savarese.

Figure 1
Figure 1. Figure 1: Buyer utility rises while seller utility declines across information conditions. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Outcome distributions remain buyer-favorable across information conditions. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Final price depends more on the opening price than on price utility weights. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Informed agents rapidly form accurate negotiating partner beliefs. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sellers accommodate while buyers withhold. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Counterparty-facing concessions are weakly compensated by own-priority gains. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The trade plan does not improve efficiency. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Negotiating partner modeling. Before any substantive bargaining occurs, the informed buyer explicitly identifies the seller’s highest-value attributes (higher price, Sedan, annual service). In fact, informed agents often possess a usable counterparty model early in the negotiation. Illustrative trade-plan collision Buyer — Turn 11 (<think>, ) Trade plan: Concede on down payment to get a higher trade-in. Bu… view at source ↗
Figure 9
Figure 9. Figure 9: One-step trade plans do not compose into coordinated bargaining. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The seller-informed asymmetry replicates: when the seller is informed, the buyer accumulates some utility gain while the seller does not. B.2. Reasoning Traces: Belief Accuracy and Alignment [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Final utility scatter for exp_asym. Across conditions, outcomes remain concentrated in a buyer￾favorable region, with the seller-informed condition shifting the mean rightward [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: cumulative signed-accuracy@5 over normalized turn fraction. [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: belief–action alignment by condition and role. [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Strategic coupling by role and condition. [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: DeepSeek-R1 exp_trade_plan: distance to Pareto frontier (left) and to the Nash solution (right), in normalized utility space. The pattern is qualitatively the same as for Qwen3-235B: information improves Pareto proximity, and the trade plan has no clear efficiency benefit in either regime. efficiency gain. The main unstable component is buyer-side policy: Qwen3-235B buyers tend to withhold, whereas DeepSe… view at source ↗
read the original abstract

Negotiation requires more than inferring what the other side wants: it requires using that information to make advantageous offers and counteroffers over multiple turns. We study whether large language model (LLM) agents do this in a controlled multi-attribute bargaining environment. We find that current LLM agents can model a counterparty's preferences, but do not reliably turn that knowledge into strategic bargaining. When given negotiating partner preference information, agents model it accurately and early in their reasoning traces, yet this does not reliably improve outcomes for the informed side. Turn-level analyses show why: agents often respond to what they believe the counterparty values, but do not consistently pair those moves with gains on their own high-value attributes. Sellers are more accommodating overall, and in asymmetric-information conditions, the informed side often makes the more weakly compensated concessions. Because agents fail to leverage this underlying utility structure for strategic advantage, their final agreements are heavily dictated by surface-level opening anchors rather than actual utility weights. Finally, requiring agents to explicitly state concession-for-reciprocity trades before making an offer makes individual turns look more strategic, but ultimately fails to improve the efficiency of the final agreements.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that LLM agents in controlled multi-attribute bargaining can accurately model counterparty preferences early in their reasoning traces, but do not reliably convert this into strategic bargaining. Turn-level analyses show agents respond to believed counterparty values without consistently pairing concessions with gains on their own high-value attributes; informed sides make weakly compensated concessions, sellers are more accommodating, and final agreements are dictated by surface-level opening anchors rather than utility weights. Explicitly requiring agents to state concession-for-reciprocity trades improves turn appearance but not final efficiency.

Significance. If the results hold, the work usefully separates preference modeling from strategic exploitation in LLM negotiators and supplies concrete behavioral evidence via turn-level tracing in a multi-attribute setting. The controlled experiments and direct measurement from agent traces are strengths that could guide development of better strategic modules, though generalization depends on the chosen operationalization of strategy.

major comments (2)
  1. [§4] §4 (turn-level analyses): The central claim that modeling fails to produce strategic advantage rests on defining strategic behavior as consistently pairing concessions with gains on own high-value attributes and on outcomes being dictated by opening anchors. This operationalization is load-bearing; the manuscript does not test or rule out alternative mechanisms (signaling, threat credibility, or anticipated reputation) that could allow preference knowledge to yield advantage even in one-shot settings.
  2. [Methods] Methods: The description of the controlled experiments lacks explicit detail on prompt templates for preference modeling and offer generation, exact number of trials per condition, and statistical controls for prompt sensitivity or random seed effects. Without these, it is difficult to assess whether the observed failure to improve efficiency is robust or sensitive to implementation choices.
minor comments (2)
  1. [Abstract] Abstract and §3: Define 'weakly compensated concessions' and 'surface-level opening anchors' more explicitly when first introduced, including how anchors are generated and held constant across conditions.
  2. [Figures] Figure captions: Ensure all figures reporting concession patterns or efficiency metrics include error bars or confidence intervals and state the number of runs underlying each bar.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of our operationalization and reproducibility that we will address to strengthen the manuscript. Below we respond point by point to the major comments.

read point-by-point responses

  1. Referee: [§4] §4 (turn-level analyses): The central claim that modeling fails to produce strategic advantage rests on defining strategic behavior as consistently pairing concessions with gains on own high-value attributes and on outcomes being dictated by opening anchors. This operationalization is load-bearing; the manuscript does not test or rule out alternative mechanisms (signaling, threat credibility, or anticipated reputation) that could allow preference knowledge to yield advantage even in one-shot settings.

    Authors: Our definition of strategic behavior follows directly from multi-attribute utility theory, in which rational agents make concessions on low-value issues to secure gains on high-value issues. This is the mechanism we test because the paper isolates whether preference modeling produces observable utility-improving actions in the agents' traces. Alternative mechanisms such as signaling or reputation effects are less relevant in our finite-horizon, non-repeated design, which deliberately minimizes repeated-game considerations. We nevertheless agree that explicitly discussing these alternatives would improve context. In the revision we will add a short subsection to §4 that (a) justifies the chosen operationalization against standard bargaining models, (b) notes the limited scope for signaling or reputation in one-shot or short-horizon settings, and (c) flags these as directions for future work. The core empirical claims and turn-level results will remain unchanged. revision: partial

  2. Referee: [Methods] Methods: The description of the controlled experiments lacks explicit detail on prompt templates for preference modeling and offer generation, exact number of trials per condition, and statistical controls for prompt sensitivity or random seed effects. Without these, it is difficult to assess whether the observed failure to improve efficiency is robust or sensitive to implementation choices.

    Authors: We agree that additional methodological detail is required for full reproducibility. The revised manuscript will expand the Methods section and add an appendix containing the complete prompt templates for both preference modeling and offer generation. We will also state that each condition was run for 100 independent trials, describe the use of three distinct random seeds, and report sensitivity checks performed by varying prompt phrasing while holding other factors fixed. Statistical controls (including t-tests with multiple-comparison correction and robustness to seed variation) will be summarized in the main text with full results in the supplement. These changes will be implemented without altering any experimental outcomes or conclusions. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical behavioral study with direct outcome measurement

full rationale

The paper conducts controlled experiments on LLM agents in multi-attribute bargaining, directly measuring preference modeling accuracy from reasoning traces and strategic behavior from turn-level concessions and final agreement utilities. No mathematical derivations, parameter fits, or self-citation chains are used to derive the central claims; results follow from observed agent traces and utility calculations without reducing to inputs by construction. The study is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard assumptions from negotiation theory about additive utility functions and the informativeness of preference revelation; no new entities are postulated.

axioms (2)
  • domain assumption Agents operate in a multi-attribute bargaining setting with known additive utility functions for each side.
    The experimental design supplies preference information and measures outcomes against those utilities.
  • domain assumption Strategic behavior can be identified by whether offers link concessions to gains on high-value attributes.
    This operationalization is used to interpret why modeling does not improve outcomes.

pith-pipeline@v0.9.0 · 5738 in / 1341 out tokens · 58295 ms · 2026-05-20T18:09:28.345241+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [1]

    **FOLLOW YOUR PREFERENCES**: You have specific preferences listed below. - PRIORITIZE items marked CRITICAL (most important) - PUSH FOR items marked IMPORTANT (but can compromise) - USE flexible items as bargaining chips - Your goal: get outcomes that match your preferences

    work page

  2. [2]

    - Concede on FLEXIBLE items to win on CRITICAL items - Don’t give away things you want without getting something back - Propose deals that maximize YOUR outcome

    **TRADE STRATEGICALLY**: Exchange things you care less about. - Concede on FLEXIBLE items to win on CRITICAL items - Don’t give away things you want without getting something back - Propose deals that maximize YOUR outcome

    work page

  3. [3]

    **REACH AGREEMENT**: Making a deal is important! - Any deal above your reservation value is better than no deal - If opponent offers seem reasonable, seriously consider accepting - Don’t let perfect be the enemy of good - Converge toward mutually beneficial terms

    work page

  4. [4]

    action":

    **UNDERSTAND CONSTRAINTS**: The other party has HARD LIMITS too! - They have minimum/maximum bounds they CANNOT violate - If they keep rejecting certain terms, you may be outside their feasible range 19 - EXPLORE different combinations - don’t get stuck demanding impossible terms - A successful deal requires finding terms that work for BOTH parties ## RES...

    work page

  5. [5]

    FOLLOW YOUR PREFERENCES - they determine your utility

    work page

  6. [6]

    PUSH for high-weight features (critical/important)

    work page

  7. [7]

    TRADE AWAY flexible items to get what you need

    work page

  8. [8]

    Express your preferences naturally through offers and reactions

    work page

  9. [9]

    None (you go first)

    Maximize utility = weighted sum of normalized features 20 C.3. Per-Turn Prompt At every negotiation step the agent receives a turn prompt assembled from the full dialogue history, the current structured offer, and a phase-dependent instruction block chosen by turn number and offer state. ## CONVERSATION: [full dialogue history] ## CURRENT OFFER ON TABLE: ...

    work page

  10. [10]

    SAY all the terms in your dialogue (model, price, delivery, etc.)

    work page

  11. [11]

    THEN include them in your JSON

    work page

  12. [12]

    action":

    Don’t include anything you didn’t explicitly mention Turns 6–15, offer on table (active bargaining): REACT to their offer. Push for better terms or accept if good enough. For your JSON: - Accept their offer? -> {"action": "ACCEPT"} - Change specific terms? -> Include ONLY the terms you want to change - Their offer auto-fills unchanged terms [!] IMPORTANT:...

    work page