Pith. sign in

REVIEW 3 major objections 5 minor 64 references

The paper introduces Bazaar, a benchmark showing that frontier LLM merchants in a dynamic auction market capture less than a third of hindsight-optimal profit, with profit driven by margin discipline rather than win rate.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 01:16 UTC pith:T42UYOE5

load-bearing objection A useful new benchmark for LLM pricing, but the headline oracle-efficiency number is inflated by an integer-price mismatch; the qualitative findings hold. the 3 major comments →

arxiv 2608.00102 v1 pith:T42UYOE5 submitted 2026-07-30 cs.AI cs.MA

Can LLM Agents Price Competitively? A Dynamic Multi-Attribute Auction Benchmark for Agentic Commerce

Shimaa Ahmed , Yiwei Cai , Mohsen Minaei , Rahul Rachuri This is my paper
classification cs.AI cs.MA
keywords agentic commerceLLM pricingmulti-attribute auctionregret decompositiondemand shockmargin disciplinebenchmarksealed-bid auction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether LLM agents can act as merchants—learning hidden customer preferences, setting prices, and revising beliefs when demand shifts—in a repeated competitive auction. It introduces Bazaar, a sealed-bid multi-attribute auction benchmark with closed-form utilities, and evaluates 11 frontier LLMs. The central finding is that even the best agent captures less than a third of the profit a hindsight-optimal seller would earn. Profit is almost perfectly explained by margin per win, not win rate; more thinking effort helps models win more but can leave money on the table. A demand-shock experiment reveals that the fastest pre-shock learners are often the slowest to revise, a failure mode static benchmarks cannot see.

Core claim

The paper's central claim is that current frontier LLMs, when placed in a dynamic market loop, can learn to win customers but not to price profitably: total profit across models spans an 11x range, and the strongest agent reaches only 0.321 regret efficiency, about a third of the hindsight-optimal profit. Profit correlates with margin per win at r=0.99, versus r=0.88 with win rate, so the decisive skill is charging enough, not merely winning. Increasing inference effort moves models along a two-dimensional failure surface: it reduces lost-auction regret but increases under-pricing regret. When preferences shift without warning, models that learned fastest before the shock tend to adapt slowe

What carries the argument

The central object is Bazaar's repeated sealed-bid multi-attribute auction: each round, an LLM merchant submits a configuration over three attributes and a price to each of 24 customers with hidden additive utilities; the customer picks the offer with highest utility. Three rule-based specialist bots adapt their prices, and 12 customers undergo an unannounced pairwise swap of their value curves mid-game. Because utilities and costs are closed-form, the paper can compute a hindsight-oracle profit and decompose regret into loss regret (auctions the oracle would have won) and underprice regret (margin left on the table in won auctions). The margin-per-win statistic is the diagnostic that carrie

Load-bearing premise

The load-bearing assumption is that 30 rounds give every model enough time to saturate before the preference shock; if some models are still learning when the shock hits, the measured learning-speed and recovery rankings could be artifacts of where the shock is placed.

What would settle it

Re-run Bazaar with the pre-shock window extended from 30 to 60 rounds, keeping the same shock schedule; if the fastest pre-shock learners still show the largest post-shock degradation after full saturation, the strong-learner/weak-adapter claim is supported, and if they recover as well as the fastest adapter, the reversal was an artifact of the window.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central finding holds, inference-time compute is a first-order determinant of agentic-commerce performance: GPT-5.4 earns 7.3x more profit at high effort than at none.
  • Rankings are objective-dependent: the model that wins the most customers (77.9% win rate) is not the profit leader, so reporting coverage alone can mislead.
  • Fast pre-shock learning does not imply fast post-shock adaptation; agentic-commerce benchmarks should include unannounced preference shifts to surface this failure mode.
  • Classical bandits set a non-trivial floor: per-customer Thompson Sampling beats six of eleven base LLMs on profit.
  • Newer model versions are not reliably better: a newer model in three of four families regresses against an earlier version at default settings, and thinking effort can reverse those gaps.

Where Pith is reading between the lines

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

  • If margin discipline is the binding constraint, evaluation and monitoring of deployed pricing agents should track surplus extraction (margin per win relative to a ceiling), not just win rate or revenue.
  • The strong-learner/weak-adapter pattern suggests that confident exploitation strategies may trade away adaptivity; a testable extension is to add explicit hypothesis-revision prompts after consecutive losses and measure recovery.
  • The benchmark pits one LLM against rule-based bots; running full LLM-vs-LLM competition could compress or amplify the observed gaps, and would test whether the bots' asymmetric price adjustment makes the environment easier or harder than live markets.
  • Because customer preferences are synthetic archetypes, a natural follow-up is to ground Bazaar in real purchase or preference data; the regret-decomposition machinery transfers directly, but the absolute profit gaps may change.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces Bazaar, a repeated sealed-bid multi-attribute auction benchmark in which an LLM merchant competes with three adaptive specialist bots for 24 customers with hidden additive preferences. A staggered preference shock tests adaptation. The environment uses closed-form utilities and costs, so surplus and hindsight regret can be computed exactly. The authors evaluate eleven frontier LLMs and their variants, reporting that total profit spans an 11x range, correlates strongly with margin per win (r=0.99) rather than win rate (r=0.88), that thinking effort shifts models from Loser to Underpricer failure archetypes, and that strong pre-shock learners are typically weak post-shock adapters. The best reported efficiency is η=0.321, i.e., less than a third of hindsight-optimal profit, leaving substantial headroom.

Significance. Bazaar is a valuable evaluation instrument: it combines learning, pricing, and belief revision in a dynamic competitive setting, uses exact oracle-based diagnostics rather than learned proxies, includes Thompson Sampling and EXP4 baselines, reports token/API costs, and states that prompts and logs are released. The margin-per-win vs. win-rate dissociation and the loss/underprice regret decomposition are genuinely useful diagnostics. The main central claims, however, rest on the exactness of the oracle and on the chosen pre-shock window. The integer-price issue and the empirical window selection are fixable and do not invalidate the benchmark design, but they affect the headline efficiency number and the adaptation findings.

major comments (3)
  1. [§2.5, Eq. (4); Appendix P] The oracle in Eq. (4) treats price as continuous, but the implemented game constrains bids to integer prices (Appendix P: 'price': <integer >= minimum cost>; §3.3 examples). For a winnable customer-round, the continuous oracle profit is V_i,t(x)-q_i,t-c_f(x), while the best integer-feasible profit is floor(V_i,t(x)-q_i,t)-c_f(x) (or one less if ties are not guaranteed wins). The oracle is therefore inflated by the fractional part of V-q on every winnable round. Across 24 customers and ~80 rounds this can be hundreds of dollars. For Opus 4.6 (adaptive/xhigh), reported profit $2,976 and oracle ~$9,282 give η=0.321; a correction of a few hundred dollars moves η above 1/3, so the headline 'less than a third of hindsight-optimal profit' is not robust. The underprice-regret component in §4.4 is likewise measured against a continuous optimum. Please recompute all η values, tier boundaries, and
  2. [§3.5, §4.2] The 30-round pre-shock window is stated to have been 'chosen empirically across our model sweep.' Learning speed is measured as the rise in win rate to the R26–30 peak, and shock recovery is measured relative to that peak. If some models had not saturated by round 30, their post-shock trajectories conflate continued pre-shock learning with adaptation; selecting the window after observing model trajectories also risks overfitting the reversal. The strong-learner/weak-adapter reversal is a load-bearing contribution. Please provide a principled saturation criterion and sensitivity analyses over window lengths (e.g., 20, 30, 40) and shock timings, and report whether the reversal and recovery rankings persist.
  3. [Table 3; Table 6; §3.5] Table 3 places GPT-5.5 xhigh in Tier 1 with a footnote '‡3 seeds rather than 10', and Table 6 again marks this row as 3-seed. This is inconsistent with the text's claim that all models are averaged over ten seeds (with ten seeds fixing bot randomness and shock schedules). The permutation-test tier boundaries in §4.1 are computed on seed-level outcomes; a 3-seed row cannot be placed in a tier with the same confidence. The same row also supports the efficiency claim 0.307 and the statement that GPT-5.5 xhigh 'clears GPT-5.3.' Please either run that cell to 10 seeds or explicitly mark the row as preliminary and exclude it from tier and efficiency comparisons. Also state the tie-breaking rule for equal-utility offers; Eq. (4) appears to assume that a tie is a win.
minor comments (5)
  1. [§3.7] Typo: 'met:ics' should be 'metrics.'
  2. [§2.5 / Appendix P] The bid-stage prompt states that ties reveal nothing because 'the winner was random.' If tie-breaking is random, the oracle in Eq. (4) should account for strict preference rather than silently assuming a tie wins, or the tie rule should be defined and enforced in the environment.
  3. [§4.5 vs. Appendix J] The main text says estimated API cost varies 14× across models and that Gemini 3.1 Pro is the most cost-effective; Appendix J says cost varies ~50×, with GPT-5.3 Chat the cheapest at $26/seed. Please reconcile these numbers and clarify which subset each figure refers to.
  4. [Appendix O] The EXP4 update uses K both for the number of arms and for the attribute count in the main text; use distinct symbols or clarify the notation.
  5. [Appendix L] The anchor probes use 3 seeds and 30 rounds. The caveat says they are confirmatory; this is acceptable, but the main text should not cite them as evidence beyond their stated power.

Circularity Check

0 steps flagged

No significant circularity: the benchmark's oracle and regret metrics are computed from hidden ground-truth values and independent baselines, not from the agents' own outputs.

full rationale

The paper's central evaluation chain is self-contained. The oracle in Eq. (4) is defined directly from hidden customer values V_{i,t}, the best competitor utility q_{i,t}, and the focal merchant's cost c_f(x): π⋆_{i,t} = max_x [V_{i,t}(x) − q_{i,t} − c_f(x)]_+. As the paper states, 'Because these quantities depend on hidden values and costs, merchants cannot observe them; we use them only for evaluation.' Regret efficiency η = realized/oracle profit is therefore a genuine external comparison against a ground-truth upper bound, not a quantity fitted to or derived from the agent's behavior. The bandit baselines (Thompson Sampling, EXP4) are independent algorithms from the cited literature and are run on the same action space with the same win/loss feedback; they are not constructed from the LLM results. The 30-round pre-shock window in §3.5 was chosen empirically across the model sweep, but it is a protocol parameter that defines measurement windows for learning-speed and recovery analysis; it does not enter the oracle, the profit metric, or any quantity presented as a prediction. No load-bearing self-citations appear, and no 'uniqueness theorem' or prior-work ansatz is invoked to force the framework's choices. The skeptic-identified integer-price versus continuous-oracle discrepancy is a metric-construction concern about the tightness of the oracle upper bound, not a circularity: it does not reduce the headline result to its own inputs. Accordingly, there are no circular steps to report.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

All absolute profit and efficiency numbers live inside a hand-built synthetic environment. The free parameters are environment constants, not fitted to real market data. The pre-shock window is the one parameter selected with knowledge of model behavior, so it carries more selection risk than the others. No new physical or theoretical entity is introduced; Bazaar is a benchmark protocol, not a postulated object of nature.

free parameters (4)
  • Base cost and attribute cost ladder = c0=70; C=[0,1,3,6,12]; cost multipliers: focal 0.75, specialists 0.4/1.0
    Hand-chosen in §3.1. Defines the structural cost advantages and the minimum-price floor; all absolute profit and efficiency numbers are denominated in this scale.
  • Customer value curve shapes and noise = flat=[0,1,1.5,1.8,2], mid=[0,2,6,10,11], late=[0,2,4,7,13], noise U[-0.3,0.3] on nonzero entries
    Hand-constructed archetype clusters in §3.2/Appendix D determine what the agents must learn and what the oracle knows. Different curves would change absolute profit and recovery numbers.
  • Pre-shock window length and shock schedule = 30 pre-shock rounds; shocks staggered uniformly in {31,...,40}; 12 of 24 customers shocked; 4 per swap type
    §3.2/§3.5. The 30-round window was chosen empirically across the model sweep, so it directly shapes the learning-speed and recovery metrics that support the strong-learner/weak-adapter claim.
  • Bot margin update dynamics = open margin $3; raise ~U(0.5,1.5) after a win, lower ~U(0.25,0.75) after a loss, floor $1
    §3.4/Appendix E. Designed to prevent trivial underbidding and shapes the competitive pressure, win rates, and absolute profit levels experienced by every model.
axioms (5)
  • domain assumption Customer utility is additive and separable: Ui,t(x,p)=Σk Vi,t,k[xk] − p.
    §2.2. This is what makes the oracle and regret decomposition exact. If real customer preferences were non-additive, the computed 'hindsight-optimal profit' would not be exact for real markets.
  • domain assumption Merchant costs are additive: cm(x)=cm,0+Σk Cm,k[xk], with bids constrained to p≥cm(x).
    §2.2. Defines profit and the cost-based structural advantage; the specialist-bot design relies on it.
  • domain assumption Opponent bots always offer their specialty configuration and only adapt price with the specified margin dynamics.
    §3.4. Creates the controlled competitive environment. As the Limitations state, one LLM versus three identical bots is not the same as multi-LLM competition.
  • domain assumption Structured reasoning and strategy text produced by the LLM is a faithful proxy for its internal belief state and strategy.
    §4.3/Appendices M/N. The behavioral archetypes and the 'revision is the bottleneck' conclusion rest on reading these texts; the paper itself calls this analysis qualitative.
  • ad hoc to paper The empirically chosen 30-round pre-shock window is long enough for all models to reach saturation before the first shock.
    §3.5. Chosen across the model sweep. If any model had not saturated, its pre-shock peak and post-shock recovery numbers are not directly comparable.

pith-pipeline@v1.3.0-alltime-deepseek · 24877 in / 14910 out tokens · 161753 ms · 2026-08-04T01:16:47.542819+00:00 · methodology

0 comments
read the original abstract

Agentic commerce is moving from concept to deployed infrastructure: payment networks, retailers, and AI platforms are setting the stage for agents to transact on behalf of merchants and consumers. Yet whether the LLMs behind these agents can price competently in real markets, where customer preferences are hidden, competitors adapt in real time, and demand can shift without warning, has not been systematically tested. We introduce Bazaar, a dynamic sealed-bid benchmark for multi-attribute auction under these conditions. Despite its dynamics, the benchmark is grounded in closed-form customer utilities, enabling exact evaluation. Across 11 frontier LLMs from four providers, the leading agents on customer acquisition (e.g. Gemini 3.1 Pro) are often not the leading agents on profit (e.g. Opus 4.6). The ranking shifts again under demand shocks: agents that learned fastest pre-shock are typically the slowest to revise their beliefs afterwards, while Gemini 3.1 Pro recovers fastest despite not leading on profit. However, even the strongest agent captures less than a third of hindsight-optimal profit, suggesting current LLMs are progressing in agentic commerce but leave substantial headroom.

Figures

Figures reproduced from arXiv: 2608.00102 by Mohsen Minaei, Rahul Rachuri, Shimaa Ahmed, Yiwei Cai.

Figure 1
Figure 1. Figure 1: Abstract overview of Bazaar. adaptive competitors, and preference shifts. De￾spite the dynamics, closed-form customer utilities and merchant costs enable an exact evaluation. Its surplus and oracle-regret diagnostics separate per￾formance into preference learning (configuration choice), margin extraction (price), and post-shift re￾vision. This lets us classify agents into diagnostic failure archetypes (Los… view at source ↗
Figure 2
Figure 2. Figure 2: Profit is driven by margin, not coverage. Each [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The regret plane. Top-left Losers forfeit auc￾tions the oracle would win; bottom-right Underpricers win but leave surplus on the table. base, Grok NR, Opus 4.7, Sonnet 4.6) forfeit the majority of their regret by losing auctions the oracle would have won: they bid too low or on the wrong configuration, and competitors take the customer entirely. Underpricers (Gemini 3.1 Pro, GPT-5.4 high, GPT-5.5 high) win… view at source ↗
Figure 4
Figure 4. Figure 4: Concrete instantiation of Bazaar used in our experiments. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Learning curves, rounds 1–30 (pre-shock). All models cold-start below 40% win rate and rise sharply; the [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Win rate on shocked customers, aligned by the per-customer shock round. Pre-shock is the R26–30 peak [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mean profit vs. API cost. Bubble color matches the model family [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

64 extracted references · 2 canonical work pages

  1. [1]

    2024 , eprint=

    Measuring Bargaining Abilities of LLMs: A Benchmark and A Buyer-Enhancement Method , author=. 2024 , eprint=

    2024

  2. [2]

    2025 , howpublished =

    Project Vend 2 , author =. 2025 , howpublished =

    2025

  3. [4]

    2025 , eprint=

    Evaluating Multi-Turn Bargain Skills in LLM-Based Seller Agent , author=. 2025 , eprint=

    2025

  4. [5]

    2025 , eprint=

    Magentic Marketplace: An Open-Source Environment for Studying Agentic Markets , author=. 2025 , eprint=

    2025

  5. [6]

    2024 , eprint=

    NegotiationToM: A Benchmark for Stress-testing Machine Theory of Mind on Negotiation Surrounding , author=. 2024 , eprint=

    2024

  6. [7]

    2024 , eprint=

    Game-theoretic LLM: Agent Workflow for Negotiation Games , author=. 2024 , eprint=

    2024

  7. [10]

    2023 , eprint =

    Put Your Money Where Your Mouth Is: Evaluating Strategic Planning and Execution of LLM Agents in an Auction Arena , author =. 2023 , eprint =

    2023

  8. [11]

    How Well Can

    Bianchi, Federico and Chia, Patrick John and Yuksekgonul, Mert and Tagliabue, Jacopo and Jurafsky, Dan and Zou, James , booktitle =. How Well Can. 2024 , volume =

    2024

  9. [12]

    2024 , eprint =

    Simulation-Based Benchmarking of Reinforcement Learning Agents for Personalized Retail Promotions , author =. 2024 , eprint =

    2024

  10. [13]

    2024 , month =

    Algorithmic Pricing: Implications for Marketing Strategy and Regulation , author =. 2024 , month =. doi:10.3386/w32540 , url =

    work page doi:10.3386/w32540 2024

  11. [14]

    2025 , eprint =

    Learning from Synthetic Labs: Language Models as Auction Participants , author =. 2025 , eprint =

    2025

  12. [15]

    2024 , howpublished =

    Evidence from the Synthetic Laboratory: Language Models as Auction Participants , author =. 2024 , howpublished =

    2024

  13. [16]

    2025 , eprint =

    InfoBid: A Simulation Framework for Studying Information Disclosure in Auctions with Large Language Model-based Agents , author =. 2025 , eprint =

    2025

  14. [18]

    2024 , eprint =

    Peng, Bo and Ling, Xinyi and Chen, Ziru and Sun, Huan and Ning, Xia , booktitle =. 2024 , eprint =

    2024

  15. [19]

    2022 , eprint =

    Yao, Shunyu and Chen, Howard and Yang, John and Narasimhan, Karthik , booktitle =. 2022 , eprint =

    2022

  16. [21]

    2026 , eprint =

    Shopping Companion: A Memory-Augmented LLM Agent for Real-World E-Commerce Tasks , author =. 2026 , eprint =

    2026

  17. [22]

    American Economic Review , volume =

    Artificial Intelligence, Algorithmic Pricing, and Collusion , author =. American Economic Review , volume =. 2020 , doi =

    2020

  18. [23]

    The RAND Journal of Economics , volume =

    Properties of Scoring Auctions , author =. The RAND Journal of Economics , volume =. 2008 , doi =

    2008

  19. [24]

    Biometrika , volume =

    On the Likelihood that One Unknown Probability Exceeds Another in View of the Evidence of Two Samples , author =. Biometrika , volume =

  20. [25]

    SIAM Journal on Computing , volume =

    The Nonstochastic Multiarmed Bandit Problem , author =. SIAM Journal on Computing , volume =

  21. [26]

    44th Annual IEEE Symposium on Foundations of Computer Science (FOCS) , pages =

    The Value of Knowing a Demand Curve: Bounds on Regret for Online Posted-Price Auctions , author =. 44th Annual IEEE Symposium on Foundations of Computer Science (FOCS) , pages =. 2003 , publisher =

    2003

  22. [27]

    Operations Research , volume =

    Dynamic Pricing Without Knowing the Demand Function: Risk Bounds and Near-Optimal Algorithms , author =. Operations Research , volume =. 2009 , doi =

    2009

  23. [28]

    Surveys in Operations Research and Management Science , volume =

    Dynamic Pricing and Learning: Historical Origins, Current Research, and New Directions , author =. Surveys in Operations Research and Management Science , volume =. 2015 , doi =

    2015

  24. [29]

    Proceedings of the 29th International Conference on Machine Learning (ICML) , pages =

    Online Bandit Learning Against an Adaptive Adversary: from Regret to Policy Regret , author =. Proceedings of the 29th International Conference on Machine Learning (ICML) , pages =. 2012 , url =

    2012

  25. [30]

    Proceedings of the 16th ACM Conference on Economics and Computation (EC) , pages =

    Econometrics for Learning Agents , author =. Proceedings of the 16th ACM Conference on Economics and Computation (EC) , pages =. 2015 , doi =

    2015

  26. [31]

    Management Science , volume =

    Learning in Repeated Auctions with Budgets: Regret Minimization and Equilibrium , author =. Management Science , volume =. 2019 , doi =

    2019

  27. [32]

    2004 , publisher =

    The Theory and Practice of Revenue Management , author =. 2004 , publisher =

    2004

  28. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Regret Minimization in Games with Incomplete Information , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =. 2007 , url =

    2007

  29. [34]

    2020 , publisher =

    Bandit Algorithms , author =. 2020 , publisher =

    2020

  30. [35]

    Journal of Political Economy , volume =

    Auctions versus Posted Prices in Online Markets , author =. Journal of Political Economy , volume =. 2018 , doi =

    2018

  31. [36]

    American Economic Review , volume =

    Internet Advertising and the Generalized Second-Price Auction: Selling Billions of Dollars Worth of Keywords , author =. American Economic Review , volume =. 2007 , doi =

    2007

  32. [37]

    International Journal of Industrial Organization , volume =

    Position Auctions , author =. International Journal of Industrial Organization , volume =. 2007 , doi =

    2007

  33. [38]

    Management Science , volume =

    Optimal Dynamic Pricing of Inventories with Stochastic Demand over Finite Horizons , author =. Management Science , volume =. 1994 , doi =

    1994

  34. [39]

    Operations Research , volume =

    Dynamic Pricing with an Unknown Demand Model: Asymptotically Optimal Semi-Myopic Policies , author =. Operations Research , volume =. 2014 , doi =

    2014

  35. [40]

    2024 , eprint =

    Algorithmic Collusion by Large Language Models , author =. 2024 , eprint =

    2024

  36. [41]

    Anthropic . 2025. https://www.anthropic.com/research/project-vend-2 Project vend 2 . Anthropic research blog

    2025

  37. [42]

    Raman Arora, Ofer Dekel, and Ambuj Tewari. 2012. https://arxiv.org/abs/1206.6400 Online bandit learning against an adaptive adversary: from regret to policy regret . In Proceedings of the 29th International Conference on Machine Learning (ICML), pages 1747--1754

    Pith/arXiv arXiv 2012

  38. [43]

    John Asker and Estelle Cantillon. 2008. https://doi.org/10.1111/j.1756-2171.2008.00004.x Properties of scoring auctions . The RAND Journal of Economics, 39(1):69--85

    arXiv 2008

  39. [44]

    Schapire

    Peter Auer, Nicol \`o Cesa-Bianchi, Yoav Freund, and Robert E. Schapire. 2002. The nonstochastic multiarmed bandit problem. SIAM Journal on Computing, 32(1):48--77

    2002

  40. [45]

    Balseiro and Yonatan Gur

    Santiago R. Balseiro and Yonatan Gur. 2019. https://doi.org/10.1287/mnsc.2018.3174 Learning in repeated auctions with budgets: Regret minimization and equilibrium . Management Science, 65(9):3952--3968

    arXiv 2019

  41. [46]

    Hofman, Brendan Lucier, Chinmay Singh, Markus Mobius, Akshay Nambi, Archana Yadav, Kevin Gao, David M

    Gagan Bansal, Wenyue Hua, Zezhou Huang, Adam Fourney, Amanda Swearngin, Will Epperson, Tyler Payne, Jake M. Hofman, Brendan Lucier, Chinmay Singh, Markus Mobius, Akshay Nambi, Archana Yadav, Kevin Gao, David M. Rothschild, Aleksandrs Slivkins, Daniel G. Goldstein, Hussein Mozannar, Nicole Immorlica, and 5 others. 2025. https://arxiv.org/abs/2510.25779 Mag...

    arXiv 2025

  42. [47]

    Omar Besbes and Assaf Zeevi. 2009. https://doi.org/10.1287/opre.1080.0640 Dynamic pricing without knowing the demand function: Risk bounds and near-optimal algorithms . Operations Research, 57(6):1407--1420

    arXiv 2009

  43. [48]

    Federico Bianchi, Patrick John Chia, Mert Yuksekgonul, Jacopo Tagliabue, Dan Jurafsky, and James Zou. 2024. https://proceedings.mlr.press/v235/bianchi24a.html How well can LLM s negotiate? N egotiation A rena platform and analysis . In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Resea...

    2024

  44. [49]

    Emilio Calvano, Giacomo Calzolari, Vincenzo Denicol \`o , and Sergio Pastorello. 2020. https://doi.org/10.1257/aer.20190623 Artificial intelligence, algorithmic pricing, and collusion . American Economic Review, 110(10):3267--3297

    doi:10.1257/aer.20190623 2020

  45. [50]

    Chunkit Chan, Cheng Jiayang, Yauwai Yim, Zheye Deng, Wei Fan, Haoran Li, Xin Liu, Hongming Zhang, Weiqi Wang, and Yangqiu Song. 2024. https://arxiv.org/abs/2404.13627 Negotiationtom: A benchmark for stress-testing machine theory of mind on negotiation surrounding . Preprint, arXiv:2404.13627

    Pith/arXiv arXiv 2024

  46. [51]

    Haibin Chen, Kangtao Lv, Chengwei Hu, Yanshi Li, Yujin Yuan, Yancheng He, Xingyao Zhang, Langming Liu, Shilei Liu, Wenbo Su, and Bo Zheng. 2025. https://arxiv.org/abs/2502.20196 ChineseEcomQA : A scalable e-commerce concept evaluation benchmark for large language models . Preprint, arXiv:2502.20196

    Pith/arXiv arXiv 2025

  47. [52]

    Jiangjie Chen, Siyu Yuan, Rong Ye, Bodhisattwa Prasad Majumder, and Kyle Richardson. 2023. https://arxiv.org/abs/2310.05746 Put your money where your mouth is: Evaluating strategic planning and execution of llm agents in an auction arena . Preprint, arXiv:2310.05746

    Pith/arXiv arXiv 2023

  48. [53]

    den Boer

    Arnoud V. den Boer. 2015. https://doi.org/10.1016/j.sorms.2015.03.001 Dynamic pricing and learning: Historical origins, current research, and new directions . Surveys in Operations Research and Management Science, 20(1):1--18

    work page doi:10.1016/j.sorms.2015.03.001 2015

  49. [54]

    Gonczarowski, and Ran I

    Sara Fish, Yannai A. Gonczarowski, and Ran I. Shorrer. 2024. https://arxiv.org/abs/2404.00806 Algorithmic collusion by large language models . Preprint, arXiv:2404.00806

    arXiv 2024

  50. [55]

    He He, Derek Chen, Anusha Balakrishnan, and Percy Liang. 2018. https://doi.org/10.18653/v1/D18-1256 Decoupling strategy and generation in negotiation dialogues . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2333--2343, Brussels, Belgium. Association for Computational Linguistics

    doi:10.18653/v1/d18-1256 2018

  51. [56]

    Wenyue Hua, Ollie Liu, Lingyao Li, Alfonso Amayuelas, Julie Chen, Lucas Jiang, Mingyu Jin, Lizhou Fan, Fei Sun, William Wang, Xintong Wang, and Yongfeng Zhang. 2024. https://arxiv.org/abs/2411.05990 Game-theoretic llm: Agent workflow for negotiation games . Preprint, arXiv:2411.05990

    Pith/arXiv arXiv 2024

  52. [57]

    Gurusha Juneja, Jayanth Naga Sai Pasupulati, Alon Albalak, Wenyue Hua, and William Yang Wang. 2025. Magpie: A benchmark for multi-agent contextual privacy evaluation. arXiv preprint arXiv:2510.15186

    arXiv 2025

  53. [58]

    Robert Kleinberg and Tom Leighton. 2003. https://doi.org/10.1109/SFCS.2003.1238232 The value of knowing a demand curve: Bounds on regret for online posted-price auctions . In 44th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 594--605. IEEE

    Pith/arXiv arXiv 2003

  54. [59]

    Bo Peng, Xinyi Ling, Ziru Chen, Huan Sun, and Xia Ning. 2024. https://arxiv.org/abs/2402.08831 eCeLLM : Generalizing large language models for e-commerce from large-scale, high-quality instruction data . In Proceedings of the 41st International Conference on Machine Learning (ICML), volume 235 of PMLR

    Pith/arXiv arXiv 2024

  55. [60]

    Wang, Arif Kerem Dayi, John J

    Anand Shah, Kehang Zhu, Yanchen Jiang, Jeffrey G. Wang, Arif Kerem Dayi, John J. Horton, and David C. Parkes. 2025. https://arxiv.org/abs/2507.09083 Learning from synthetic labs: Language models as auction participants . Preprint, arXiv:2507.09083

    Pith/arXiv arXiv 2025

  56. [61]

    Eilam Shapira, Omer Madmon, Itamar Reinman, Samuel Joseph Amouyal, Roi Reichart, and Moshe Tennenholtz. 2024. Glee: A unified framework and benchmark for language-based economic environments. arXiv preprint arXiv:2410.05254

    arXiv 2024

  57. [62]

    Talluri and Garrett J

    Kalyan T. Talluri and Garrett J. van Ryzin. 2004. https://doi.org/10.1007/b139000 The Theory and Practice of Revenue Management . Springer

    doi:10.1007/b139000 2004

  58. [63]

    Thompson

    William R. Thompson. 1933. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3--4):285--294

    1933

  59. [64]

    Issue Yishu Wang, Kakam Chong, Xiaofeng Wang, Xu Yan, DeXin Kong, Chen Ju, Ming Chen, Shuai Xiao, Shuguang Han, and jufeng chen. 2025. https://arxiv.org/abs/2509.06341 Evaluating multi-turn bargain skills in llm-based seller agent . Preprint, arXiv:2509.06341

    Pith/arXiv arXiv 2025

  60. [65]

    Tian Xia, Zhiwei He, Tong Ren, Yibo Miao, Zhuosheng Zhang, Yang Yang, and Rui Wang. 2024. https://arxiv.org/abs/2402.15813 Measuring bargaining abilities of llms: A benchmark and a buyer-enhancement method . Preprint, arXiv:2402.15813

    Pith/arXiv arXiv 2024

  61. [66]

    Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022. https://arxiv.org/abs/2207.01206 WebShop : Towards scalable real-world web interaction with grounded language agents . In Advances in Neural Information Processing Systems, volume 35, pages 20744--20757

    Pith/arXiv arXiv 2022

  62. [67]

    Yue Yin. 2025. https://arxiv.org/abs/2503.22726 Infobid: A simulation framework for studying information disclosure in auctions with large language model-based agents . Preprint, arXiv:2503.22726

    Pith/arXiv arXiv 2025

  63. [68]

    Zijian Yu, Kejun Xiao, Huaipeng Zhao, Tao Luo, and Xiaoyi Zeng. 2026. https://arxiv.org/abs/2603.14864 Shopping companion: A memory-augmented llm agent for real-world e-commerce tasks . Preprint, arXiv:2603.14864

    Pith/arXiv arXiv 2026

  64. [69]

    Shenzhe Zhu, Jiao Sun, Yi Nian, Tobin South, Alex Pentland, and Jiaxin Pei. 2025. https://doi.org/10.18653/v1/2025.nllp-1.2 The automated but risky game: Modeling agent-to-agent negotiations and transactions in consumer markets . In Proceedings of the Natural Legal Language Processing Workshop 2025. Association for Computational Linguistics

    doi:10.18653/v1/2025.nllp-1.2 2025