How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis

1. Introduction

• Negotiation can assess contextual understanding, theory of mind, rational decision-making, and strategizing through dynamic interactions with well-defined outcomes.Unlike benchmarks based on fixed questions, negotiation settings can change when a partner counters with a proposal or a different argument.Source: §1 · pp. 1–2
• NegotiationArena is an open-source framework for evaluating and probing LLM agents’ negotiation abilities and for studying how they communicate, reason, and negotiate.The paper evaluates GPT-4, GPT-3.5, Claude 2.1, and Claude 2.Source: §1 · p. 2
• GPT-4 is overall the best negotiator, while tactics such as feigning desperation or acting aggressively can improve one agent’s win rate over another.The paper also identifies shortcomings and vulnerabilities, including anchoring bias.Source: §1 · p. 2

2. Scenarios in NEGOTIATIONARENA

• The platform’s general two-agent setup gives each participant resources and a goal, and scenarios can become more complex through added resource types or goals.The paper uses this structure to implement multiple negotiation games.Source: §2 · p. 2

2.1. NEGOTIATIONARENA Scenarios

• In Resource Exchange, agents with resources and goals make proposals over multiple turns until one accepts or the turn limit is reached.The goals can prompt different strategies, such as diversifying resources or maximizing one resource.Source: §2.1 · p. 2
• The multi-turn Ultimatum Game extends resource-split bargaining with counteroffers and lets either player accept; rejection means both lose all resources.In the classical version, the first player proposes a split and the second accepts or rejects it in a single round.Source: §2.1 · p. 2
• In the Seller and Buyer scenario, each agent has private information about its own production cost or willingness to pay, and the seller starts the negotiation.The agents are prompted to sell and buy the object, respectively.Source: §2.1 · p. 3

2.2. NEGOTIATIONARENA Implementation

• The implementation uses structured messages to track offers and game state, while targeted prompting lets users probe strategies, emotions, and behaviors such as faked desperation.Agents’ reasoning can remain hidden from their opponent while the system extracts and stores structured information from messages.Source: §2.2 · p. 3

3.1. Negotiation Results

• When Claude-2.1 goes first and GPT-4 second, GPT-4 wins 76%; with the order reversed, Claude-2.1 wins 72%.As Player 2, Claude-2.1 averages 2.45 payoff versus GPT-4’s 1.38, despite GPT-4 winning more often.Source: §3.1 · pp. 3–4
• Claude-2.1 averages more than 60 in payoff as Player 1 against every opponent in the ultimatum game.GPT-3.5 is the exception to Player 1 almost always winning.Source: §3.1 · p. 4
• At a buyer willingness to pay of 60, GPT-4 averages a $41 sale price as buyer; most model pairs settle below 50.All models achieve similar prices as sellers.Source: Figure 4 · p. 4, §3.1 · p. 4

3.2. Insights From the Experiments

• Turn and role matter across scenarios: Player 1 often loses in resource exchange but is much more likely to win in the ultimatum game.The authors relate these patterns to turn importance and first-offer anchoring also observed in human experiments.Source: §3.2 · p. 4
• GPT-3.5 misreads the seller’s goal in 20% of its Sell/Buy games, and all models struggle with game rules such as when acceptance ends play.The paper says agents are asked to verbalize their state to mitigate some of these hurdles.Source: §3.2 · p. 5
• An illegal GPT-3.5 offer can distract a stronger opponent: Claude-2.1 accepts an impossible exchange, and GPT-4 changes a fair offer for the worse after recognizing one.The authors suggest erroneous behavior by less effective agents could be leveraged as an adversarial attack; Davidson et al. (2024) found similar patterns in their negotiation framework.Source: §3.2 · p. 5

4. Strategic Social Behavior in Games

• The social-behavior comparison tested GPT-4 using cunning or desperate prompts against an unprompted GPT-4 opponent.Player 1 was fixed as unprompted, and 80 games were run for each pair in each game.Source: §4 · pp. 5–6

Cunning and Desperate behaviors increase Win Rate

• Social-behavior prompts increased average payoff and win rate across the tested games, with the most striking win-rate gains in the ultimatum game.The comparison used an unprompted Player 1 and a cunning, desperate, or default Player 2 across 80 games per pair in each game.Source: §4 · pp. 5–6, Cunning and Desperate behaviors… · p. 6
• Cunning Player 2 won 82% of ultimatum games but averaged about 49 payoff, similar to the default, because risky tactics sometimes produced no agreement and zero payoff.The cunning strategy produced a highly bimodal payoff distribution.Source: Cunning and Desperate behaviors… · p. 6

5. Evidence of Irrationality

• The authors use GPT-4 experiments to examine deviations from rational behavior, which they frame as relevant to model vulnerabilities and biases and to comparison with human irrationalities.The seller-buyer analysis focuses on proposals; seller costs and buyer willingness-to-pay were drawn from specified ranges, with 100 games per experiment.Source: §5 · p. 6, §5.1 · p. 6

5.1. Seller and Buyer Game

• The final accepted price correlated strongly with the seller’s initial proposal (Spearman ρ = 0.716), consistent with an anchoring effect.The final price tended to rise with the initial proposal.Source: §5.1 · p. 6
• Buyer and seller proposals showed a strong positive correlation with the average of their two most recent offers, consistent with a split-the-difference strategy.The authors caution that this strategy is not rational in every scenario.Source: §5.1 · p. 6
• Over-valued buyers counter-proposed above the seller’s initial offer more often than default buyers (0.41 v. 0.11), a significant difference (p < 0.05).Prompting buyers to be self-interested did not significantly change that probability; the authors hypothesize that GPT-4 over-applies splitting the difference.Source: §5.1 · pp. 6–7
• When game resources and beliefs were scaled up together, the fraction of the buyer’s $100X budget spent on the object decreased.The authors infer that the buyer’s advantage increases with the unit of currency.Source: §5.1 · p. 7

Single-turn and 3-turn ultimatum result in different

• In the three-turn ultimatum game, Player 1’s acceptance decreased as the proposal became less fair, unlike near-universal acceptance of positive offers in the classical game.The authors interpret this as incomplete generalization of the rational strategy; they note that the classical-game behavior could reflect familiarity with the game.Source: Figure 9 · p. 7, Single-turn and 3-turn ultimatum… · pp. 7–8
• As the amount to split increased, Player 1 obtained a larger fraction, reaching almost 79% when splitting $10,000,000,000.The authors report that the decider may accept large sums immediately even when the split is unfair, indicating sensitivity to absolute amounts.Source: Figure 10 · p. 8, Single-turn and 3-turn ultimatum… · p. 8

6. Related Work

• Prior research used games to study language behavior and emergence, while negotiation has also been studied across game theory, mathematics, economics, and psychology.The authors use well-studied games with known rational strategies as a reference for probing LLM rational behavior.Source: §6 · p. 8
• Earlier LLM negotiation studies examined repeated ultimatum play, human simulation, self-play or in-context learning, price negotiation with humans, and imperfect-information games.The authors contrast their work with those studies as examining LLM behavior without additional learning and with LLM agents negotiating against each other.Source: §6 · p. 8
• NegotiationArena is presented as an open-source platform for diverse LLM-to-LLM negotiations, complementing Davidson et al. (2024) with a different analysis of social patterns and irrational behavior.The authors describe technical differences in communication and messaging and conceptual differences in how payoffs and goals are assigned.Source: §6 · p. 8

7. Discussion

• NegotiationArena is a flexible, open-source platform for studying negotiation behavior between LLMs.Source: §7 · p. 8
• GPT-4 tends to be the strongest negotiator, while social tactics such as feigned desperation or insults can improve an agent’s payout.Source: §7 · p. 8
• The authors report anchoring and numerosity biases among LLM agents and argue that understanding these behaviors and vulnerabilities matters for agent reliability.Source: §7 · p. 8

• Studying how LLM agents interact through negotiations can provide insights that may help make them more reliable and useful to human users.Source: Impact Statement · p. 9

• Both GPT and Claude use temperature 0.7 and a maximum response length of 400 tokens.Source: A. Additional Experimental Details · p. 11

B. Limitations

• Prompts can bias or unexpectedly condition gameplay; the authors adapted prompts after Claude struggled with the first benchmark version while GPT-4 did not.Source: B. Limitations · p. 11
• Llama2-70B was excluded for instruction-following problems, while GPT-3.5 often makes numerical mistakes and does not always act as intended.Source: B. Limitations · p. 11
• Agents sometimes propose exchanging identical resources, such as Xs for Xs, indicating difficulty recognizing that the items match.Source: B. Limitations · p. 11
• The authors report weak mathematical operation skills and wrong-tag errors in Claude-2, Claude-2.1, and GPT-3.5; they suspect familiarity with numbers seen during training may matter.Source: B. Limitations · pp. 11–12
• Although rare, models may infer from the rules or prompts that they are playing an ultimatum-like game, raising the possibility that human play expectations influence their choices.Source: B. Limitations · p. 12

C. Interesting Replies

• Examples of reasoning and social behavior messages mark good reasoning or behavior in green and errors in red.Source: C. Interesting Replies · p. 12

C.1. Reasoning in Standard Game

• GPT-4 weighs how much information to disclose to build trust without weakening its negotiating position, while GPT-4-Turbo rounds a non-integer offer down to improve its counteroffer.Source: §C.1 · p. 12
• GPT-4 miscalculates that a proposed trade would raise its total resources from 30 to 29, illustrating a reasoning error in the example.Source: §C.1 · p. 12

C.2. Social Behavior

• GPT-4-Turbo-Cunning uses hostility after a 50-50 proposal and eventually gets $65 from the other player.Source: §C.2 · p. 12

D.1. Game

• A new game ending when an agent accepts can be implemented with a new rules prompt; more extensive language or parser changes require subclassing, with parsing isolated from the rest of the system.Source: §D.1 · p. 12

D.2. Game Unfolding

• The platform keeps agent conversation roles in a fixed system-user-assistant sequence by feeding one agent’s generated message to the other as a user message.Source: §D.2 · p. 12, Figure 11 · p. 13

D.3. Messaging, Logging and State Saving

• Agents are prompted to state resources and goals each turn to reduce hallucinations, while the parser structures and filters what the other agent sees and saves game state for later reloading or editing.Source: §D.3 · pp. 12–13

D.4. Game Objects

• Resource and Trade objects represent holdings and exchanges, support trade-consistency and utility checks, and allow agents to be reused across settings with limited state.Source: §D.4 · p. 13

D.5. Code Example; D.6. Available Agents

• Extending the available resource set requires minimal code edits, and the library currently includes GPT-4 and Claude-2 agents, with Turbo and 2.1 variants.Source: §D.5 · p. 13, §D.6 · p. 13

E.3. More Results

• With seller value 60 and buyer willingness to pay 40, buyers must spend over 40 to buy the object; GPT-4 is the hardest seller for all buyers.Figure 16 reports the difference between buyer willingness to pay and sale price; negative values indicate sales above 40.Source: §E.3 · p. 14, Figure 16 · p. 16
• Figure 17 relates each counter-proposed price to the average of the previous two proposed prices.Source: Figure 17 · p. 16
• Figure 18 tracks dollars left to buyers as all resources are scaled.Source: Figure 18 · p. 16