acceptodds
Under review as a conference paper at ICLR 2027

Finding Kaiji Among LLM Agents: When Everyone Yields

Abstract

When LLM agents compete for scarce resources, do they turn on each other, and if their societies fail, why? This study asks where betrayal should pay: restricted rock–paper–scissors, the elimination game of the manga Kaiji, with a closed deck. Twenty agents message each other freely; the rules never mention cooperation, and every readout comes from the engine log with no LLM judge. Agents negotiate tie pacts before 93.5% of matches and break only 10.2%, but the breaks are aimed: 70.0% play the one card that beats the agreed card (chance is 50%). Their limit is the division of roles: who proposes and who accepts. Removing memory removes proposers, and coordination falls to chance. Removing partner choice leaves agreements at 97.7% yet cuts coordination by two thirds: faced with conflicting proposals, both yield at once in 53% of cases and collide on the swapped cards, as two people do in a corridor. Under scarcity, eliminations rise and betrayals are aimed no better. Reasoning changes what the pact is for: with thinking enabled, the same model keeps the pact but breaks it three times as often, mostly with the beating card, and a single reasoning agent profits at its non-reasoning neighbours' expense. The convention is not universal. Seven open-weight models run without reasoning almost never form it: they play to win, fewer survive, and several raise the stakes; the one re-run with reasoning forms it in most matches.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multi-agent LLM, LLM agents, coordination, emergent conventions, evaluation methodology, game-theoretic environment, deception, reasoning models, open-weight models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vH6AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces a Kaiji-inspired card game to study negotiation, cooperation, and betrayal among LLM agents across different models and conditions. For the main model, proposals to tie are 93.5% of matches, yet agents break agreements 10.2% of the time, 70% of which are to win at the other models' expense through deception. The paper also does various extension analyses, such as removing memory, testing the game in which opponents are assigned, which creates a new failure mode (53% of the time) in which both agents sometimes accommodate the other’s conflicting proposal, preventing the intended tie. Scarcity worsens survival without a clear increase in betrayal rates, and reasoning makes the main model betray more often and can help it profit at non-reasoning agents’ expense, but its effects vary across models. Overall, the paper contributions a new environment and an initial investigation of how model choice, memory, and interaction protocols shape cooperation and betrayal.

Critical strengths

The main strength of the paper is that it tackles a very topical and important subject: the collaboration, deception, and interaction dynamics across LLM agents of different types in multi-agent settings. As LLM agents become more proliferated in the real world, this becomes a very important topic. The paper contributes a creative new environment which tests this and does extensive and thorough analyses across various ablations and experiments. The thoroughness of the experiments, in terms of the various ablations and extension analyses, is the main strength of the paper.

Critical weaknesses

The main weakness of the paper is that the evaluation is limited to one game, which makes it unclear which findings generalize beyond Kaiji’s particular rules and incentives vs. LLM agents in general. Also several findings deserve deeper investgation:

  • why do some agents choose the losing alternative when breaking a pact?
  • why does adaptation under scarcity fails to improve outcomes?
  • why agents strongly favor ties despite the financial rewards for additional stars?

In particular, it remains unclear whether agents explicitly calculate interest and expected returns or respond only loosely to these incentives. Qualitative analysis of these decisions, targeted incentive experiments, and replication in another environment would strengthen the contribution. Some passages also feel AI-generated and are difficult to parse, especially the definitions and financial mechanics.

Questions for authors
  • Do the same findings hold in other games, or do they depend on Kaiji’s set-up specifically?
  • Why do some agents break a pact by playing the losing card? Are they making mistakes, forgetting the agreement, or intentionally helping their opponent?
  • Why does scarcity increase eliminations without making betrayal more strategically effective?
  • Why do agents favor ties despite the ¥4 million reward for each extra star? How do they weigh safe escape against potential profit?
  • Do agents explicitly calculate compound interest and expected returns, or only consider money qualitatively? How does this differ between reasoning and non-reasoning models?
Reviewer_DFncOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The authors study LLM behavior on a game from the manga Kaiji called "restricted rock paper scissors", specifically motivated by studying deception in games.

Critical strengths

The topic of deception is important. Games can provide a way to study this.

Critical weaknesses

I'm sorry to say this but the writing of the paper is really poor, and for me immediately sufficient reason to reject the paper. Ideally there should have been a desk rejection process that rejected this paper. (I suspect that this is due to heavy LLM use; in any case it's poor wherever it came from.) I don't mean that the English is poor, it is fine (though it sounds like an LLM); I'd much rather read a paper in broken human non-native English that is well structured.

First of all, obviously you cannot assume that people know this game. If you don't know the game, then for the first few pages (and the abstract) as a reader you're just lost, you have no idea what this is all about. There is absolutely no point in reporting statistics about "colliding on swapped cards" and so on in the abstract; I imagine 99% of readers have no idea what this means at this point. The AI-generated images in figure 1 are perhaps cute but clearly this is not the best way to help the reader understand the game. It takes way too long to get to a description of the game, nor is the motivation for why this game, among all possible games, is a great game to study.

The discussion does not feel like a scientific conclusion to a paper and rather an LLM-generated list of policy recommendation type of things.

Overall, I get the sense that the authors just want to have a paper, as opposed to that they actually want to teach the reader something.

Questions for authors

The paper is not ready to ask questions. It needs to be rewritten from scratch to focus on creating a paper that a human being can read from beginning to end and understand what is going on.

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

This paper implements a restricted Rock–Paper–Scissors society as a testbed for evaluating LLM agents' emergent cooperation. The authors find that the cooperation depends partly on the harness design, and thus it is unfair to compare cooperation rates across different harness protocols.

Critical strengths

S1. The writing is good and the paper is well organized.

S2. The proposed testbed resolves three limitations in exisiting LLM-emergent-cooperation benchmarks.

S3. The authors identify an interesting coordination failure and convincingly argue that it should be accounted for when evaluating cooperation rates among LLM agents.

Critical weaknesses

W1. The experiments in Sections 3 and 4 only consider the DeepSeek-V4.1-Flash model, which makes the generalizability of the findings questionable.

W2. Although an interesting coordination failure is identified, it is unclear whether and how it changes the main findings in prior work.

W3. The range of the evaluated models is limited, and the selected models are outdated.

Questions for authors

Q1. I wonder why the experiments mainly focus on open-source LLMs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.