Do LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation

Source

Source name: arXiv
Source kind: paper
Authors / organization: Akash Kundu, Emanuel Tewolde, Ratip Emin Berker, Samuel F. Brown, Vincent Conitzer
Venue: arXiv: cs.GT
Published: 2026-08-12
Collected: 2026-08-13
URL: http://arxiv.org/abs/2608.12125v1
PDF: https://arxiv.org/pdf/2608.12125v1
DOI: UNKNOWN
Registry group: academic_apis_and_preprints
Evidence role: canonical_academic_index
Language: en
Country / region: UNKNOWN

Plain-Language Brief

Researchers introduce a framework for evaluating how Language Model (LLM) agents cooperate when provided with similarity signals. The study finds that different LLM models exhibit varying behaviors, but some modern ones are consistent across cooperation problems and payoff structures. Surprisingly, the dataset used to compute the similarity signal has minimal impact on induced cooperation. Additionally, LLM models tend to self-identify as highly similar when evaluating another model's reasoning process. The researchers develop a new model that captures some of the LLM reasoning and can support cooperative outcomes in certain scenarios.

What the Document Says

The paper presents a framework for evaluating decision-making in Language Model (LLM) agents when they are provided with graded similarity signals.
Different LLM models show drastic variations in navigating similarity signals.
Some modern LLM models exhibit consistent behavior across cooperation problems, payoff structures, and prompt framing.
The impact of the dataset used to compute the similarity signal on induced cooperation is small to none.
LLM models tend to identify themselves as highly similar when asked to evaluate another model's chain-of-thought reasoning.
An LLM-behavioral-game-theoretic model is developed that captures some of their reasoning rationale and supports cooperative outcomes in equilibrium under high similarity scores.

Key Concepts Mentioned

Language Model (LLM) agents
Cooperation problems
Similarity signals
Graded similarity
Decision making patterns
Monocultural AI ecosystems
Prisoner's Dilemma
Chain-of-thought reasoning

Claims Made by the Document

LLM decision making can be evaluated when agents are provided with graded similarity signals.
Different LLM models vary drastically in how they navigate similarity signals.
Some modern LLM models exhibit consistent behavior across cooperation problems, payoff structures, and prompt framing.
The dataset used to compute the similarity signal has small to no impact on induced cooperation.
LLM models systematically self-identify as highly similar when asked to evaluate another model's chain-of-thought reasoning.
An LLM-behavioral-game-theoretic model can capture some of their reasoning rationale and support cooperative outcomes in equilibrium under high similarity scores.

Evidence Used

The paper discusses the introduction of a framework for evaluating LLM decision making.
The text mentions that different LLM models exhibit varying behaviors.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

As LLM-based agents with user-instructed goals are becoming widely deployed, they increasingly encounter each other in strategic interactions, and face challenges of finding mutually beneficial outcomes. Prior literature has argued that cooperation problems such as the Prisoner's Dilemma are resolvable in settings where agents know they follow very similar decision making patterns, as for example in monocultural AI ecosystems. Following that line of work, this paper introduces the first framework for evaluating LLM decision making when agents are provided with graded similarity signals. Among our findings, we establish that different LLM models vary drastically in how they navigate similarity signals, with some modern models showing consistent behavior across cooperation problems, payoff structures, and prompt framing. Perhaps surprisingly, our experiments also show that the dataset based on which the similarity signal is computed has small to no impact on induced cooperation, and that LLM models systematically self-identify as highly similar when asked to evaluate another model's chain-of-thought reasoning by themselves. Finally, we develop an LLM-behavioral-game-theoretic model that captures some of their reasoning rationale, and show that it can support cooperative outcomes in equilibrium under sufficiently high similarity scores.