acceptodds
Withdrawn from ICLR 2027

Mutual Recursive Improvement: Toward Co-evolving Multi-Agent Societies in Games

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

How can LLM agent societies adapt their cooperation as conditions change, and how are the resulting benefits distributed? Strong individual reasoning does not ensure effective coordination among agents with distinct interests, particularly when changing environments make existing strategies and commitments inappropriate. We introduce **Mutual Recursive Improvement (MRI)**, a framework connecting individual learning with institutional adaptation through an interaction–reflection–revision cycle. Human-inspired interventions help agents interpret outcomes, reassess experience, and revise private rules and shared commitments. These revisions shape subsequent interactions, whose outcomes inform further adjustment, while model parameters remain frozen and agents receive no explicit instruction to maximize collective profit. We evaluate eight LLMs over 50 simulated quarters in market-conditioned Dictator, Trust, and Public Goods games, using disclosed historical SPDR S&P 500 ETF Trust (SPY) returns. The final configuration improves allocation quality and collective surplus over baseline in 23 of 24 model–game combinations each, with both improving in 22. Mean allocation quality increases by , and mean regime-aligned cooperation rises from to . These findings demonstrate the potential of structured interventions to support adaptive cooperation through revisable experience and commitments. The code will be open sourced upon acceptance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Multi-Agent Systems, Recursive Improvement, Game Theory

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

This paper introduces Mutual Recursive Improvement (MRI), a framework for adapting cooperation among frozen LLM agents through an interaction-reflection-revision cycle. The framework exposes agents to five human-inspired interventions: Counterparty Review (CR), Self-Reflection (SR), Public Contracting (PC), Case-Based Reasoning (CBR), and Experience Discounting (ED). The authors evaluate cumulative configurations of these interventions across eight LLMs, three market-conditioned economic games, and 50 rounds derived from quarterly SPY returns. The evaluation considers allocation quality, collective surplus, regime-aligned cooperation, and payoff dispersion.

Critical strengths
  • The paper studies a meaningful problem: how LLM agents with different interests adjust their cooperation as external conditions change. This matters for multi-agent deployment because strong individual agents do not necessarily coordinate well as a group.

  • The paper is clearly written. The framework, intervention levels, metrics, and experiments are easy to follow, and the authors report negative results such as the degradation caused by routine self-reflection rather than presenting every intervention as beneficial.

Critical weaknesses
  • MRI combines familiar ideas such as reflection, memory, case support, communication, and contracting, but the experiments do not isolate what the full recursive loop adds beyond these components. This concern is important because CBR is the highest-scoring standalone intervention in 22 of 24 model-game rows, while standalone and composed CBR are applied at different points in the decision process. Without matched ablations and budget-controlled baselines, the claimed method-level novelty remains unclear.

  • The empirical scope is limited. Each configuration is reported on one fixed 50-quarter market sequence without uncertainty across repeated executions, and the heterogeneous evaluation is confined to one Trust-game setting. It is therefore unclear whether the gains are stable across runs and market paths, or whether they support the broader claim of co-evolving agent societies.

Questions for authors
  • Do the authors have results that separate the contribution of the full recursive loop from CBR, such as L5 without CBR, L4 with ED while retaining SR, or L4 without SR? If not, what existing evidence supports attributing the gains to the recursive loop rather than the fixed case library?

  • Were the main B0/L1/L4/L5 comparisons repeated under the same settings? If so, please report run-level variation or uncertainty. If not, how sensitive might the reported 23/24 and 22/24 counts be to repeated execution or a different market path?

  • How do model-call count, context length, and token usage differ across B0-L5? Have the authors tested a budget-matched deliberation baseline or a simple regime-rule prompt baseline that could rule out additional prompting budget as the main explanation?

For revision, the paper would also benefit from reporting the construction criteria and full contents of the CBR case library, distinguishing standalone CBR from composed CBR whenever their results are discussed, using “allocation quality,” “collective surplus,” and “mutual benefit” consistently, and defining B0-L5 once in a compact table with consistent notation across the main text, captions, and appendix. These are presentation and reproducibility suggestions rather than separate reasons for rejection.

Reviewer_mvLrOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Mutual Recursive Improvement (MRI), a coordination framework for LLM cooperation that uses reflection, review, and experience reassessment to support the revision of cooperation. MRI is a set of prompting-style interventions, i.e. Counterparty Review (CR), Self-Reflection (SR), Public Contracting (PC), and these intervention can be composed and replaced into different configurations, i.e. L1 to L5. Configurations was evaluated on a set of market-conditioned cooperation games, on eight LLMs over 50 rounds. Evaluation outcomes include allocation quality (RBOAS, denoted by OTO_T), collective surplus (RBCS, denoted by STS_T), adaptive cooperation (RACS, denoted by KTK_T), and payoff dispersion (CV, denoted by vTv_T).

Critical strengths
  • The paper develops its evaluation environment from classic experimental-economics games (Camerer, 2004; Ledyard, 1995), and conditions their payoffs on realized SPY returns.
  • The set of evaluation outcome cover allocation quality, collective surplus, adaptive cooperation and payoff dispersion, reflecting economic notions.
Critical weaknesses
  • Limited novelty. There is already a series of works on LLM agents cooperating in sequential social dilemma games, several of which the paper itself cites in Section. 2 and Table 1. The individual interventions, although inspired from real world case, also exist in the literature, such as Reflexion and ExpeL. The contribution is therefore largely the game setting and the composition of known components, as well as a series of observations.
  • Observation without hypothesis testing. The paper reports adding each intervention separately to the baseline reveals substantial differences in effectiveness in Section 4.3, but offers no hypothesis for why and no targeted test of one. This is addressable with hypothesis-driven ablations or analysis of agent reasoning and retained rules.
  • nit: The paper uses multiple terms such as allocation quality, RBOAS, and OTO_T interchangeably, making it difficult for readers to follow throughout the paper. This is easy to fix by fixing one name and using them consistently.
Questions for authors
  • The paper states that agents receive no explicit instruction to maximize collective profit. What motivates the agents to interact and reflect on their actions? What is the objective told to each agent? Without an objective, there is no “recursive improvement” to begin with.
Reviewer_xGnJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a new framework, MRI, for helping multi-agent groups coordinate effectively. They experiment on 5 different interventions for helping agents collaborate (L1-5), all of which help agents implement different interaction, reflection, revision procedures. They evaluate these procedures using 8 different LLMs in 3 different games (Dictator, Trust, Public Goods). They use real-world data from the SPY ETF as part of their evaluations to simulate realism.

Critical strengths
  • The MRI is seem easy to apply to any multi-agent system.
  • The paper is mostly easy to read.
  • The problem setting and approach are interesting.
Critical weaknesses
  • minor issue: this sentence starting on line 22 is awkward: "The final configuration improves allocation quality and collective surplus over baseline in 23 of 24 model–game combinations each, with both improving in 22."
  • L from equation 1 is never defined. I'm assuming this is the model, but it would be good to explicitly define it.
  • H from equation 1 is not well-defined. It is defined as the "accessible history" but this is underspecified. E.g., what is contained within the history? Just observations? actions? both? etc.
  • the definition of Z^{dec}{i,t} from equation 1 and Z{i,t} from equestion 2 are also unclear to me. Z^{dec}{i,t} is defined as the "intervention context available at decision time" but it is unclear to me what an "intervention context" is. Z{i,t} is defined as "retained resources" "contribut[ing] to Z^{dec}{i,t} together with newly available information and any analysis performed before the next decision." From this definition I cannot clearly understand what Z{i,t} looks like (e.g., is it a vector, string, etc.) and it is also unclear to me how exactly Z^{dec}{i,t} and Z{i,t} interact. E.g., is Z^{dec}{i,t} simply a list of Z{i,t} values? This lack of clarity is an issue when defining a formal model.
  • The experiments are based on real world data from the SPY ETF (the S&P 500) between 2014 and 2026. This is concerning to me as between 2014 and 2026, the US market experienced a historically bullish market. This makes me question the validity of the results in different market conditions. Furthermore, the LLM models used where likely trained on some of this data and certainly trained on data talking about the 2014-2026 S&P 500 behaviour (e.g., professional analysts talking about how to behave in this market). This is a confounding factor as this training data could be influencing the models. This confounding factor further makes me question the validity of the results on new data and differing market conditions.
  • The baseline used in Section 4 seems far too weak. This is one of my most major concerns. As I understand it the baseline, B0, is simply seeing what models do when presented with the game observations and interaction history. Additionally, as with the other agents, I understand that B0 "agents receive no instruction to maximize system profit W_t." Hence, I understand B0 as simply seeing how well agents performed when simply dropped into the game with minimal direction. While this is an interesting baseline, I find it far too weak to be the only baseline to compare against. For example, given the existing experiments, it is unclear to me whether MRI can even beat the weak baseline of an agent who is prompted with "do your best to collaborate and increase collective profitability." Furthermore, Table 1 shows many prior works which were evaluated in similar settings. I would expect to see a comparison to those works as well to understand how MRI improves on prior work.
Questions for authors
  • Why are the 3 chosen environments good test environments? Why these 3 environments and not any other environments? I read lines 277-279, but I found the argument unsubstantiated.
  • How do we know that 50 rounds is sufficiently long for the experiments?
  • I don't understand why L1-5 constitute a progression. To me these simply seem like different methodologies.

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

Related papers

Discussion (0)

Sign in to comment.