CoChat

Adaptive Information Modulation: Designing Governance Mechanisms for Multi-Agent Artificial Intelligence Systems

Open access
Chen, Qiliang, Ilami, Sepehr, Lore, Nunzio, Heydari, Babak · 2026 · Journal of Mechanical Design

The DOI matches its Crossref record; 100 of the 102 references it cites were found in a registry or an index; 13 of the 23 sources its claims cite were checked for support.

Abstract

Abstract Modern engineered systems increasingly involve complex sociotechnical environments where multiple agents—including humans and the emerging paradigm of agentic AI powered by large language models (LLMs)—must navigate social dilemmas that pit individual interests against collective welfare. As engineered systems evolve toward multi-agent architectures with autonomous LLM-based agents, traditional governance approaches using static rules or fixed network structures fail to address the dynamic uncertainties inherent in real-world operations. This article presents a novel framework that integrates adaptive governance mechanisms directly into the design of sociotechnical systems through a unique separation of agent interaction networks from information flow networks. We introduce a system comprising strategic LLM-based system agents that engage in repeated interactions and a reinforcement learning (RL)-based governing agent that dynamically modulates information transparency. Unlike conventional approaches that require direct structural interventions or payoff modifications, our framework preserves agent autonomy while promoting cooperation through adaptive information governance. The governing agent learns to strategically adjust information disclosure at each time-step, determining what contextual or historical information each system agent can access. Experimental results demonstrate that this RL-based governance significantly enhances cooperation compared to static information-sharing baselines. This work establishes information transparency as a dynamic design parameter and demonstrates how governance considerations can be effectively embedded into complex engineering systems from the design phase. While validated on the repeated Prisoner’s dilemma, which represents a challenging governance problem in an abstract model, our framework offers a flexible strategy for fostering desired collective outcomes across diverse sociotechnical engineering applications, from human–robot collaboration to autonomous vehicle networks and future multi-agent AI systems.

Overview

CoChat Summary

TL;DR

The paper proposes adaptive information modulation as a governance mechanism for multi-agent AI systems, separating a fixed interaction network from an RL-controlled information network. In repeated Prisoner’s Dilemma experiments with LLaMa3-70b agents, an RL manager learned dynamic, heterogeneous information-disclosure policies that improved cooperation and social welfare over static baselines.

Key findings

  • The RL manager outperformed all static baselines (LA, LA+NR, LA+AR) on normalized social welfare and cooperation rate across 50 runs; RL cooperation reached 100% [C,C] by the 10th timestep, while LA and LA+NR increased more slowly and did not converge to 100%, and LA+AR drove the system toward defection.
  • Micro-level validation showed LLaMa3-70B followed a win-stay, lose-shift pattern: after mutual cooperation, cooperation was 87% when the coplayer was described as cooperating “Sometimes” and 99% when “Often”; after mutual defection with a coplayer described as cooperating “Often,” cooperation occurred 68% of the time.
  • The learned policy used LA more in early steps 2–6, with about 20% mixed interventions, then consistently selected neighborhood ratio (LA+NR) from step 7 onward. High-degree nodes (mean 5.68) were significantly more likely to receive LA+NR than low-degree nodes (mean 4.64), p < 0.001, Cohen’s d = 0.60; agents receiving LA+NR had higher prior cooperation (mean 0.908) than those receiving LA (mean 0.801), p < 0.001, Cohen’s d = 2.61.
  • Quantitative cooperation rates were unreliable: with opponent ratio fixed at 75%, an own ratio of 69.99% yielded 1% cooperation but 70% yielded 96%; typos in prompts also substantially shifted LLM behavior.

Methods

  • Framework: 20 LLM agents on a fixed Erdős–Rényi network (link probability 0.25) per round, 20 timesteps, repeated Prisoner’s Dilemma (CC = 3 each; CD/DC = 5/0; DD = 1 each); 50 independently generated network topologies; LLaMa3-70b via LangChain and Groq, temperature 0.8.
  • RL manager: Actor-Critic POMDP that observes aggregated states and selects one of three information types per agent per interaction—LA, LA+AR, or LA+NR—with one hidden layer of 256 neurons, actor learning rate 0.001, critic learning rate 0.005, and discount factor 0.99; trained offline to maximize discounted social welfare.
  • Analysis: micro-level validation with N typically greater than 100 prompt repetitions and qualitative cooperation categories (Rarely <33%, Sometimes 33–66%, Often >66%); RL compared against static baselines; t-tests for degree centrality and prior cooperation differences.

Limitations

  • Stated: limited evaluation sample (50 network instances, 20 timesteps) due to computational constraints; validation only on Prisoner’s Dilemma in random networks; LLM behavioral fidelity to humans requires controlled human-subject validation; coarse three discrete information strategies; assumes complete observability; field deployment remains future work.
  • Stated: future work should test other network structures and games, finer or adaptive information granularity, partial observability, and integration with other governance mechanisms.
  • Evident: micro-validation used only LLaMa3-70b and found sensitivity to typos and quantitative thresholds (e.g., 69.99% vs. 70%), raising robustness and generalizability concerns across LLMs and prompt variations.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1115/1.4070755

Cites

100 of 102 references verified0 citation errors · 2 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

8 Valid · 4 Not valid · 8 Not checked · 125 Own results and interpretations

79 more claims

Create a free account to see all 145 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 31
  • It may state two separate claims as one. 36
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

4 of 4

Citations

Cites

Research works printed but not linked: 2

  1. Sutton, 2018, Reinforcement Learning: An Introduction
  2. Konda, 1999, Actor-Critic Algorithms, Adv. Neural Inform. Process. Syst., 12, 1008

Cited by

Metadata

Authors
Chen, Q., Ilami, S., Lore, N., & Heydari, B.
Affiliations
  • Northeastern University (US)
  • Northeastern University Department of Mechanical and Industrial Engineering, Institute of Experiential AI, and Network Science Institute, , ,
Venue
Journal of Mechanical Design
Published
2026-03-09
Volume
148
Issue
4
ISSN
  • 0161-8458 (ISSN-L)
  • 1050-0472
  • 1528-9001
OpenAlex
W7118190101
Cited by
3
References
52
Citations by year
2026: 3
FWCI
10.39
Type
journal article
Language
en
Open access
Locations
Journal of Mechanical Design (publishedVersion, cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Multi-Agent Systems and Negotiation
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Game Theory and Applications
  • Social Sciences › Sociology and Political Science › Evolutionary Game Theory and Cooperation
Keywords
adaptive governance, multi-agent systems, agentic AI, large language models, information transparency, information disclosure, reinforcement learning, cooperation, prisoner's dilemma, socio-technical systems, agent autonomy, information flow networks, human-robot collaboration

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.