Collaborative Multi-Agent Reflection

• Collaborative multi-agent reflection is a framework where teams of LLM-driven agents apply theory of mind and structured critique to achieve cognitive synergy.
• It employs iterative cycles of contribution, evaluation, and integration using specialized roles and tools like Neo4j and Clingo.
• Empirical results show enhanced argument quality and risk resolution when agents combine anticipatory reasoning with rigorous critique.

Collaborative multi-agent reflection is a set of principled mechanisms enabling teams of artificial agents—typically powered by LLMs—to achieve adaptive, higher-order collective intelligence not through simple output aggregation, but by actively modeling peers' perspectives (theory of mind), systematically critiquing each other's reasoning, and refining their own outputs based on structured interaction protocols. This paradigm is motivated by cognitive science insights into human team synergy: successful groups do not merely pool facts or ideas, but recursively engage in critique, anticipation of others’ reasoning, and iterative revision, yielding cognitive synergy where the group output surpasses the sum of its parts (Kostka et al., 29 Jul 2025).

1. Foundations: Theory of Mind and Structured Critique

Collaborative multi-agent reflection is grounded in the integration of explicit Theory of Mind (ToM) modeling and systematic critical evaluation:

• Theory of Mind (ToM): Each agent is required to anticipate and model the beliefs, intentions, and likely contributions of its peers. Practically, ToM is embedded at the prompt level: before providing an argument or proposal, an agent "pre-commits" with a statement such as, “Anticipating the Realist will focus on budget constraints, I argue...”, then adapts its argument to maximize complementarity and reduce redundancy. This operationalizes agent-level mentalizing, leading to distributed anticipation and tighter, more non-redundant coordination.
• Critical Evaluation: A distinct Critic Agent is tasked with structured critique, detecting logical inconsistencies, unsupported claims, biases, and missed risks within peers’ outputs. The Critic’s role is strictly evaluative (not generative): it does not produce its own solution, but acts as a formalized peer reviewer, triggering revision cycles without advancing an alternative proposal.

These elements are coordinated by explicit workflow components—Orchestrator and Integrator agents, coupled to an external knowledge base (Neo4j) and logical reasoning backend (Clingo, Answer Set Programming)—that support rigorous integration and gap-filling (Kostka et al., 29 Jul 2025).

2. Structured Framework for Collaborative Reflection

The multi-agent reflection framework is formally realized as an orchestrated, iterative process:

• Specialized Expert Agents: Each embodies distinct reasoning roles (e.g., Data and Logic Specialist, Visionary Strategist, Implementation Realist).
• Critic Agent: Engages in round-based, post hoc critique.
• Integrator: Aggregates agent outputs using a graph knowledge base (Neo4j) and formalizes referential/logic gaps using Clingo.
• Orchestrator: Controls process flow, triggers further refinement based on Critic feedback and integration status.
• Prompts: ToM is enforced via prompt scaffolding, requiring agents to predict peer contributions and adapt their own output accordingly.

Mathematical formalization:

Let the system state at each cycle be

$$\mathcal{C} = \{\text{Agents}, \text{Critic}, \text{Orchestrator}, \text{Integrator}, \mathcal{K}\}$$

where $\mathcal{K}$ is the knowledge base. Each agent receives scenario and history, computes peer anticipations ($T_{pred}^i$), and produces a current response ($T_{out}^i$). The Critic aggregates these responses and generates a critique ($C_t$); the Orchestrator then determines whether another round is needed or if integration should proceed.

For ToM-enabled agent output: $$\mathsf{Response}_i = f_i(\text{Scenario}, \mathit{History}, \mathsf{Anticipation}_{j \neq i})$$ with critique generation: $\mathsf{Critique} = g(\{\mathsf{Response}_i\})$ The process is an iterative state machine: cycles of contribution, critique, and possible revision, terminating on satisfaction of coverage/consistency criteria.

3. Mechanism Interplay and Emergent Cognitive Synergy

The effectiveness of collaborative reflection rests on the synergistic interplay between ToM and Critique:

• ToM-driven anticipation enables agents to minimize redundancy, align complementary perspectives, and foreground distinctions between contributions; model runs with ToM alone show improved referential cohesion.
• Critique deepens agent reasoning, surfaces hidden risks or unsupported conclusions, and triggers needed self-corrections and position revisions. Critique alone enhances depth but does not resolve coordination misalignments.
• Emergence of Synergy: When both mechanisms are enabled (TT), agents preemptively adapt their output in anticipation of critique, while critique prompts further adaptive iteration—mirroring the recursive, mutually corrective dynamics of effective human teams. Empirically, this leads to significant gains in argument quality, critical engagement, risk resolution, and adaptivity (e.g., TT condition achieves risk resolution of 48.75% and 1.75 substantive position revisions per conversation vs. poor integration and revision in baseline).

The mechanism interplay is not simply additive: ToM and Critique together yield cognitive gains unattainable by either in isolation (Kostka et al., 29 Jul 2025). This emergent property is analogous to human cognitive synergy and is a defining feature of collaborative reflection.

4. Empirical Evaluation and Metrics

Empirical validation centers on complex decision-making scenarios—e.g., agents tasked with allocating R&D budgets across candidate technologies, balancing feasibility, market, and operational constraints.

• Configurations evaluated: Four settings—Base (FF: no ToM, no Critic), ToM only (TF), Critic only (FT), and ToM+Critic (TT).
• Core metrics: Argument quality (1–5 scale), critical engagement (substantive turn proportion), reference cohesion (1–3), risk resolution (fraction of risks addressed), revision triggers (number of substantive position revisions per conversation).
• Key results: TT consistently outperforms all other configurations across metrics (argument quality, dynamic revision, risk coverage), substantiating that only the joint operation of both mechanisms produces robust, adaptive collective reasoning.

Table: Empirical Performance Summary (selected metrics, see (Kostka et al., 29 Jul 2025) for complete data):

Setting	Arg. Quality	Crit. Engagement	Risk Resolution (%)	Revision Triggers
TT	Highest	Highest	48.75	1.75
TF	High	Moderate	Lower	Fewer
FT	Moderate	Moderate	Moderate	Some
FF	Low	Low	Lowest	Minimal

Agents in the TT regime not only generate higher-quality arguments but revise their stances more often, indicating a dynamic, iterative collaborative reflection cycle.

5. Comparison to Prior Multi-Agent Systems and Future Directions

The proposed framework establishes a new baseline for cognitive synergy in LLM-based multi-agent systems:

• Distinctiveness: Unlike “MoA”/“SMoA” and static role-based systems, this approach achieves dynamic, perspective-aware reasoning, systematic critical feedback, adaptive self-correction, and greater resilience to overlooked risks and logical flaws.
• Formal integration of ToM and Critique: Prior frameworks lack recursive, explicit modeling of peer reasoning and a formalized, iterative critique-refinement loop as central workflow constructs.
• Generality and limitations: Although shown effective in strategic decision-making scenarios—with support from tools like Clingo and Neo4j for formal reasoning and memory—the approach requires further validation in larger, more complex teams and a broader set of domains.

A plausible implication is that scalable, adaptive cognitive synergy in MAS will depend on designing architectures where anticipation of criticism and explicit role reasoning are deeply entangled in the reflection cycle.

6. Theoretical and Algorithmic Formalization

The core theoretical structure is an iterative collaborative state machine, with explicit representation of agent states, role-specific anticipation, and critical evaluation flows:

• System state:

$$\mathcal{C} = \{\text{Agents}, \text{Critic}, \text{Orchestrator}, \text{Integrator}, \mathcal{K}\}$$

• Agent output:

$$\mathsf{Response}_i = f_i(\text{Scenario}, \mathit{History}, \mathsf{Anticipation}_{j \neq i})$$

• Critic output:

$\mathsf{Critique} = g(\{\mathsf{Response}_i\})$

• Iteration condition: Cycles proceed until all referential/logical gaps are closed, and the Orchestrator deems consensus/integration satisfactory.

This formalization does not rely on direct mathematical optimization of ToM/critique, but rather on explicit prompt structuring and workflow orchestration; however, it is compatible with further mathematical or learning-driven instantiation.

7. Broader Implications and Research Trajectory

Collaborative multi-agent reflection, as defined and instantiated in this framework, represents a paradigm shift toward cognitively inspired MAS. Through the explicit combination of Theory of Mind and structured critical evaluation, orchestrated in an iterative, integrative process, these systems approach the essential features of human team reasoning: recursive anticipation, mutual critique, and dynamic self-correction. Empirical evidence demonstrates substantial gains over non-reflective or uni-mechanism alternatives in domains demanding complex, adaptive reasoning.

This line of work suggests a roadmap for future multi-agent research: enabling ever-more scalable, transparent, and cognitively plausible group reasoning machinery—critical for real-world, high-stakes decision making, strategic planning, and AI alignment tasks where collective intelligence and self-correction are paramount (Kostka et al., 29 Jul 2025).