Teacher–Critic–Student Socratic Loop

• The Teacher–Critic–Student Socratic Loop is a modular, dialogic framework that uses distinct AI agents to generate open-ended questions, critique responses, and refine answers iteratively.
• It decomposes complex reasoning tasks through iterative substeps with specialized roles that enhance prompt optimization, lead to higher accuracy, and provide transparent, interpretable workflows.
• The framework integrates constructivist and Vygotskian principles to scaffold critical thinking in both AI systems and human learners, demonstrating measurable improvements in educational outcomes and research reliability.

The Teacher–Critic–Student Socratic Loop is a modular, dialogic framework underpinning recent advances in AI-driven prompt optimization, educational reinforcement learning, and critical thinking scaffolding. In this paradigm, the iterative interaction among specialized agents—namely, a Teacher, a Critic, and a Student—realizes Socratic dialogue patterns originally conceptualized for human pedagogy, now instantiated in multi-agent AI architectures and human–AI collaborative systems. This loop is used to drive prompt optimization for LLMs (Zhang et al., 21 Mar 2025), enhance critical, independent reasoning in education (Degen et al., 7 Aug 2025, Jiang et al., 12 Dec 2025), and scaffold logical rigor in AI-assisted writing (Hugenroth et al., 8 Apr 2026). It enables decomposition of complex reasoning or optimization tasks, enforces interpretability, and ensures persistent engagement with foundational concepts through interrogative, rather than merely instructive, exchange.

1. Socratic Loop Architectures Across Domains

The Teacher–Critic–Student Socratic Loop has been operationalized in various multi-agent and human–AI systems, notably in MARS (Zhang et al., 21 Mar 2025), ERL4SIIP (Jiang et al., 12 Dec 2025), orchestrated AI tutors (Degen et al., 7 Aug 2025), and Critical Inker (Hugenroth et al., 8 Apr 2026). Each implementation assigns differentiated, interacting responsibilities to the Teacher, Critic, and Student agents, as summarized below:

System	Teacher Role	Critic Mechanism	Student Role
MARS	Generates open-ended, progressive	Boolean approval/suggestion for questions	Refines prompt via question
	questions per sub-step	to ensure adherence to Socratic style	grounded learning
ERL4SIIP	RL-driven, persona-diverse policy	Hierarchical dense reward for safety	Simulated agent with STEM
	generating Socratic actions	and adaptive, multi-level feedback	mastery and misconceptions
Beyond Automation	LLM tutor posing Paul-style	Targeted, framework-driven critique	Human learner refines
(AI Tutor)	questions calibrated to ZPD	(e.g., PICOT logic, empirical tractability)	research questions
Critical Inker	Human/system-authored prompts,	LLM-mediated argument analysis	Human author iteratively
	model configuration	and Socratic questioning/visual feedback	revises and self-explains

This agent-based partition allows for scalable, interpretable, and pedagogically aligned orchestration of multi-step reasoning, supporting both human and autonomous learning scenarios.

2. Formal and Algorithmic Foundations

The loop is generally represented as an iterative process, in which the Teacher issues epistemically open prompts or questions, the Critic enforces Socratic principles or evaluates fit, and the Student produces (and justifies) a revised hypothesis, prompt, or response. In MARS (Zhang et al., 21 Mar 2025), this workflow is formalized as:

• Problem sub-decomposition: $$\mathbf{ST} = [st_1, \ldots, st_n] = \mathrm{Planner}(x; p_0)$$
• Teacher question: $$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$
• Critic filtering: Accepts or revises $q[st_i]$ until Socratic quality is met.
• Student update: $$p_i = \mathcal{M}_{\text{student}}(q[st_i]; p_{i-1})$$
• Target evaluation: $$\text{accuracy} = f(\mathcal{M}_{\text{tar}}(x; p_n), y)$$

In ERL4SIIP (Jiang et al., 12 Dec 2025), the Teacher is a LoRA-Division parameterized RL agent, updated via hierarchical reward signals encompassing gatekeeping (safety/ethics), process-level Socratic adaptivity, and outcome-driven learning gains. The Student is a dynamic simulator with explicit latent mastery, misconceptions, and affect vectors, producing text responses and triggering state transitions:

$$m_i^{(t+1)} = m_i^{(t)} + \alpha\, \mathbb{I}_{\mathrm{ZPD}}(d_{\mathrm{action}})\,(1 - m_i^{(t)})$$

The Critic provides multi-layer dense reward signals, and the policy is trained via EA for global diversity and PPO for local refinement.

Algorithmic realization in Critical Inker (Hugenroth et al., 8 Apr 2026) involves:

• Extraction of argumentation graph $G = (V, E)$ from student text.
• Logical evaluation and identification of flawed inference edges.
• Socratic questioning for each unresolved logical flaw, with stepwise revisiting until all invalidities are addressed.

3. Empirical Performance and Benchmarks

The Socratic loop implementation in MARS demonstrates state-of-the-art performance in prompt optimization benchmarks (Zhang et al., 21 Mar 2025):

• On 12 general-task datasets (BBH, MMLU), MARS achieves $85.11\%$ average accuracy, exceeding OPRO by $+6.04$ points.
• On five domain-specific datasets, MARS attains $75.81\%$ accuracy, improving on previous methods by $$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$0 points.
• Ablation studies show a pronounced drop ($$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$1 on BBH) when removing the entire Socratic guidance loop, with smaller—but still impactful—losses for eliminating the Critic or Planner agents.
• Efficiency metrics reveal that MARS achieves higher accuracy with fewer LLM API calls relative to generate-and-search methods, and converges within $$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$2–$$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$3 iterations.

Critical Inker attains $$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$4 argument-structural overlap and $$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$5 logical validity accuracy on annotated writing datasets (Hugenroth et al., 8 Apr 2026).

In educational intervention, orchestrated Socratic AI tutors yield significant improvement in perceived critical, independent, and reflective thinking compared to uninstructed chatbots, with effect sizes $$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$6 to $$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$7 on standardized scales (Degen et al., 7 Aug 2025).

4. Interpretability, Scalability, and System-Level Properties

The Teacher–Critic–Student Socratic Loop is characterized by strong process and result interpretability:

• The explicit sequence of questions, critiques, and revisions produces transparent chains of reasoning and changes (see Figure 1 and Appendix D in (Zhang et al., 21 Mar 2025)).
• Each agent’s actions (Teacher’s questioning, Critic’s gating, Student’s revisions/explanations) are auditable and can be traced to specific subgoals or reasoning steps.
• Scalability is supported by decomposing tasks into substeps, parallelizing Socratic rounds across subgoals (as in MARS’s Planner), and isolating semantic dependencies.

Institutional and curriculum-level evaluation reveals favorable cost-effectiveness ($$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$8 $$q[st_i] = \mathcal{M}_{\text{teacher}}(st_i; p_{i-1})$$9 per student per session for AI-tutors), reduced faculty rote workload, and new roles for orchestrating agent-based instruction and assessment (Degen et al., 7 Aug 2025).

5. Pedagogical and Theoretical Implications

The loop operationalizes epistemic agency and metacognitive engagement principles from constructivism and Vygotskian ZPD theory (Degen et al., 7 Aug 2025). The Teacher’s questioning is calibrated to the learner’s proximal development, while the Critic’s feedback enforces logical and empirical rigor (e.g., PICOT framework for research formulation).

In reinforcement learning settings, dynamic simulators and hierarchical reward decomposition align AI instruction with educationally valid progressions, addressing traditional reward sparsity and policy collapse (Jiang et al., 12 Dec 2025). In writing support systems, iterative Socratic probing is used to scaffold the revision of inference structures without offloading critical judgment to the system (Hugenroth et al., 8 Apr 2026).

6. Limitations and Future Directions

Current instantiations of the loop face open challenges in adapting Socratic guidance to highly heterogeneous or ill-defined task domains. The potential for universalizing Socratic decompositions or integrating environmental/user feedback is underexplored (Zhang et al., 21 Mar 2025). System-level open problems include dynamic personalization of Socratic pacing, orchestration architectures for cross-agent memory coherence, and delegation heuristics for distributing agent responsibilities (Degen et al., 7 Aug 2025). The evolution of regulatory, ethical, and curricular frameworks for agent-mediated co-authorship and assessment remains an active area of investigation, as does the measurement of higher-order transfer in longitudinal educational deployments.

---

The Teacher–Critic–Student Socratic Loop constitutes a foundational pattern in agentic AI systems for both prompt engineering and pedagogy, distinguished by its capacity to enforce reflective, transparent, and interpretable reasoning or optimization via dialogic, iterative exchange across modular roles (Zhang et al., 21 Mar 2025, Degen et al., 7 Aug 2025, Jiang et al., 12 Dec 2025, Hugenroth et al., 8 Apr 2026).