Dual Refinement Cycle Learning (DRCL)

• DRCL is a learning paradigm defined by iterative, bidirectional refinement cycles between complementary modules to mutually enhance predictions.
• It leverages dual-phase mechanisms that alternate between generation and evaluative feedback to update pseudo-labels and representations in applications like re-ID, text clustering, and depth estimation.
• DRCL improves performance by enforcing mutual consistency through cycle-consistent, memory-efficient feedback loops that reduce noise and sharpen feature discrimination.

Dual Refinement Cycle Learning (DRCL) denotes a class of iterative, bidirectional learning frameworks in which two modules—often representing distinct modalities, supervision signals, or levels of abstraction—mutually refine each other's predictions or representations in a cyclic manner. The strategy is characterized by embedding a dual feedback loop between paired subsystems (e.g., label and feature, semantic and structural, generation and reconstruction, or perception and reasoning), with the refinement at each stage governed by newly synthesized or repurposed pseudo-labels, consistency constraints, or regularizers. The paradigm has been instantiated in diverse domains including unsupervised domain adaptive person re-identification, cycle-consistent prompt engineering for large multimodal models, unsupervised text community detection, deep equilibrium self-supervised geometry, and dual-system GUI automation.

1. Fundamental Principles and Dual-Phase Mechanism

Every DRCL instantiation operates by alternating between two complementary phases in either an offline/online or inner/outer loop structure:

• Phase 1: Module-A generates, clusters, or predicts. This module processes either raw data or refined representations, yielding candidate labels, features, textual prompts, or policy actions.
• Phase 2: Module-B consumes outputs from A (potentially in batch), providing feedback via re-clustering, semantic evaluation, or cycle-consistent loss. This feedback is used to update the module-A’s parameters, pseudo-labels, or representations.

This cyclical exchange enforces mutual consistency and drives the two modules toward a fixed point representing a consensus between their competing or complementary objectives. Modality-specific implementations differ in how pseudo signals are formed, which consistency objectives are enforced, and what forms of regularization are applied.

For example:

• In UDA re-ID, label refinement via hierarchical clustering alternates with feature refinement via memory-based regularization (Dai et al., 2020).
• For prompt engineering and multimodal models, a forward generator and reverse generator enforce cycle-consistency on prompts or images, with each round refining the input context (Diesendruck et al., 2024).
• In graph-based text community detection, GCN-driven structure and transformer-based text semantics cross-inform via successive pseudo-label propagation (Wang et al., 8 Dec 2025).
• In deep geometric perception, depth and pose estimates are alternately refined based on updated local evidence, iterated to equilibrium (Bangunharcana et al., 2023).
• For GUI agents inspired by dual-system cognition, perception (System 1) and policy optimization (System 2) iteratively explore and refine via structured reward cycles (Wei et al., 22 Jun 2025).

2. Core Methodologies and Mathematical Formulations

While implementation specifics vary, modern DRCL frameworks exhibit advanced coupling mechanisms:

Domain/Task	Module-A	Module-B / Feedback Mechanism	Cycle Objective
UDA re-ID (Dai et al., 2020)	Hierarchical Cluster (labels)	IM-Spread-Out Feature Memory	Label/Feature Consistency, Metric Losses
Prompt Refinement (Diesendruck et al., 2024)	Forward Generator f(X→Y)	Backward Generator g(Y→X), Discriminative Hints	Cycle-Consistency Loss in Semantic Space
Text Graph (Wang et al., 8 Dec 2025)	GCN-based Detector (structure)	Transformer/Mamba (semantics)	Clustering & Cross-Entropy
Depth-Pose Estimation (Bangunharcana et al., 2023)	Depth/Pose Networks	Deep Equilibrium Feedback Loop	Photometric, Smoothness, Equilibrium Loss
GUI Agent (Wei et al., 22 Jun 2025)	Visual Omni-Parser (SPOT)	Group-based Policy Optimization (GRPO)	Group-wise Relative/Format Rewards

Mathematically, DRCL recursions typically follow:

• Pseudo-label refinement:

$$\hat{y}_i = \arg\max_{l} \frac{1}{R} \sum_{r=1}^R f_i^\top c_{l,r}$$

where $f_i$ is a feature, $c_{l,r}$ are cluster prototypes.

• Memory-based feature regularization:

$$\mathcal{L}_{\text{spread}} = \frac{1}{N} \sum_{i=1}^N \log\Big[1 + \sum_{k \in \mathcal{K}_i} \sum_{n \notin \mathcal{K}_i} \exp(f_i^\top v_n - f_i^\top v_k + m)\Big]$$

• Cycle-consistency for prompt refinement:

$$\mathcal{L}_{\text{cycle}}(s_0; t) = \mathcal{D}[s_0, g(f(t + s_0))]$$

with hint-based updates $s_{i+1} = s_i + \text{hint}_{i+1}$.

• Alternating structural-semantic update:

$$X^{(I)} \leftarrow X^{(t)}, \qquad C \leftarrow \begin{cases} C^{(s)} & \text{if odd or early epoch}\ C^{(t)} & \text{otherwise} \end{cases}$$

where $X^{(t)}$ are semantic features, $C^{(s)}$ and $C^{(t)}$ structural and semantic communities.

• Equilibrium modeling for geometry:

$f_i$0

3. Domain-Specific Instantiations

3.1 Unsupervised Domain Adaptive Person Re-Identification

A two-phase DRCL cycle alternates:

• Offline: DBSCAN for coarse clustering of image features, followed by K-means for prototype extraction and label refinement.
• Online: Joint metric loss over both original and refined pseudo-labels, plus instant memory spread-out regularization that disperses embedding space, maintained by a no-momentum instant memory bank. Losses are weighted and blended, driving representations and pseudo-labels toward mutual consistency. Iteration continues until label assignments stabilize or epochs are exhausted (Dai et al., 2020).

3.2 Cycle-Consistent Prompt Refinement in Multimodal Models

CyclePrompt applies DRCL by composing f (e.g., image→caption, text→code) and g (e.g., caption→image, code→text), using a discriminator to identify and apply semantic corrections to the free prompt or context in each iteration. Each cycle applies forward generation, backward reconstruction, hint extraction, and prompt concatenation. The process is repeated until a cycle-consistency or semantic similarity threshold is met, without any fine-tuning of the LLM or vision model (Diesendruck et al., 2024).

3.3 Unsupervised Text Classification and Community Detection

For text-attributed graphs, DRCL bridges GCN-driven structure and transformer-based semantic modeling. Initial communities are constructed by running Louvain modularity on the graph, with scale-quality filtering. The GCN module propagates these signals structurally, updating node representations via soft and hard pseudo-labeling. Alternately, a Mamba backbone encodes node texts, with a classifier trained on GCN pseudo-labels. The community assignments produced by each module are cyclically swapped as supervision, supporting mutual reinforcement (Wang et al., 8 Dec 2025).

3.4 Deep Equilibrium Geometry (Depth and Pose Estimation)

DRCL refines geometric predictions by alternating updates to depth estimates, generated by analyzing epipolar line matching costs with current pose, and pose estimates, informed by alignment residuals with the current depth. An implicit deep equilibrium model solves for the fixed point jointly, using Anderson/Broyden updates and implicit differentiation for scalable, memory-efficient training (Bangunharcana et al., 2023).

3.5 Dual-System GUI Agents

CogniGUI utilizes DRCL within a dual-system framework:

• System 1 parses all candidate screen/UI elements (SPOT) based on a learned relevance score combining clickability and instruction match.
• System 2 executes and refines action policies (GRPO) over the filtered candidate set, utilizing groupwise rewards for both format and efficiency. Failed candidates are iteratively eliminated, and the policy is updated using group-baseline gradients, aligning exploration and mastery phases of interaction (Wei et al., 22 Jun 2025).

4. Representative Algorithms and Pseudocode

Each DRCL approach provides a procedural algorithm; a generalized DRCL pseudocode cycle is as follows:

$f_i$1

Each concrete domain replaces module_A/B and feedback mechanisms with its domain-specific operations (e.g., clustering, pseudo-label assignment, semantic feedback, cycle prompting).

5. Quantitative Performance and Ablation Studies

DRCL consistently yields substantial improvements versus prior one-pass or unimodal/decoupled baselines. Notable empirical gains substantiated in peer-reviewed evaluations include:

• UDA re-ID (Dai et al., 2020):
  • Duke→Market1501: Baseline mAP 67.9%, DRCL 78.0% (+10.1 pp); rank-1 85.7%→90.9%.
  • Market1501→Duke: Baseline mAP 56.4%, DRCL 67.7% (+11.3 pp).
  • MSMT17 transfer: DRCL outperforms MMT by up to +4 pp on mAP.
  • Ablations: Removing hierarchical label refinement or IM-spread-out independently reduces performance; combined they yield the largest boost.
• Prompt refinement (Diesendruck et al., 2024):
  • HumanEval code synthesis: Zero-shot GPT-4 80.5%, CyclePrompt DRCL 87.2% (+6.7 pp).
  • VQAv2/FigureQA caption accuracy: CyclePrompt captions consistently outperform zero-shot GPT-4V captions.
• Text community detection (Wang et al., 8 Dec 2025):
  • Cora: DRCL NMI 0.604 (MAGI 0.590), ACC 0.768 (MAGI 0.750).
  • Citeseer: DRCL NMI 0.433 (CommDGI 0.425), ACC 0.678 (CommDGI 0.623).
  • Ablation: Removing structural or semantic branch drops ACC by 2–5 points.
• Geometry (Bangunharcana et al., 2023):
  • Outperforms published self-supervised baselines in KITTI depth and odometry, with improved geometric consistency and reduced error.
• GUI grounding/navigation (Wei et al., 22 Jun 2025):
  • ScreenSpot accuracy: CogniGUI 98.5–94.3% ACC (vs. 91.8–83.5% for baselines).
  • ScreenSeek navigation (CPQ): CogniGUI 0.62–0.75 (vs. 0.23–0.51 for other agents).
  • SPOT ablation reduces CPQ by ~50%, confirming dual-cycle synergy.

6. Theoretical and Practical Implications

DRCL demonstrates that cross-informing refinement cycles can suppress noise in pseudo-labels, sharpen feature discrimination, yield more truthful semantic or structural assignments, and efficiently explore complex action spaces. Across instantiations, DRCL obviates the need for full supervision or heavy cross-domain fine-tuning, instead exploiting intrinsic consistency and mutual correction between dual components. The architecture is robust against label noise and modality shift, and scales to large-scale or memory-constrained domains via equilibrium or memory-efficient cycles.

While most examples focus on paired module induction, extensions to n-way cycles, hierarchical multi-phase cycles, or more tightly-coupled equilibrium systems are plausible directions. DRCL embodies a general template for unsupervised or self-supervised learning via cyclic mutual refinement, and is now evidenced in diverse fields including computer vision, graph learning, natural language processing, and autonomous decision making (Dai et al., 2020, Diesendruck et al., 2024, Wang et al., 8 Dec 2025, Bangunharcana et al., 2023, Wei et al., 22 Jun 2025).