IRIS Framework Overview

• IRIS framework is a multidisciplinary system encompassing iterative causal discovery, visible-light iris recognition, heterogeneous task scheduling, and more.
• It leverages modular, extensible architectures with integrated LLM and deep learning methods to ensure rigorous benchmarking and reliable performance.
• Applications span biomedical data analysis, robotics-driven data collection, interpretable action assessment, and high-resolution imaging in digital pathology.

The term “IRIS Framework” encompasses several distinct systems across scientific, engineering, computational, and biometric domains. The primary interpretations in contemporary literature include: (1) iterative frameworks for verifiable causal discovery from unstructured sources, (2) end-to-end platforms for visible-light iris recognition and anti-spoofing, (3) modular runtimes for classical-quantum heterogeneous task scheduling, (4) systems for interpretable action quality assessment, (5) XR-based data collection environments for robotics, and (6) optimized infrastructure for streaming and interacting with high-resolution biological images. Each addresses domain-specific challenges but generally adheres to principles of modularity, extensibility, and rigorous benchmarking.

1. Iterative and Integrated System for Verifiable Causal Discovery

The IRIS framework as introduced by Cao, S.H., Ding, J., and Lee, S. (Feng et al., 10 Oct 2025) targets causal discovery in the absence of tabular datasets. The system expands an initial variable set $V_0$ to a larger variable set $V$ for which it iteratively constructs a causal graph $G=(V,E)$ via automated document collection, variable extraction, and a hybrid discovery pipeline.

Each iteration $t$ comprises four integrated modules:

• Document Retrieval and Ranking: Queries all nonempty subsets $S \subseteq V_t$ (and synonyms), retrieving $k_S$ top documents per subset. Documents are ranked by $R(d;V)=\sum_{X \in V} 1_{X \in d}$.
• Variable Extraction: LLM-based extraction of variable values $v_{ij}$ (from doc $d_i$ for variable $X_j$), with chain-of-thought prompts and confidence scoring. A matrix $X_t \in \mathbb{R}^{T\times|V_t|}$ is constructed with thresholding on $c_{ij}$ for reliability.
• Hybrid Causal Discovery: Combines traditional structure learning (PC, GES, NOTEARS) on $X_t$ with LLM-verified edge claims across document subsets. Edges are included if statistical and/or LLM-based veracity thresholds are met, and strongly refuted edges are excluded.
• Missing Variable Proposal: LLMs prompt for additional relevant variables per document. Selection is via co-occurrence-based PMI and LLM veracity measures. The most supported variables are added, and the process repeats until convergence.

Formal guarantees on correctness follow from sufficient retrieval coverage and LLM accuracy bounds; the hybrid merge recovers the true graph $E^*$ with high probability as $T \to \infty$. This approach relaxes classical assumptions of acyclicity and causal sufficiency and enables variable discovery outside known ontologies.

Empirical evaluation on biomedical, epidemiological, and insurance datasets outperforms zero-shot LLM, CoT, RAG, and classical statistical methods in both expanded causal recall and normalized Hamming distance ratio. Individual modules demonstrate statistically significant precision, recall, and F1 improvements over prior art.

Limitations include dependency on external search APIs (potential for source bias), LLM inference latency, and assumed independence and coverage among retrieved documents. Suggested extensions include active learning, integration of do-calculus queries, and scalable graph partitioning (Feng et al., 10 Oct 2025).

2. IRIS for Visible-Light Iris Recognition

In the biometrics domain, IRIS denotes an open-source, quality-assured smartphone iris recognition stack focused on the visible spectrum (VIS) (Venkataswamy et al., 17 Dec 2025). The system standardizes every acquisition stage:

• Data Acquisition: Android capture application leverages YOLOv3-Tiny for real-time iris/eye detection with automated ISO/IEC 29794-6 compliance, enforcing sharpness, contrast, concentricity, and margin quality for each image. Frames passing quality checks are cropped and saved with minimal user intervention.
• Segmentation: LightIrisNet, a MobileNetV3-Large-based multi-task segmentation head, predicts dense iris and pupil masks, boundaries, signed distance transforms, and elliptic fit parameters for robust normalization.
• Normalization and Matching: Masks feed a Daugman “rubbersheet” normalizer and contrast enhancer; normalized strips are matched by: (a) OSIRIS pipeline (Gabor features, phase quantization, Hamming distance), or (b) IrisFormer (VIS-trained Vision Transformer with rotation-invariant patch embeddings and cosine similarity).
• Evaluation: The CUVIRIS dataset of 752 ISO-compliant VIS iris samples from 47 subjects serves as the main benchmark. OSIRIS achieves TAR=97.9% at FAR=0.01, EER=0.76%; IrisFormer achieves EER=0.057%. CNN baselines (DeepIrisNet2, SCNN, DSF) lag behind in error rates.

All modules are released with full source, pretrained weights, Docker reproducibility, and standardized protocols for community benchmarking (Venkataswamy et al., 17 Dec 2025).

3. IRIS as a Heterogeneous Task-Based Runtime (Classical–Quantum Workflows)

IRIS, as extended in Q-IRIS (Miniskar et al., 15 Dec 2025), represents an asynchronous, device-agnostic runtime for scheduling compute tasks across CPUs, GPUs, FPGAs, and quantum accelerators. The core primitives are:

• Task Graph Model: User describes computation as a DAG $(T, E)$; each task $t$ specifies a kernel, inputs, outputs, and preferred device.
• Multi-Device Scheduling and DMEM: Scheduler dispatches ready tasks to available devices, tracking data residency transparently.
• Quantum Kernels as Native Tasks: Via the QIR-EE (Quantum Intermediate Representation Execution Engine), IRIS can schedule and orchestrate quantum kernels and simulators alongside classical tasks.
• Circuit Cutting: Demonstrated with four-qubit GHZ circuits, quantum workloads are decomposed into independent subcircuits using quasi-probability decomposition, distributing subcircuit execution for improved throughput and minimized resource contention.

The architecture is extensible, supporting new device backends and resource types via plugin APIs. Open challenges include latency hiding between classical and quantum domains, error propagation through quantum tasks, and large-scale data movement optimizations (Miniskar et al., 15 Dec 2025).

4. Interactive Research Ideation and Hypothesis Generation

IRIS is also used as an “Interactive Research Ideation System” for scientific discovery, integrating LLM agents, a Monte Carlo Tree Search (MCTS) controller, researcher-in-the-loop feedback, and structured literature retrieval (Garikaparthi et al., 23 Apr 2025).

• The multi-agent system (ideation, review, retrieval agents) explores research hypotheses as trajectories in an MCTS search tree, guided by fine-grained reviewer feedback.
• The workflow includes steerable node expansion, feedback acceptance/rejection, and explicit user override channels, offering transparency and control absent from automated multi-agent systems.
• Empirical user studies report absolute and ELO-based gains over state-of-the-art GPT and cloud LLM baselines in research creativity and quality metrics.

Steerable, aspect-wise review mitigates reward hacking, and in-context literature retrieval (with manual PDF ingestion) promotes domain-relevance. All infrastructure, prompts, and pipelines are open source (Garikaparthi et al., 23 Apr 2025).

5. Frameworks for Segmentation, Anti-Spoofing, and Whole-Slide Imaging

Iris Segmentation and Anti-Spoofing

Multiple IRIS-branded frameworks are prominent in iris segmentation:

• DeepIrisNet2: End-to-end, segmentation-agnostic, scale-robust CNN with spatial transformer layers for NIR and VIS biometrics; requires no classical normalization or precise segmentation (Gangwar et al., 2019).
• U-Net+VGG16 for Bovine: Pixel-level segmentation on minimal annotation with U-Net/VGG16, optimized for livestock traceability (Dice=98.35%, Acc=99.5%) (Yoon et al., 2022).
• IAS-CDT and Masked-MoE: Unifying anti-spoofing evaluation with a generalized cross-domain protocol (IAS-CDT), and introducing a Masked-MoE architecture with expert-masked regularization, supporting superior cross-sensor and cross-race generalization ($\approx$4.85% ACER for cross-device) (Zou et al., 2024).

Whole-Slide Imaging and Tile Streaming

In digital pathology, “Iris” refers to a low-overhead microservice for high-resolution whole slide image streaming in IFE format. The Iris RESTful Server:

• Uses a lock-free, dual-queue C++ architecture (Asio networking/Boost Beast, file system workers) for tile-level streaming at $\geq$5,000 req/s (median 21 ms).
• Offers DICOMweb WADO-RS and custom REST APIs for tile and metadata access.
• Provides seamless drop-in for OpenSeaDragon viewers via the IrisTileSource, requiring no client modification beyond the tile source initialization (Landvater et al., 8 Aug 2025).

6. IRIS for Interpretable Video-Based Action Quality Assessment

In action quality assessment (AQA), IRIS denotes a rubric-informed segmentation system with explicit interpretability:

• Architecture: 3D-CNN for video embedding, multi-stage TCN for temporal segmentation aligned with event rubrics, conditional CNNs for technical-program scoring (TES, PCS), additive aggregation for final output (Matsuyama et al., 2023).
• Interpretability: All intermediate outputs, including segment labels, GOE adjustments, and PCS subscores, are surfaced for user traceability.
• Evaluation: Outperforms black-box AQA models in both regression (Spearman’s $\rho$, Pearson’s $r$) and segmentation (Dice).

7. Cross-Domain Trends and Limitations

IRIS systems consistently exhibit:

• Modular, extensible architectures enabling adaptation to new tasks, devices, or sensors.
• Hybridization of classical and learned (deep or LLM-based) methodologies, often for robustness or interpretability.
• Explicitly defined benchmarks, datasets, and open-source code for reproducibility.
• Quantitative evaluation under diverse, often challenging real-world or cross-domain conditions.

Reported limitations include model size and training complexity in deep models, real-time and inference latencies in LLM and anti-spoofing systems, data acquisition bottlenecks, and assumptions on annotation or device/sensor availability.

• "IRIS: An Iterative and Integrated Framework for Verifiable Causal Discovery in the Absence of Tabular Data" (Feng et al., 10 Oct 2025)
• "An Open-Source Framework for Quality-Assured Smartphone-Based Visible Light Iris Recognition" (Venkataswamy et al., 17 Dec 2025)
• "IRIS: Interactive Research Ideation System for Accelerating Scientific Discovery" (Garikaparthi et al., 23 Apr 2025)
• "Iris RESTful Server and IrisTileSource: An Iris implementation for existing OpenSeaDragon viewers" (Landvater et al., 8 Aug 2025)
• "IRIS: Interpretable Rubric-Informed Segmentation for Action Quality Assessment" (Matsuyama et al., 2023)
• "Novel Deep Learning Framework For Bovine Iris Segmentation" (Yoon et al., 2022)
• "IRIS: An Immersive Robot Interaction System" (Jiang et al., 5 Feb 2025)
• "Q-IRIS: The Evolution of the IRIS Task-Based Runtime to Enable Classical-Quantum Workflows" (Miniskar et al., 15 Dec 2025)
• "A Unified Framework for Iris Anti-Spoofing: Introducing Iris Anti-Spoofing Cross-Domain-Testing Protocol and Masked-MoE Method" (Zou et al., 2024)
• "Learning scale-variant features for robust iris authentication with deep learning based ensemble framework" (Zheng et al., 2019)
• "DeepIrisNet2: Learning Deep-IrisCodes from Scratch for Segmentation-Robust Visible Wavelength and Near Infrared Iris Recognition" (Gangwar et al., 2019)
• "An approach to human iris recognition using quantitative analysis of image features and machine learning" (Khuzani et al., 2020)