AIRA-Design: Structured Design Methods

• AIRA-Design is a term for structured design methods applied in AR explanations, AI agent scaffolds, neural architecture discovery, and risk assessment.
• It uses explicit guidelines such as the 'When/What/How' framework and diversity cues to enhance system performance and design adaptability.
• The approach operationalizes layered detail and context-aware adaptability, validated through surveys, expert workshops, and benchmark testing.

Searching arXiv for papers explicitly referring to “AIRA-Design” and closely related AIRA usages to ground the article. AIRA-Design is a research designation used for several design-oriented frameworks rather than a single unified system. In the available arXiv literature, it denotes a practical application of the XAIR framework to explainable augmented-reality interfaces, a set of design recommendations for AI research agents centered on ideation diversity, the low-level mechanistic half of the dual-framework “agentic discovery” paradigm paired with AIRA-Compose, and a blueprint for concrete and connected AI risk assessment. Across these uses, the term consistently concerns structured design procedures, but the object of design differs substantially: AR explanations, agent scaffolds, neural architectures and training scripts, or governance workflows (Xu et al., 2023, Audran-Reiss et al., 19 Nov 2025, Pepe et al., 15 May 2026, Xia et al., 2023).

1. Scope of the term

The main uses of the designation can be summarized as follows.

Usage of “AIRA-Design”	Domain	Core design object
XAIR applied in practice	Explainable AR	“When / What / How” explanation design
Design recommendations for AIRA-Design	AI research agents	Ideation-diversity mechanisms
AIRA-Design in agentic discovery	Foundation-model engineering	Attention mechanisms, encoders, training loops
AIRA-Design blueprint	AI risk assessment	Layered, connected assessment workflow

In the explainable-AR literature, AIRA-Design is presented as “XAIR applied in practice,” with explicit guidance on when to explain, what to explain, and how to explain AI outputs in AR. In AI research-agent work, the same label is used for a design program that systematically injects ideation diversity into agent scaffolds. In “Agentic Discovery of Neural Architectures,” AIRA-Design is the low-level code-writing complement to AIRA-Compose, targeting mechanistic implementation of novel attention mechanisms and training scripts. In the C$^2$AIRA mapping study, “AIRA-Design” denotes a blueprint for a concrete and connected risk-assessment framework (Xu et al., 2023, Audran-Reiss et al., 19 Nov 2025, Pepe et al., 15 May 2026, Xia et al., 2023).

The designation should not be conflated with several unrelated AIRA or Aira systems in the literature: Aira as remote sighted assistance for people with visual impairments, AIRA as an infrared platform for autonomous drone landing, AIRA as AI-Induced Risk Audit for code inspection, AIRA as Automated Impulse Response Analysis for personalized well-being advice, and Aira as an AI-powered parallelization adviser for SMT processors (Lee et al., 2018, Liu et al., 2024, Parris, 19 Apr 2026, Blaauw et al., 2017, Los et al., 31 Aug 2025).

2. AIRA-Design for explainable augmented reality

In "XAIR: A Framework of Explainable AI in Augmented Reality" (Xu et al., 2023), AIRA-Design is the practical application of XAIR to explainable AR interfaces. Its organizing structure is the triplet “When / What / How.” “When to Explain?” is split into availability and delivery: explanations should always be prepared, user-triggered delivery is the default, and auto-triggering is reserved for just-in-time cases where user capacity is sufficient and either surprise or unfamiliarity is present or model confidence is low. “What to Explain?” covers seven explanation types—Input/Output, Why / Why-Not, How, Certainty, Example, What-If, How-To—together with global versus local scope and a concise-by-default detail policy with an expandable “more…” view. “How to Explain?” governs modality and visual paradigm: the modality should match the AI outcome, text is primary, graphics are supplementary only when self-explanatory, and explanations should be implicitly embedded when compatible with the environment and otherwise rendered explicitly.

The framework is grounded in a multidisciplinary methodology. The paper reports a review of approximately 100 papers across XAI, HCI, and AR; an end-user survey with $N=506$; and three expert workshops involving 12 experts. The survey found that 89.7% demand explanations and 63.4% prefer context-dependent delivery, with top content preferences given as Input/Output, How, Why, and Certainty. The design guidelines are formalized as G1–G8, including “Always prepare explanations for any AI outcome,” defaulting to user-triggered delivery, selecting content by intersecting system goals, user goals, and user AI literacy, highlighting Why first, and exposing a “More…” affordance for deeper detail (Xu et al., 2023).

Validation is reported at both designer and end-user levels. In the designer workshops ($N=10$), the framework obtained a System Usability Scale score of $74 \pm 6$, and the Creativity Support Index subscales were Exploration 7.9, Expressiveness 7.1, Enjoyment 8.0, Results Worth Effort 7.3, Collaboration 6.8, and Immersion 4.4. In the end-user AR evaluation ($N=12$), explanations increased intelligibility by approximately $+2.1$, transparency by approximately $+2.3$, trustworthiness by approximately $+2.5$, and approval of the When, What, and How dimensions by at least $+1.8$ on 7-point Likert measures; the AR system with explanations achieved SUS scores of $86 \pm 3$ and $N=506$0. Concrete case studies include jogging-route detours, plant-care alerts, movie-night food recommendation under driving conditions, cooking-timer advice under high load, and smart do-not-disturb settings (Xu et al., 2023).

3. AIRA-Design as a research-agent design program

In "What Does It Take to Be a Good AI Research Agent? Studying the Role of Ideation Diversity" (Audran-Reiss et al., 19 Nov 2025), AIRA-Design appears as a set of design recommendations for building stronger AI research agents. The central construct is ideation diversity, defined over the diversity of ML architectures proposed during the ideation phase. If $N=506$1 are the distinct high-level architectures extracted from the first $N=506$2 Draft nodes and $N=506$3 their empirical frequencies, diversity is quantified by Shannon entropy:

$N=506$4

A second measure, tree-level diversity, is the number of distinct architectures among the first $N=506$5 drafts:

$N=506$6

The empirical study is large-scale. It evaluates 75 Kaggle problems from MLE-bench, spanning vision, NLP, tabular, time-series, and multimodal tasks, over 11,000 runs and approximately 1.2 million tree nodes, with a compute budget of 24 hours per run and 264,000 GPU-hours total. The paper compares three scaffolds—AIDE, AIRA$N=506$7, and AIRA$N=506$8—all sharing Draft, Debug, and Improve operators. On MLE-bench lite with an o3 backbone, the reported average architecture entropy is 1.1 bits for AIDE, 1.8 bits for AIRA$N=506$9, and 1.7 bits for AIRA$N=10$0; mean $N=10$1 is 2.2, 3.5, and 3.3, respectively. Across seeds, the paper reports Pearson correlations of approximately $N=10$2–$N=10$3 between $N=10$4 and medal rate, with high-performing backbones clustering around $N=10$5–2.2 bits and medal rate around 50% (Audran-Reiss et al., 19 Nov 2025).

The controlled ablation makes the causal claim explicit. Baseline agents include sibling memory, prompt-adaptive complexity, and a “be diverse” cue; low-diversity variants remove the complexity cue and replace it with an instruction to give similar solutions. Baseline agents propose at least 3 distinct architectures in 60% of tasks, versus 30% for low-diversity variants, and mean $N=10$6 drops from approximately 3.3 to approximately 1.9. Medal rate falls from $N=10$7 to $N=10$8 for AIRA$N=10$9 and from $74 \pm 6$0 to $74 \pm 6$1 for AIRA$74 \pm 6$2. The same degradation appears in Valid Submission Rate, Average Normalized Score, Percentile, and ELO (Audran-Reiss et al., 19 Nov 2025).

The resulting AIRA-Design recommendations are concrete. They include explicit diversity cues in the system prompt, prompt-adaptive complexity across the 5 drafts, sibling-memory sharing to de-duplicate ideas without overwriting diversity, at least 5 initial ideas, and a target of $74 \pm 6$3 bits, corresponding to approximately 3 distinct architectures out of 5. The paper also states that prompt-based diversity control is more reliable than merely tuning temperature, while suggesting a sampling range of $74 \pm 6$4 if temperature is used, and recommends balancing novelty with feasibility by biasing 20% of drafts toward riskier novel ideas and 80% toward safe variations (Audran-Reiss et al., 19 Nov 2025).

4. AIRA-Design in agentic neural architecture discovery

In "Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design" (Pepe et al., 15 May 2026), AIRA-Design is the low-level mechanistic half of a dual-framework paradigm. AIRA-Compose searches over a fixed palette of Transformer, MLP, and SSM (Mamba-2) blocks, whereas AIRA-Design requires agents to write from scratch the code implementing new attention mechanisms, encoders, and training loops. Its objectives are twofold: first, to implement subquadratic, long-range attention or SSM mechanisms in JAX/Flax for the Long Range Arena benchmark; second, to engineer GPT training scripts optimizing validation bits-per-byte under a fixed 5-minute GPU budget for the Autoresearch benchmark.

The agent architecture is explicitly multi-agent and multi-model. The study deploys up to 20 agents, pairing 12 different reasoning LLMs with two search scaffolds in the AIRA-dojo harness: one-shot and greedy. The operator set is Draft, Analyze, Debug, and Improve. Each run can last up to 24 GPU hours on LRA or 5 minutes on Autoresearch, with each step evaluated by an isolated evaluate.py harness. The tracked metrics are Raw Score, Valid Submission Rate, and Normalized Score, the last defined through the “march-of-9s” transform

$74 \pm 6$5

followed by normalization between the minimum and the state of the art (Pepe et al., 15 May 2026).

AIRA-Design’s most distinctive outputs are the agent-written mechanisms themselves. The paper highlights three modules. Bi-SVR (“Bidirectional Selective Vector Recurrence”) for ListOps projects tokens to receptance, key, value, forward-decay, and backward-decay channels, runs a two-vector recurrence in both directions, and fuses the result with GroupNorm; its per-token cost is stated as $74 \pm 6$6. Cross-Covariance Text Attention for IMDB sentiment replaces $74 \pm 6$7 with

$74 \pm 6$8

normalizes $74 \pm 6$9 and $N=12$0 over the token dimension, and computes

$N=12$1

with a learnable per-head temperature $N=12$2. Hybrid Local–Global Attention for document retrieval combines block-local causal softmax with global linear attention using $N=12$3, and mixes the two paths with a sigmoid gate. The paper also attributes smaller but important tweaks to the agents, including partial rotary positional embeddings, per-head learnable output scaling, tanh soft-caps on logits, and very small standard deviations in embedding initializations (Pepe et al., 15 May 2026).

The training scripts are equally central. Agent-written train.py files use ClimbMix web-text shards, a GPTConfig with custom block modules, precomputed rotary embeddings up to 20 times sequence length, and a MuonAdamW hybrid optimizer. The learning-rate schedule is warmup, then constant, then linear warmdown over the final 55% of time, with final LR fraction equal to 0.0. Gradient accumulation continues until TOTAL_BATCH_SIZE = 2^21 tokens per step, and no checkpoints are used because training runs to a 5-minute single-GPU wall clock (Pepe et al., 15 May 2026).

Quantitatively, the abstract reports that agent-designed architectures reach within 2.3% and 2.6% of human state-of-the-art on document matching and text classification, and that Greedy Opus 4.5 achieves 0.968 validation bits-per-byte under a fixed time budget, surpassing the published minimum. The detailed Long Range Arena table reports best configurable accuracies of 0.5050 on ListOps, 0.8792 on Text, and 0.7943 on Retrieval, with configurable performance substantially closer to human SOTA on Text and Retrieval than on ListOps. The paper also emphasizes that one-shot scaffolds fail entirely on mechanistic tasks, that valid submission rates are poor for one-shots and only moderate for many agents, and that wider hyperparameter exposure can hurt weaker agents unless accompanied by stronger search policies (Pepe et al., 15 May 2026).

5. AIRA-Design as a blueprint for concrete and connected AI risk assessment

In "Towards Concrete and Connected AI Risk Assessment (C$N=12$4AIRA): A Systematic Mapping Study" (Xia et al., 2023), AIRA-Design is a blueprint rather than a deployed software artifact. It arises from a systematic mapping of 16 non-academic AI risk-assessment frameworks published between 2018 and 2022, covering governments, industry, and NGOs. The study evaluates those frameworks against RAI principles, stakeholder coverage, lifecycle stages, geographical scope, targeted domains, assessment types, risk factors, and mitigation guidance. The derived deficiencies are eightfold: incomplete principle coverage, narrow stakeholder scope, inconsistent lifecycle coverage, limited contextual adaptation, subjective and unchecked Q&A, weak connectedness, minimal mitigation planning, and poor process clarity.

The proposed blueprint is governed by eight essential design principles. C1 is comprehensiveness over principles, stakeholder levels, lifecycle stages, domains, regions, and risk factors. C2 is connectedness, linking each high-level question $N=12$5 to concrete assessment or mitigation patterns $N=12$6 through a dependency graph $N=12$7. C3 is clarity through an explicit process model

$N=12$8

where each step specifies inputs, activities, outputs, resources, and responsible roles. C4 is layering, C5 is extensibility and adaptivity through a context model, C6 combines objective and subjective evidence through

$N=12$9

C7 structures mitigations as a catalog with effect, cost, and residual risk, and C8 embeds the framework in governance processes (Xia et al., 2023).

The blueprint is operationalized as a ten-step procedure. It begins with context initialization and stakeholder mapping, then sets up the principle and risk taxonomy, decomposes the lifecycle into stages from planning and requirements through operation, builds layered question banks and their dependency graph, configures an adaptive Q&A toolchain with objective tests, executes the assessment, retrieves mitigation patterns above threshold $+2.1$0, generates structured reports, and feeds deployment incidents back into the pattern library. The adopted risk-factor taxonomy is

$+2.1$1

The paper’s comparison matrix positions C$+2.1$2AIRA as fully supporting dimensions that are only partially supported or absent in representative frameworks, especially connectedness, objective metrics, mitigation structuring, and context extensibility (Xia et al., 2023).

6. Common design motifs, limits, and scholarly interpretation

Across its distinct usages, AIRA-Design consistently names a structured response to design under constraint. In explainable AR, the constraint is cognitive load, modality, and context-dependent delivery. In research-agent design, it is the tendency of scaffolds to collapse into a narrow architectural prior unless diversity is explicitly engineered. In mechanistic self-design, it is the difficulty of producing valid low-level code under tight evaluation budgets. In AI risk assessment, it is the fragmentation of existing frameworks and the lack of connections between principles, stages, stakeholders, and concrete mitigation patterns (Xu et al., 2023, Audran-Reiss et al., 19 Nov 2025, Pepe et al., 15 May 2026, Xia et al., 2023).

Several motifs recur. First, all variants are staged: “When / What / How” in AR, Draft/Debug/Improve or Draft/Analyze/Debug/Improve in agent systems, and stepwise context-to-report workflows in risk assessment. Second, each variant gives explicit treatment to state or context: user capacity and task load in AR, architecture entropy and scaffold memory in research agents, benchmark budgets and validity constraints in mechanistic design, and region/domain/organizational context in governance. Third, layered detail is a shared device: concise defaults with expandable “More…” in AR, multiple draft branches in research agents, configurable versus non-configurable hyperparameter exposure in neural architecture discovery, and $+2.1$3 to $+2.1$4 assessment layers in C$+2.1$5AIRA (Xu et al., 2023, Audran-Reiss et al., 19 Nov 2025, Pepe et al., 15 May 2026, Xia et al., 2023).

The limits are equally consistent. The explainable-AR framework notes trade-offs between auto-triggering and interruption, and between rich graphics and clutter. The ideation-diversity study shows that stronger performance depends not simply on more exploration, but on diversity mechanisms that preserve implementability. The mechanistic AIRA-Design work states that the discoveries are largely engineering recombinations of existing ideas rather than new theory, and that one-shot scaffolds fail on low-level code-generation tasks. The C$+2.1$6AIRA blueprint, for its part, is prescriptive and comparative, but still formulated as a blueprint whose value depends on later implementation and governance adoption (Xu et al., 2023, Audran-Reiss et al., 19 Nov 2025, Pepe et al., 15 May 2026, Xia et al., 2023).

This suggests that AIRA-Design is best understood not as a single canonical architecture but as a recurring research label for design methods that combine explicit structure, adaptation to context, and operational evaluation. The shared name signals a family resemblance in methodology, while the technical substance remains domain-specific.