Program in detail
ICAITD 2026
Oct 12 2026, Monday
Time
ID
| Title |
Authors
14:15
PA-001
LLM Hallucination in High-Stakes Professional Domains: A Comparative Audit of Seven Models on Legal, Medical, and Educational Query Sets
Suvendu Sekhar Mohanty
Abstract: Despite extensive hallucination benchmarking on general knowledge tasks, little systematic work compares hallucination rates and error types across professional domains where errors have high consequences. We construct three domain-specific query sets (n=150 each): legal reasoning (contract clauses, case citations), medical Q&A (drug interactions, diagnostic criteria), and educational content (curriculum standards, learning objective design). We evaluate seven models (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3.3 70B, Mistral Large, Qwen2.5, DeepSeek-R1) against expert-verified ground truth. We find hallucination rates 2.3-4.1× higher in domain tasks vs. general benchmarks, with citation fabrication as the dominant failure mode, and propose a domain-weighted hallucination score (DWHS) for practitioners.
14:30
PA-002
AI Catalyst for Quantum and Computational Chemistry
Ephraim Eliav
Abstract: The integration of artificial intelligence (AI) and machine learning (ML) into computational chemistry is revolutionizing our ability to model complex molecular systems by overcoming traditional trade-offs between accuracy and computational cost. Graph neural networks and E(3)-equivariant models (such as Uni-HamGNN and DeepH-E3) successfully parameterize complex density functional theory (DFT) and spin-orbit coupled Hamiltonians [1]. These universal models preserve essential physical symmetries and allow for the rapid, high-throughput screening of large-scale quantum materials and supercells without the need for resource-intensive, systemspecific retraining. Furthermore, AI is unlocking new frontiers in strongly correlated many-body physics. Generative Transformers and Neural Network Quantum States (NNQS), such as QiankunNet, can effectively learn the structure of many-body states, dynamically selecting relevant electronic configurations to achieve near full configuration interaction (FCI) accuracy [2]. For systems containing heavy elements, where relativistic effects traditionally demand staggering computational resources, ?-machine learning models ?-SO-ML) efficiently predict spin-orbit and electron correlation corrections for NMR chemical shifts [3]. This AI-driven approach recovers up to 85% of spin-orbit contributions, halving the deviation from experimental values at virtually no extra computational cost. Finally, combining these accelerated electronic structure methods with generative algorithms enables autonomous material discovery and complex reaction network generation. Surrogate models like CAS-ESM learn reduced density matrices to drive multireference excited-state dynamics at near force-field costs, while scaled graph networks have recently identified millions of new, stable crystal structures [4]. Ultimately, the strategic integration of AI provides a predictive, scalable, and highly efficient framework that bridges the gap between fundamental quantum mechanics and the design of next-generation functional materials.
15:00
LT-001
Lecture on a research paper: ”Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems” in Japanese
Yuki Shimizu
15:50
DS-001
Discussion
Chair
Oct 13 2026, Tuesday:
Time
ID
| Title |
Authors
10:10
OP-001
Opening talk
Kazuo Ohzeki
Keynote 1:
10:30
KN-001
Artificial Intimacy and Intelligence Illusions: Navigating Trust, Bias, and Sustainability in AI.
Veena Menezes and Amberlynn Menezes
Keynote 2:
Chair: Veena Menezes and Amberlynn Menezes
11:15
KN-002
to be added
Daisuke Ishii
Lunch break
Session: AI for Science
Chair: Veena Menezes and Amberlynn Menezes
14:00
FP-001
SSE: A Parameter-Efficient Training Recipe for Static Retrieval Embeddings
Rikka Botan
Abstract:Static retrieval embeddings provide fast inference but often trail contextual encoders in accuracy. We present SSE (Stable Static Embedding), a complete training recipe that combines contrastive ranking, Matryoshka prefix supervision, a 15-source data mixture, large-batch one-epoch optimization, and a lightweight post-pooling transform. The released checkpoint has 15.6M parameters, approximately half the 31.3M parameters of the prior static reference. Its 256-dimensional prefix reaches NanoBEIR (English) nDCG@10 of 0.5036, matching the reference's native 1024-dimensional score of 0.5030 with one quarter of the vector storage; the full 512-dimensional representation reaches 0.5159. The model encodes 119,832 queries per second on CPU, has the highest measured quality within the throughput-comparable static family, and lies on the measured speed-quality Pareto frontier. These results support SSE as a parameter- and storage-efficient recipe for static dense retrieval.
14:30
PP-001
Benchmarking and Rethinking Evaluation of Pre-trained Language Models for Automated Vulnerability Patching
Zanis Ali Khan and Seifeddine Bettaieb
Abstract:nd effective mitigation strategies. While advancements in Automated Program Repair (APR) have primarily targeted general software bugs, the domain of vulnerability patching, which is a security-critical subset of APR, remains underexplored. This paper investigates the potential of pre-trained language models, CodeBERT and CodeT5, for automated vulnerability patching across diverse datasets and five programming languages. Rather than reporting accuracy alone, we examine three factors that determine how such results should be interpreted: how preprocessing-induced information loss affects accuracy, how well commonly used evaluation metrics agree with one another, and how patch complexity relates to model performance. We find that preprocessing loss significantly degrades BLEU scores for both models, while structural and exact-match metrics remain largely unaffected, indicating that preprocessing artifacts disrupt surface-form similarity more than deeper correctness. Evaluation metrics themselves show weak, model-dependent agreement, with no dataset showing consensus across metrics on which model performs better, revealing that metric choice alone can change reported conclusions. Patch complexity is consistently associated with lower accuracy across both models, though this relationship is only statistically robust for CodeBERT, and no significant difference in complexity sensitivity between the two models could be confirmed. Together, these findings benchmark model performance while highlighting key limitations in both the models and the metrics used to evaluate them, offering guidance for more reliable evaluation of automated vulnerability patching in practical security applications.
15:00Coffee break
15:30
PP-002
Think in English, Output in Japanese: Mitigating GPU Workloads through Translation Offloading
Kairi Nishio
Abstract: The widespread adoption of Large Language Models (LLMs) has led to a significant increase in the power consumption of AI accelerator servers, presenting a critical environmental and economic challenge. While upgrading to the latest high-performance GPUs is a common countermeasure, the massive embodied carbon footprint from manufacturing often negates the lifecycle energy savings. To address this, we propose a softwarebased approach to reduce GPU power consumption by offloading a portion of the generation load to the CPU. Specifically, leveraging the difference in token counts between languages, our method instructs the LLM to process Japanese prompts and generate outputs in English, which are subsequently translated back into Japanese using an external API on the CPU. Experimental evaluations using the Qwen2.5-7B-Instruct model and a dataset of 1,000 tasks demonstrated an average power reduction of 36.3%, with the median power consumption dropping from 0.00045 kWh to 0.0002 kWh per inference. These results indicate that optimizing the output token representation can effectively minimize GPU load. This approach not only curbs operational energy use but also reduces the required number of GPUs, thereby contributing to the preservation of semiconductor resources and substantial reductions in overall infrastructure costs.
16:00
FP-002
A Self-Reinforcing Neuroimmune Clearance–Pruning Loop and Its EEG Signature: A Spatially Graded Hypothesis for Autism Spectrum Disorder and Developmental Dyslexia, A Perspective
Günet Eroğlu
Abstract: Autism spectrum disorder (ASD) and developmental dyslexia share a common low-frequency (theta-range) EEG shift, yet differ systematically in its spatial extent - diffuse in ASD, left-temporoparietal and focal in dyslexia. This conjunction - a shared, direction-invariant low-frequency signature co-occurring with a systematic spatial dissociation - is the single observation this framework is built to explain, and no existing account, I argue, predicts it as a unit. I propose that it arises from a self-reinforcing loop between two systems usually studied apart: microglial synaptic pruning and the brain's glymphatic-meningeal lymphatic clearance. Each drives the other - reactive microglia raise the clearance burden, and impaired clearance sustains the inflammatory tone that keeps microglia reactive - so the pair behaves as one bistable system whose clinical expression is set by two parameters: the direction of pruning error and its spatial extent. I support the proposal with a cause-and-effect molecular pathway (AQP4-cytokine-C1q/C3/CR3/C4b-CX3CR1), a biophysical account of why theta specifically (thalamocortical dysrhythmia from reduced recurrent excitation), a regionspecific rationale for left-temporoparietal vulnerability in dyslexia, and a network-level EEG framework (connectivity, graph topology, phase-locking, entropy, microstates, crossfrequency coupling, source localization) rather than power alone. The chain is formalized as a compact dynamical model exhibiting bistability and hysteresis, and yields falsifiable predictions with an explicit falsification experiment. Evidence remains largely correlational and preclinical; the aim is a testable organizing framework, not a proven mechanism. The novelty lies in the coupling - treating pruning and clearance as one bistable system - not in any individual mechanism, and the framework is deliberately bounded: it aims to explain one EEG conjunction across two disorders, not either disorder in full.
16:30
DS-002
Discussion
Chair
Oct 14 2026, Wednesday:
Time
ID
| Title |
Authors
09:55
OP-002
Opening message
Kazuo Ohzeki
Keynote 3:
Chair: Veena Menezes and Amberlynn Menezes
10:00
KN-003
Technological Singularity: The Ultimate Trajectory of Recursive Self Improvement in Automated Science
Yatima Kagurazaka
Keynote 4:
Chair: Veena Menezes and Amberlynn Menezes
10:45
KN-004
Building Trustworthy AI Systems: Reliability, Safety, and Human-AI Collaboration
Phani Rohitha Kaza
Session: AI Business
11:30
PP-003
Selling Connection, Selling Dependence: The Business Model of AI Companions
Veena Menezes and Amberlynn Menezes
Session: AI Assistant Research
12:00
PP-004
Reproducible, Traceable, and Robust, Yet Still Unwarranted: Toward Claim-Typed Falsification in AI-Assisted Research
Truong Xuan Khanh
Abstract: AI research agents increasingly automate literature synthesis, code generation, experiment execution, analysis, and manuscript writing. Several verification layers have developed in response: claim-level auditability links reported statements to documentary sources; metamorphic and mutation testing probe program behaviour under transformation; workflow provenance records execution lineage; and multiverse and specification-curve methods assess robustness across declared analysis choices. Each layer evaluates a distinct object and answers its own question well. We argue that a computational result can satisfy all of these layers and still fail to warrant the claim reported in a paper. The reason is that verification is organised by the objects it inspects, whereas falsification obligations are organised by what a claim asserts. What would challenge a threshold claim differs from what would challenge a transfer claim, and neither is captured by the null-crossing criterion appropriate to some effect claims. The mapping between the two organisations is not absent; it is tacit, performed by whoever holds the claim in mind. That arrangement was adequate while claims were produced at the pace of human attention. We identify the selection and composition of obligations given a claim as a distinct layer of the verification problem: one that is performed tacitly in current practice and represented in none of the four layers we survey. We set out a preliminary multi-label account of claim types and the obligations each carries, and state what would have to be settled before such a layer could be built. We propose no validator and claim no deficiency in existing methods
Chair: Veena Menezes and Amberlynn Menezes
Lunch break
Session: ADAS, Test, Traffic, Cut-in
Chair: Veena Menezes and Amberlynn Menezes
Keynote 5:
14:30
KN-005
Beyond Black-Box AI: Closed-Loop Simulation and Open Standards for NextGeneration Systems
Stefan Alexander Schneider
Abstract: Recent advances in Artificial Intelligence—particularly large language models and data-driven approaches—are rapidly transforming engineering workflows. However, when applied to safety-critical systems such as Advanced Driver Assistance Systems (ADAS), these methods face fundamental challenges, including limited traceability, lack of physical consistency, and difficulties in validation. This keynote presents a complementary perspective: combining AI-based methods with structured, model-based system development through closed-loop simulation. Building on the classical sense–plan–act paradigm, simulation environments based on open standards (e.g., ASAM OSI and scenario-based approaches) enable reproducible and interpretable system behavior. Using the FASim simulation framework, the talk illustrates how ADAS functions can be modeled, executed, and systematically evaluated in controlled environments. A particular focus is placed on emerging concepts such as resimulation and reconstruction, where AI-generated or real-world data is integrated into standardized simulation loops. The keynote outlines how this integration can form the basis of a future activities, where AI and simulation are combined into a unified methodology for scalable, transparent, and trustworthy ADAS development.
15:15
FP-003
Evaluating Prompt Engineering Techniques for LLM-Based Test Case Generation in Automotive ADAS Development
Anastasios Zosimidis
Abstract: The growing complexity of Advanced Driver Assistance Systems (ADAS) demands efficient methods for generating test cases from natural language requirements. Systematic surveys of prompt engineering [12, 13] document a rapidly growing catalogue of techniques. This paper evaluates eight of them, applied to three LLMs (GPT-4o, Claude 3.5 Sonnet, Claude 3 Haiku), for automated test case generation from industrial ADAS requirements, and reports the reliability of its own evaluation panel alongside the comparative results. Using Latin Hypercube Sampling across five dimensions (prompt technique, requirement, model, temperature, and top_p), we generate 50 experimental configurations and evaluate them in a fully crossed design: four automotive domain specialists each rated all 50 test cases on five quality criteria, yielding 1,000 individual ratings with no missing values. All 50 configurations returned a complete response without a recorded execution failure. Technique means span a narrow band of 0.30 points (3.39 to 3.69) around an overall mean of 3.60, with Few-Shot obtaining the highest descriptive mean, and no comparison across techniques, models, or sampling parameters reached statistical significance. The ratings carry little resolvable signal at the level on which these comparisons operate: the between-test-case variance component is not distinguishable from the residual test-case-by-evaluator variation (F = 0.98, p = 0.52), and the observed spread between test cases (SD = 0.22) is fully accounted for by measurement noise alone (0.23). The evaluators additionally differed systematically in how leniently they used the scale, accounting for an estimated 71.8% of the variance-component total. The evaluators did not agree on which test cases were better than others (ICC(2,4) = -0.01, effectively zero agreement). We conclude that this study did not detect differences between prompt techniques, models, or sampling parameters under the evaluation design used, and that the uncalibrated four-rater panel used here could not resolve differences of the size observed. The study was not designed as an equivalence test, so the absence of detected differences should not be read as evidence that they do not exist.
15:45Coffee break
16:15
IN-001
DETERMINISTIC LOGICAL CORE ARCHITECTURE (DLCA) Decoupling Deterministic Logic from Probabilistic Language Interface
Daniel Thomas Price
Abstract: Current Large Language Models(LLMs) combine probabilistic generation with limited explicit reasoning capacity, producing systems that are fluent but impossible to audit. The Deterministic Logical Core Architecture (DLCA) proposes deterministic reasoning processes which can be separated by Human-In-The-Loop (HITL) processes from necessary probabilistic processes. A Pre-Processing Unit (PPU) refines propositions to rectify language, cultural and logical factors before the Logical Core (LC) mathematical expressions evaluate the proposition. A probabilistic Social Interface(SI) following the LC is restricted to the role of governed translator producing human digestible text from formalized LC output. An Audit File(AF) records meta data describing every event from raw input to rendered output. LC content is intended to be derived from expert knowledge and foundational documents using public domain sources. Using meta data, the logic will thereby be traceable to foundational documents or experts. A core architectural proposition, which is deterministic symbolic evaluation is more efficient, has been empirically demonstrated in peer-reviewed literature. DLCA is realisable in hardware or software. Hardware implementation provides stronger immutability guarantees and software implementation is a deployable first stage or within an LLM. The architecture is intended to facilitate precision, reproducibility, auditability, efficiency, cultural specificity, and HITL control
16:45
PP-005
V2X-Based Cut-in Negotiation with Micro-Tip Protocol: A Cooperative Approach to Equitable Lane-Changing in Mixed Autonomous-Human Traffic
Kazuo Ohzeki
Abstract: Autonomous vehicles (AVs) are engineered to yield whenever safety cannot be guaranteed, and human drivers can exploit this deference to gain a unilateral time advantage - a structural inequity we formalise as Unilateral Dominance-Equilibrium Collapse (UDEC), for which no unilateral AV policy admits a Nash equilibrium. We reframe the cut-in interaction as a cooperative negotiation with three contributions. First, a cut-in is permitted if and only if the detected distance headway is at least a physically-derived stopping distance times an explicit safety factor; under exact perception this guarantees zero dangerous cutins permitted by construction, so the only empirical question is how well distance and target speed can be estimated from a camera. Second, we show that single-frame camera estimation breaks this guarantee, and that temporal integration over a short frame window - not a new recognition algorithm - substantially restores it, with the safety factor providing an explicit, transparent margin against residual perception error. Third, a V2X micro-tip protocol frames the resulting negotiation as a single-round ultimatum game, in which the yielding AV prices the concession and settles it through lightweight, identitytraceable signing rather than blockchain consensus, and derives the condition under which a dynamic tip restores the equilibrium that fixed AV compliance otherwise forecloses. The system requires no legal reform and operates as a voluntary bilateral agreement.
17:15
LT-002
AI Generated music
to be added
18:00
WD-001
Working Dinner
Chair
Oct 15 2026, Thursday:
Time
ID
| Title |
Authors
10:10
OP-003
Opening message
Kazuo Ohzeki
Session: Education 1
Chair: Veena Menezes and Amberlynn Menezes
10:15
PP-006
Development of a Method for Detecting the Illusion of Understanding Using LLM-Based Simulated Learners
Chengyi Hu and Shigeru Hosono
Abstract: In programming education, learners often subjectively believe they have mastered a concept before they have truly understood it - a phenomenon known as the Illusion of Understanding (IOU), which impedes effective intervention by adaptive learning support systems that typically rely on accuracy alone and lack a mechanism for detecting this metacognitive state in real time. This study proposes and develops a two-axis detection method based on the average calibration error (CE_avg) and the Area Under the ROC Curve (AUROC), and uses eight LLM-based simulated learner personas to validate the method prior to human-subject experiments. Through three iterative development cycles, we identified two methodological issues: non-computable AUROC values caused by single-class response patterns, and spuriously perfect AUROC values caused by small but systematic confidence gaps. We addressed these issues through persona refinement and a two-stage agent architecture that separates answer generation from confidence estimation. The final design recovered the intended IOU cases more consistently than the intermediate version, suggesting that LLM-based simulated learners can support the methodological development of IOU detection prior to human-subject experiments.
10:45
PP-007
BDI Faithfulness Audit: Testing the Necessity of Stated Beliefs in LLM Decisions
Koyo Matsuoka and Shigeru Hosono
Abstract: Large language model (LLM) agents increasingly provide explanations for their decisions, yet a plausible explanation may not reflect the factors that influenced the outcome. This paper proposes the BDI Faithfulness Audit (BFA), a framework for testing whether beliefs stated in a Belief-Desire-Intention (BDI) explanation were necessary for an LLM's decision. BFA grounds each stated Belief in a verbatim citation from the input, counterfactually modifies the cited span, and measures whether the decision changes across repeated trials. Based on the resulting flip rates, BFA classifies explanations as Faithful, Partial, or Unfaithful; cases that cannot be settled are withheld as Undecidable. The same procedure also evaluates whether supplied candidate omitted factors may have influenced the decision. An illustrative example shows how the verdict rule sorts flip-rate patterns into the three verdicts and the withheld outcome without relying on subjective judgments of plausibility. BFA thus provides an externally verifiable operationalization of explanation faithfulness based on the necessity of stated beliefs, while empirical validation remains future work.
11:15
PP-008
A Decision-Making Framework for Shared Local LLM Infrastructure
Yuki Ono and Shigeru Hosono
Lunch break
Abstract: Local large language models (LLMs) are increasingly operated on shared GPU-equipped computers in research laboratories. In such environments, users must decide whether a workload should be executed in a shared state or under exclusive use; however, these decisions are often based on individual experience rather than explicit operational criteria. This study proposes a decision-making framework that standardizes shared/exclusive usage decisions by organizing the information required for judgment, including current resource status, standalone resource consumption, performance impact during concurrent execution, reservation status, and processing requirements. To support practical operation, we implemented monitoring, reservation, notification, and shared/exclusive control mechanisms on a ThinkStation PGX and evaluated 15 local LLMs. Although all models executed successfully and sufficient memory remained available, concurrent execution reduced the generation speed of an existing process by 51.5-68.8%. These results indicate that memory availability alone is insufficient for determining shareability and that performance interference must be considered independently. The study demonstrates that executability and shareability are distinct operational concerns and presents a practical decision-making framework for standardizing usage decisions in shared local LLM infrastructures.
Session: Education 2, IRT
Chair: Veena Menezes and Amberlynn Menezes
13:30
PP-009
Bayesian Item Response Analysis of Data Science Reskilling Assessment Items: Regression Module, A Two-Year Comparison (2022–2023)
Ryota Kozakai
Abstract: In this study, we administered a supervised learning (regression) assessment as part of an internal reskilling program in 2022 (? ?=?23) and 2023 (? ?=?25). All participants were technical employees of AKKODiS Consulting Ltd., predominantly engaged in IT-related work. Using a 2-Parameter Logistic (2PL) Multidimensional Bayesian item response model, we estimated and visualized the test's item structure, participants' latent proficiency, and the difficulty and discrimination parameters of each item. The objective is to obtain evidence to inform future instructional design and to support the ongoing evaluation of corporate reskilling initiatives.
14:00
FP-004
L2-Regularized Nonlinear Regression-Based Fuzzy Clustering for Satisfaction Survey Analysis in the Academy Services
Ryota Kozakai
Abstract: 08 users of AKKODiS Academy's talent development services over the period from June 2024 to January 2026. In the survey, respondents rated six items-likelihood to recommend, program content, overall quality, contribution to skill improvement, support system, and cost performance-on ordinal scales, and provided free-text responses describing what they found positive and what they would like to see improved. Using keywords extracted from the free-text responses as explanatory variables, we constructed an L2-regularized nonlinear fuzzy cluster regression model to represent the ordinal scale ratings, classified respondents into five latent segments based on the regression coefficients, and visualized the characteristics of each segment as well as the differences across evaluation items. The analysis revealed that the cost performance rating is structurally lower than the other five items, and that segment membership tends to become ambiguous at the boundary between the group that prioritizes expertise and quality and the group that prioritizes relationships. These findings are intended to inform future improvements to the service.
14:30Coffee break
Session: Decolonalization
Chair: Veena Menezes and Amberlynn Menezes
14:50
PP-010
Decolonizing African Curricula Through AI-Enhanced Pedagogy: Reclaiming Indigenous Knowledge and Future Skills in Postcolonial Education Systems
Emmanuel Dumbuya
Abstract: This paper explores the intersection between artificial intelligence (AI) and curriculum decolonization in sub-Saharan Africa, arguing for a transformative educational paradigm that integrates indigenous knowledge systems with AI-enhanced learning. As African nations grapple with postcolonial curriculum models rooted in Eurocentric frameworks, the emergence of AI offers an unprecedented opportunity to reimagine pedagogical tools that honor cultural identity while preparing learners for the demands of a digital future. Through a critical review of literature and analysis of current curriculum practices, this study proposes an Afrocentric-AI curriculum model grounded in contextual relevance, epistemic justice, and future-ready competencies. The implications for curriculum reform, teacher training, and educational policy are discussed, with practical recommendations for implementation.
15:20
PP-011
Algorithmic Colonialism and the Appropriation of Indigenous Data: Safeguarding Cultural Epistemologies in the Digital Age
Nouridin Melo
Abstract: This paper examines the emerging phenomenon of algorithmic colonialism and its implications for Indigenous data sovereignty in Cameroon. Through critical analysis of case studies across Cameroon's diverse ethno-linguistic communities, we demonstrate how contemporary data extraction practices perpetuate colonial power dynamics by appropriating, commodifying, and misrepresenting Indigenous knowledge systems. The research employs a mixed-methods approach combining qualitative interviews with Indigenous knowledge keepers (n=32), computational analysis of algorithmic systems operating in Cameroon, and participatory action research within affected communities in the Northwest, West, Southwest, and East regions. Findings reveal systematic patterns of epistemological violence enacted through algorithmic systems that fail to recognize Cameroonian Indigenous data governance frameworks. We propose a decolonial framework for algorithmic justice that centers Cameroonian Indigenous data sovereignty principles while offering practical guidelines for ethical engagement with Indigenous data. This framework contributes to broader discussions on digital ethics by prioritizing relational accountability, contextual integrity, and cultural sustainability in technological development within the Cameroonian context.