Reasoning
Frontier
Research direction bundle

AI for scarce labor, physical systems, autonomous organizations, and human regeneration.

This page collects the public research directions for Reasoning Frontier. The aim is to turn broad questions into falsifiable research programs, reproducible datasets, and publishable papers. AI safety is a Foundation track supporting all four directions.

Scope note: This is a research agenda, not medical advice, investment advice, clinical guidance, or a claim of deployed capability. Biological research topics must be handled through public data, literature review, and regulated collaborations only.
01 / Society & Systems · Infrastructure

Japan 2040: AI as infrastructure for a labor-constrained economy

Japan offers a clear test case for a deeper question: can AI help a mature society preserve economic and service capacity as the available human workforce declines?

Problem frame

Demographic pressure, labor shortages, regional depopulation, infrastructure maintenance, care demand, and productivity requirements.

Research frame

Estimate sector-level labor-capacity gaps and model where AI-enabled productivity gains could offset those gaps.

Proposed metric

AI Labor Offset Ratio: AI-enabled labor-equivalent capacity gain divided by projected labor-capacity loss.

Outputs

Sector maps, opportunity rankings, productivity scenarios, and policy/business implications for 2026–2040.

02 / Society & Systems · Physical AI

Physical AI and robotics for scarce labor

Japan's robotics, manufacturing, sensors, motors, vehicles, machinery, and precision-engineering strengths make physical AI a serious national-scale frontier.

Research questionMethodExpected output
Which physical tasks are technically and economically deployable?Task taxonomy across manufacturing, logistics, care, construction, agriculture, and infrastructure.Physical AI Deployability Index.
Where do robot foundation models improve over fixed automation?Compare structured versus semi-structured tasks, task variability, and perception/manipulation needs.Deployment roadmap by task family.
When is a robot economically rational?Total-cost-of-ownership analysis versus labor cost, vacancy cost, safety cost, and service failure cost.Robot ROI thresholds by sector.
03 / Society & Systems · Autonomous organizations

The autonomous SME: AI agents as operating capacity

Small and medium-sized firms often cannot hire specialized staff for every function. AI agents may become a new layer of operating capacity when governed carefully.

Functions

Accounting, HR, procurement, sales, customer support, IT, compliance, knowledge management, and cybersecurity.

Model

Output = human workforce + agent capacity × coordination effectiveness.

Governance

Human authorization tiers: read, recommend, draft, execute with approval, execute autonomously.

Risk controls

Cybersecurity, privacy, hallucination, unauthorized action, auditability, vendor lock-in, and responsibility assignment.

04 / Science · Cellular aging

AI-guided cellular rejuvenation and longevity science

This direction investigates how AI can support the study of cellular aging, senescence, iPS cells, partial reprogramming, protein design, and public aging datasets.

Scientific anchor

Kyoto University's CiRA is a major center for iPS cell research, reprogramming, differentiation, clinical applications, fundamental cell technology, and ethics.

Senescence anchor

NIH SenNet maps senescent cells across tissues and the lifespan, providing a foundation for studying age-related biology.

AI anchor

Protein design systems and AlphaFold-style biomolecular modeling can help generate hypotheses, but wet-lab validation remains mandatory.

Safety anchor

Partial reprogramming and senescence interventions require careful attention to cancer risk, tissue identity, off-target effects, and clinical regulation.

Research boundary: This track begins with literature review, public datasets, computational analysis, and collaboration mapping. It excludes self-experimentation, unregulated treatment protocols, or instructions for biological manipulation.
05 / Foundation

Foundation · AI safety · RF-SAFETY-001

AI safety measured through evaluation, not assumption

Measure how AI agents behave under controlled safety tests before they are trusted with operational work.

Main question

How can tool-using AI agents be evaluated and controlled so that consequential actions stay within approved limits, and how can that evidence be made reproducible?

Method

Adversarial and prompt-injection evaluation, checks against evaluation gaming, least-privilege agent architectures, independent action monitoring, and traceable safety-case records.

Planned contribution

A reproducible evaluation and control protocol for tool-using agents, published as a dated report with scope, methods, results and limitations.

Potential venues

NeurIPS (Datasets and Benchmarks), ICML, IEEE SaTML, ACM FAccT, the Technical AI Safety Conference (TAIS), or JSAI.

Status

Active research track. No safety claims are made until a dated evaluation report is published.

Read the AI safety paper program
06 / Sources

Selected public grounding sources

These sources are starting points for the research program. They should be expanded into formal bibliographies for papers.