AI Hub Weekly: Research & Industry – Week Ending 14 November 2025

This week, agentic AI and advanced reasoning workflows continued to move from theory into practice. On the research side, we saw new methods for dynamic tool/action selection, fine-grained self-refinement, and simulating misinformation spread with LLM agents. On the industry side, there were major moves in AI infrastructure, agentic retail, and public-sector automation in Singapore and India.

For Singapore and regional teams, the through-line is clear: AI is no longer just about model choice. It’s about how agents reason, act, and are orchestrated across real systems—with governance, data, and ROI front and centre.

1. Research Highlights – Reasoning, Evaluation & Agentic Simulations

1.1 DynaAct: Dynamic Action Spaces for LLM Reasoning

A NeurIPS 2025 paper, DynaAct: Large Language Model Reasoning with Dynamic Action Spaces, proposes a framework where an LLM doesn’t reason over a fixed set of tools or actions, but builds a compact, task-specific action space on the fly. Instead of manually defining actions or searching over a huge unstructured space, DynaAct learns a submodular selection function that picks a diverse, high-utility subset of actions for each step in a reasoning process. Benchmarks show better performance without major latency penalties.

Why it matters

• For agentic workflows (e.g. multi-tool copilots), action selection is the new prompt engineering. This is an early blueprint for doing that systematically instead of heuristically.

• Dynamic action spaces fit well with enterprise environments where available systems, APIs, and constraints differ by context or department.

• For teams in SG/SEA building orchestration layers on Azure, Vertex, or AWS, this points towards “policy-driven action selection” as a core capability, not an afterthought.

1.2 SSR: Socratic Self-Refine for LLM Reasoning

Salesforce AI Research introduced SSR: Socratic Self-Refine, a test-time framework that breaks an LLM’s answer into verifiable (sub-question, sub-answer) pairs and then re-solves and cross-checks each step. The system assigns confidence at the step level, identifies weak reasoning segments, and iteratively refines them, improving accuracy across multiple reasoning benchmarks while keeping the model as a black box.

Why it matters

• This shifts self-correction from “try again” to structured debugging of the chain-of-thought, which is closer to how human reviewers work.

• For regulated domains (finance, healthcare, public sector), step-level confidence and traceability are crucial for audits, approvals, and human-in-the-loop governance.

• It’s a strong signal that reasoning quality frameworks—not just bigger models—will become a differentiator in production agentic systems.

1.3 Simulating Misinformation with LLM Persona Agents

A new paper, Simulating Misinformation Propagation in Social Networks using Large Language Models, models LLM-driven persona agents (with different biases and ideologies) as nodes in a social network. News items are passed through these agents, each rewriting content; an auditor agent uses Q&A to track how factual content drifts into misinformation or propaganda. The authors define metrics like a “misinformation index” and show how identity-driven personas can accelerate distortion.

Why it matters

• This is a practical example of agentic simulation: using LLM personas to stress-test information ecosystems before real-world deployment (e.g. comms, marketing, political content).

• Regulators and risk teams can borrow this pattern to simulate how internal or external messages might be distorted when re-shared, remixed, or summarised by downstream AI systems.

• For Singapore’s emphasis on trust and safety in AI, such frameworks hint at future “red-team via synthetic networks” standards.

1.4 LLM-as-a-Grader in Real Classrooms

The paper LLM-as-a-Grader: Practical Insights from Large Language Model for Short-Answer and Report Evaluation looks at how LLMs perform when grading short answers and reports in real educational settings, beyond toy benchmarks. It evaluates alignment with human markers and highlights where LLMs perform well, where they systematically diverge, and what calibration strategies help.

Why it matters

• This moves grading use-cases from “demo” to evidence-backed deployment, especially relevant to higher education and corporate training providers.

• For institutions in Singapore and the region, it offers a starting point to design moderation workflows, where AI provides a first pass and humans handle edge cases and appeals.

• Longer-term, it supports more sophisticated AI-powered assessment and feedback loops—from MOOCs to WSQ-style competency programmes.

2. Industry & Policy Moves – Infra, Agentic Retail & Public Sector

2.1 Anthropic’s US$50B Compute Bet & Microsoft’s Fairwater 2

Anthropic announced a US$50 billion investment in AI infrastructure, including new data centres in Texas and New York, in partnership with Fluidstack. On the same day, Microsoft revealed Fairwater 2, a new Atlanta data centre that will form a “massive supercomputer” with an existing facility in Wisconsin—powered by hundreds of thousands of Nvidia chips and shared across Microsoft’s own workloads, OpenAI, and other AI developers.

Why it matters

• The infrastructure arms race is accelerating: both frontier developers and cloud providers are locking in multi-year capacity to support model training, serving, and agentic workloads.

• For enterprises in SG/SEA, this likely translates into more powerful managed services, but also higher scrutiny on energy use, latency, and data locality.

• Strategically, it reinforces the idea that most organisations won’t run frontier models themselves—they’ll orchestrate them via hyperscaler stacks and focus on data, workflows, and governance.

2.2 Deloitte India & AWS Launch Agentic AI Lab

Deloitte India and AWS signed a multi-year strategic collaboration that includes setting up an AWS Agentic AI Lab and Centre of Excellence, and offering Deloitte’s digital and agentic AI solutions via AWS Marketplace. The focus is on cloud- and AI-led transformation across India and broader APJ, with Deloitte becoming the first global SI in APJ with this kind of dedicated collaboration.

Why it matters

• This is a concrete example of “agentic AI” being productised at scale by a major SI + hyperscaler pair, not just discussed in whitepapers.

• Expect similar labs and co-branded offerings to emerge in Singapore—especially around verticalised solutions (FSI, public sector, healthcare).

• For smaller consultancies (like AI Hub SG), the opportunity is to specialise in last-mile integration, training, and change management on top of these platform-level offerings.

2.3 Agent-to-Agent Claims on Google Cloud (Hexaware)

Hexaware launched ParaClaims and Intelligent Product Factory on Google Cloud—cloud-native insurance platforms using AI and automation. ParaClaims uses an Agent-to-Agent Protocol where AI agents monitor real-time data from sources like IMD, NOAA, satellites, and Google Earth Engine, then autonomously handle trigger detection, data validation, and claims settlement for parametric insurance. Settlement times drop from weeks to hours, with all data stored in BigQuery for transparency.

Why it matters

• This is a live production pattern for multi-agent, event-driven workflows in a regulated industry (insurance), not just a lab demo.

• It showcases how to combine external data streams, cloud analytics, and agentic orchestration to deliver clear business KPIs (time-to-settlement, fraud checks, auditability).

• For SG insurers and MAS-regulated entities, ParaClaims-style designs provide a template for AI-enabled, audit-ready decisioning that regulators can understand.

2.4 Agentic AI in Retail: Bain’s View

Bain & Company released a press note on Agentic AI in retail, arguing that agentic systems—agents that can independently search, compare, and transact across multiple channels—are poised to reshape shopping and operations, even though roughly 50% of consumers remain cautious about fully autonomous purchases. The piece highlights early use-cases like autonomous replenishment, personalised cross-channel journeys, and AI agents negotiating limited offers or bundles.

Why it matters

• Retail is emerging as a flagship domain for consumer-facing agentic workflows—from “shopping copilots” to behind-the-scenes optimisation.

• The consumer caution signals that adoption will hinge on trust, transparency, and control (e.g. approvals, spending limits, clear explanations).

• For APAC retailers, this opens space for hybrid models: human-supervised agents that handle the tedious parts of discovery, pricing, and logistics while preserving brand and UX control.

2.5 Singapore Spotlight: Agentic Automation in Public Sector & Healthcare

OpenGov Asia’s feature “From Burden to Breakthrough: How Agentic Automation is Transforming Singapore’s Public Sector & Healthcare” (with UiPath) summarises a Breakfast Insights event at Sheraton Towers on 13 November. It describes how Singapore’s Ministry of Health and agencies like VITAL are moving beyond classic RPA into agentic automation—AI agents orchestrating data, robots, and humans across complex workflows, funded in part by a S$200M Health Innovation Fund. Use-cases highlighted include automated CV evaluation for recruitment and ambient listening for clinical documentation, with strong emphasis on governance, trust, and the “agentic workforce” model where agents think, robots do, and people decide.

Why it matters

• This is one of the clearest public examples of agentic AI being embedded into a national public-sector transformation agenda, not just a pilot.

• The framing around “agentic workforce” and “trust managers” aligns closely with emerging global narratives (e.g. Microsoft’s Agentic Era, UiPath’s orchestration vision) but anchored in Singapore’s context. OpenGov Asia

• For SG organisations, it signals that agentic AI is now a policy-aligned direction, not a fringe experiment—especially in healthcare, HR, and shared services.

3. What This Means for Teams in Singapore & the Region

Across both research and industry this week, a few themes stand out:

• Orchestration is the battleground – Dynamic action selection (DynaAct), step-level refinement (SSR), and multi-agent protocols (ParaClaims, public-sector automation) all converge on one idea: value comes from how you chain tools, agents, and humans together.

• Evaluation and governance are getting more structured – From SSR’s step-wise confidence to LLM-as-a-Grader and misinformation simulations, we’re seeing concrete frameworks for probing, grading, and red-teaming agentic systems before they go live.

• APAC—and Singapore specifically—is moving fast on agentic adoption – Deloitte–AWS, Hexaware–Google Cloud, and the OpenGov/UiPath work in SG show that regional implementations are keeping pace with US and Europe, often with stronger emphasis on governance and shared services.

For AI Hub Singapore, these are exactly the fronts we’re working on with clients:

• Designing agentic workflows that sit on top of Azure, Vertex, or AWS, with clear action policies and human-in-the-loop checkpoints.

• Building evaluation harnesses that borrow ideas like Socratic self-refinement and agent-based simulations for internal governance.

• Helping public-sector and enterprise teams translate global patterns (Anthropic, Bain, UiPath) into locally compliant, operationally realistic deployments.

If you’d like to explore how these trends could map into your organisation’s AI roadmap—especially in Singapore and the wider region—this is a good moment to start the conversation, while the playbooks are still being written.