Keen Dang, PhD

Struggling to benchmark your agentic system against a standard ML model? Lambert Leong's article includes 6 ways to create continuous scores from agentic outputs for a fair AUC-based evaluation.

5

to make a RAG system work well one has to tweak parameters 🎛️ Follow our AAAI  2026 paper (even better - meet Matan Orbach in Singapore this week) to learn how this could be done automatically and efficiently. Check out the Paper - Hyper-Parameter Optimization for RAG: https://lnkd.in/dFr-MNeD

8

[AI Paper] We are finally treating Memory as an "Action Space," not just a "Storage Bucket." 🧠 As AI Architects, we usually treat Long-Term Memory (LTM) and Short-Term Memory (STM) as separate engineering problems. We use Vector DBs for LTM and Context Window sliding for STM. The agent has no real control over what it remembers; it's just passive retrieval. A new paper from Alibaba Group & Wuhan University, "Agentic Memory (AgeMem)," proposes a unified framework that changes this paradigm. The Architectural Shift: Memory as a Tool: Instead of a hidden background process, memory operations (Store, Retrieve, Update, Forget) are exposed as Tool Calls in the agent's action space. The model decides when to write to its own disk. Unified Policy: They don't use a separate "Memory Controller" model. They train the main agent (via a 3-stage RL process) to manage its own context, balancing LTM and STM dynamically. Progressive RL: The training recipe is key. It starts with "Memory-Free" warmups, moves to "Passive Memory" (learning to read), and ends with "Active Memory" (learning to write/manage). The Result: It outperforms standard RAG and heuristically managed memory systems (like MemGPT) on long-horizon tasks (HotpotQA) because the agent learns what is worth remembering. If you are building autonomous agents that need to run for days or weeks, "Passive Memory" isn't enough. You need "Active Memory Management." 👇 Link to the paper in the comments. #AIArchitecture #AgenticAI #MemorySystems #ReinforcementLearning #LLM #Alibaba #Research #AgeMem

18

1 Comment

🚀 ML PM - Senior Deep Learning Research Engineer (Hedge Fund) We're looking for a rare profile at the intersection of deep learning research and portfolio management. This is not your typical ML role - it's about building the revenue-generating core of the ML/AI stack. 🔹 Applying neural networks to tick data and fast data 🔹 Leveraging reinforcement learning on time series 🔹 Constructing non-linear portfolios using trained models 🔹 Driving alpha through model-driven trading strategies You’ll be training models, deploying them in production, and directly influencing PnL. Ideally, you’re a PM who can think in tensors and trade in basis points. The bar is high. The impact is real. If this sounds like you — or someone you know — let’s talk. #MachineLearning #DeepLearning #ReinforcementLearning #QuantFinance #PortfolioManagement #AI #MLPM #Hiring

18

1 Comment

As always, zero hype and just clear, practical insights from Andrew Ng on agentic systems, autonomy levels, evals, voice apps, AI coding, MCP, agent-to-agent, and why coding still matters. Truly useful advice on building smarter AI systems. https://lnkd.in/gPCFMiJM

15

🪄With distributed parallelism, #MindSpore helps you scale #AI training through data, model, optimizer, and pipeline parallelism—built to handle large models and datasets efficiently and smoothly. Learn the how-to in the tutorials 👇 https://lnkd.in/guU_Ez5W #MindSporeTutorials

1

✈️ Just arrived to AAAI 2026 in Singapore to present our RAG hyper-parameter optimization paper! 🤔 What is our work about ? Its about squeezing the best performance out of a RAG system with hundreds of possible configurations. But how? With so many components in a modern RAG system… Where do you even start? Which hyper‑parameters matter most? And how can you avoid endless trial‑and‑error? That’s exactly the challenge we tackled in work .. 🔍 What did we do? A systematic, realistic evaluation of hyper‑parameter optimization (HPO) methods across: • 5 diverse RAG datasets • 162 distinct RAG configurations • 3 evaluation metrics • Multiple HPO algorithms And we even established upper bounds via full grid search, so we can see how close contemporary algorithms get to the best possible solution. 💡 What did we learn? What do we recommend to RAG practitioners? - Start simple, a random search over a handful of configurations delivers surprisingly good gains. - Go smarter, with a greedy, model‑first optimization, it will find high‑quality settings with far less work. - Scale efficiently, for large datasets, follow our development‑set sampling strategy to cut compute cost without sacrificing quality. 📦 Want to learn more, or experiment yourself? Check out the Paper: https://lnkd.in/dFr-MNeD And our open‑sourced full grid results: https://lnkd.in/dmrht35w Build upon our results to further explore new RAG metrics, optimization algorithms and optimization setups. Curious to dive deeper? lets chat.. #AAAI2026 #RAG #HPO #IBMResearch Ohad Eytan Benjamin Sznajder Ariel Gera Odellia Boni Yoav Kantor Gal Bloch Omri Levy Hadas Abraham Nitzan Barzilay Eyal Shnarch Michael Factor Shila Ofek-Koifman Paula Ta-Shma Assaf Toledo  

46

1 Comment

I'm glad to notice that our resesrch team just published a big piece of recent works on 'Verifiable PIR' . Title: Publicly Verifiable Private Information Retrieval Protocols Based on Function Secret Sharing Authors: Lin Zhu, Lingwei Kong, Xin Ning, Xiaoyang Qu, Jianzong Wang Comments: Accepted by the 21st International Conference on Information Security  and Cryptology (Inscrypt2025 Link: http://t.cn/AXhs9AWA Abstract: Private Information Retrieval (PIR) is a fundamental cryptographic primitive that enables users to retrieve data from a database without revealing which item is being accessed, thereby preserving query privacy. However, PIR protocols also face the challenge of result verifiability, as users expect the reconstructed data to be trustworthy and authentic. In this work, we propose two effective constructions of publicly verifiable PIR (PVPIR) in the multi-server setting, which achieve query privacy, correctness, and verifiability simultaneously. We further present three concrete instantiations based on these constructions. For the point query, our protocol introduces minimal computational overhead and achieves strong verifiability guarantees with significantly lower communication costs compared to existing Merkle tree-based approaches. For the predicate query, the communication complexity of our scheme remains stable as the database size increases, demonstrating strong scalability and suitability for large-scale private query applications.

1

How do you scale inference and RL on petabytes of autonomy data 🤖? At Ray Summit 2025, Yi Sheng Ong and Eric Higgins detailed how Applied Intuition leverages Ray on K8s to orchestrate inference and reinforcement learning at petabyte scale . Full talk ➡️ https://lnkd.in/grsQfNpr

144

2 Comments

Scicom MSC Berhad AI open source the biggest agentic dataset distilled from combination of big open source models, up to 142k rows with 15 subcategories dataset, the pipeline automatically generate multi-functions, multi-turn, multi-lingual, verify user intents, multi-tools, generate mock function outputs including handling various scenarios, 1. Single and extended function generated based on domains and complexities, including examples to trigger the function. 2. Use (1) to generate ms-en, ta and zh examples to trigger the function. 3. Generate multifunctions up to 20 functions based on 301 domains, 3 complexities, and 26 workflows. Because we do BPO for call centre, we also generate premium multifunctions specific for TaaS. 4. Use (3) to generate multi-turns user assistant responds with auto tools at least 5 turns up to 10 turns. 5. Use (3) to generate multi-turns user assistant responds with auto tools including mid conversation language switching between ms, en, ta and zh with extra, - Out-of-context handling, user asks irrelevant questions - Error/failure scenarios, API failures, invalid inputs, rate limits - Multi-intent handling, user asks multiple things at once - Correction/backtracking, user changes mind mid-conversation - Clarification dialogs, assistant asks for missing information - Frustrated/confused user scenarios - Implicit context references, "use the same address", "that order" - Chit-chat handling, greetings, thanks, small talk - Confirmation flows, assistant confirms before destructive actions - Partial/incomplete information provided incrementally Special thanks to ScitiX for the GPU access!

122

2 Comments

Building Near Real-Time Bus Tracking App with Aiven Kafka Tan Ze Han Dylan UPM and I built near real-time pipeline for tracking Rapid KL Bus, powered by Kafka on Aiven #FreeKafka and a Lambda Architecture pipeline. 📍 Kafka (Aiven) for real-time ingestion 🤖 Spark Streaming + Kafka Connector for processing 🚦 AWS S3 + Iceberg for the lakehouse  ⚡Cache & serve near real-time state with Redis ⚙️ Kubernetes + Prefect for orchestration 🗺️ LeafletJS + Next.js for live map visualization This will be our submission for Kafka competition hosted by Aiven. Enjoy the Demo!

225

12 Comments

Reimagining Science with AI Agents — with Ashton Teng & Quinn Leng of Kepler AI Z Potentials sat down with Ashton Teng and Quinn Leng, co-founders of Kepler AI, who are building the “central nervous system” for research organizations. In a field where sequencing costs have collapsed and data has exploded by over a trillion-fold, scientists still wait weeks for analysis. Kepler AI wants to change that with AI Agents that act more like end-to-end scientists than tools. From casual conversations while skiing and diving to simultaneously quitting their jobs to start a company, Ashton and Quinn embody the founder’s leap from curiosity to conviction—proving that timing and trust can ignite breakthroughs. 👉 Read the full article here: https://lnkd.in/eijKe3xR 💡 In this exclusive interview, Ashton & Quinn share: “After an experiment is done, scientists often wait days or weeks to get bioinformatics results. That slows the entire cycle of iteration dramatically.” “What we’re building is a form of human–AI interaction science has never seen before.” “General platforms can’t solve the ‘last mile’ problem. They won’t dive into enterprises and integrate with workflows. That’s where Kepler stands apart.” “By optimizing for biology-specific needs, we outperformed GPT-5, Claude, and others on the authoritative bioinformatics benchmark Bixbench, with 33.4% accuracy.” “In the past, big pharma had little interest in startups. Now the window has opened. AI transformation is becoming a key measure of competitiveness.” “Our ambition goes beyond biology. Materials, climate, agriculture—anywhere R&D is needed, AI will be essential.” 🧠 Read the full story from Z Potentials to see how Kepler AI is shaping the future of AI in science.

18

In collaborative machine learning (CML), early contributors take on more risk and encourage participation from wait-and-see parties, but existing CML frameworks treat every party as if they join simultaneously. The #NeurIPS2025 paper of Jiangwei Chen, Nguyen Pham, Rachael Sim, Arun Verma, Zhaoxuan Wu, Chuan Sheng Foo, and Bryan Kian Hsiang Low proposes a time-aware CML framework that rewards parties not just for what they contribute, but when they contribute: •⁠ ⁠8 incentives extending fairness with time-aware notions that incentivize early participation while preventing low-quality early submissions. •⁠ ⁠2 reward distribution methods that theoretically satisfy all incentives and recover the #ShapleyValue in time-agnostic cases. •⁠ ⁠A practical recipe for constructing valuation functions and reward realization, with empirical validation on synthetic and real-world data. Find our more at our posters! #NeurIPS2025 Thu, Dec 4, Exhibit Hall C, D, E #1104 #EurIPS2025 Wed, Dec 3, Poster Stand #56 Paper: https://lnkd.in/g_yM6j39 #FederatedLearning

50

As someone deeply invested in agentic systems and reinforcement learning, I find the Chain-of-Agents (CoA) approach both refreshing and practical, it takes the chaos of multi-agent orchestration and distills it into something elegant: a single model with native agent-like behavior. Here’s what I see: Traditional multi-agent setups have been great for modularity and reasoning, but they come with a high tax, too much inter-agent communication, redundant memory syncing, and unnecessary tool calls. What CoA proposes is radical in its simplicity: distill the behaviors of successful multi-agent runs into one model, then train that model to act as a team, internally. They start with multi-agent distillation: successful traces from existing orchestrators like OAgents are transformed into CoA-style traces, complete with planning steps, tool usage, reflections, and execution paths. But they don’t stop at replay, they filter hard cases and focus on high-quality trajectories that demonstrate useful agent behaviors like tool efficiency and coherent planning. This becomes the foundation of their SFT phase. Once that base is established, Agentic RL kicks in. The model is tasked with solving complex reasoning problems and importantly, ones where tools make or break the outcome (web QA, coding, math). Rewarding is done via exact-match or LLM-as-Judge strategies, making it possible to scale reward signals across noisy tasks. This setup tunes the model to reason reflectively, use tools judiciously, and respond with stability under diverse challenges. The kicker: by collapsing the roles of multiple agents into a single coherent model, CoA delivers an 84.6% reduction in inference cost. It’s a major unlock for production-grade agent systems where latency, memory, and tool limits matter. What excites me most is how CoA generalizes the spirit of ReAct/TIR, but upgrades it, dynamically invoking roles, using tools only when needed, and preserving a unified memory state that avoids the “who’s speaking now?” confusion. To me, Chain-of-Agents is a promising new frontier: a single-model, multi-mind architecture that captures the power of coordination while optimizing for real-world deployment constraints. And that’s where agentic AI needs to go. https://lnkd.in/exTxvsDD

63

2 Comments

Professor Jian Sun Joins AltDataTech's Academic Committee! We are excited to announce that Professor Jian Sun from Singapore Management University has joined AltDataTech's Academic Committee. This strategic appointment brings world-class academic expertise to our R&D and innovation efforts in financial technology, while further strengthening the bridge between academic research and industry practice. Professor Sun holds a PhD in Finance from MIT Sloan School of Management and currently serves in both the Lee Kong Chian School of Business and the School of Economics at Singapore Management University. His research focuses on market microstructure and market design, with influential work published in top-tier journals such as American Economic Review, Journal of Finance, and Review of Financial Studies. His academic excellence has earned him multiple accolades, including the Best PhD Paper Award at the 2022 Annual Conference on Digital Economics, the Best Paper Award at the 7th Vietnam International Conference in Finance (Asset Pricing, Investment, and FinTech), and the WRDS-TME Best Paper Award. He also serves as an ad-hoc reviewer for leading journals including Journal of Finance, Review of Financial Studies, Management Science, Journal of Corporate Finance, and Journal of Economic Dynamics and Control. As a pioneer in applying satellite data to enterprise-level services in China, AltDataTech continues to advance multi-modal data capabilities through cutting-edge remote sensing analytics and big-data cloud technologies. Our goal is to deliver high-value data products and integrated alternative-data solutions for investment and commercial decision-making. During the Committee meeting, members unanimously agreed that Professor Sun's academic depth in finance and economics aligns closely with AltDataTech's development strategy, significantly enhancing the Committee's leadership in the FinTech domain. Professor Sun shared that he looks forward to contributing actively through academic guidance, technology-innovation reviews, and standards development—supporting AltDataTech in building a robust ecosystem that connects academic insights with technological breakthroughs and real-world applications. His appointment represents a significant step in strengthening our academic foundation and innovation capabilities, bringing the expertise needed to refine our R&D roadmap, enhance the scientific rigor of our solutions, and ensure our technologies stay aligned with evolving market needs. Moving forward, AltDataTech will continue to strengthen collaboration between academia and industry through the Academic Committee platform, advancing innovation in financial technology and supporting high-quality growth across the industry. #AltDataTech #FinTech #AcademicCommittee #FinancialResearch #AlternativeData #SatelliteData #DataScience #IndustryResearchCollaboration #SingaporeManagementUniversity 

12

What began as a research effort for Prof Kwok Yan Lam to make sense of vast volumes of unstructured data has evolved into an AI venture with global enterprise impact. With support from NTUitive’s Gap Fund for POCs, this research journey gave rise to Tau Express and Tau Intelligence, translating academic innovation into enterprise-ready solutions adopted across multiple industries. With over S$5 million raised and international clients on board, TAU is scaling beyond Singapore and preparing for its Series A. From idea to impact - the 19th Grant call for NTUitive’s Gap Fund is now open. More details here: https://lnkd.in/gV7XhJ22

74

1 Comment