Ziliang (Daniel) Lin

Ever wondered what agentic collusion looks like in practice - not theory? Following on from this week’s earlier post... here’s how researchers at King’s College London (KCL) actually built their multi-agent simulation to test it 👇 Architecture: - A multi-agent reinforcement learning market built with RLlib and PettingZoo Learning setup: - PPO (policy-gradient) and DDPG (actor-critic) for continuous control - Centralised training, decentralised execution - agents share gradients during training but act independently in the market - Reward shaping: baseline = pure PnL; hybrid = PnL + inventory penalty + market-stability term Compute: - ~8 × A100 GPUs - >10 million training steps with parallel rollouts via Ray - Convergence to stable quoting policies after ~500 GPU hours Findings: - Pure-PnL agents learned aggressive quoting behaviour - Hybrid-reward agents converged to smoother, more stable policies - In game-theory terms, a correlated equilibrium emerged - spontaneously 📄 Wang, Ventre & Polukarov (ICAIF 2025) - Multi-Agent Reinforcement Learning for Market Making: Competition without Collusion ...

1

In working with LLMs, a consistent source of performance gaps comes from limitations in memory and learning over time. In my own work, this gap shows up repeatedly: amazing performance in isolated tasks, followed by brittleness once interactions extend over time or require the system to build on prior experience. Memory and self-learning may not be the only way to frame the problem, but they are currently useful lenses for understanding where today’s LLM architectures break down. What makes this gap interesting is the contrast: these systems can be super-human in narrow capabilities, yet frustratingly wrong when even modest persistence or learning over time is expected. You don’t need to be a cognitive scientist to notice that the human brain is not just a very large context window. From a practical standpoint, it can be useful to think architecturally in terms of adding or improving memory and learning—not as the only way to frame the problem or its solution, but as architecturally useful working models. Seen this way, the issue is not access to more text via larger context windows, but the absence of persistent, updateable processes—mechanisms that govern what is retained, revised, and abstracted across interactions. I recently looked at Hindsight (link to repo in the comments), an open approach to agent memory that treats memory as more than a flat retrieval layer. Instead of storing everything in a single vector space, it introduces explicit memory categories that resemble familiar cognitive distinctions: 🔹World knowledge — relatively stable facts 🔹Experiences — what the agent itself has done and observed 🔹Opinions — beliefs that carry confidence or uncertainty 🔹Observations — higher-level models formed by reflecting across prior memory What stands out is not just the taxonomy, but the separation of memory operations: 🔹Retain information intentionally 🔹Recall relevant past context when needed 🔹Reflect to synthesize new insights from accumulated history This structure makes it easier to reason about how an agent’s knowledge evolves—particularly how beliefs change, how experience accumulates, and how abstractions form over longer horizons. Many existing RAG-style systems struggle here once interactions extend beyond short-term context. The approach has also been evaluated on LongMemEval, where it reports leading accuracy among open systems, with independent verification by external research groups. Whether or not this specific architecture becomes standard, it points toward a direction that feels necessary: systems that can accumulate experience, revise beliefs, and build internal models over time. That seems like a prerequisite for anything we would reasonably call self-learning. How are you thinking about memory and persistent learning in your own AI or LLM projects? And how much improvement do you think this could bring to LLM or AI systems’ performance? #AIAgents #LLMs #AgentMemory #MachineLearning #AIArchitecture

2

3 Comments

In the early 1980s, Gerry Bamberger, a recent graduate of Columbia University's computer science program, joined Morgan Stanley. His task was to write simple software that updated trading results. He noticed that traders often executed large “pair trades”, for example, buying millions in Coca-Cola shares while simultaneously selling the same amount in Pepsi. As Bamberger observed these trades, he realized that the spread between the two stocks would temporarily diverge due to market impact, then revert to the mean, a pattern he saw as an opportunity. With a $500K budget from the bank, he built a computer-driven statistical arbitrage system and eventually grew it into a $30M strategy. I wrote an article on how to apply cointegration with Python to uncover such opportunities, feel free to check it out. https://lnkd.in/dQAjVZCx

8

Ever wondered why a single AI model struggles to analyze stocks like a professional Wall Street analyst? 🤔 Or why traditional LLMs hallucinate financial metrics when asked to provide comprehensive investment analysis? The answer lies in understanding the complexity overload problem—and that's exactly what multi-agent collaboration solves. Recently, I looked into AI in finance and stock trading and then built RiskNavigator AI using Google's Agent Development Kit (ADK), orchestrating 5 specialized agents to deliver institutional-grade stock analysis and investment decision with risk assessment. I implemented the complete sequential agent pattern—from data gathering through MCP to risk assessment—testing it on Google Cloud Run's serverless infrastructure. The results were eye-opening: - 20% improvement in output consistency compared to single-LLM approaches - 85% reduction in hallucinations through specialized agent roles - <60 second end-to-end analysis with 99.9% uptime - Production-ready deployment on serverless infrastructure with zero GPU costs But the most surprising finding? State-based communication (shared memory) should be your default choice for sequential agent workflows, not message-passing as many assume. TL;DR I wrote a comprehensive blog breaking down: ✅ What multi-agent systems are and why they outperform monolithic LLMs for complex tasks ✅ The 12 fundamental agent design patterns—and how to choose the right one ✅ How Model Context Protocol (MCP) connects agents to 60+ financial APIs securely ✅ Agent-to-Agent Communication (A2A) implementation using Google's AgentTool pattern ✅ Different types of agent memory (short-term, long-term, shared, episodic, semantic) ✅ Complete code implementation with all 5 specialized agents (Data, Trading, Execution, Risk, Summary) ✅ Production deployment on Google Cloud Run with FastAPI + Docker ✅ Real experimental results showing the power of agent specialization Whether you're building financial AI systems or just curious about multi-agent architectures, this guide takes you from fundamentals to production deployment with real code and architecture diagrams. 💻 GitHub: https://lnkd.in/gkewN_7y Special thanks to Google Cloud for the incredible "5-Day AI Agents Intensive Course" that inspired this deep dive into multi-agent systems, and to Anthropic for introducing the Model Context Protocol that makes tool integration so elegant! 🙏 What's your experience with multi-agent systems? Have you used Google ADK, LangGraph, or CrewAI in production? I'd love to hear your insights in the comments! 👇 📖 Read the full blog here: https://lnkd.in/gGYw65jP #AI #MultiAgentSystems #FinancialAI #GoogleCloud #AgenticAI #MCP #ProductionAI #MachineLearning #MLOPS

34

2 Comments

A fantastic deep dive by William Arnold (NVIDIA) into Dynamo, an open-source system that rethinks high-performance LLM inference. If you're building or hosting models, this talk is a must-watch. Learn about smart cache routing, disaggregated serving, and real production-ready design. At AER Labs, we’re passionate about efficient AI + system-level innovation. Do subscribe to our channel for more such expert sessions. Our Batch V1 applications are now open. If you want to contribute to open-source AI research (part-time or full-time) or building a cool AI project, do check it out. Free and open to all. Fill this form to connect: https://lnkd.in/dJG8bXtt

12

Many of us were holding our breath for Nvidia’s Q3 earnings call last week. It’s no secret that resource availability impacts downstream products, so how do we ensure that the “growth” of AI doesn’t outpace its supply stream? Last weekend, I attended #PlanetAction at Massachusetts Institute of Technology in a quest for some answers. I learned that society is inextricable from its environment, and economy laces between the two – connecting supply to demand and tugging at both landscapes. That transactional lattice ought to be pliable. When a material resource runs low, good products reconnect to new inputs to maintain consistent outputs. Likewise, if production pipelines adversely impact surrounding communities, alternate methods ought to be tested to regenerate social decay. These methods need no discovery; They need adoption ahead of the endpoints they were designed to avoid. Nvidia is safe for now. Do we breathe a sigh of relief? Or, do we push for an economy that feels a little less like a game of jenga?

32

3 Comments

Most existing #lean4 datasets contain only correct proofs. Models learn error correction with RL, that's expensive. With UW Math AI lab we release a dataset of 260k erroneous Lean proofs with - compiler feedback - reasoning trace - corrected proof Improvements in Error Correction: - Goedel 8B: 2x - Kimina 8B: 3x Paper: https://lnkd.in/gaybt4bd

118

2 Comments

Excited to share that our team’s paper on aspect-guided review summarization has been accepted to EMNLP 2025’s industry track! 📈🧠🤓 In this paper, we describe a live system at Wayfair that leverages LLMs alongside classical NLP techniques to summarize tens of millions of customer reviews, making it easier for shoppers to discover what matters most to them on site. To further support the research community, we’re also open-sourcing a dataset of ~12M anonymized reviews and their extracted aspects which we hope accelerates progress in aspect extraction, sentiment analysis, and summarization tasks. This milestone reflects the hard work of many folks across Wayfair, including our co-authors Ilya Boytsov, Misha Balyasin, Joe Walt, Caitlin Eusden, Marie-Claire Rochat, and Margaret Pierson. https://lnkd.in/ezJMi3g7 #NLP #EMNLP #EMNLP2025 #LLMs #Summarization #ABSA #OpenSource #Data #Dataset #ecommerce #wayfair #machinelearning #ai #publication

74

3 Comments