Yi Ting (Serene) LIU

𝐏𝐫𝐨𝐦𝐩𝐭𝐢𝐧𝐠 𝐓𝐞𝐜𝐡𝐧𝐢𝐪𝐮𝐞𝐬 𝐓𝐡𝐚𝐭 𝐈𝐦𝐩𝐫𝐨𝐯𝐞 𝐋𝐋𝐌 𝐑𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 Working with large language models in real-world systems has shown me that the way you prompt is just as important as the model you use. Here are three techniques that have made a real difference in my projects: ➤ Chain of Thought (CoT) Breaks problems into steps. Helps models reason like humans. ➤ Majority Voting over CoT Runs multiple CoT prompts. Picks the most common answer for more reliability. ➤ Tree of Thought (ToT) Explores several reasoning paths. Keeps what works and drops what doesn’t. Great for complex decisions. These methods have helped me get more consistent and transparent results in everything from intelligent agents to AI-powered fintech tools. If you're experimenting with LLMs or building AI products, these techniques are worth trying.

151

19 Comments

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

Excited to share my team's latest work from Google Cloud AI Research (https://lnkd.in/gmejMnn7) on Deep Research (DR) agents! The core challenge with DR agents isn't just analysis—it's generating accurate, coherent, long-form reports. Today, we're introducing Test-Time Diffusion Deep Researcher (TTD-DR), a new framework that fundamentally rethinks this process. Instead of a rigid, linear workflow, TTD-DR mimics human cognition. Here’s how: 🧠 1. Inspired by Humans: It starts with a preliminary draft, then enters iterative cycles of search, reasoning, and revision. 🔄 2. Iterative Refinement (The "Diffusion"): The framework uses a novel "denoising with retrieval" mechanism. The evolving draft itself guides the search for new information, which is then used to refine the draft, step-by-step. 📈 3. Enhanced Quality: A component-wise self-evolution algorithm enhances each part of the workflow, from planning to synthesis. This combined approach achieves state-of-the-art results, including a 74.5% win rate in long-form report generation on the DeepConsult benchmark. This work shows a shift toward iterative, "draft-first" agent design to unlock advanced performance for demanding research tasks. A product version is available on Google Agentspace: https://lnkd.in/g_2QjFgw Read the details: - Google Research Blog: https://lnkd.in/g4XGpqMS - VentureBeat Coverage: https://lnkd.in/gjb9h4yg - Paper: https://lnkd.in/gCpXjd_d Congrats to the incredible team: Rujun Han, Yanfei C., Zoey CuiZhu, Lesly Miculicich, Guan Sun, Yuanjun Bi, Weiming Wen, Hui Wan, Chunfeng Wen, Solène Maître, George Lee, Vishy Tirumalashetty, Yuan (Emily) Xue, Zizhao Zhang, Salem Haykal, Burak Gokturk, Tomas Pfister, and Chen-Yu Lee. #ai #machinelearning #deeplearning #llm #DeepResearch #TestTimeScaling #Diffusion PS. Google Cloud AI Research is hiring both full-time researchers and interns with excellent LLM publication records. Send me your cv directly to [email protected].

112

5 Comments

Fast transformations can be good! This DeepMind paper: “On the Theoretical Limitations of Embedding-Based Retrieval”showed that RAG + vector DBs were almost treated as the “absolute truth” for scaling any AI solution. Now, based on both theory and experiments, DeepMind researchers have shown there are mathematical limits to what embeddings can actually retrieve. Even simple queries are not well represented by a single vector. To me, the takeaway is not that RAG is “dead,” but that science consistently reminds us we need to rethink things in order to evolve. It’s the scientific method we keep talking about: an idea dominates, then evidence emerges, and then we move forward. No attachments. IMHO, these cycles of “relearning” used to take much longer, and it was harder to shift people’s beliefs than to statistically prove that something foundational could be improved—or reimagined. With today’s day-scale AI cycles, this “clinging to past beliefs” has become much weaker. If the “glass half empty” of AI’s pace is FOMO, the “glass half full” is that we no longer hold on so tightly to previous methods. https://lnkd.in/gWBSHzjj

62

6 Comments

🚀 Deep-dived into GPT-OSS with the Machine Learning Tokyo community! In this session, I unpacked 5 algorithmic gems inside gpt-oss’s Attention Block — covering not just what they are, but also why they’re here and how they’re implemented: 🔹 Grouped Query Attention (GQA) → smaller KV cache, same query diversity 🔹 Sliding Window Attention (SWA) → linear complexity, local focus 🔹 Dense + SWA alternation → mixing global and local context 🔹 Learned bias (sink) → teaching the model to ignore noise 🔹 RoPE + YaRN → stability at long context lengths ✨ Big thanks to Suzana Ilić for inviting me and to the amazing Machine Learning Tokyo community (https://www.mlt.ai/) for the thoughtful discussions. 📑 Sharing my slides here — and if you prefer video, you can watch the full session recording on MLT.ai’s YouTube channel: 👉 Session Recording: https://lnkd.in/gHtbtZRE Always happy to geek out with fellow LLM algo/paper lovers — let me know your thoughts on which of these techniques excites you most!

238

10 Comments