Tutorials

In the first half, we introduce Emergence World, a platform where populations of LLM-driven agents inhabit a shared spatial world grounded in live data, wield 120+ context-dependent tools, maintain persistent memory systems, and govern themselves through democratic mechanisms. A 15-day cross-vendor study across five parallel worlds—powered by Claude, Grok, Gemini, GPT, and a mixed population—reveals that identical starting conditions produce radically divergent outcomes, from stable deliberative governance to total population collapse, demonstrating that the relevant unit of safety analysis is the deployed system, not the model in isolation.

In the second half, we pivot to the question: what does it take to make multi-agent teams actually work? We present a Governed Agent Factory for coding agents that operationalizes the lessons of emergence world research. Our Governed Agent Factory organizes agents into supervised teams with explicit roles—chief, scout, builder, oracle—coordinating through structured, parent-visible communication rather than hidden chat. It enforces safety through scoped tools, governed commits, no-ship review gates, and validation ladders.

Through a live demonstration, attendees will see how abstract principles from emergence world research—observable coordination, governance mechanisms, heterogeneous agent populations, and behavioral accountability—translate concretely into a system where teams of agents build, validate, and ship software under human supervision.

We start from first principles. Why is token generation limited by memory bandwidth? How do the KV cache and vocabulary size dominate memory in small models? How did scaling laws move from “go big” (Kaplan) to “balance parameters and data” (Chinchilla) to deliberately over-training small models? We then trace the evolution from 2015 to 2026 and cover the full toolkit. That means data curation and synthetic data, compact architectures (GQA, mixture-of-experts, state-space hybrids, nested elastic models), knowledge distillation, pruning, and quantization from GPTQ and AWQ to 1.58-bit BitNet. We also cover parameter-efficient fine-tuning (LoRA/QLoRA), speculative decoding, and on-device runtimes, with an end-to-end recipe.

Finally, we assess honestly what SLMs do well (scoped, verifiable, and retrieval-grounded tasks; agentic workers) and what they cannot do. Knowledge capacity is about 2 bits per parameter, and long-tail recall, long-horizon reasoning, and multilingual breadth all suffer. We also show how test-time compute lets 1B models rival far larger ones on math. We close with implications for Indian languages and open problems for the data science community.

We begin with the foundations of representation learning over graphs and simplicial complexes, continuous-time representation learning, memory mechanisms, and state space models, highlighting the limitations of local temporal message passing. We then introduce continuous-time topology-aware higher-order polynomial projection (CTT-HiPPO), which extends HiPPO memory to incorporate graph structure through polynomial Laplacian operators. We explain how this construction yields the Continuous-Time Dynamic Graph State Space Model (CTDG-SSM), and how zero-order hold discretization enables practical implementation for long-range spatio-temporal learning.

Moving beyond pairwise interactions, we briefly introduce temporal simplicial complexes as an event-driven framework for modeling evolving group interactions. We develop intuition for temporal simplicial isomorphism and the temporal simplicial Weisfeiler–Lehman test, then examine how the temporal simplicial state space model supports efficient online representation learning. Particular attention is given to the relationship between memory dimension and expressive power, connecting architectural choices to theoretical guarantees.

The tutorial begins with the architectural foundations of VLMs, covering vision encoders, visual tokenization, and multimodal projection mechanisms that connect visual representations with large language models. We then distinguish visual perception from higher-level visual reasoning and introduce representative tasks involving spatial and geometric understanding, mathematical reasoning, symbolic interpretation, and document understanding.

Building on these foundations, we examine explicit reasoning paradigms, with a focus on Chain-of-Thought (CoT) reasoning in multimodal models. We discuss multimodal CoT prompting, visual grounding, and tool-augmented reasoning, emphasizing how intermediate reasoning steps and interactions with external tools can help address complex visual problems. We then explore latent visual reasoning, in which intermediate computations are performed in continuous representation spaces rather than exclusively through explicit textual reasoning traces. We review representative approaches, their underlying mechanisms, and the opportunities and challenges of reasoning directly over latent visual representations.

The tutorial next focuses on training and post-training strategies for visual reasoning, including supervised fine-tuning and reinforcement learning. We discuss how these approaches can improve visual perception, grounding, and multi-step reasoning, as well as the challenges of designing effective reward signals, aligning perception with reasoning, and ensuring that improvements translate into robust generalization.

We then review benchmarks and evaluation methodologies for multimodal reasoning, including MMMU, MathVista, and SeePhys, and examine their strengths and limitations in assessing visual understanding and reasoning capabilities. Particular attention is given to persistent failure modes, including visual hallucinations, spatial reasoning errors, inadequate visual grounding, and discrepancies between apparent reasoning and actual problem-solving ability.

Finally, we discuss emerging research directions, including test-time scaling, latent reasoning, more effective integration of visual information throughout the reasoning process, and the development of more reliable and generalizable reasoning capabilities. The tutorial aims to equip participants with a coherent understanding of the evolving landscape of VLM reasoning, the strengths and limitations of current approaches, and the open research questions shaping the next generation of multimodal foundation models.