UC Berkeley

Large language models (LLMs) have advanced rapidly in recent years, evolving from the conversational assistants of just a few years ago into general-purpose agentic systems that perform multi-step reasoning, generate and execute code, and act through software interfaces. As these capabilities mature, the central object of study has shifted from the model in isolation to the broader agent: a system in which an LLM is combined with perception, memory, reasoning, and tool-use components and asked to pursue goals over long horizons. The binding constraints on agent performance are therefore increasingly operational rather than purely model-internal: how much context can be processed in a single trajectory, how compute is allocated across parallel and serial reasoning, and how reliably an agent can judge the correctness of its own outputs. These constraints suggest three complementary dimensions along which an agent system must scale: sequential, parallel, and recursive. An efficiency layer underpins all three, determining the marginal cost of every additional unit of compute.In this thesis, we develop techniques across these four directions: efficient model foundations, sequential scaling, parallel scaling, and recursive scaling. For efficient foundations, we introduce Q-Diffusion, a post-training quantization framework that pushes large diffusion models to 4-bit weights without retraining, and SparseLoRA, which accelerates LLM fine-tuning by dynamically selecting per-token active subnetworks at runtime through a lightweight contextual-sparsity estimator. For sequential scaling, we present LLoCO, which combines offline context compression with parameter-efficient fine-tuning to extend the effective context window of a 4k LLaMA model to 128k tokens using roughly 30x fewer tokens at inference, and STORM, a token-efficient long-video understanding model that integrates spatiotemporal information into compressed visual tokens through a Mamba-based temporal projector. For parallel scaling, we develop Adaptive Parallel Reasoning, in which a single LLM learns through supervised training and end-to-end reinforcement learning to dynamically distribute reasoning across parallel and serial threads via a parent--child threading mechanism. For recursive scaling, we present V1, a framework that unifies generation and self-verification by training one model to both produce candidate solutions and judge pairs of them, showing that joint training compounds test-time compute into reliable capability gains. Taken together, these contributions outline a path toward LLM agent systems that are both scalable, in context, compute, and efficiency, and self-improving, able to convert their own judgments into training signal.