Thesis Proposal - Yuxin Yang

Title:   Controlled Test-Time Scaling for Long-Horizon LLM Agents 
 
 
Date and Time: Tuesday, July 7, 2026, 9:30–11:00 AM
 
 
Location: Hughes Aircraft Electrical Engineering Center (EEB), Room 132           
 
 
Committee Members: Viktor K. Prasanna (Chair), Professor Bistra Dilkina, Professor Sandeep Gupta, Professor Sai Praneeth Karimireddy, Professor Mengyuan Li
 
 
Abstract:  Long-horizon LLM agents increasingly rely on system-level test-time scaling: branching, refinement, search, verification, and multi-agent coordination. These mechanisms can improve performance, but additional computation often saturates or even degrades task outcomes when the generated artifacts are not consolidated into a usable state. Deeper trajectories may accumulate stale or incorrect context, wider exploration may produce redundant evidence or solutions, and parallel agents may introduce contradictions or integration conflicts. Test-time scaling therefore turns inference into a resource-allocation problem: additional computation can be spent on generating, evaluating, or integrating intermediate artifacts, but its value depends on the state those artifacts create.This thesis formalizes controlled test-time scaling as a budgeted sequential decision problem over scaling operators and a consolidation map that turns intermediate artifacts into the information state used for subsequent decisions. It tests three hypotheses. H1: state-conditioned scaling improves the quality–cost tradeoff over strong fixed scaling schedules. H2: a central failure mode of scaling is insufficient consolidation, where expansion produces more intermediate artifacts than the system can select, reconcile, or integrate. H3: the consolidation mechanism required for effective scaling becomes more structured as artifact coupling increases: weakly coupled settings may require only order-insensitive selection, moderately coupled settings require consistency-maintaining reconciliation, and tightly coupled settings require concurrency-controlled integration.These hypotheses are evaluated across three domains with increasing artifact coupling: retrieval-augmented generation, deep research, and long-horizon coding. H1 and H2 are tested within each domain using matched-budget fixed-schedule baselines, state-conditioned controllers, and consolidation ablations. H3 is tested both across domains and through mechanism transfers, asking when a consolidation mechanism that suffices in a weakly coupled setting fails in a more tightly coupled one. The completed system SParC-RAG provides initial evidence for H1 and H2 in RAG. Two proposed systems, COMPASS-DR and HYDRA-Code, extend the framework to deep research and long-horizon coding. Together, these studies reframe test-time scaling as a control problem: the bottleneck in long-horizon LLM agents is not merely model capability or inference budget, but the system’s ability to transform expanded computation into a compact, actionable information state that determines what computation should be spent next.