Not Worth Another Token: Marginal Value Estimation for Efficient Deep Research Agents

Long-horizon research agents often suffer from rapidly accumulating context that increases token costs, latency, and noise. This paper investigates how to manage this context by estimating the marginal value of information at different stages of the research pipeline. The authors define a stage-aware framework with three intervention points: Pre-Retrieval (filtering subqueries), Post-Retrieval (filtering retrieved evidence), and Pre-Synthesis (compressing the final context). They compare various pruning strategies—ranging from lightweight heuristics like Maximal Marginal Relevance (MMR) to learned models and LLM-based judges—to determine the optimal trade-offs between report quality, efficiency, and faithfulness.

The study reveals that where you prune is more important than how you prune. Post-retrieval pruning is the most effective strategy for reducing operational costs, with lightweight heuristics like MMR reducing token usage by up to 73% while maintaining high report quality. In contrast, pre-synthesis pruning is less effective at reducing total costs because most of the computational expense has already been incurred, but it is highly effective at refining the final report's quality. The authors find that no single method dominates across all metrics; instead, the best performance is often achieved by combining different strategies across stages, such as using novelty-sensitive heuristics early and quality-focused methods late.

As agentic systems become more complex, the ability to manage context efficiently is critical for scalability. This work provides a systematic guide for developers to optimize their pipelines, demonstrating that significant cost reductions are possible without sacrificing the quality or grounding of the final output. By identifying that early-stage intervention is the key to cost control, the paper offers a practical roadmap for designing more sustainable and responsive research agents.