B-PASTE: Beam-Aware Pattern-Guided Speculative Execution for Resource-Constrained LLM Agents

A new research paper introduces B-PASTE, an extension of Pattern-Aware Speculative Tool Execution (PASTE) designed to optimize LLM agent performance. LLM agents operate through interleaved reasoning-and-action loops where serial dependencies between reasoning steps and tool calls create latency issues and idle model time. While PASTE addressed this by speculating future tool invocations based on control-flow and data-flow patterns, it remained tool-centric and limited to individual invocations. B-PASTE expands this approach by speculating entire local branch hypotheses under strict resource constraints. The system maintains a bounded beam of future execution subgraphs, ranking them by expected critical-path reduction rather than raw execution probability. High-value branch prefixes are scheduled on transient slack resources, explicitly modeling resource constraints. This beam-aware extension lifts speculation from single tools to bounded future branches, aiming to reduce end-to-end latency more effectively. The research was announced on arXiv with identifier 2604.16469v1 as a cross announcement. The work builds on prior research that identified how serial dependencies inflate latency and leave models idle during tool execution.