Overview
CORVUS (Context Optimization and Reduction Via Underlying Synchronization) introduces a novel trajectory architecture for LLM-based coding agents. Unlike conventional append-only systems that permanently fix file-read snapshots within a chronological history—leading to stale data, reasoning errors, and redundant re-reads—CORVUS decouples file-read actions from their observations. By maintaining a synchronized registry of relevant files and injecting only their current contents at each reasoning cycle, the architecture ensures trajectories remain lightweight and consistently aligned with the actual codebase state.
Key Innovation
Traditional LLM coding agents rely on trajectories that accumulate reasoning steps, tool calls, and results for multi-step decision-making. However, when files change due to agent edits or concurrent human modifications, the fixed snapshots in the trajectory become outdated. This forces agents to re-read files, each time appending another copy and bloating the trajectory with redundant information. CORVUS eliminates this inefficiency by:
- Decoupling file-read actions from their observations
- Maintaining a dynamic, synchronized file registry
- Injecting only the current file contents at each reasoning cycle
This structural change not only reduces trajectory bloat but also prevents stale snapshots, allowing agents to reason more accurately with up-to-date context.
Evaluation & Results
The authors evaluated CORVUS on two benchmarks—SWE-POLYBENCH_VERIFIED and SWE-BENCH PRO—across four different LLMs. The results demonstrate substantial efficiency gains:
- 9–50% reduction in average input tokens per task
- 15–32% shorter final prompts
- Up to 37% fewer reasoning cycles
- Comparable pass rates to baseline methods
These improvements highlight how CORVUS maintains task performance while significantly reducing computational overhead and token usage, making it a promising approach for deploying LLM agents in dynamic software engineering environments.
Relevance to 2026
As LLM-based coding agents become increasingly integrated into real-world, collaborative development workflows in 2026, the ability to handle concurrent edits and maintain context efficiently is critical. CORVUS addresses a key limitation of current systems: trajectory bloat and stale context. Its synchronization mechanism is particularly relevant for agents operating in live repositories or continuous integration pipelines, where file states change frequently. By optimizing context usage and reducing redundant tokens, CORVUS supports scalable, cost-effective deployment of coding agents in production settings.
Subjects & Citation
- Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Software Engineering (cs.SE)
- Cite as: arXiv:2607.22711 [cs.LG] (or arXiv:2607.22711v1 [cs.LG] for this version)
- DOI: https://doi.org/10.48550/arXiv.2607.22711
Authors
Mingwei Zheng, David OBrien, Siwei Cui, Pardis Pashakhanloo, Rajdeep Mukherjee, Myeongsoo Kim, Sachit Kuhar
Submitted on 20 Jul 2026
via ArXiv LG