Research map / Multi-agent systems and games / LLM agents

CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CORVUS, a coding-agent trajectory architecture that replaces inline file reads with a synced file registry whose current contents are re-injected before each reasoning cycle, reducing tokens, cycles, and duplicate reads while keeping pass rates comparable.

Problem

Append-only agent trajectories freeze file-read snapshots in history, which become stale as files change and cause redundant re-reads and stale-context edit failures, inflating token cost and reasoning cycles.

Approach

Decouple file-read actions from observations: a sync_file tool only registers a file in a synced set and leaves a marker in history, while a context-sync phase injects the current contents of all registered files into a dedicated block at the end of each prompt, so at most one current copy per file appears.

Key results

• Input tokens reduced 9-50% per task across four models, e.g. 50.42% for Claude Sonnet 4.5 on SWE-PolyBench_Verified.
• Reasoning cycles reduced by up to 37.33% (45.03 to 28.22) for Claude Sonnet 4.5 on SWE-PolyBench_Verified.
• Pass@1 remains comparable: Sonnet 4.5 improves from 28.10% to 29.49%, while Sonnet 4 drops from 25.83% to 24.61%.

Surprising findings

• Sonnet 4 pass rate slightly decreases under CORVUS (25.83% to 24.61%), so gains are not uniformly positive on success.
• Efficiency gains are smaller on SWE-Bench Pro for some models, e.g. Qwen 480B refactoring shows a 0.6% token increase.

Limitations

• Evaluation is limited to four models and two benchmarks, with SWE-Bench Pro restricted to a sampled subset of 191 instances.
• Synchronization refreshes whole files rather than regions, and the paper leaves desync and partial sync as future work.
• Pass rate differences are small and no significance testing is reported.

Experimental setup

Datasets: SWE-PolyBench_Verified SWE-Bench Pro

Benchmarks: SWE-PolyBench_Verified SWE-Bench Pro

Metrics: pass@1 input tokens output tokens cost reasoning cycles duplicate file reads final request length end-to-end runtime

Baselines: Minimum Viable Agent with read_file (append-only baseline) Sliding Window Summarization Observation Masking AgentDiet

Builds on: Strands Agents SWE-Bench SWE-PolyBench AgentDiet

Closest papers by meaning

• When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding (NeurIPS 2026 · similarity 0.88)
• TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering (NeurIPS 2026 · similarity 0.87)
• The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements (NeurIPS 2025 · similarity 0.87)
• Executable Code Actions Elicit Better LLM Agents (ICML 2024 · similarity 0.87)
• EvoCodeBench: Evaluating Coding Agents in Multi-Turn Iterative Interactions (NeurIPS 2026 · similarity 0.87)
• SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks (NeurIPS 2026 · similarity 0.87)