Taiqiang Wu

Memristor-based analog compute-in-memory (CIM) architectures provide a promising substrate for the efficient deployment of Large Language Models (LLMs), owing to superior energy efficiency and computational density. However, these architectures suffer from precision issues caused by intrinsic non-idealities of memristors. In this paper, we first conduct a comprehensive investigation into the impact of such typical non-idealities on LLM reasoning. Empirical results indicate that reasoning capability decreases significantly but varies for distinct benchmarks. Subsequently, we systematically appraise three training-free strategies, including thinking mode, in-context learning, and module redundancy. We thus summarize valuable guidelines, i.e., shallow layer redundancy is particularly effective for improving robustness, thinking mode performs better under low noise levels but degrades at higher noise, and in-context learning reduces output length with a slight performance trade-off. Our findings offer new insights into LLM reasoning under non-ideality and practical strategies to improve robustness.

Multi-agent systems for software engineering often suffer from a \textit{delegation tax}, where the coordination overhead of sub-agents degrades performance compared to single-agent baselines. We trace this penalty to failures at the inter-agent \textit{handoff boundary}, defined by \textit{context loss} during task delegation and \textit{weak verification} upon return. We propose \ours{}, a framework that transforms this boundary into a governed interface through three \textit{handoff interventions}: (1) \textit{Delegation Review}, a pre-dispatch gate ensuring task specifications are self-contained; (2) \textit{Retrospective Reflection}, a mandatory self-certification by the agent; and (3) \textit{Verification Fork}, an independent validation of the return by the orchestrator. Averaged over three independent runs, \ours{} resolves \textit{47.2\%} of SWE-bench-Pro tasks with Qwen3.5-397B-A17B (+4.0 over a matched SWE-Agent baseline) and, on the longer-horizon Commit0-lite, reaches a \textit{61.6\%} pass rate with Claude-4.5-Sonnet (+11.3) and \textit{59.8\%} with Qwen3.5-397B-A17B (+8.0), while consistently reducing run-to-run variance. Ablations on the full SWE-bench-Pro confirm that ungoverned delegation degrades performance below the single-agent baseline and that each intervention is necessary. Our claims are deliberately scoped to multi-agent substrates in long-horizon settings: governing the handoff removes the delegation tax where decomposition is necessary, rather than replacing single-agent test-time scaling on atomic tasks.