LLM agents increasingly coordinate through persistent shared memory: one agent's write becomes another agent's premise, and eventually a tool call with real side effects. Current agent memory systems treat every accepted write as immediately actionable truth, so a polluted tool result, a stale update, or a teammate's half-finished note can silently drive an irreversible action. We argue that a memory write is not a belief commit. We present MemTX, a transactional belief-commit protocol. Each record carries evidence, permissions, provenance, and validity. Writes are staged inside snapshot-isolated transactions and admitted by a validate-and-commit pipeline, irreversible tool calls are gated on in-flight belief state, and retracting a belief triggers typed cascading repair of its derived records and tool side effects. Two invariants, action-safety gating and cascade-repair completeness, are machine-checked by property-based testing and bounded exhaustive enumeration of 5.5 million protocol states, with zero violations. Across five backbones from three model families, MemTX leads all eight baselines with paired-McNemar significance on four backbones and statistically ties the best baseline on the fifth and strongest, while remaining the only method with zero downstream harm on every backbone. Backbone capability does not substitute for commit discipline.
Xiaoyang Li, Yiqi Wang, Haohui Lu et al.· 4 citations
Large Language Models (LLMs) have demonstrated remarkable potential in automated code generation. However, existing test-driven code generation and refinement frameworks are often hindered by the tests' quality: they typically treat self-generated tests as ground truth, leading to ineffective debugging loops where code is modified to satisfy erroneous tests. To address this, we propose BiVCoder, a diagnosis-driven multi-agent framework featuring a novel bidirectional code-test diagnosis mechanism. BiVCoder coordinates three specialized agents—Coding Agent, Test Agent, and Review Agent. Central to this architecture is the Review Agent, which serves as a diagnosis and decision-making hub. By integrating an MCP-based code-test execution tool, the Review Agent rigorously executes programs to diagnose failure root causes, distinguishing between implementation bugs and test case deficiencies, and subsequently triggers targeted repairs (Coding Agent or Test Agent). Furthermore, we introduce BiVCoder-SFT, a role-specific instruction fine-tuning scheme. We construct high-quality datasets to fine-tune the Qwen3-4B base model into specialized agents for coding, testing, and reviewing. Extensive experiments on HumanEval, MBPP, and their rigorous ''ET'' variants demonstrate the superiority of our approach. With GPT-3.5, BiVCoder achieves a Pass@1 of 77.3%. Notably, the specialized BiVCoder-SFT achieves an average score of 79.5% across the four datasets, not only outperforming its base model (69.7%) but also surpassing larger general-purpose models such as Qwen3-8B (77.8%). Additionally, BiVCoder boosts the performance of more powerful models, improving DeepSeek-V3.2 from 87.2% to 92.1% on HumanEval.
Xiaoyang Li, Jinhao Dong, Wenhang Shi et al.· Proceedings of the 32nd ACM...· 0 citations
A one-round study provides initial evidence for PRD-guided self-evolution, motivating validation at larger scales and in industrial settings, and presents AgentOmnia, a framework coordinating task-space definition, data synthesis, post-training, evaluation, and improvement across To-Consumer (ToC), To-Business (ToB), and To-Employee (ToE) applications.
Hao Jiang, Gangtao Xin, Ying Huang et al.· 0 citations