EVENT DOSSIER
The researchers proposed a method called “Decision-Sensing Memory Cards” aimed at addressing the issue where large language model agents rely solely on long contexts and lack relevant decision-making evidence. This method uses counterfactual heuristics to construct an instance context graph, sorts candidates based on the expected utility of their next actions rather than semantic similarity, and compresses selected evidence into typed memory cards. Experiments were conducted on 50 SWE-bench Verified document retrieval instances, and results showed that using the Qwen3.6-Plus model along with this method to re-rank the top 50 candidates for BM25 improved the hit@1 metric from 0.58 to 0.78, and MRR@10 from 0.634 to 0.790. This solution supports托管 LLM judges, local agents, or lightweight rankers, achieving auditable cross-model interoperability. Controlled diagnostic tests demonstrated that this method can effectively identify key evidence for action.
现代大型语言模型（LLM）智能体需要决策相关证据而非仅长上下文。研究人员提出决策感知上下文选择方法，按其对智能体下一步行动的预期效果而非语义相似度对检索文件、测试、痕迹、规则和记忆进行排序。他们展示了反事实启发的上下文层（CICL），构建实例上下文图，估算候选单元的决策导向效用，并将选定证据压缩为类型化记忆卡片。该协议支持托管 LLM 裁判、本地代理或轻量级排名器，实现跨模型的可审计性。在 50 个 SWE-bench Verified 文件检索实例上，Qwen3.6-Plus 对 BM25 前 50 候选的重新排序将 hit@1 从 0.58 提升至 0.78，MRR@10 从 0.634 提升至 0.790，所有 2,500 个判断均可解析。受控诊断显示 CICL 识别行动关键证据：移除最高效用语义单元使…