SCoRe:在 ExploitBench 上通过上下文压缩延展智能体漏洞利用能力
Stopping Is Not the Limit: Understanding and Eliciting Agent Capability in Long-Horizon Vulnerability Exploitation
两个模型在漏洞利用任务里老提前放弃,这篇论文发现让它们继续跑、把"做不到"的判断标记为未证实,MiMo 的分数从 32.4% 拉到 67.0%,超过了 GPT-5.5 的 58.6%。做智能体上下文管理的开发者能直接借鉴这套 SCoRe 做法。
论文研究语言模型智能体在漏洞利用任务中过早停止的问题。作者在 ExploitBench 上用 DeepSeek-V4-Flash-0731 和 MiMo-V2.6-Pro 做实验,发现智能体首次给出停止响应后,若继续执行仍能取得经验证的额外收益,但收益随执行变长而递减。
作者识别出两种与低效延续相关的模式:同一上下文中停止响应比例不断上升,以及过早停止时摘要里保留了对未达成能力的"不可能"判断。据此提出 SCoRe,在停止响应密集出现时压缩上下文,并把这类判断标记为"未经证实"。
在 2,000 轮执行下,SCoRe 让 MiMo-V2.6-Pro 的能力覆盖分达到 67.0%,远高于原停止策略下的 32.4%,也超过按 ExploitBench 300 轮参考结果折算的 GPT-5.5 的 58.6%,说明停止并非智能体的能力上限,配合上下文管理继续执行可以释放可观潜力。
Stopping Is Not the Limit: Understanding and Eliciting Agent Capability in Long-Horizon Vulnerability Exploitation
Vulnerability exploitation tests whether language model agents can develop working exploits from known software flaws. Sequential test-time scaling can improve exploitation performance, but larger budgets do not ensure sustained exploration: agents often stop with budget remaining. Yet what happens when execution continues beyond stopping remains underexplored. We investigate this gap on ExploitBench using DeepSeek-V4-Flash-0731 and MiMo-V2.6-Pro. We find additional verified gains after the first stop response, but gains diminish over longer execution. We identify two patterns associated with less productive continuation: an increasing proportion of stop responses within the same context and earlier stopping associated with summaries retaining impossibility judgments about unachieved capabilities. To test whether context management can elicit further capability, we propose SCoRe, which compacts context at stop bursts and marks these judgments as unproven. SCoRe improves verified capability over summary-based continuation and other evaluated context-management baselines. With SCoRe, MiMo reaches a capability coverage score of 67.0% at 2,000 turns, exceeding both 32.4% under the original stopping policy and the 58.6% computed for GPT-5.5 from ExploitBench's 300-turn reference results on the same tasks. These findings show that stopping need not mark an agent's exploitation limit, and that continued execution with context management can elicit substantial further capability.
This report is categorized under AITOP's "Papers" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include ExploitBench, SCoRe, MiMo-V2.6-Pro, DeepSeek-V4-Flash-0731.
Primary source verification: published directly by arXiv: DeepSeek with first-party authority.