[2026年7月15日提交]

标题：Multi-Head Latent Control：LLM 智能体决策的统一接口

查看论文《Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making》的 PDF，作者：Amirhosein Ghasemabadi 及其他 3 位作者

查看 PDF HTML（实验性）

摘要：大语言模型（LLM）越来越多地被部署为智能体（agent），但可靠的智能体行为需要的不仅仅是下一个词元预测。在推理时，更希望智能体能够决定是继续当前推理、转交给更强的模型、请求额外信息、调用外部工具，还是在给定设置下弃权。现有方法通过提示级路由、外部编排或任务特定微调来处理这些决策，这些方法主要依赖输入端信号，且通常成本高昂，难以随着模型基座（backbone）的演变进行维护。我们探讨了这些控制决策是否可以直接从模型潜在生成过程中推断出来。我们提出了 Multi-Head Latent Control，这是一种轻量级层，它从冻结的 LLM 或 VLM 中读取隐藏状态轨迹，以生成部署时的控制信号。能力头（Capability Head）预测当前模型是否能够解决该实例，或者应该转交给更强的协作者；解析头（Resolution Head）则预测合适的解析决策：澄清（Clarification）、工具使用（Tool Use）、弃权（Abstention）或直接回答（Direct Answering）。两个头仅基于同一冻结 LLM 基座上的潜在轨迹进行训练，从而能够在不修改模型的情况下进行事后适配。在语言和视觉-语言设置中，Multi-Head Latent Control 持续改善了多模型系统的质量-成本权衡，实现了基于部分生成的提前切换以及更准确的干预决策。在路由执行（小型模型 + 大型模型）中，它在 AndroidWorld 上减少了高达 90.7% 的大型模型使用量，在各项基准测试中平均减少了 27-53% 的大模型使用量，同时保留了大部分大型模型的性能。此外，学习到的控制信号改善了工具使用决策质量，相对得分提升最高达 +158%，并且漏掉所需工具调用的情况减少了 65.5%。

提交历史

来自：Amirhosein Ghasemabadi [ 查看邮件 ]

[v1] 2026年7月15日，星期三，18:36:54 UTC（1,940 KB）

[Submitted on 15 Jul 2026]

Title:Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making

View a PDF of the paper titled Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making, by Amirhosein Ghasemabadi and 3 other authors

View PDF HTML (experimental)

Abstract:Large language models are increasingly deployed as agents, but reliable agentic behavior requires more than next-token prediction. At inference time, it is preferred that an agent can decide whether to proceed with its current reasoning, defer to a stronger model, request additional information, invoke external tools, or abstain under the given setup. Existing approaches address these decisions through prompt-level routing, external orchestration, or task-specific fine-tuning, which primarily rely on input-side signals, and are often costly and difficult to maintain as model backbones evolve. We ask whether such control decisions can be inferred directly from a model's latent generation process. We introduce Multi-Head Latent Control, a lightweight layer that reads hidden-state trajectories from a frozen LLM or VLM to produce deployment-time control signals. A Capability Head predicts whether the current model can solve the instance or should defer to a stronger collaborator, while a Resolution Head predicts appropriate resolution decision Clarification, Tool Use, Abstention, or Direct Answering. Both heads are trained only on latent traces from the same frozen LLM backbone, enabling post hoc adaptation without modifying the model. Across language and vision-language settings, Multi-Head Latent Control consistently improves the quality-cost tradeoff of multi-model systems, enabling early handoff from partial generations and more accurate intervention decisions. In routed execution (small + large model), it reduces large-model usage by up to 90.7 percent on AndroidWorld and 27-53 percent on average across benchmarks, while retaining most of large-model performance. Additionally, the learned control signals improve tool-use decision quality, yielding up to +158 percent relative score gain and 65.5 percent fewer missed-required tool calls.

Submission history

From: Amirhosein Ghasemabadi [

view email

]

[v1]

Wed, 15 Jul 2026 18:36:54 UTC (1,940 KB)