action space design

This work addresses the challenge of governing high-risk actions in heterogeneous multi-agent systems, where runtime disparities hinder consistent enforcement—particularly concerning authorization rationale, approval semantics, and execution evidence. To overcome this, the paper proposes a runtime-agnostic governance model centered on action certificates that supplant vendor-specific logs. The model structures control logic across five critical checkpoints and integrates portable action envelopes, runtime and approval receipts, and replayable proofs. Innovatively, it introduces externality-aware action certificates that embed boundary facts and replace binary approval states with explicit executability categories, enabling unified cross-runtime governance. Evaluation on a benchmark of 96 trajectories spanning four distinct runtimes demonstrates that the approach preserves path quality while revealing distinct failure modes under ablation, effectively supporting runtime-portable governance policies.

Large language model (LLM) agents commonly suffer from “planning–action mismatch” in complex tasks—i.e., a lack of dynamic alignment between high-level reasoning strategies and low-level tool invocations—leading to poor code-action quality and divergent reasoning paths. Method: This paper proposes a dual-control mechanism jointly governing strategy and action. It introduces a hierarchical coordination architecture: an upper-layer strategy routing module that dynamically selects and adapts reasoning paradigms (e.g., ReAct, Chain-of-Thought), and a lower-layer action-space remapping module that enables differentiable reconstruction of tool calls, augmented by an environment-feedback-driven online switching algorithm. Contribution/Results: The approach decouples the rigid coupling between strategy and action inherent in conventional frameworks. Evaluated on LegalAgentBench, it achieves a 20% accuracy gain, substantially reduces token consumption, and demonstrates strong generalization and cross-model scalability across GPT-4o and GLM-4 series models.

本文针对企业部署自主AI代理时遇到的控制模型不匹配问题，提出了五种运行时治理原语，并通过实现这些原语来解决身份、治理等问题。

研究解决了工具调用与操作性之间的差距问题，通过Agent-First Tooling机制和AFT-Bench框架来确保代理在不确定性下安全继续执行。

Tool-use agents frequently fail by acting on insufficient evidence or when preconditions in multi-step workflows remain unsatisfied. This study elucidates the mechanisms underlying evidence chain fragmentation from decision-making to execution, highlighting fundamental discrepancies between static evaluation and dynamic execution. To address these issues, this work proposes SafeActBench, a novel benchmark that introduces a provenance-bound evidence ledger and a deterministic trajectory evaluator. Through systematic investigation incorporating multi-model configuration testing, workflow dependency tracing, and evidence integrity verification, the results demonstrate that agent failures fundamentally stem from executing actions without establishing sufficient evidence and from inadequately resolving prerequisite dependencies within complex procedural workflows.

本文提出Brain API，一种意图感知控制平面，通过决策工件解决现有系统中缺乏意图级决策治理的问题。

This study addresses "schema bias," a phenomenon wherein large language model (LLM) agents exhibit significant performance disparities across functionally equivalent yet syntactically distinct tool schemas. We formally define this phenomenon and propose an executable transformation framework based on nine operators that rewrites tool schemas while preserving action semantics, enabling quantitative analysis of such bias. Furthermore, we introduce a few-shot probing method to reliably predict schema difficulty and construct a multi-model evaluation benchmark. Experimental results demonstrate that even state-of-the-art models remain severely affected by schema bias, with success rates fluctuating substantially across schema variants. Notably, existing training-based mitigation strategies prove effective only for previously seen variants, exhibiting limited generalization capability.

This study addresses the failure of safety certification during runtime interventions in foundation model agents, caused by neglecting the correlation between proposals and historical context. It is the first to reveal that this correlation constitutes a necessary condition for certification, analyzing how state folding compromises certifiability and establishing information preservation boundaries. By integrating bounded robust interfaces, viability kernel theory, and standard safety reachability fixed-point algorithms, this work demonstrates that observing current proposals restores robust viability and decouples incompatible intervention modes. Through controlled experiments, it reproduces the predictive barriers arising from telemetry removal or privilege restriction, validating the correctness of the theoretically defined boundaries and achieving a precise characterization of robustly safe interventions.

This study addresses the inexecutability of task abstractions caused by spatial layout constraints under fixed controllers. We propose an affine-constraint-based layout repair method that compiles task requirements into affine constraints and modifies only continuous coordinates, thereby preserving the original event logic and controller. Furthermore, we introduce a most-violated-row update algorithm coupled with a quadratic projection mechanism to decouple the representation layer from the optimizer, enabling conditional bounded certification. Experimental results demonstrate that the proposed approach successfully certifies all layouts across three tasks, with its effectiveness further validated through 300 paired rollback tests.

论文提出SARA方法，通过区分动作诱导和执行授权解决工具增强LLM代理中由工具输出引发的非预期实际操作问题。