Tag
该论文评估了Lean编译成功是否等同于忠实的自然语言到Lean语句形式化，发现多个系统在编译通过与语义正确之间存在1.3至29.5个百分点的显著差距，且工具增强型代理的大部分差距源于真实的翻译错误。
X-Ragent 是一个面向胸片（CXR）证据驱动推理的后训练框架，通过工具增强强化学习解耦辅助证据获取与有效性门控的量化测量；在 Chest AgentBench 等三个基准上取得最佳结果，其中完全未见基准提升 8.6 分。
This paper identifies tool-mediated recovery as a failure mode in LLM unlearning and proposes Agentic Tool Unlearning (ATU) to reduce both parametric recall and tool-based recovery while preserving normal tool use.
Introduces PHREEQC-MCQ-200, a diagnostic benchmark for evaluating tool-augmented LLM agents on deterministic aqueous-geochemistry simulations, revealing that simulator access improves accuracy but also causes regressions in items correctly answered without tools.
FinAcumen is a framework that accumulates reasoning experience from prior trajectories into a persistent memory bank for financial multimodal reasoning, improving performance across four benchmarks while maintaining a frozen 8B vision-language model.
This paper empirically tests whether LLM agents with GNN tools exercise judgment or blindly obey the tool, finding that agents agree with the GNN 97.6–99.2% of the time and that stronger backbones defer even more. The cost of this deference does not shrink with capability, and selective invocation remains an open problem.
SpatialClaw is a training-free framework that uses code as an action interface to enable flexible, stateful spatial reasoning in vision-language models, achieving superior performance across diverse 3D/4D spatial reasoning tasks.
This paper presents SPADER, a reinforcement learning framework for multi-answer QA that uses step-wise peer advantage for credit assignment and diversity-aware exploration rewards to improve recall of long-tail entities, achieving better performance on several benchmarks.
This paper introduces CoCoDA, a framework that uses a co-evolving compositional Directed Acyclic Graph (DAG) to manage tool libraries for augmented agents. It enables small language models to efficiently retrieve and compose tools, allowing an 8B model to match or exceed the performance of a 32B model on reasoning benchmarks.