ObserverBench is a benchmark framework that evaluates whether internal mechanistic estimators—called observers—are suitable for guiding interventions, control, or safety actions in language models. It separates estimation accuracy from the loss incurred by the chosen action, demonstrating that accurate average estimates can still lead to poor decisions. Experiments on GPT‑2‑small, Qwen2.5‑7B, Gemma‑2‑9B‑it, and Qwen3.5‑9B show that observers trained on action loss tend to select lower‑loss actions, while traditional metrics like AUROC can rank monitors differently from deployment loss, highlighting the need for task‑specific evaluation.
ObserverBench is a benchmark framework that evaluates whether internal mechanistic estimators—called observers—are suitable for guiding interventions, control, or safety actions in language models. It separates estimation accuracy from the loss incurred by the chosen action, showing that accurate predictions do not always lead to better decisions. Experiments on GPT‑2‑small, Qwen2.5‑7B, Gemma‑2‑9B‑it, and Qwen3.5‑9B demonstrate that observers trained on action loss can reduce deployment loss, while traditional metrics like AUROC may rank monitors differently from actual performance.
By Vijay Erramilli
arXiv:2609.07944v1 Announce Type: new
Abstract: Existing causal-inference benchmarks for LLMs mostly score method descriptions or whether generated code runs, not whether the executed workflow recove...
By Yonghong Zhang, Ricardo Correia, Isabel M. Parra, Yong Xie
arXiv:2606. 02965v1 Announce Type: new Abstract: Benchmarks for autonomous agents measure whether agents complete tasks, yet this framing is systematically blind to whether an agent should have proceeded at all.
By Victor Ojewale, Suresh Venkatasubramanian
arXiv:2609.35889v1 Announce Type: cross
Abstract: Tool-using language-model agents select and execute third-party artifacts. Different implementations can return the requested output while producing...
By Xiaoyu Xu, Zi Liang, Minxin Du, Qipeng Xie, Qingqing Ye, Yuyuan Li, Haibo Hu
arXiv:2607. 08065v1 Announce Type: new Abstract: LLM-as-judge (Zheng et al.
By Kaihua Ding
Agent evaluations tell us that a model picked the wrong tool, but rarely why. We introduce canary tools: diagnostic probe tools planted in an agent's Model Context Protocol (MCP) tool set, each engineered to probe one specific tool-selection weakness.
arXiv:2608. 12489v1 Announce Type: new Abstract: Organizations decide whom to treat under a budget and want to know what a targeting rule would have earned before deploying it.
By Binshuang Li
CASEL (Consequence-Aware Selective Execution Layer) is a runtime allocation policy for tool‑using language agents that chooses among direct execution, verification, sandboxing, clarification, approval, and abstention for each proposed tool call. It combines calibrated failure probability, state‑dependent consequence, reversibility, ambiguity, and evidence freshness to minimize a per‑mode cost. In controlled simulations, CASEL cuts harmful‑side‑effect trajectories from 14.75 % to 7.02 % and reduces mean normalized harm from 0.184 to 0.123 compared to an uncertainty‑ranked gate.
By Yanyan Zhang, Fei Shen, Ting Yan, Hao Zhang
arXiv:2608. 04719v1 Announce Type: new Abstract: Agent evaluations tell us that a model picked the wrong tool, but rarely why.
By Atul Anand, Sourav Chattaraj
arXiv:2608. 11727v1 Announce Type: new Abstract: When a coding agent obeys a rule, it may simply have been going to do that anyway.
By Zining Huang, Haoran Que, Hong Zeng, Ge Zhang, Zuo Wang, Jin Chen, Haodong Wang, Zhongfei Hou, Changxin Pu, Shen Yan, Wenhao Huang
The paper introduces a denoised counterfactual protocol to accurately measure the selectable headroom of world action models (WAMs) and uses these measurements to build a fail‑gated selector. Across four WAMs and two benchmarks, the study shows that the standard estimate overstates selectable headroom, that structured action perturbations can add up to 16 points of headroom, and that imagined futures retain candidate‑outcome information without loss. The selector improves closed‑loop task success from 76.9 % to 81.6 % on held‑out training episodes and from 62.6 % to 69.8 % on unseen tasks.
By Haojia Gao, Yue Cao, Yumin Liu, Haohua Que, Mingkai Liu, Guanqiu Guo, Zhuo Zou, Bokui Chen, Xueqian Wang