The paper investigates how language‑model judges can make version‑dependent errors when evaluating upgraded agents. Using 35 public coding‑agent submissions, two customer‑service agents, and over a thousand expert‑labeled trajectories, the authors show that fixed judges often reject task‑conditioned error invariance and can incorrectly approve failed patches, especially as agent capability increases. Paired audits of current outputs reduce interval width only marginally, and the study concludes that independent human patch review is still necessary.
By Jiapeng Li
arXiv:2609.08618v1 Announce Type: new
Abstract: Benchmark scores describe what a checkpoint can do now, but they do not determine how it will respond to the next training episode. We measure this mis...
By Zhongxuan Liu, Sicheng Zhou, Hongzhi Wang
The paper introduces the concept of intervention fidelity in latent world models, measuring whether a model’s open‑loop transitions align with actual environment interventions. Experiments on TD‑MPC2, Cheetah, and DreamerV3 show that high reward fit does not guarantee fidelity, and that self‑supervised models can outperform task‑anchored ones in preserving intervention effects. The authors propose a capture‑gated audit to localize failures and argue that fidelity must be directly audited on the model’s native interface.
By Donna Vakalis
The paper introduces Cros, a risk‑constrained stopping layer for sequential clinical diagnosis agents that determines when to stop testing and make a diagnosis. Cros combines state‑wise error ranking, policy design on disjoint development splits, and exact tests of selective diagnostic error to provide finite‑sample guarantees. On a MIMIC‑derived abdominal‑pain benchmark, Cros achieves higher state‑error AUROC and lower selective error rates compared to baseline stopping methods, though its performance varies across development resplits.
By Yuexin Wu, Vasile Rus
arXiv:2608. 07914v1 Announce Type: new Abstract: Behavioral contamination detectors can return "no evidence" either because a benchmark is clean or because the audit has little power.
By Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma
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.
LLM agents are increasingly evaluated on multi-week decision tasks in which the state that drives cost is never directly observed. On such tasks the final cost cannot say why an agent failed: it may have misread the world, or read it correctly and still failed to act (the knowing-doing gap).
arXiv:2605. 22949v3 Announce Type: replace Abstract: Foundation-model pools are increasingly used as black-box responders in coordinated systems where a coordinator must decide which response to trust.
By Joss Armstrong
The paper introduces a schema‑adaptive action‑conditioned Joint‑Embedding Predictive Architecture (SAAC‑JEPA) for cross‑machine CNC transfer when only a subset of sensors overlap between source and target machines. Experiments show that pretraining does not improve source‑only forecasting, but a carefully selected action‑conditioned JEPA model achieves a zero‑shot RMSE of 0.546 on the target, outperforming persistence but falling short of certain baseline models. Ablation studies reveal that adding RevIN improves RMSE but harms calibration, and limited post‑lock adaptation can further reduce error.
By Ayoub Louaye Bouaziz, Matthieu Ostertag, Anton Demasles
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:2608. 03222v1 Announce Type: cross Abstract: Software engineering (SWE) agents resolve repository-level issues through long trajectories that grow increasingly expensive as context accumulates.
By Chenyu Wang, Yunbo Lyu, Junda He, Zhou Yang, Chenxing Zhong, Yaniv Harel, David Lo
arXiv:2602. 11619v2 Announce Type: replace Abstract: Running the same LLM agent on identical inputs yields 2.
By Aman Mehta