arXiv:2607. 11183v2 Announce Type: cross Abstract: Large language models increasingly operate as tool-using agents, where small format, argument, or function-call errors can invalidate otherwise plausible responses.
By Sheng Xu, Junhua Wang, Boyuan Huang, Ke Jia, Jiadun Zhu, Zhen Chen
Large language models increasingly operate as tool-using agents, where small format, argument, or function-call errors can invalidate otherwise plausible responses. We study inference-time feed-forward network (FFN) intervention for improving structured outputs without retraining model weights.
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
The paper investigates how the upper spectral tails of weight matrices in decoder‑only transformer language models influence reasoning behavior. By performing controlled interventions on the query–key product and comparing them to factor‑level surgeries, the authors find that edits targeting the spectral tail more strongly affect model performance across multiple checkpoints and reasoning benchmarks. The study also explores how inverse participation ratios predict accuracy transitions and shows that tail‑aware low‑rank adaptations converge faster than standard methods.
By Ibne Farabi Shihab, Sanjida Akhter, Md Najmus Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Anuj Sharma
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 entry-state sharpening, a data‑free pre‑training step that prepares a language model’s checkpoint in a sharper, lower‑entropy state before test‑time reinforcement learning (TTRL). By reducing policy entropy, the model can more efficiently use its limited adaptation budget, leading to higher endpoint conversion efficiency across tasks such as MATH, GPQA, and AMC. Experiments with different data‑free objectives (e.g., R‑Zero vs. SPIRAL) demonstrate that the choice of pre‑training objective strongly influences the checkpoint’s readiness for TTRL, and a label‑free self‑distillation intervention can further sharpen the entry state.
By Zhanming Zhang, Vinoth Selvendran
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
arXiv:2606. 05263v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards improves reasoning and tool use, yet long-horizon language agents still learn unsupported evidence chains, belief drift, and shortcut actions that satisfy terminal checks.
By Renwei Meng
The paper investigates why constrained updates in supervised fine‑tuning (SFT) still lead to forgetting by developing a functional‑path analysis that decomposes changes in old loss into controlled direction, gradient staleness, probe mismatch, and finite‑step effects. It introduces a conditional refresh budget and identifies a restoring direction provided by informative replay, showing that the placement of constraints relative to Adam and the granularity of constraints significantly affect retention across code‑label audits. The study also explores checkpointed generation, replay schedules, and sequential‑task comparisons, demonstrating that subspace regularization, projection, and supervised replay differentially influence backward transfer and content retention.
By Qingsen Ma, Dianyun Wang, Jiahong Liu, Irwin King
arXiv:2609.14825v1 Announce Type: cross
Abstract: Large language models (LLMs) are often deemed unsafe for clinical question answering because of their tendency to hallucinate. Retrieval augmentation...
By Zeyu Dong, Benjamin Wang, Joyee W. Jin
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.
arXiv:2606. 04296v1 Announce Type: new Abstract: As autonomous AI agents move from conversational systems to long-horizon software execution, runtime safety layers that decide when to interrupt an agent have become essential.
By Manvendra Modgil