A 0.6B language model consistently answers YES to 1,200 logical tests, yet its behavior shows no discrimination. Linear probes reveal the correct verdict with high AUC (0.96) and transfer to unseen structures, but a single scalar readout fails due to a saturated decision threshold offset by +4.6 σ. Adjusting this threshold restores behavior accuracy from 50 % to 81 % and improves higher‑scale models, demonstrating that miscalibrated readouts, not hidden knowledge loss, drive performance gaps.
By Gnaneswar Villuri, Hashmath Shaik, Alex Doboli
The paper introduces an action‑class diagnostic framework for multi‑turn tool‑calling in large language model agents, breaking failures into action‑class miscalibration and action‑execution failure across a four‑class action space (TOOL_CALL, ASK, REFUSE, CONFIRM). It defines a self‑revealing upper bound (Acc GAR) to expose state‑grader masking of miscalibration and shows that miscalibration is a significant, previously hidden failure mode, especially for heavily tool‑trained families. The study demonstrates that calibration can be reshaped by context‑only perturbations, but the effects vary widely across models and perturbation mechanisms, underscoring the need for diagnostics beyond aggregate accuracy.
By Kangjia Zhao, Jiajun Li, Haozhan Shen, Wei Chow, Linfeng Li, Hang Song, Lingdong Kong, Chen Zhi, Tiancheng Zhao, Songhua Liu, Jianwei Yin
The study examines how two small instruction‑tuned language models, Qwen2.5‑1.5B and Llama‑3.2‑1B, respond to user pushback on TriviaQA. When initially correct, the models flip to a wrong answer in about 42–43% of cases, with the effectiveness of different pushback styles varying by model. Attempts to decode capitulation from the pre‑response residual stream fail under a rigorous validation protocol, revealing overfitting and a measurement hazard that underestimates capitulation by 18–24 percentage points.
By Saad Aamir, Muhammad Awais Bin Adil
When a tool-using agent is given the same task in a different language, does it still take the same steps? Multilingual evaluation rarely asks: it compares final answers and discards the actions.
The paper presents a new multi‑turn benchmark of 423 conversations with 1,661 labeled turn‑states to study when language models should refrain from answering. It shows that probes for unanswerability transfer well across datasets that share the same underlying signal, but fail to generalise to other forms of epistemic uncertainty. While a calibrated probe can identify underspecified turns more accurately than chance, it does not consistently improve overall generation quality compared to standard methods.
By Jerzy Kami\'nski, Ilya Galyukshev, Artem Kuznetsov, Danil Fedorov, Kirill Redko, Sergey Chuprin, Aidar Shumbalov, Stanislav Chumakov, Anna Kalyuzhnaya
The paper investigates whether accuracy and order sensitivity in large language model (LLM) multiple‑choice question (MCQ) evaluation move together. Using proprietary and open‑weight models on MMLU and ARC‑Challenge, the authors measure order sensitivity with flip rate and recall standard deviation (RStd) and find that these metrics often diverge from accuracy and from each other. They conclude that accuracy alone is insufficient and recommend reporting both flip rate and RStd alongside accuracy.
By Karl Hanna, Chen Feng
arXiv:2607. 08456v1 Announce Type: cross Abstract: A model should refuse two different things: answers it would get wrong, and questions it should not answer at all, such as unanswerable ones or ones resting on a false premise.
By Benedikt J. Wagner
arXiv:2609.13308v1 Announce Type: cross
Abstract: A companion evaluation found that naming the target part in a manipulation prompt increased action accuracy by 0.32-0.63 across eight vision-language...
By Sarthak Sattigeri
arXiv:2606. 30449v1 Announce Type: new Abstract: Probes on model internals could help monitor agentic systems if they identify harmful text or tool actions before those actions are generated.
By Max Fomin, Elad David, Amit LeVi
arXiv:2607. 14111v1 Announce Type: cross Abstract: Can small language models detect and report on perturbations their own internal activations?
By Ely Hahami, Ishaan Sinha, Lavik Jain
The paper evaluates safety monitors by measuring recall only on prompts that the target model actually answers, rather than on all harmful prompts. Across several guard systems, recall at a 1% false‑positive rate drops sharply when focusing on answered prompts, with monitors catching refused requests 1.1–6.4 times more often than answered ones. Rewriting prompts to be less explicit dramatically increases compliance and reveals that many harmful requests slip past monitors, especially when phrasing is softened. Fine‑tuning guards on these rewritten prompts improves recall from 0.24 to 0.89 on answered requests and generalizes to unseen benchmarks.
By Sripad Karne
The paper introduces a benchmark for evaluating whether off‑the‑shelf small language models (SLMs) can reliably perform microtasks that support a large language model (LLM) planner, such as auto‑approving shell commands, writing memory, selecting tools, and ranking past turns. Using fixed prompts and confidence‑interval‑aware eligibility thresholds, the authors test several Qwen3 models (0.6/1.7/4/8 B) in FP16 with no tuning and find that none of the 16 configurations meet the eligibility criteria. Quantization to 4‑bit precision further degrades performance, with the eligibility gap tracking model size rather than precision, and the issue persists across different models (e.g., Llama‑3.x) and prompt variations.
By Jundong Hu, Shekar Ramachandran