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.
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.
arXiv:2609.21996v1 Announce Type: new Abstract: Large language models can hold knowledge they do not report. A model may sandbag on a capability evaluation, or answer against what it internally knows...
arXiv:2610.15415v1 Announce Type: cross Abstract: Before an agent acts, it must decide whether action is warranted. A single accuracy cannot distinguish failure to recognize a warranted action from r...
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.
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.
arXiv:2608. 13329v1 Announce Type: new Abstract: A model that behaves differently when it senses it is being tested would undermine the evaluations we rely on, so recent work has sought to read that sense directly from a model's activations.
The paper introduces an audit layer for vision‑language model benchmarks that reuses per‑question artifacts from evaluation runs to assess the benchmark itself. Six detectors analyze saved outputs, and an additional text‑only rerun removes visual inputs to produce per‑question reports and aggregate statistics, including permutation‑null excess and interval estimates. Applying this method to the MMStar benchmark reveals that 20.2% of its questions show no accuracy gain from images, with a small but significant excess in mathematics and science categories, and identifies questions where all models chose the same non‑gold option.
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.
arXiv:2607. 28639v1 Announce Type: cross Abstract: We show that knowledge distillation in small instruction-tuned language models has asymmetric effects on bias.
arXiv:2608. 13591v1 Announce Type: new Abstract: High-confidence errors in large language models are often treated as evidence of fragile internal inference.
The study compares two ways of obtaining predictions from language models fine‑tuned on customer behavior: scoring answer tokens directly versus generating a written rationale and then scoring the resulting answer. Across 13 model‑domain cells covering four retail tasks, scored readouts consistently rank outcomes more accurately than generated readouts, with an AUC improvement ranging from 1.5 to 14.5 points. The authors also find that a third readout—eliciting a probability before any verdict—improves calibration but only when outcome rates are represented in training, and they recommend using generated rationales for interpretability while relying on scored heads for ranking.
The study investigates how false content in training data can influence language models’ downstream decisions even without explicit triggers. By comparing models trained on misleading versus truthful documents in a controlled decision task and a real‑world bushfire case, the authors find a gap between factual answers and the decisions derived from them: correct facts do not always lead to correct decisions, and removing an injected number does not eliminate the misleading narrative. The research highlights that data poisoning can subtly alter model behavior beyond what is detectable through direct probing.