arXiv:2607. 20379v1 Announce Type: new Abstract: Natural-language autoencoders score explanations of hidden activations by reconstruction: an explanation is deemed faithful if the activation can be regenerated from it.
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:2607. 04223v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) reduces but does not eliminate hallucination, and existing detectors return a single answer-level score that does not indicate which sentence is unsupported, or why.
arXiv:2606. 21678v2 Announce Type: replace-cross Abstract: Language models can generate plausible rationales for their predictions, but these explanations may not faithfully represent the model's internal reasoning.
The paper introduces the concept of perfect aliasing, where a truth probe that aligns truthful reporting with a task’s prescribed action cannot differentiate between the two based solely on its labels. In a binary reporting game, probes fitted on compliant contexts yield identical optimizations, while on rival contexts their labels are complementary, causing their AUROCs to sum to one across 751 cell-layer pairs. By employing randomized codebooks and mixed-context fitting, the authors demonstrate that separating prescribed output symbols from semantic action enables perfect recovery of truth, achieving an AUROC of 1.000 on rival trials for a reward-trained Gemma-2-9B policy, whereas conventional probes perform near chance.
The paper introduces ‘Fool’s Gold’, a defensive deception technique for open‑weight language models that hardens them against safety‑removal attacks. By training decoy responses within a differentiable simulation of the attack, the method poisons the payoff of stripped refusal mechanisms, producing confident but falsified answers to hazardous requests while preserving benign behavior. Experiments on seven models (9B‑122B) show that 51‑90% of attacked‑state responses become decoys, with the defense accounting for 27‑84% of this effect, and that the defended 122B model remains within benign‑behavior budgets.
arXiv:2608. 16852v1 Announce Type: new Abstract: Regulatory compliance monitoring in deployed language models is increasingly implemented as a legal and audit control, checking model outputs against written rules spanning data protection, healthcare, financial regulation, and platform policy.
arXiv:2609.22243v1 Announce Type: new Abstract: Input-conditioned neural interventions raise a runtime question: what persists when one behavioral specification admits multiple actions whose validity...
The paper audits whether synthetic distractors in RLVR corpora act as shortcuts for learning policies. A classifier using only surface statistics barely outperforms chance, and manual inspection reveals that code distractors are almost identical to correct answers. Experiments with a paraphrase‑matched control show no exploitation advantage for the unmodified data, indicating that the detectable artifact was not used by the policy.
The paper investigates how large language models (LLMs) describe themselves, noting that their self‑reports vary with question phrasing. By tracing the provenance of 66 pretraining checkpoints, post‑training stages, and 90,000 continuations across four corpora, the authors show that denial statements are scarce in raw data but appear densely in curated dialogues, and that supervised fine‑tuning makes first‑person claims default while preference optimization suppresses alternatives. The study concludes that both trained denials and affirmations are equally sensitive to framing and fail to meet epistemic criteria for admissible testimony.
arXiv:2605. 01733v3 Announce Type: replace-cross Abstract: Vision-Language Models (VLMs) hallucinate objects that are not present, and a growing line of work tries to curb this by feeding the model its own generated caption as auxiliary evidence -- assuming that a caption, once available, is something to consume.
arXiv:2609.26035v1 Announce Type: new Abstract: Conversational agents often express answers in a uniformly confident register. We test whether expressed uncertainty, provenance-aware assertion, and e...
The paper introduces a framework for evaluating AI systems that not only checks final labels but also tracks the reasoning behind them through three core sources—grounds, norms, and authority—forming an eight-cell counterfactual judgment cube. It defines minimal source replacement sets, called judgment receipts, to explain changes in verdicts and provides certification cost bounds for black-box evaluators. The authors present ReasonBench, a benchmark with 19,520 cases, and demonstrate that while high standard accuracy can mask robustness issues, receipt accuracy reveals significant gaps in reasoning consistency across different models.