An arXiv preprint introduces self-rewriting: an already trained 27B model uses its own wei…
The study models an expert human researcher as a proxy saboteur and asks auditors to deter…
This study audits SAR vehicle recognition on MSTAR and ATRNet-STAR and finds that near-99%…
The paper stress-tests whether a model can deliberately make its chain of thought difficul…
The paper separates three claims often conflated in confidence-based LLM-agent systems: di…
This arXiv preprint defines sycophancy as inappropriate agreement: a model confirms a user…
MInfoVisQA introduces a parallel question-answering benchmark for information-rich visual …
Sleep2Word frames EEG-informed sleep analysis without EEG hardware as a cross-modal discre…
RAW reframes the user-side impact of LLM refusals as an observational causal-inference pro…
Margin Dynamics for Large Language Model Alignment frames pairwise preference optimization…
The paper turns a layerwise learning-rate profile into a single, computable effective rate…