arXiv:2608. 08212v1 Announce Type: new Abstract: In-context learning (ICL) can induce emergent misalignment (EM), where narrow misaligned examples alter answers to unrelated questions.
arXiv:2605. 14473v4 Announce Type: replace-cross Abstract: Retrieval-Augmented Generation (RAG) is usually evaluated by whether the final answer is correct.
Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment.
arXiv:2606. 27383v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used as research assistants, yet it remains unclear whether they can calibrate research takeaways to the strength and scope of the supporting evidence.
arXiv:2606. 24267v1 Announce Type: cross Abstract: While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing.
DeflectBench is a new benchmark that evaluates how large language models (LLMs) generate rhetorical fallacies when prompted. The study tests 23,990 generations from four leading models using three deflection strategies (whataboutism, ad hominem, red herring), seven prompt framings, and 80 claims across four controversy levels. Results show that refusal to produce fallacies depends mainly on request structure, with prompt framing and fallacy type dramatically affecting compliance rates.
arXiv:2606. 24267v2 Announce Type: replace-cross Abstract: While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing.
arXiv:2606. 12747v1 Announce Type: new Abstract: Safety-relevant studies of language models, including alignment and jailbreaking evaluations and AI control protocols, often rely on prefilling model outputs.
The study evaluates 12 instruction‑tuned open‑weight LLMs on six causal‑graph benchmarks, testing five prompting strategies and four confidence sources. Findings show that LLMs tend to over‑predict edges, misclassify indirect or reversed edges as direct, and exhibit high over‑confidence, while conventional confidence estimates are unreliable and agreement signals offer limited improvement. The results suggest LLMs should be used as externally validated soft causal priors rather than definitive causal‑structure evidence.
arXiv:2602. 23971v4 Announce Type: replace-cross Abstract: Sycophancy, the tendency of large language models to favour user-affirming responses over critical engagement, has been identified as an alignment failure, particularly in high-stakes advisory and social contexts.
arXiv:2606. 11502v1 Announce Type: cross Abstract: Language models can state that "the Earth orbits the Sun" and, when role-playing Aristotle, assert the opposite.
As large language models (LLMs) grow more capable, they are increasingly deployed in context-rich settings where task inputs are often accompanied by long, partially irrelevant context. In a controlled setting, we find that state-of-the-art models often appear robust to task-irrelevant context at the aggregate level: prepending it to benchmark questions causes little change in overall accuracy.
arXiv:2607. 26929v1 Announce Type: cross Abstract: The same diagnostic result can support or challenge one causal claim yet fail to address another when the claims concern different populations, outcomes, estimands, pathways, or identifying assumptions.