An arXiv preprint released October 16 2026 finds that a small language model fine-tuned with LoRA in 2.5 hours on an Apple M5 Pro reaches 69.9 percent severity-tier accuracy…
New preprint shows that fine-tuning on carefully constructed benign examples can substantially degrade safety alignment in LLMs by inducing a refusal-to-compliance behavioral…
An arXiv preprint from October 19 2026 shows that simply stating a legal role such as "I am a lawyer" can increase refusal rates on benign legal-domain prompts by 2-20x in four…
An October 19 2026 arXiv preprint presents a large-scale study of parameter-efficient fine-tuning design choices on three MoE model backbones, evaluating module selection…
A September 2026 arXiv preprint applies sparse crosscoders and a new swap readout technique to examine what on-policy distillation actually changes inside student LLMs during…
A September 2026 arXiv preprint hypothesizes that LLM layers divide labor across conceptualization, reasoning and textualization stages and introduces Layer-Informed…
An arXiv preprint dated September 28 2026 introduces PrivacySkills a framework that measures how system instructions and skill level labels affect whether LLM agents select…
New arXiv research on eight instruction-tuned LLMs shows that refusal can sometimes increase safety risks in suicidal ideation contexts. Clinician-guided protocols reduced both…
arXiv preprint from October 16 2026 introduces JEPA-Prune a post-hoc framework that uses a joint-embedding predictive architecture trained on PII-free surrogate data to guide…
A preprint introduces user-side pseudo-anonymization that replaces PII with realistic surrogates, keeping real names, numbers, and details on your device while still letting…
New preprint analyzes fMRI data from real dialogue to compare how large language models align with human brain activity during language production and comprehension, linking…
arXiv preprint dated 2026-09-28 examines compensatory feature injection on Qwen3.5 models using sparse autoencoders. It finds sycophancy cuts do not consistently strengthen…