Researchers are developing new approaches to address safety and alignment challenges in large language models, with multiple papers published on arXiv in March 2026 examining different aspects of the problem.
According to a paper on explainable LLM unlearning (arXiv:2603.09980), a new method called targeted reasoning unlearning (TRU) aims to remove undesirable knowledge from pre-trained LLMs while preserving general capabilities. The researchers argue that previous gradient ascent methods resulted in “unintended degradation of general capabilities, incomplete removal of knowledge, and the generation of incoherent responses.” TRU introduces reasoning-based unlearning targets to provide “explicit guidance on what and how models should unlearn,” and the authors report it “achieves more reliable unlearning while preserving general capabilities.”
Meanwhile, research on defensive refusal bias (arXiv:2603.01246) identifies a significant issue with current safety alignment approaches. Based on 2,390 examples from the National Collegiate Cyber Defense Competition, the study found that LLMs refuse defensive cybersecurity requests containing security-sensitive keywords at “2.72× the rate of semantically equivalent neutral requests (p < 0.001).” The highest refusal rates occurred in “system hardening (43.8%) and malware analysis (34.3%),” with the researchers concluding that “current LLM cybersecurity alignment relies on semantic similarity to harmful content rather than reasoning about intent.”
A separate study on offline LLM evaluation for Turkish heritage language education (arXiv:2603.09996) found that “anomaly resistance is not solely dependent on model scale” and that reasoning-oriented models in the 8B-14B parameter range represent the “most balanced segment in terms of cost-safety trade-off.”