https://uncensoredainews.com/news/mechanistic-interpretability-reveals-sycophancy-reduction-does-not-guarantee-stronger-model-refusals-3562026-10-19T07:35:48ZUncensored AI Newsen2026-10-19T07:35:48ZPreprint Explores Sycophancy Reduction and Refusal Behavior via Mechanistic Interpretabilityhttps://uncensoredainews.com/news/vaccinebooster-hybrid-defense-shows-trade-off-between-safer-outputs-and-firmer-refusals-3402026-10-19T04:34:43ZUncensored AI Newsen2026-10-19T04:34:43ZVaccineBooster Preprint Reveals Trade-off in Defenses Against Harmful Fine-tuninghttps://uncensoredainews.com/news/jarq-refines-group-wise-quantization-by-alternating-grid-and-code-optimization-3232026-10-19T01:33:45ZUncensored AI Newsen2026-10-19T01:33:45ZJARQ Preprint Proposes Alternating Refinement to Improve Group-wise LLM Quantizationhttps://uncensoredainews.com/news/new-multilingual-privacy-tagger-matches-encoder-speed-with-strong-redaction-accuracy-3212026-10-18T22:33:08ZUncensored AI Newsen2026-10-18T22:33:08ZarXiv Preprint Presents Multilingual Named-Entity Tagger for Privacy Redactionhttps://uncensoredainews.com/news/tsda-track-preprint-aligns-features-for-cross-modal-object-tracking-3202026-10-18T19:31:33ZUncensored AI Newsen2026-10-18T19:31:33ZTSDA-Track Preprint Aligns Features for Cross-Modal Object Trackinghttps://uncensoredainews.com/news/new-preprint-proposes-dual-constrained-erasure-to-strengthen-llm-unlearning-3122026-10-18T16:30:33ZUncensored AI Newsen2026-10-18T16:30:33ZNew Preprint Proposes Dual-Constrained Erasure to Strengthen LLM Unlearninghttps://uncensoredainews.com/news/ready2blend-maps-natural-language-to-composable-alignment-prompts-3112026-10-18T13:29:38ZUncensored AI Newsen2026-10-18T13:29:38ZReady2Blend Maps Natural Language to Composable Alignment Promptshttps://uncensoredainews.com/news/new-faithfulness-benchmark-tests-if-open-weight-llms-actually-use-injected-geographic-context-3092026-10-18T10:28:38ZUncensored AI Newsen2026-10-18T10:28:38ZFaithfulness Benchmark Probes If Open-Weight LLMs Override Injected Geographic Contexthttps://uncensoredainews.com/news/distilling-real-reports-beats-synthetic-data-for-fine-tuning-open-models-on-radiology-entity-extraction-3042026-10-18T07:27:43ZUncensored AI Newsen2026-10-18T07:27:43ZDistilling Real Reports Outperforms Synthetic Data for Fine-Tuning Open Models on Radiology Label Extractionhttps://uncensoredainews.com/news/vane-descent-aligned-routing-scores-lora-updates-instead-of-tokens-3032026-10-18T04:26:44ZUncensored AI Newsen2026-10-18T04:26:44ZVANE Preprint Proposes Scoring LoRA Updates via Descent Alignmenthttps://uncensoredainews.com/news/new-geodesic-tools-help-llms-navigate-grids-without-heavy-reasoning-2992026-10-18T01:26:08ZUncensored AI Newsen2026-10-18T01:26:08ZVector-Quantized Geodesics Enable Efficient LLM Tool Selection for Spatial Taskshttps://uncensoredainews.com/news/reconstruction-loss-scale-causes-optimization-imbalance-in-llm-quantization-2942026-10-17T22:24:39ZUncensored AI Newsen2026-10-17T22:24:39ZReconstruction Loss Scale Drives Optimization Imbalance in LLM PTQhttps://uncensoredainews.com/news/new-shad-benchmark-targets-explainability-gap-in-time-series-anomaly-detection-2912026-10-17T19:24:10ZUncensored AI Newsen2026-10-17T19:24:10ZarXiv Preprint Proposes SHAD Benchmark for Time Series Anomaly Detectionhttps://uncensoredainews.com/news/preprint-shows-modern-llms-already-act-as-fast-decision-models-without-extra-training-2792026-10-17T16:22:48ZUncensored AI Newsen2026-10-17T16:22:48ZPreprint Explores How Readily LLMs Serve as Fast Jev-Style Decision Modelshttps://uncensoredainews.com/news/brain-alignment-scores-persist-without-true-structure-new-preprint-finds-5142026-10-17T13:21:38ZUncensored AI Newsen2026-10-17T13:21:38ZPreprint Questions Brain Alignment and Cross-Lingual Scores in Language Models