https://uncensoredainews.com/news/dualquant-delivers-calibration-free-4-bit-weight-and-activation-quantization-for-llms-8202026-10-12T12:39:10ZUncensored AI Newsen2026-10-12T12:39:10ZDualQuant: Dual-scale Calibration-free Weight-and-Activation Quantization for LLMshttps://uncensoredainews.com/news/structural-trajectory-scoring-improves-coding-agent-fine-tuning-over-outcome-filtering-8092026-10-12T09:39:01ZUncensored AI Newsen2026-10-12T09:39:01ZTQS Structural Scoring Filters Better Coding-Agent Trajectories for Fine-Tuninghttps://uncensoredainews.com/news/cast-preprint-improves-cross-tokenizer-distillation-for-smaller-models-8082026-10-12T06:38:11ZUncensored AI Newsen2026-10-12T06:38:11ZCAST Preprint Proposes Credit Alignment for Cross-Tokenizer Distillationhttps://uncensoredainews.com/news/can-rubric-tweaks-make-llm-judges-agree-more-without-matching-human-judgments-8072026-10-12T03:35:44ZUncensored AI Newsen2026-10-12T03:35:44ZLabel-Free Rubric Repair Boosts LLM Judge Consensus but Not Human Agreementhttps://uncensoredainews.com/news/safeevo-preprint-maps-how-refusal-circuits-evolve-during-llm-alignment-6852026-10-12T00:34:31ZUncensored AI Newsen2026-10-12T00:34:31ZSafeEvo Preprint Maps How Refusal Circuits Evolve During LLM Alignmenthttps://uncensoredainews.com/news/safety-gaps-across-recurrent-depths-in-looped-language-models-7432026-10-11T21:34:12ZUncensored AI Newsen2026-10-11T21:34:12ZLoopLMs Show Varying Safety Under Jailbreaks Across Recurrent Depths, Preprint Reportshttps://uncensoredainews.com/news/fine-tuned-0-8b-qwen3-5-slm-outperforms-prompted-gpt-5-on-grammar-annotation-7422026-10-11T18:32:46ZUncensored AI Newsen2026-10-11T18:32:46ZFine-Tuned 0.8B Qwen3.5 SLM Outperforms Prompted Frontier Models on Grammar Annotationhttps://uncensoredainews.com/news/narrative-wrapping-bypasses-llm-safety-new-cross-language-benchmark-7412026-10-11T15:31:33ZUncensored AI Newsen2026-10-11T15:31:33ZNarrative Wrapping Bypasses LLM Refusals Across Languages, Preprint Showshttps://uncensoredainews.com/news/fedalphaedit-aligns-null-space-editing-and-merging-for-private-collaborative-knowledge-updates-7402026-10-11T12:30:31ZUncensored AI Newsen2026-10-11T12:30:31ZFedAlphaEdit Aligns Null-Space Editing and Merging for Private Collaborative Knowledge Updateshttps://uncensoredainews.com/news/ranking-prior-alignment-improves-cold-start-credit-scoring-in-low-data-regimes-7382026-10-11T06:28:29ZUncensored AI Newsen2026-10-11T06:28:29ZRanking Prior Alignment Improves Cold-Start Credit Scoring in Low-Data Regimeshttps://uncensoredainews.com/news/preprint-links-kv-cache-quantization-to-linear-attention-via-ram-net-migration-7372026-10-11T03:27:33ZUncensored AI Newsen2026-10-11T03:27:33ZPreprint Links KV-Cache Quantization to Linear Attention via RAM-Net Migrationhttps://uncensoredainews.com/news/lower-reconstruction-loss-can-hurt-low-bit-llm-quantization-new-refinement-method-proposed-7362026-10-11T00:26:40ZUncensored AI Newsen2026-10-11T00:26:40ZLower Reconstruction Loss Can Degrade Low-Bit LLM Quantization, DRQ Refinement Proposed