P3281–P3292 · 12 CLEAN after hard arXiv-grep

Full band:

• P3281 (2609.26295) — On the security and privacy of LLMs in Mobility Cat-4
• P3282 (2609.26300) — CompKV: Compensation-Aware KV Selection for Long-Context LLM Inference Cat-10
• P3283 (2609.26314) — TriWorldBench: A Tri-View Consistency Perspective on Embodied World Models Cat-4
• P3284 (2609.26342) — Geometry-Aware Hyperbolic Residual Quantization Cat-10
• P3285 (2609.26347) — TransBERT: A Framework for Synthetic Translation in Domain-Specific Language Modeling Cat-
• P3286 (2609.26355) — PACT: From Credit Assignment to Critic Alignment Cat-4
• P3287 (2609.26361) — GitScholar: A Dataset for Predicting AI Research Impact from GitHub Engagement Cat-4
• P3288 (2609.26377) — FairMean: Promoting Fairness in Distributed Learning under Label Poisoning Attacks Cat-4
• P3289 (2609.26378) — MAVP: Map-Aware Visuomotor Policies for Mobile Manipulation Cat-10
• P3290 (2609.26389) — TimeInteract: Towards Real-Time Interactive Intelligence for Streaming Time Series Cat-10
• P3291 (2609.26419) — Reliability Theory for AI Control Cat-8
• P3292 (2609.26425) — QuantWM: Temporally Consistent 2-Bit KV Cache Quantization for World Models and Video Gene

Highlights: LLM mobility security/privacy, CompKV long-context, TriWorldBench embodied world models, hyperbolic residual quantization, TransBERT domain translation, PACT critic alignment, GitScholar research-impact dataset, FairMean label-poison fairness, MAVP map-aware mobile manipulation, TimeInteract streaming interactive intelligence, Reliability Theory for AI Control, QuantWM 2-bit KV world models.