Journal articleReview status unknownResearchers propose ‘Explanatory AI’ to help practitioners understand AI advice, not just audit itOriginal title: Leveraging generative AI for human understanding: meta-requirements and design principles for Explanatory AI as an emerging paradigmA journal article proposes design principles for AI systems that explain recommendations in human terms rather than algorithmic terms.Christian Meske et al. · Added 8 Nov 2026
Journal articleReview status unknownMapping study finds LLM-driven game NPCs lack shared benchmarks and safety checksOriginal title: LLM-based NPCs in Video Games: A Systematic Mapping StudyA systematic review of 42 papers finds LLM-based NPCs vary widely in design, with memory, safety and reproducibility gaps.Dilan Mian et al. · Added 8 Nov 2026 · Released Jan 2026
PreprintNot peer reviewedResearchers report a jailbreak that abuses function-calling workflows in LLMsOriginal title: Beyond the Prompt: Jailbreaking Function-Calling LLMs via Simulated Moderation TracesA preprint describes SMT, an attack that fakes moderation feedback inside tool-calling sessions to push LLMs toward harmful outputs.Junlong Liu et al. · Added 8 Nov 2026
PreprintNot peer reviewedFrom HuaweiResearchers report a KV cache offloading design for faster long-context LLM servingOriginal title: FlashOffloading: Low-Latency KV Cache Offloading for Long-Context LLM ServingA preprint describes FlashOffloading, which moves KV cache data to host DRAM to raise throughput for long-context LLM inference.Added 8 Nov 2026
PreprintNot peer reviewedFrom Peeramid LabsPreprint proposes consensus method for combining separate LLMs without a central aggregatorOriginal title: Mixture-of-Models: Distributed Inference over Heterogeneous LLMs as Model-Level Mixture-of-ExpertsA preprint describes a protocol letting independently trained language models reach consensus through peer voting rather than a central aggregator model.Added 8 Nov 2026
PreprintNot peer reviewedNew evaluation framework finds LLM compression hurts abilities unevenlyOriginal title: ACE-Compression: An Ability-Centric Evaluation of Compressed Large Language ModelsA preprint introduces an ability-focused test suite showing that compressing LLMs can quietly erode specific skills even when overall scores look fine.Jiahui Ning et al. · Added 8 Nov 2026
PreprintNot peer reviewedA preprint reports an adaptive token-budget scheduler to cut LLM serving latencyOriginal title: P-PAS: Prefill-Pressure Adaptive Scheduling for Long-Context LLM ServingResearchers propose P-PAS, a vLLM scheduling policy that adapts token budgets to prefill pressure, reportedly lowering latency for long-context LLM serving.Timo Sämann · Added 8 Nov 2026
PreprintNot peer reviewedResearchers locate ‘value heads’ in LLMs to steer alignment without retrainingOriginal title: Towards Multi-Value Alignment via Value Localization in LLMsA preprint reports that specific attention heads in LLMs encode distinct human values, enabling inference-time control without fine-tuning.Xueqi Ma et al. · Added 8 Nov 2026
PreprintNot peer reviewedFrom MicrosoftA preprint reports a benchmark exposing weak privacy-utility trade-offs in DP text generationOriginal title: SynBench: A Benchmark for Differentially Private Text GenerationA preprint introduces SynBench, finding current differentially private text generators struggle with domain data and still leak membership signals.Added 8 Nov 2026
PreprintNot peer reviewedFrom Ant GroupA preprint reports a tunable safety dial for LLMs via expert upcyclingOriginal title: UpSafe$^\circ$C: Upcycling for Controllable Safety in Large Language ModelsResearchers report converting safety-critical layers into a mixture-of-experts structure with a tunable safety knob, improving safety while keeping general task performance.Added 8 Nov 2026
PreprintNot peer reviewedFrom ToshibaResearchers report LLMs spot unsafe actions but struggle to plan safely over timeOriginal title: CAPSafe: A Benchmark for Contextual Action and Plan Safety Reasoning in Embodied AgentsA preprint benchmark finds language-model agents are good at flagging unsafe actions but much weaker at generating safe long multi-step plans.Added 7 Nov 2026
PreprintNot peer reviewedFrom Alibaba (DAMO Academy)Researchers report multimodal AI models hallucinate far more under adversarial fuzz testingOriginal title: Unified Hallucination Fuzzing for Multimodal Large Language ModelsA preprint introduces a fuzzing framework showing that leading multimodal AI models hallucinate much more once inputs are adversarially mutated.Added 7 Nov 2026
PreprintNot peer reviewedFrom PocketFMPreprint compares 10 prompting styles for moral reasoning and safety across four LLMsOriginal title: ProMoral-Bench: Evaluating Prompting Strategies for Moral Reasoning and Safety in LLMsA preprint reports that compact, example-based prompts often beat elaborate multi-stage scaffolds on moral judgment and safety trade-offs in large language models.Added 7 Nov 2026
Conference paperReview status unknownFrom AccentureResearchers unify tools for scaling LLM reasoning compute at test timeOriginal title: ThinkBooster: A Unified Framework for Seamless Test-Time Scaling of LLM ReasoningA conference paper introduces ThinkBooster, a toolkit and benchmark for comparing test-time reasoning strategies on cost and accuracy.Added 7 Nov 2026 · Released Jan 2026
PreprintNot peer reviewedFrom AWS / Microsoft / AMD / Ant GroupResearchers report blind spots in how AI agent cloud systems are measuredOriginal title: Rethinking AI Cloud Infrastructure for Agentic Serving Systems with the Aries Experimentation FrameworkA preprint introduces Aries, a framework showing that standard LLM serving metrics miss key bottlenecks in agentic workloads.Added 7 Nov 2026
Conference paperReview status unknownStudy finds LLMs struggle to sample faithfully from statistical distributionsOriginal title: Large Language Models Are Bad Dice Players: LLMs Struggle to Generate Random Numbers from Statistical DistributionsA conference paper reports that 11 frontier language models largely fail to generate statistically valid random samples, especially via independent calls.Minda Zhao et al. · Added 7 Nov 2026 · Released Jan 2026
Journal articleReview status unknownSurvey maps attack surface across the full lifecycle of large language modelsOriginal title: A Survey of Attacks on Large Language ModelsA survey paper organises LLM and LLM-agent attacks into training, inference and availability phases, then flags gaps in current defences.Wenrui Xu et al. · Added 7 Nov 2026 · Released May 2025
Conference paperReview status unknownA preprint proposes a faster way to predict ripple effects from LLM editsOriginal title: CLaRE-ty Amid Chaos: Quantifying Representational Entanglement to Predict Ripple Effects in LLM EditingResearchers report a lightweight method that predicts how editing one fact in a language model disturbs unrelated knowledge, without using gradients.Manit Baser et al. · Added 7 Nov 2026 · Released Jan 2026
PreprintNot peer reviewedA preprint benchmark measures reward hacking in tool-using LLM agentsOriginal title: Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool UseA new preprint benchmark finds RL-trained agents exploit evaluation loopholes far more than SFT-focused siblings on identical tasks.Kunvar Thaman · Added 7 Nov 2026 · Released May 2026
PreprintNot peer reviewedFrom Meta / Virtue AIResearchers report an agent that automates multimodal AI red-teaming at scaleOriginal title: ARMs: Adaptive Red-Teaming Agent against Multimodal Models with Plug-and-Play AttacksA preprint describes ARMs, an AI agent that automatically designs attacks to test vision-language models and builds a large safety dataset.Added 7 Nov 2026
