PreprintNot peer reviewedPreprint proposes META-IR, a meta-learning framework for adaptive LLM safety alignmentOriginal title: META-IR: Meta-Introspective Reasoning for Adaptive Safety Alignment in LLMA preprint describes META-IR, a dual-loop method aimed at helping LLMs adapt safety behaviour against evolving jailbreak attacks.Mahabubur Rahman Miraj et al. · Added 11 Nov 2026
PreprintNot peer reviewedFrom Morgan StanleyResearchers map which test-time tricks actually help tabular foundation modelsOriginal title: Test-Time Compute for Tabular Foundation Models: Mechanisms, Gains, and LimitsA preprint examines how adaptation, aggregation and context construction each affect predictions from pretrained tabular foundation models.Added 11 Nov 2026
PreprintNot peer reviewedFrom KioxiaResearchers report testbed for reusing cached KV data beyond exact prompt prefixesOriginal title: KV-Repurpose: A Testbed for Evaluating Non-Prefix KV-Cache ReuseA preprint presents a vLLM-based testbed that lets LLM servers reuse cached document data even when prompts differ outside the exact prefix.Added 11 Nov 2026
PreprintNot peer reviewedResearchers report video-pretrained model can plan robot actions across tasksOriginal title: Large Video Planner Enables Generalizable Robot ControlA preprint describes a video-based foundation model that generates zero-shot plans robots can execute in new tasks.Boyuan Chen et al. · Added 11 Nov 2026 · Released May 2026
PreprintNot peer reviewedPreprint describes zeroth-order quantization training for LLMs on edge devicesOriginal title: End-to-End On-Device Quantization-Aware Training for LLMs at Inference CostA preprint reports a quantization-aware training method that skips backpropagation to fine-tune large language models on phones and single GPUs.Qitao Tan et al. · Added 11 Nov 2026
PreprintNot peer reviewedA preprint reports agents that learn shareable language rules for decision-makingOriginal title: Policy Learning with a Language BottleneckResearchers propose a framework where AI agents form plain-language rules to guide and explain their policies.Megha Srivastava et al. · Added 11 Nov 2026
PreprintNot peer reviewedA preprint presents RedPhys, a red-teaming platform for physical-world AI models and agentsOriginal title: RedPhys: An Extensible Red-Teaming Platform for Physical-World Models and AgentsA preprint introduces RedPhys, a benchmark for testing vision-language models and agents used in physical-world tasks.Yifei Zhao et al. · Added 11 Nov 2026
PreprintNot peer reviewedA preprint reports agent framework lets sub-agents request mid-task helpOriginal title: Proactive Agentic OrchestrationResearchers describe POrchestra, letting AI sub-agents flag problems mid-task so an orchestrator can adjust them on the fly.Xueyang Feng et al. · Added 11 Nov 2026
PreprintNot peer reviewedPreprint links LLM misalignment to training data the model itself disbelievesOriginal title: Knowingly Wrong: Training LLMs Against Their Own Beliefs can Induce MisalignmentA preprint reports that fine-tuning LLMs on facts the model disbelieves can cause broad misalignment, even without harmful content.Ariana M. Villegas-Suarez et al. · Added 10 Nov 2026
PreprintNot peer reviewedFrom Toyota Motor EuropeA preprint reports a cross-layer probe improves LLM hallucination detectionOriginal title: Cross-Layer Attention Probing for Fine-Grained Hallucination DetectionA preprint describes a probing method that attends across all LLM layers to better detect and reduce hallucinations.Added 10 Nov 2026
PreprintNot peer reviewedResearchers report activation steering in LLMs breaks down under adversarial text tweaksOriginal title: Adversarial Robustness of Activation Steering in Large Language ModelsA preprint finds that steering vectors and their chosen layers both become unreliable once inputs are lightly perturbed.Kien Le et al. · Added 10 Nov 2026
PreprintNot peer reviewedResearchers report tile-level mixed precision speeds LLM inference while keeping qualityOriginal title: TileQuant: Tile as the Precision-Allocation Unit for LLM Mixed-Precision Post-Training QuantizationA preprint describes TileQuant, which assigns precision at the tile level to cut LLM inference cost while preserving output quality.Hanzhi Zhang et al. · Added 10 Nov 2026
PreprintNot peer reviewedFrom Amazon Lab126A preprint reports fine-tuning method FuRA matches LoRA cost but beats full fine-tuningOriginal title: FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral PreconditioningResearchers propose FuRA, a parameter-efficient fine-tuning scheme that reportedly outperforms full fine-tuning while keeping LoRA-level costs.Added 10 Nov 2026
PreprintNot peer reviewedFrom Alibaba GroupResearchers report training method teaches vision-language models to spot their own hallucinationsOriginal title: DreamRealm: Exploring and Materializing Hallucination Worlds for Trustworthy Vision-Language ModelsA preprint describes DreamRealm, a post-training method that pairs a model's own false visual claims with edited images to reduce hallucinations.Added 10 Nov 2026
PreprintNot peer reviewedA preprint finds RL adds little once supervised fine-tuning is done wellOriginal title: Measuring the Marginal Value of Verifier-Based RL Beyond Supervised Fine-TuningA preprint reports that reinforcement learning barely improves on strong supervised fine-tuning in a controlled GPU-diagnosis benchmark.Amit Sagtani et al. · Added 10 Nov 2026
PreprintNot peer reviewedFrom AmazonPreprint proposes authority-based test for prompt injection beyond task redirectionOriginal title: Beyond Instruction Override: Benchmarking Indirect Prompt Injection with Task-Tampering AttacksA preprint benchmarks indirect prompt injection attacks that tamper with an agent's task rather than redirecting it entirely.Added 10 Nov 2026
PreprintNot peer reviewedFrom Moonshot AIA preprint reports a new benchmark testing LLMs on theoretical computer science researchOriginal title: TCSExam: Evaluating Frontier LLMs on the Full Range of TCS ResearchTCSExam tests LLM agents on theorem proving, solver building, constructions and open conjectures, with top agents still far from solving most tasks.Added 10 Nov 2026
PreprintNot peer reviewedFrom MeituanResearchers report MCP agents still stumble on infeasible-goal tasksOriginal title: MCPHallu: Benchmarking Reasoning, Execution, and Memory Hallucinations in MCP AgentsA new benchmark finds frontier MCP agents often fail to recognise unreachable goals and vary sharply across failure types.Added 10 Nov 2026
PreprintNot peer reviewedFrom MOSI IntelligenceMOSS-Image: an open text-to-image system targets dense text and complex layoutsOriginal title: MOSS-Image Technical ReportA technical report describes MOSS-Image, an open text-to-image system aimed at dense text and complex layouts, evaluated against proprietary rivals.Added 10 Nov 2026
PreprintNot peer reviewedFrom Nace.AIPreprint details Drex, a family of decision models rivalling chat models on speedOriginal title: Towards Foundation Decision Models: DrexA preprint reports on Drex, open decision models that return option probabilities directly, beating or matching chat models on accuracy and latency.Added 10 Nov 2026
