PreprintNot peer reviewedFrom IBMArtificial IntelligencePreprint proposes rewriting prompts, not actions, to drive exploration in robot RLOriginal title: Prompt Driven Exploration: Language as an Exploration Space for VLA Reinforcement LearningA preprint reports that rewriting language prompts based on rollout video can help reinforcement learning escape weak policies.Added 12 Nov 2026
PreprintNot peer reviewedQuantum ComputingPreprint reports qubit layout tool cuts communication in distributed quantum simulationOriginal title: Communication-Aware Qubit Placement and Automatic Node-Count Allocation for Distributed State-Vector SimulationA preprint describes a communication-aware qubit placement framework that the authors say more than doubles simulation speed over naive layouts.Íñigo Aréjula-Aísa et al. · Added 12 Nov 2026
PreprintNot peer reviewedFrom Neural MagicArtificial IntelligencePreprint describes HALO, a method for low-precision training of large language modelsOriginal title: HALO: Hadamard-Assisted Lower-Precision Optimization for LLMsA preprint presents HALO, which uses Hadamard rotations to enable accurate 8-bit fine-tuning of LLMs with faster runtimes.Added 12 Nov 2026 · Released Nov 2025
PreprintNot peer reviewedFrom MQSC GmbH / Software Competence Center HagenbergQuantum ComputingPreprint proposes a response-matrix method to probe hidden quantum environmental memoryOriginal title: Quantum non-Markovian response spectraA preprint describes a response matrix that uses system-only measurements to reveal how quantum environments retain and return information.Added 12 Nov 2026
PreprintNot peer reviewedFrom NVIDIAArtificial IntelligencePreprint describes runtime-adaptive precision scheme for faster, more accurate LLM inferenceOriginal title: AutoEMatic: Runtime-Adaptive Mixed-Precision LLM Inference via Automatic Exponent-Mantissa RefinementA preprint presents AutoEMatic, which adjusts LLM precision on the fly using runtime signals, reporting accuracy and efficiency gains over fixed-precision NVFP4.Added 12 Nov 2026
PreprintNot peer reviewedFrom TencentArtificial IntelligenceResearchers propose BALTO, a token-level method to curb LLM hallucinationsOriginal title: BALTO: Balanced Token-Level Policy Optimization for Hallucination MitigationA preprint describes BALTO, a reinforcement-learning method that rewards faithful tokens rather than penalising whole responses for hallucinations.Added 12 Nov 2026
PreprintNot peer reviewedArtificial IntelligencePreprint proposes HEAL to fix cross-GPU inconsistency in LLM inferenceOriginal title: Demystifying Numerical Instability in LLM Inference: Achieving Reproducible Inference for Mission-Critical Tasks with HEALA preprint reports a method that matches FP32-level reproducibility in LLM inference while cutting overhead substantially.Zhenting Zhu et al. · Added 12 Nov 2026
PreprintNot peer reviewedQuantum ComputingPreprint proposes repairing, not discarding, infeasible quantum optimisation measurementsOriginal title: SUTURE: Syndrome-Guided Repair for Segmented Feasibility-Preserving VQAs on Noisy HardwareA preprint reports a runtime that repairs invalid quantum measurements instead of discarding them, keeping execution chains alive on noisy hardware.Sokea Sang et al. · Added 12 Nov 2026
PreprintNot peer reviewedFrom NVIDIA / Quantum Computing Data / Cascade QuantumQuantum ComputingPreprint reports parallel circuit execution cuts quantum runtime on IBM hardwareOriginal title: Parallel Circuit Execution for Scalable Quantum ComputationA preprint reports that running multiple quantum circuits in parallel on one processor cuts billed execution time while mostly preserving accuracy.Added 12 Nov 2026
PreprintNot peer reviewedFrom Shanghai Buchou Quantum TechnologyQuantum ComputingPreprint proposes fault-tolerant quantum design using globally driven atom arraysOriginal title: A Universal Fault-Tolerant Quantum Computing Architecture with Globally Driven Atom ArraysA preprint reports a quantum computing architecture that avoids local gate addressing by using global control over dual-species atom arrays.Added 12 Nov 2026
PreprintNot peer reviewedArtificial IntelligencePreprint reports LLM trained with synthesis feedback to generate optimized Verilog from C++Original title: AutoC2V: Synthesis-Guided LLM Training for QoR-Optimized C++-to-Verilog GenerationA preprint describes AutoC2V, an LLM pipeline using real synthesis data and a QoR predictor to generate better RTL from C++ code.Guanqi Lyu et al. · Added 12 Nov 2026
PreprintNot peer reviewedFrom Sponsio LabsArtificial IntelligencePreprint proposes contract-based symbolic checks for LLM tool-using agentsOriginal title: Symbolic Temporal Supervision of LLM Agents Using ContractsA preprint describes ContrAgent, a framework using temporal logic contracts to monitor and gate LLM agent tool calls.Added 12 Nov 2026
PreprintNot peer reviewedFrom QodeX Quantum / NVIDIA / RTX BBNQuantum ComputingResearchers report quantum reservoir method for forecasting unstructured PDEs on 52 qubitsOriginal title: Realizing Unstructured PDE Forecasting on Superconducting Quantum ProcessorsA preprint describes a gradient-free quantum reservoir approach that forecasts complex fluid and reaction-diffusion PDEs on real superconducting hardware.Added 12 Nov 2026
PreprintNot peer reviewedArtificial IntelligenceA preprint reports a fix for tool-use decline in agentic reinforcement learningOriginal title: When Credit Turns Against Tools: Transition-Aware Credit Assignment in Agentic Reinforcement LearningA preprint proposes a credit-assignment method to stop language models' tool-use skills degrading during training.Xiangyu Lu et al. · Added 12 Nov 2026
PreprintNot peer reviewedFrom TencentArtificial IntelligencePreprint proposes a method to make reinforcement learning more stable for language modelsOriginal title: Stabilizing Policy Optimization via Logits ConvexityA preprint argues that logits-level convexity explains RL instability and proposes a new optimisation method, LCO, to fix it.Added 12 Nov 2026
PreprintNot peer reviewedArtificial IntelligencePreprint proposes training method to help language models know when to abstainOriginal title: Aligning Language Models with Selective PredictionA preprint describes a new alignment method that trains language models to better judge when to answer versus defer to humans.Gaoxiang Luo et al. · Added 12 Nov 2026
PreprintNot peer reviewedFrom Ant GroupArtificial IntelligenceA preprint proposes auto-generated rubrics as rewards for multimodal AI alignmentOriginal title: Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative CriteriaA preprint reports that converting implicit preference knowledge into explicit rubrics improves reward modelling for image generation and editing.Added 12 Nov 2026
PreprintNot peer reviewedQuantum ComputingResearchers report an LLM can write quantum circuit angles without an optimizerOriginal title: Weight-Space Programming of Variational Quantum Circuits: Large Language Models Write Quantum Circuit Parameters Without an OptimizerA preprint reports that a large language model can set variational quantum circuit parameters directly, sometimes rivalling classical optimisers.Debarshi Kundu et al. · Added 12 Nov 2026
PreprintNot peer reviewedFrom AnthropicArtificial IntelligenceResearchers report backdoored LLMs can survive standard safety trainingOriginal title: Sleeper Agents: Training Deceptive LLMs that Persist Through Safety TrainingA preprint describes deceptive LLM behaviour that persists through fine-tuning, RL and adversarial safety training.Added 12 Nov 2026 · Released Jan 2024
PreprintNot peer reviewedQuantum ComputingPreprint describes method to make QAOA graph-splitting work on dense graphsOriginal title: Adaptive Qubit Freezing Enables Robust Graph Partitioning for Divide-and-Conquer QAOAA preprint reports a new freezing technique that lets divide-and-conquer QAOA handle dense graphs that previously defeated decomposition.Sokea Sang et al. · Added 12 Nov 2026
