PreprintNot peer reviewedA survey reframes LLM long-context compression as memory lifecycle managementOriginal title: From Long Context to Selective Memory:A Survey of Memory-Centric Compression for Large Language ModelsA preprint proposes a unified framework treating context compression and memory handling in large language models as one cost-budgeted lifecycle.Chaogui Gou · Added 3 Nov 2026
PreprintNot peer reviewedFrom HuaweiPreprint proposes debugging near-miss NPU kernels instead of regenerating them from scratchOriginal title: Don't Regenerate, Debug: A Domain-Specific Agent for Repairing Near-Miss Accelerator KernelsA preprint reports that repairing near-miss accelerator kernels beats regenerating them, with higher success and far fewer tokens.Added 3 Nov 2026
PreprintNot peer reviewedFrom GooglePreprint proposes subspace tuning to cut LLM forgetting during fine-tuningOriginal title: Towards Continual Learning: Mitigating Forgetting With Subspace TuningA preprint reports that restricting fine-tuning updates to non-principal weight directions reduces catastrophic forgetting in large language models.Added 3 Nov 2026
PreprintNot peer reviewedFrom GoogleResearchers report language models rarely reuse representations they learn in-contextOriginal title: Language Models Struggle to Use Representations Learned In-ContextA preprint finds open-weight and frontier models often fail to use semantics they learn from context to solve related tasks.Added 3 Nov 2026
PreprintNot peer reviewedResearchers propose random projection to protect privacy in RAG datastoresOriginal title: Private Retrieval Augmented Generation via Random ProjectionA preprint reports that random projection can protect RAG datastores from data extraction while keeping answer quality higher than differential-privacy methods.Dixi Yao et al. · Added 3 Nov 2026
PreprintNot peer reviewedFrom GoogleResearchers propose formally verified self-evolving LLM agents with zero constraint violationsOriginal title: SEVerA: Verified Synthesis of Self-Evolving AgentsA preprint presents SEVerA, a framework binding LLM agent calls to formal contracts so fine-tuning cannot break correctness.Added 3 Nov 2026
PreprintNot peer reviewedFrom AmazonResearchers report a small 8B model beats larger LLMs at GPU kernel tuningOriginal title: How to Train Your Little Dragon: Teaching a Small Language Model to Construct and Refine GPU Optimization ScriptsA preprint reports an 8-billion-parameter model trained on compiler search data outperforms far larger language models at writing GPU optimisation scripts.Added 3 Nov 2026
PreprintNot peer reviewedFrom AmazonResearchers report training multimodal LLMs without real paired audio-vision dataOriginal title: MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal DataA preprint describes training multimodal models using text proxies instead of real paired audio-vision examples.Added 3 Nov 2026
PreprintNot peer reviewedPreprint proposes an evidence-first AI agent for diagnosing technical faultsOriginal title: LLM-as-an-Investigator: Evidence-First Agentic Reasoning for Robust Interactive Problem DiagnosisA preprint describes an LLM agent that tracks rival hypotheses before diagnosing technical problems, reportedly outperforming simpler baselines.Fabrizio Marozzo et al. · Added 3 Nov 2026
PreprintNot peer reviewedResearchers report sparse circuit tuning boosts Hindi-to-Urdu and Hindi-to-Sindhi translationOriginal title: Translation Circuit Discovery and Targeted Optimization for Low-Resource Multilingual TranslationA preprint describes TITAN, a method that finds and tunes small translation circuits in an LLM to lift low-resource translation quality.Atai Waris Khan et al. · Added 3 Nov 2026
PreprintNot peer reviewedFrom Alibaba GroupA preprint reports ReFlow, a replay-based system for faster LLM reinforcement learningOriginal title: ReFlow: Replay-Driven Decoupling for Efficient RL Post-TrainingA preprint describes ReFlow, which reuses recent rollout trajectories to cut GPU idle time during LLM reinforcement learning post-training.Added 3 Nov 2026
PreprintNot peer reviewedPreprint proposes unified reward model to train memory agents jointly, not in isolationOriginal title: All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory AgentsA preprint presents AiO-Mem, which trains retrieval and memory writing together using one shared reward signal.Bingbing Wang et al. · Added 3 Nov 2026
PreprintNot peer reviewedFrom Multiverse ComputingResearchers report fixes for slow remote KV caching in long-context LLM servingOriginal title: When the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel PrefillA preprint explains why a shared KV cache tier can be slower than recomputation, and reports fixes that cut tail latency sharply.Added 3 Nov 2026
PreprintNot peer reviewedResearchers report multi-model system beats single-LLM pipelines for code generationOriginal title: Multi-LLM Orchestration for High-Quality Code Generation: Exploiting Complementary Model StrengthsA preprint describes PerfOrch, which assigns different LLMs to generate, debug and refine code, reporting higher accuracy and faster programs than single-model approaches.Huashan Chen et al. · Added 3 Nov 2026
PreprintNot peer reviewedFrom IBMA preprint reports a KV-cache eviction method with speed and accuracy gainsOriginal title: DropKV: Decoupling Residual-Output Perturbation for Near-Optimal KV-Cache EvictionA preprint describes DropKV, a per-token KV-cache eviction scheme that the authors say gives near-optimal accuracy with faster inference.Added 3 Nov 2026
PreprintNot peer reviewedFrom Prentis AIResearchers report a new GUI grounding dataset and training method for screen-interaction modelsOriginal title: GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement LearningA preprint introduces a large GUI grounding dataset and a two-stage training method, reporting strong results for an 8-billion-parameter model.Added 3 Nov 2026
PreprintNot peer reviewedFrom TencentResearchers report video AI models often trust misleading on-screen captions over the footageOriginal title: VisualTrust: Video Reasoning under On-Screen Claim FramingA preprint finds video language models often accept a false on-screen claim over visual evidence, and proposes an agent that largely restores accuracy.Added 3 Nov 2026
PreprintNot peer reviewedSurvey maps how LLM agent actions can be traced and audited for trustOriginal title: From Agent Traces to Trust: A Survey of Evidence Tracing and Execution Provenance in LLM AgentsA preprint surveys methods for tracing evidence and execution steps in LLM agents, finding major gaps in current benchmarks.Yiqi Wang et al. · Added 3 Nov 2026
PreprintNot peer reviewedFrom Valeo.aiResearchers report self-supervised tasks sharpen MLLM visual groundingOriginal title: V-GIFT: Boosting Visual Instruction Tuning with Self-Supervised GuidanceA preprint reports that mixing a small share of self-supervised visual tasks into instruction tuning improves vision-centric performance in multimodal language models.Added 3 Nov 2026
PreprintNot peer reviewedFrom SAPA preprint reports TabFORGE, a foundation model for generating synthetic tabular dataOriginal title: TabFORGE: A Tabular Foundation Model for Generative ModellingThe authors report that a new pretrained generator for tabular data better preserves feature dependencies than prior foundation models and dataset-specific baselines.Added 3 Nov 2026
