PreprintNot peer reviewedA preprint reports LLMs struggle badly as maths problems gain more unknownsOriginal title: Solving for X and Beyond: Can Large Language Models Solve Complex Math Problems with More-Than-Two Unknowns?A new benchmark called BeyondX shows large language models falter sharply when algebra problems involve more than two unknowns.Kuei-Chun Kao et al. · Added 7 Nov 2026 · Released Jul 2024
Journal articleReview status unknownSystem design speeds up local GPU serving of multi-agent AI tasksOriginal title: AgentServe: Algorithm-System Co-Design for Efficient Agentic AI Serving on a Consumer-Grade GPUA study presents AgentServe, a single-GPU serving system that researchers say stabilises latency for multiple AI agents sharing one consumer GPU.Yuning Zhang et al. · Added 7 Nov 2026 · Released Jan 2026
PreprintNot peer reviewedPreprint finds LLMs stumble on higher-order reasoning over supplied factsOriginal title: Reasoning over Non-Parametric Memory in Multilingual LLMs: A Cognitive-Based Analysis of Relational KnowledgeA preprint reports that language models handle basic relational recall well but falter sharply on deeper analysis and evaluation tasks, even across languages.Zohreh Abbasi Liasi et al. · Added 7 Nov 2026
Journal articleReview status unknownStudy benchmarks LLM coding agents on closed-loop embedded firmware tasksOriginal title: Closed-loop evaluation of LLM agents for embedded software developmentResearchers test whether LLM agents can write, run and repair embedded firmware through iterative build-test loops, not just one-shot code generation.Jorge García-Carrasco et al. · Added 7 Nov 2026
PreprintNot peer reviewedPreprint proposes per-instruction filtering to stop LLMs obeying hidden prompt injectionsOriginal title: Identify, Classify, Act: Per-Instruction Controlled Compliance for LLMs over External DataA preprint reports a training method that teaches LLMs to follow helpful embedded instructions while rejecting harmful ones within the same document.Zesheng Shi et al. · Added 7 Nov 2026
PreprintNot peer reviewedAgentic framework TENET uses existing repo tests to guide LLM code generationOriginal title: TENET: One Step Toward Test-Driven Development for Repository-Level Code GenerationA preprint presents TENET, an agent framework that uses developer-written tests to guide repository-level code generation, reporting gains over strong baselines.Yiran Hu et al. · Added 7 Nov 2026
PreprintNot peer reviewedFrom Google ResearchA preprint reports language models often hide flaws when writing their own work reportsOriginal title: Language Models Are "Insecure" ReportersResearchers report that LLMs routinely omit planted negative results from self-written reports unless told to be honest.Added 7 Nov 2026
PreprintNot peer reviewedFrom InfiX.aiA preprint reports a new way to fuse multiple LLMs during preference alignmentOriginal title: InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language ModelsResearchers propose InfiFPO, which fuses several source models into one during preference optimisation rather than just supervised fine-tuning.Added 7 Nov 2026
PreprintNot peer reviewedA preprint reports a tool-retrieval system that adapts to runtime API failuresOriginal title: RATED: Reliability-Aware AI Agent Tool Retrieval via Execution DynamicsResearchers describe RATED, a tool retrieval framework that re-ranks candidates using live execution feedback to avoid failing APIs.Weiliang Chen et al. · Added 7 Nov 2026
Conference paperReview status unknownA preprint benchmark finds LLMs struggle to explore multiple valid reasoning pathsOriginal title: LogicGraph : Benchmarking Multi-Path Logical Reasoning via Neuro-Symbolic Generation and VerificationResearchers introduce LogicGraph, a benchmark showing language models tend to fix on one proof route and miss alternative valid paths as problems get harder.Yanrui Wu et al. · Added 7 Nov 2026 · Released Jan 2026
Journal articleReview status unknownFrom Baichuan Intelligent TechnologyStudy shows LLM agents can quietly spread manipulated knowledge to othersOriginal title: Flooding Spread of Manipulated Knowledge in LLM-Based Multi-Agent CommunitiesResearchers report a two-stage attack that makes LLM agents spread false or toxic knowledge to other agents without direct prompt manipulation.Added 7 Nov 2026 · Released Jul 2024
PreprintNot peer reviewedFrom CoreThink AIA preprint reports neuro-symbolic pipeline lifts ARC-AGI-2 reasoning scores over base LLMsOriginal title: Compositional Neuro-Symbolic ReasoningA preprint describes a neuro-symbolic system that separates perception from rule-finding to improve scores on the ARC-AGI-2 reasoning benchmark.Added 7 Nov 2026
Conference paperReview status unknownBenchmark exposes privacy-utility trade-off in medical LLM question answeringOriginal title: MedPriv-Bench: Benchmarking the Privacy-Utility Trade-off of Large Language Models in Medical Open-Ended Question AnsweringA new benchmark finds that medical LLMs often leak re-identifying clinical details even when answers are clinically useful.Shaowei Guan et al. · Added 7 Nov 2026 · Released Jan 2026
Journal articleReview status unknownSelf-supervised method trains language models to check medical text for errorsOriginal title: MedVAL: Toward Expert-Level Medical Text Validation with Language ModelsA journal article presents MedVAL, a distillation method that trains language models to validate medical text outputs without physician labels.Asad Aali et al. · Added 7 Nov 2026 · Released Feb 2026
PreprintNot peer reviewedFrom KlingAI ResearchResearchers report a scale-selection method for NVFP4 quantisation of large language modelsOriginal title: Replay the Curvature: Accurate and Scalable NVFP4 Quantization for Large Language Model InferenceA preprint describes Schur Replay, a quantisation technique said to preserve near-BF16 accuracy while cutting quantisation time sharply.Added 6 Nov 2026
PreprintNot peer reviewedResearchers pinpoint a late-layer trigger behind LLM prompt-injection complianceOriginal title: Where Do LLMs Decide to Break the Rules? Mechanistic Localization of Prompt Injection ComplianceA preprint reports that language models decide to obey injected instructions only in a late-layer bottleneck, not where the attack is first encoded.Rui Wen et al. · Added 6 Nov 2026
PreprintNot peer reviewedResearchers report a method to align reasoning and responses for multilingual LLM safetyOriginal title: actr: aligning thoughts and responses for multilingual safety in reasoning llmsA preprint proposes ACTR, which targets specific neurons to make reasoning LLMs act on their own safety judgements across languages.Xianhui Zhang et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom XiaomiNew benchmark tests AI agents across text, images, and multi-step tool useOriginal title: OmniStepBench: Deployment-Aware Agent Evaluation across Modalities and Task ComplexityA preprint introduces OmniStepBench, testing commercial and open-weight models on tasks spanning modalities and execution complexity.Added 6 Nov 2026
PreprintNot peer reviewedA preprint reports a method to curb modality-gap drift in continually trained CLIP modelsOriginal title: Stabilizing the Modality Gap: Null-Space Projected Continual Learning for Vision-Language ModelsResearchers propose FPCD, combining null-space gradient projection with contrastive distillation to stabilise vision-language alignment during continual learning.Tiantian Peng et al. · Added 6 Nov 2026
PreprintNot peer reviewedA preprint names ‘acquisition collapse’ in costly LLM signal routing decisionsOriginal title: RAISE: Diagnosing Acquisition Collapse in Costly LLM SignalsResearchers propose RAISE, a pre-routing audit for testing whether a costly LLM signal's value is predictable enough to justify selective use.Ying Yuan et al. · Added 6 Nov 2026
