PreprintNot peer reviewedPreprint proposes navigator model to pick replay data against LLM forgettingOriginal title: NADS: Navigator-Guided Data Selection for Mitigating Catastrophic Forgetting in Fine-TuningA preprint describes NADS, a method using a drifted reference model to choose synthetic data that helps fine-tuned LLMs keep prior skills.Jianhao Zhang et al. · Added 9 Nov 2026
PreprintNot peer reviewedFrom Waymo / Adobe Research / Bosch / NVIDIASurvey maps how generative AI could reshape autonomous driving researchOriginal title: Generative AI for Autonomous Driving: Frontiers and OpportunitiesA preprint surveys how generative models like diffusion and LLMs are being applied across the autonomous driving technology stack.Added 9 Nov 2026
PreprintNot peer reviewedFrom MicrosoftA preprint describes a multi-agent system for geospatial disaster analytics outperforming coding agentsOriginal title: GeoFaham: Domain-Specialized Multi-Agent Geospatial Analytics for Disaster ResponseResearchers report a specialised multi-agent system beat general coding agents on a 40-task disaster-analytics benchmark while using far fewer tokens.Added 9 Nov 2026
PreprintNot peer reviewedFrom Patronus AIPreprint finds masked diffusion models outperform larger autoregressive models as world simulatorsOriginal title: Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RLA preprint reports that masked diffusion language models simulate text-based agent environments more faithfully than much larger autoregressive LLMs.Added 9 Nov 2026
PreprintNot peer reviewedFrom MeituanResearchers report a scaling law predicting VLM performance from text-only LLM scoresOriginal title: What Transfers from Text to Vision? Capability Scaling Laws and Transfer Dynamics for VLMsA preprint proposes a scaling law that predicts vision-language model accuracy from a backbone's textual benchmark scores, tested on over 150 trained VLMs.Added 9 Nov 2026
Conference paperReview status unknownResearchers propose agent-based framework to counter OWASP’s top LLM security risksOriginal title: Mitigating the OWASP Top 10 For Large Language Models Applications using Intelligent AgentsA conference paper outlines an AI-agent framework meant to detect and counter the OWASP Top 10 risks for LLM applications.Mohammad Fasha et al. · Added 9 Nov 2026 · Released Feb 2024
PreprintNot peer reviewedFrom Microsoft AIOffline trajectory method rivals online RL for training code-generation LLMsOriginal title: COBALT: Contextual Bandit Learning with Offline Trajectories for Multi-Turn Code GenerationA preprint reports an offline bandit-style training method that matches online reinforcement learning for multi-turn code generation while cutting rollout costs.Added 9 Nov 2026
PreprintNot peer reviewedA preprint links positional encoding and depth to limits on LLM reasoning generalisationOriginal title: A Measure-Theoretic Analysis of Reasoning in Transformers: Structural Generalization and Approximation LimitsA preprint develops a transport-theoretic framework tying position encoding and layer depth to how well Transformers generalise reasoning to new search structures.Yuyang Zhang et al. · Added 9 Nov 2026
PreprintNot peer reviewedA preprint proposes pooling Adam’s optimizer memory for LLM trainingOriginal title: AdamPM: Memory-Efficient LLM Optimization via Moment PoolingResearchers report a modified Adam optimizer that pools second-moment values to cut memory use while keeping training performance similar.Xiaowei Ji et al. · Added 9 Nov 2026
PreprintNot peer reviewedA preprint links power sampling, self-reward RL and self-distillation in LLMsOriginal title: Power Distribution Bridges Sampling, Self-Reward RL, and Self-DistillationResearchers report that a single power distribution connects sampling, self-reward RL and distillation, offering a cheaper alternative to inference-time sampling.Akiyoshi Tomihari et al. · Added 9 Nov 2026
PreprintNot peer reviewedA preprint proposes a construction-based taxonomy for mapping prompt injection attacksOriginal title: SoK: PIAtlas: Decomposing the Unbounded Prompt Injection Attack SpaceResearchers outline PIAtlas, a taxonomy classifying prompt injection attacks by how the malicious instruction subverts the model, and test it against a new corpus.Stavros Damianakis et al. · Added 9 Nov 2026
PreprintNot peer reviewedA preprint benchmark tests multi-agent AI maths reasoning beyond final answersOriginal title: MAMath-Bench: Process-Aware and Robust Evaluation of Multimodal Multi-Agent Mathematical ReasoningA new preprint benchmark checks whether multi-agent AI systems reason reliably, not just whether they get maths answers right.Xinping Wang et al. · Added 8 Nov 2026
PreprintNot peer reviewedFrom Caura.aiPreprint proposes having LLMs grade each other to replace fixed benchmarksOriginal title: Agents-as-Peers: Bias-Controlled LLM Evaluation as Multi-Agent Peer AssessmentA preprint reports a peer-assessment protocol where AI models write, answer and judge each other's test questions, measuring bias along the way.Added 8 Nov 2026
PreprintNot peer reviewedA preprint explores using frozen language models to steer supervised vision trainingOriginal title: Language Guidance for Supervised Vision Training: An Empirical Study of GeneralizationThe authors report that pairing image classifiers with frozen language models can improve generalisation without paired image-text data.Shruthi Gowda et al. · Added 8 Nov 2026
PreprintNot peer reviewedResearchers report RLVR-trained models abandon good reasoning strategies out of distributionOriginal title: Not All Reasoning Generalizes: A Case Study on Learning Generalizable Reasoning Strategies in RLVRA preprint finds that reinforcement-trained language models learn correct step-by-step reasoning in-distribution but fall back to brute-force guessing under distribution shift.Ziqi Pang et al. · Added 8 Nov 2026
PreprintNot peer reviewedResearchers report a self-reward search method lifts SQL generation accuracyOriginal title: SQL-o1: Automatic Text-to-SQL Program Synthesis via Self-Reward Heuristic SearchA preprint describes SQL-o1, a heuristic search framework that the authors say improves text-to-SQL accuracy over prior open-source methods.Shuai Lyu et al. · Added 8 Nov 2026
PreprintNot peer reviewedFrom Bridgewater AIA LabsResearchers report jointly evolving system prompts and weights boosts LLM reasoning gainsOriginal title: Evolutionary System Prompt Learning for Reinforcement Learning in LLMsA preprint describes E-SPL, which evolves system prompts alongside RL weight updates to improve LLM reasoning and generalisation.Added 8 Nov 2026
PreprintNot peer reviewedFrom Alibaba (DAMO Academy)Researchers report a vision-centric medical AI system beating open and proprietary modelsOriginal title: ClinFusion: A Vision-Centric Multimodal LLM System for Holistic Medical UnderstandingA preprint describes ClinFusion, a medical multimodal AI combining 2D and 3D image encoders, reported to beat many rival systems.Added 8 Nov 2026
PreprintNot peer reviewedFrom IBMPreprint introduces vLLM Hook to expose and steer model internals during inferenceOriginal title: vLLM Hook v1: A Plug-in for Programming Model Internals on vLLMA preprint describes vLLM Hook, a plug-in letting developers probe or alter internal states of vLLM-served models.Added 8 Nov 2026
Conference paperReview status unknownResearchers propose four-class scheme to detect human-AI co-written textOriginal title: Towards Fine-grained and Robust Detection of Human-AI Co-authorshipA preprint reports a four-class framework for detecting AI involvement in writing and tests its robustness against adversarial paraphrasing.Ziruo Zhao et al. · Added 8 Nov 2026 · Released Jan 2026
