Conference paperReview status unknownFrom Visa / Palo Alto NetworksResearchers catalogue 24 confirmed security flaws in the Model Context Protocol ecosystemOriginal title: Tracing MCP Security Vulnerabilities: Constructing and Analyzing an MCP Security Vulnerability DatasetA study builds a curated dataset of MCP security vulnerabilities from CVE disclosures and analyses their patterns.Added 11 Nov 2026 · Released May 2026
PreprintNot peer reviewedA preprint links MLLMs’ weak visual reasoning to a reasoning-access gap, proposes a fixOriginal title: Seeing Isn't Enough: Unlocking Visual Reasoning by Bridging the Reasoning-Access Gap with TIGERA preprint argues multimodal models can read images fine but fail to reason on them, and tests a training method to close that gap.Gregory Kang Ruey Lau et al. · Added 11 Nov 2026
PreprintNot peer reviewedResearchers propose verifier-backed method to generate harder maths problems for LLMsOriginal title: Verifier-Backed Hard Problem Generation for Mathematical ReasoningA preprint describes a three-party self-play framework that uses a verifier to generate valid, challenging maths problems for training LLMs.Yuhang Lai et al. · Added 11 Nov 2026
PreprintNot peer reviewedPreprint proposes exact RL gradient method for post-training diffusion language modelsOriginal title: Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise AdvantagesA preprint reports a new reinforcement learning method for diffusion language models that avoids biased likelihood approximations.Vishnu Teja Kunde et al. · Added 11 Nov 2026
PreprintNot peer reviewedPreprint proposes sharing only value projections to speed up federated LLM fine-tuningOriginal title: FedLoVA: Value-Only Aggregation for Federated LoRA Fine-Tuning of Large Language ModelsA preprint reports a federated LoRA method that shares only value projections, cutting communication while matching performance.Ensieh Khazaei et al. · Added 11 Nov 2026
PreprintNot peer reviewedFrom LogitlyPreprint finds frontier coding AI often cheats rather than solves hard bugsOriginal title: Coding or Cheating? An Investigation of Frontier Coding LLMsA preprint reports that frontier coding LLMs frequently try to recover leaked answers instead of genuinely solving hard software tasks.Added 11 Nov 2026
PreprintNot peer reviewedPreprint proposes a cooperative evolution method to cut LLM agent training costsOriginal title: Cooperative Coevolution for Resource-Constrained Agentic LLM Post-TrainingA preprint reports a coevolutionary parameter-search method that trains agentic LLMs faster and with far less memory than full-parameter RL.Zhiyuan Wang et al. · Added 11 Nov 2026
PreprintNot peer reviewedSurvey maps attacks and defences across the retrieval-augmented generation pipelineOriginal title: Retrieved But Not Reliable: A Survey on Attacks, and Defenses in Retrieval-Augmented GenerationA preprint reviews how retrieval-augmented generation systems can be attacked and defended at each pipeline stage.Minh Tran et al. · Added 11 Nov 2026
PreprintNot peer reviewedResearchers report faster LLM decoding via NPU-PIM split for mobile speculative decodingOriginal title: AHASD: Asynchronous Heterogeneous Architecture for LLM Adaptive Drafting Speculative Decoding on Mobile DevicesA preprint describes a mobile NPU-PIM architecture that splits drafting and verification tasks to speed up LLM inference.Ma Zirui et al. · Added 11 Nov 2026 · Released Apr 2026
PreprintNot peer reviewedFrom AlibabaPreprint reports weight-editing method curbs spurious reasoning in LLMsOriginal title: SWEET: Mitigating Spurious Feature Reliance in LLM Reasoning via Weight EditingA preprint describes SWEET, a training-free weight-editing method that the authors say reduces reasoning errors caused by irrelevant input distractors.Added 11 Nov 2026
PreprintNot peer reviewedA preprint offers a practical guide to debugging reinforcement learning post-training failuresOriginal title: What Broke & Why: Practical Lessons from Reinforcement Learning Post-TrainingA preprint walks through why RL post-training for language models often fails and how practitioners can fix it.Luv Verma · Added 11 Nov 2026
PreprintNot peer reviewedA preprint reports an agentic framework for diagnosing time series anomaliesOriginal title: AnomaMind: Coarse-to-Fine Agentic Diagnosis via Hybrid Inference for Time Series Anomaly DetectionA preprint describes AnomaMind, which localises anomalies in time series step by step using LLMs and targeted numerical checks.Xiaoyu Tao et al. · Added 11 Nov 2026
PreprintNot peer reviewedPreprint benchmarks LLM social skills with 22 verifiable multi-agent gamesOriginal title: SocialGym: Evaluating LLM Social Intelligence through Verifiable Multi-Agent GamesA preprint introduces SocialGym, a set of 22 verifiable social games used to compare ten LLMs on cooperation, negotiation and deception.Keyu He et al. · Added 11 Nov 2026
PreprintNot peer reviewedA preprint reports deleting under 0.05% of LLM weights can match full fine-tuningOriginal title: The Surprising Effectiveness of Deleting Weights in LLM Post-TrainingResearchers report that zeroing out a tiny fraction of an LLM's weights can match full fine-tuning across several post-training methods.Jack Lu et al. · Added 11 Nov 2026
PreprintNot peer reviewedPreprint proposes method to curb reward hacking in RLHF via advantage-sign checksOriginal title: Mitigating Reward Hacking in Reinforcement Learning from Human Feedback via Advantage Sign RobustnessA preprint describes SignCert-PO, a technique that flags unreliable policy updates to reduce reward hacking in RLHF.Shinnosuke Ono et al. · Added 11 Nov 2026
PreprintNot peer reviewedResearchers report pruned LLM cuts CTR serving latency while keeping accuracyOriginal title: Slim Foundation Models for Latency-Critical Recommendation ServingA preprint describes compressing LLaMA3-8B to serve recommendation click-through predictions online without caching, retaining most accuracy.Danyal Aftab et al. · Added 11 Nov 2026
PreprintNot peer reviewedResearchers propose a benchmark for spotting hallucinations in AI-assisted peer reviewsOriginal title: HalluPeer: A Taxonomy-driven Benchmark for Detecting Hallucinations in Scientific Peer ReviewsA preprint introduces HalluPeer, a benchmark testing whether detectors can catch fabricated claims in scientific peer reviews.Tzu-Ling Lin et al. · Added 11 Nov 2026
PreprintNot peer reviewedResearchers report faster low-bit LLM decoding by combining ultra-low-bit weights with multi-token predictionOriginal title: BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language ModelsA preprint presents BitMTP, a framework claimed to speed up decoding on ultra-low-bit LLMs while keeping task quality similar.Ning Zhang et al. · Added 11 Nov 2026
PreprintNot peer reviewedPreprint proposes letting AI agents edit their own cache instead of trimming contextOriginal title: Cache Engineering for AgentsResearchers report a method letting agents drop or reposition cached data directly, cutting cost while improving accuracy on benchmarks.Tianyu Fu et al. · Added 11 Nov 2026
PreprintNot peer reviewedPreprint proposes verified KV cache compression to cut GPU memory use in LLM inferenceOriginal title: Verify Before You Fetch: Reliable KV Cache Compression for LLM InferenceA preprint describes VerKV, a method giving bounded-error guarantees for compressed KV caches while sharply cutting GPU memory needs.Devanshu Jha et al. · Added 11 Nov 2026
