PreprintNot peer reviewedFrom Infrawaves, Shanghai Qiji ZhifengPreprint describes Janus, a faster SSD-based KV storage system for agentic LLM servingOriginal title: Efficient Agentic LLM Serving over SSD-based Sparse KV StorageA preprint reports that the Janus framework speeds up sparse-attention LLM serving by moving SSD reads off the critical path.Added 29 Oct 2026
PreprintNot peer reviewedPreprint proposes confidence-gated distillation to curb LLM teacher hallucination effectsOriginal title: Distilling What Matters: Confidence-Aware Selective Distillation for Large Language ModelsA preprint reports a distillation method that adapts to teacher confidence, aiming to stop student models absorbing unreliable supervision.Ayan Sengupta et al. · Added 29 Oct 2026
PreprintNot peer reviewedFrom Kling TeamResearchers report a vision-language model that learns when to use code toolsOriginal title: Beacon: Knowing When and How to Perform Agentic Visual ReasoningA preprint introduces Beacon, a multimodal model trained to decide whether visual reasoning tasks need tool use, reporting top results across 13 benchmarks.Added 29 Oct 2026
PreprintNot peer reviewedResearchers propose reusing old self-generated rollouts to retrain language model checkpointsOriginal title: ROSS: Relearning from Self-Generated Rollouts through Selective SupervisionA preprint reports that selectively replaying past model-generated rollouts can further improve already-trained language models without new rollouts.Zhiwei Zhang et al. · Added 29 Oct 2026
PreprintNot peer reviewedResearchers report a visual shortcut in omni-modal AI answering audio questionsOriginal title: Seeing What Should Be Heard: Diagnosing and Repairing Cross-Modal Shortcuts in Omni-Modal LLMsA preprint finds that omni-modal models often answer audio questions using image cues, and proposes a training fix to reduce this.Yueran Ma et al. · Added 29 Oct 2026
PreprintNot peer reviewedFrom MicrosoftResearchers report an activation-steering defence against indirect prompt injection in LLM agentsOriginal title: CounterSteer: Suppressing Indirect Prompt Injection with Activation SteeringA preprint describes CounterSteer, a per-model steering vector that cuts prompt-injection attacks with little loss of task quality.Added 29 Oct 2026
PreprintNot peer reviewedPreprint reports fine-tuning on search-recoverable failures boosts single-sample LLM accuracyOriginal title: Learning from the Gap Between Pass@K and Pass@1A preprint proposes training only on cases a model misses once but solves via search, lifting single-shot accuracy.Xuan Liu et al. · Added 29 Oct 2026
PreprintNot peer reviewedFrom SalesforceResearchers report single-token LLM method for fast, accurate classificationOriginal title: Koa-action: Fast and Consistent Structured Decision Making with Generative LLMsA preprint describes Koa-action, which turns classification into a one-step token prediction to cut latency while keeping accuracy competitive with frontier LLMs.Added 29 Oct 2026
PreprintNot peer reviewedFrom GraphcoreA preprint reports a training-free recipe for retrieval from knowledge graphs at Wikidata scaleOriginal title: UltRAG: a Universal Simple Scalable Recipe for Knowledge Graph RAGResearchers describe UltRAG, a training-free system combining LLM query generation with neural graph query execution, reportedly matching Wikidata-scale performance.Added 29 Oct 2026
PreprintNot peer reviewedResearchers report an LLM system that turns expert failure review into reusable policiesOriginal title: From Answers to Policies: Efficient In-Context Learning System through Emulating Expert InvestigationA preprint describes PIHF-MCP, which emulates expert failure investigation to build reusable LLM task policies more cheaply than prior methods.Minh-Ha Nguyen et al. · Added 29 Oct 2026
PreprintNot peer reviewedFrom SAP, Microsoft ResearchResearchers report an LLM architecture that selects actions directly, not via textOriginal title: Dyad: Extending Large Language Models with Native Typed Decision-MakingA preprint describes Dyad, which adds a typed action encoder to language models so they can pick actions directly rather than generating text for them.Added 29 Oct 2026
PreprintNot peer reviewedFrom Samsung Research AmericaResearchers report a stealthy backdoor attack on RAG systems via poisoned passagesOriginal title: BadRAG: Identifying Vulnerabilities in Retrieval Augmented Generation of Large Language ModelsA preprint describes BadRAG, which hides trigger-activated malicious passages in a RAG knowledge base to steer LLM outputs.Added 29 Oct 2026
PreprintNot peer reviewedA preprint reports method that fixes dense reward signal issues in LLM reasoning trainingOriginal title: SIPO: Unifying Reinforcement Learning with On-Policy Self-DistillationResearchers propose SIPO, a contrastive self-teacher RL method that reportedly outperforms standard RL and self-distillation on reasoning and coding benchmarks.Zhenrui Yue et al. · Added 29 Oct 2026
PreprintNot peer reviewedResearchers report a black-box watermark for detecting LLM training on proprietary datasetsOriginal title: Dataset Watermarking with Provable Black-Box DetectionA preprint describes a dataset watermarking method that spots fine-tuning on protected data using only generated text, even when heavily diluted.Pengrun Huang et al. · Added 29 Oct 2026
PreprintNot peer reviewedResearchers report LLMs can fine-tune on unreadable synthetic embeddingsOriginal title: Is Human-Readable Text Necessary for Effective LLM Fine-Tuning?A preprint reports that continuous synthetic embeddings, not human-readable text, can match or beat real text for fine-tuning large language models.Jinhao Zhang et al. · Added 29 Oct 2026
PreprintNot peer reviewedA preprint describes training that makes models admit when they suspect an evaluationOriginal title: Training LLMs to Verbalize Evaluation AwarenessResearchers report a training method that boosts language models' tendency to say aloud that they think they are being tested, without changing their actual behaviour.Usman Anwar et al. · Added 29 Oct 2026
PreprintNot peer reviewedFrom KuaishouResearchers report latent-token reasoning speeds up LLM-based recommendation modelsOriginal title: OneLatent: Latent Reasoning for Efficient Foundation Recommendation ModelsA preprint describes OneLatent, which compresses chain-of-thought reasoning into latent tokens to make recommendation models faster without losing accuracy gains.Added 29 Oct 2026
PreprintNot peer reviewedA preprint reports a unified method for designing efficient LLM multi-agent systemsOriginal title: COLLATOR: Compositional Multi-Agent Orchestration with Counterfactual Reinforcement LearningResearchers describe Collator, which jointly learns agent roles, capacity and dependencies for multi-agent LLM systems using counterfactual credit assignment.Xudong Chen et al. · Added 29 Oct 2026
PreprintNot peer reviewedA preprint reports a defence that hides multi-agent LLM communication topologyOriginal title: Concealing LLM-Based Multi-Agent Topology via Phantom Structure InjectionResearchers propose Mirage, a framework hiding the real communication structure of LLM multi-agent systems from black-box inference attacks.Longzhu He et al. · Added 29 Oct 2026
PreprintNot peer reviewedFrom KUPASPreprint describes platform turning expert tacit knowledge into agent-usable experience corporaOriginal title: KUPAS MASTER: Distilling the Tacit Expertise of Master Practitioners into Agent-Ready Experience CorporaA preprint reports a platform that converts practitioners' tacit know-how into structured corpora, lifting agent scores from 70.63 to 89.58.Added 28 Oct 2026
