PreprintNot peer reviewedResearchers report compact LLM self-trains for graph reasoning without external teachersOriginal title: GraphCert: Bootstrap Agentic Graph Reasoning with Certified Evidence RubricsA preprint describes GraphCert, a framework letting a small language model generate and verify its own training data for reasoning over knowledge graphs.Weiqi Jiang et al. · Added 28 Oct 2026
PreprintNot peer reviewedFrom MemoraX AIA preprint proposes a benchmark to separate real skill reuse from code copying in AI agentsOriginal title: EngramBench: A Capability-Grounded Benchmark for Skill-Evolution HarnessesA preprint introduces EngramBench to test whether coding agents genuinely transfer skills rather than reuse memorised code.Added 28 Oct 2026
PreprintNot peer reviewedPreprint compares six RAG pipelines for scientific QA, finds late interaction winsOriginal title: Re-ranking and Late Interaction Drive Retrieval Quality: A Controlled Comparison of RAG Strategies for Scientific Question AnsweringA preprint reports that ColBERTv2's late-interaction retrieval beat five single-vector RAG strategies on scientific question answering.Bhagyesh Rathi et al. · Added 28 Oct 2026
PreprintNot peer reviewedA preprint reports serverless scheme to curb MoE LLM inference stragglersOriginal title: MoEless: Efficient MoE LLM Serving with Serverless ExpertsResearchers report a serverless framework that predicts and rebalances overloaded MoE experts to cut inference latency and cost.Hanfei Yu et al. · Added 26 Oct 2026
PreprintNot peer reviewedFrom MetaPreprint finds post-training often narrows agentic task coverage, not just boosts accuracyOriginal title: Sharpening Tax in Post-TrainingA preprint argues RL post-training sharpens LLM behaviour towards reliability while shrinking the range of agentic tasks it can solve.Added 26 Oct 2026
PreprintNot peer reviewedA preprint maps shared safety pathways that link LLM refusal across languagesOriginal title: Who Bridges Safety? Identifying and Targeting Cross-Lingual Shared Safety PathwaysResearchers report locating cross-layer safety pathways shared between high- and low-resource languages, then tuning just these to cut unsafe outputs.Shuyi Miao et al. · Added 26 Oct 2026
PreprintNot peer reviewedResearchers report LLMs become easier to jailbreak when induced to use ‘drunk’ languageOriginal title: In Vino Veritas and Vulnerabilities: Examining LLM Safety via Drunk Language InducementA preprint finds that making LLMs mimic drunken text raises jailbreak and privacy-leak rates across five models.Anudeex Shetty et al. · Added 26 Oct 2026
PreprintNot peer reviewedA preprint reports SLM ensembles beating a flagship LLM on instruction-followingOriginal title: Beyond Leaderboards: Tokenomics of Agentic Small Language Model EnsemblesA preprint finds small-model ensembles with a judge-feedback loop outperform a flagship LLM on instruction tests, at lower cost.Alexei N. Skurikhin et al. · Added 26 Oct 2026
PreprintNot peer reviewedResearchers report taxonomy-guided prompting beats fine-tuning for OSS threat detectionOriginal title: Autonomous OSS Threat Detection via Taxonomy-Aligned LLMsA preprint reports that structured prompting with GPT-4 outperforms fine-tuned models in classifying open source supply chain threats.Md. Robiul Islam Niloy · Added 26 Oct 2026
PreprintNot peer reviewedA preprint maps RAG research into four axes: efficiency, defence, interaction, reasoningOriginal title: Mapping the RAG Landscape: A Four Axis Taxonomy of Efficiency, Defense, Interactivity, and ReasoningA survey organises retrieval-augmented generation research into efficiency, security, interactivity and reasoning, highlighting gaps for future work.Meghana Sunil et al. · Added 26 Oct 2026
PreprintNot peer reviewedFrom HuaweiA preprint reports worst-case constraints fix reward hacking in LLM reasoning distillationOriginal title: The Weakest Link: Distilling LLM Reasoning with Worst-Case Constrained Reinforcement LearningResearchers propose a constrained RL method that blocks students from masking a single bad reasoning step with good ones elsewhere.Added 26 Oct 2026
PreprintNot peer reviewedResearchers report fine-tuned small models can support ED triage decisionsOriginal title: Domain-Adapted Small Language Models for Reliable Clinical TriageA preprint reports a fine-tuned 7B open-source model outperformed larger proprietary LLMs at predicting emergency triage levels.Manar Aljohani et al. · Added 26 Oct 2026
PreprintNot peer reviewedResearchers propose a Bayesian method for scoring trust in multi-LLM debate councilsOriginal title: Counting Moves, Weighing Voices: Bayesian Dialectical Argumentation for Calibrated Multi-LLM Councils under Persistent AdversariesA preprint reports a new way to weigh AI model opinions by reliability, improving calibration and resistance to persistent bad actors.Ionel Eduard Stan et al. · Added 26 Oct 2026
PreprintNot peer reviewedPreprint finds multi-agent LLM coding tools cost far more energy for little accuracy gainOriginal title: Engineering Sustainable Agents: A Systematic Comparison of Agentic LLMs for Developer WorkflowsA preprint reports that multi-agent LLM setups for coding tasks use much more energy and time than simple single queries, with limited accuracy benefit.Merve Astekin et al. · Added 26 Oct 2026
PreprintNot peer reviewedA preprint proposes a formula predicting when combining LLMs helps or hurtsOriginal title: Quantifying Diversity of Thought: A Predictive Law of Weighted LLM Ensemble LiftA preprint presents a tested formula for predicting when pooling LLM votes improves or worsens accuracy.Junade Ali · Added 26 Oct 2026
PreprintNot peer reviewedResearchers report no single AI architecture wins at every scale of red-team simulationOriginal title: No One Architecture Fits All: A Cross-Environment Evaluation of Hierarchical Red Team AgentsA preprint finds reinforcement-learning and language-model red-team agents each dominate in different cyberattack-simulation environments, not universally.Ayan Javeed Shaikh et al. · Added 26 Oct 2026
PreprintNot peer reviewedA participatory framework links AI lifecycle biases to specific anticipated harmsOriginal title: ECHO: A Participatory Framework for Bias-Anchored AI Harm AnticipationA preprint proposes ECHO, a framework that elicits how stakeholders link AI biases to potential harms in diagnosis and hiring.Nicoleta Tantalaki et al. · Added 26 Oct 2026
PreprintNot peer reviewedFrom PayPalResearchers find off-the-shelf small language models fail routine agent sub-tasksOriginal title: Measuring the Microtask Eligibility Gap: When Is an Off-the-Shelf SLM Enough for an Agent Harness?A preprint reports that none of 16 tested small-model setups met practical quality bars for agent-harness microtasks.Added 26 Oct 2026
PreprintNot peer reviewedPreprint proposes exploration method to fix regret gaps in Nash preference optimisation for LLMsOriginal title: Efficient Exploration for Iterative Nash Preference OptimizationA preprint introduces ENPO and DENPO, exploration-based methods for Nash preference learning, reporting gains over RLHF and NLHF baselines.Tianlong Nan et al. · Added 26 Oct 2026
PreprintNot peer reviewedFrom BloombergPreprint benchmarks how reliably LLMs cope with retrieval-augmented generationOriginal title: Evaluating the Retrieval Robustness of Large Language ModelsA preprint evaluates whether retrieval augmentation reliably helps large language models, finding robustness varies by task and prompting style.Added 26 Oct 2026
