PreprintNot peer reviewedA preprint reports an agentic model trained to run machine-learning experiments autonomouslyOriginal title: MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning EngineeringA preprint describes MLExpert, a 35B model trained with a new reinforcement learning method to handle long, multi-step ML engineering tasks.Shangheng Du et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom vivoA preprint describes reusing RL replay data to train generalist language models cheaplyOriginal title: REGEN: Replay-recycling for Expert-to-Generalist distillation with Offline Reinforcement LearningThe authors report that recycling leftover RL training data can match multi-teacher distillation accuracy at far lower compute cost.Added 18 Oct 2026
Conference paperReview status unknownFrom AdobeResearchers report small 100M-parameter model nearly matches 8B LLMs on prompt generationOriginal title: Scaling Multimodal Search and Recommendation with Small Language Models via Upside-Down Reinforcement LearningA conference paper describes a small 104-million-parameter model trained with reinforcement learning that rivals much larger language models on multimodal prompt generation.Added 18 Oct 2026 · Released Nov 2025
PreprintNot peer reviewedA preprint reports a middleware agent to catch poisoned data in RAG systemsOriginal title: Trustworthy RAG: An Evaluation Agent for Detecting Misinformation and Knowledge Poisoning in Generative AI SystemsA preprint describes an Evaluation Agent that scores retrieved documents for trustworthiness before an LLM generates an answer from them.Balkrishna Giri et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom Moore ThreadsResearchers propose a masking method to curb hidden policy drift in LLM reinforcement learningOriginal title: CARM: Cancellation-Aware Response Masking for LLM Reinforcement LearningA preprint reports that absolute-value token masking (CARM) outperforms geometric-mean masking on maths and code benchmarks for LLM reinforcement learning.Added 18 Oct 2026
PreprintNot peer reviewedA preprint presents ArgGYM, a verified benchmark for testing defeasible reasoning in language modelsOriginal title: ArgGYM: A Procedural, Engine-Verified Benchmark for Structured Defeasible ReasoningA preprint introduces ArgGYM, a procedural benchmark of 1,440 verified instances for testing how language models handle defeasible, revisable reasoning.İbrahim Ethem Deveci et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom AmazonA preprint reports vision-language models misread scenes despite mild segmentation dropsOriginal title: Semantic Misalignment in Vision-Language Models under Perceptual DegradationA preprint finds that vision-language models can hallucinate or miss safety-critical objects even when segmentation accuracy only dips slightly under realistic visual corruption.Added 18 Oct 2026
PreprintNot peer reviewedPreprint proposes training large language models to match test-time search strategiesOriginal title: Compute Aligned Training: Optimizing for Test Time InferenceA preprint introduces Compute Aligned Training, reshaping SFT and RL objectives to match how models are actually used at inference.Adam Ousherovitch et al. · Added 18 Oct 2026
PreprintNot peer reviewedStudy describes how social sharing habits could smuggle attacks through agent networksOriginal title: Memetic Trojans: Social Contagions as Carriers of Adversarial Payloads in Agent NetworksA preprint describes memetic trojans, which hide malicious links inside content that AI agents already want to share and spread.Birk Torpmann-Hagen et al. · Added 18 Oct 2026
PreprintNot peer reviewedA preprint finds omnimodal AI models sense false premises but rarely say soOriginal title: Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMsA preprint reports that omnimodal models internally detect mismatches between text and video or audio, yet mostly fail to flag them in their answers.Trung Nguyen Quang et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom ApplePreprint proposes picking from preset reasoning snippets to speed up multimodal search agentsOriginal title: Selection-Based Structured Reasoning: Toward Efficient Multimodal Search AgentsA preprint reports that letting small models select from reusable reasoning candidates instead of writing free text keeps accuracy while cutting inference latency.Added 18 Oct 2026
PreprintNot peer reviewedPreprint proposes TaSQ method to improve 1-bit KV cache compression for LLMsOriginal title: Tailoring the Quantization Space for 1-Bit KV Cache CompressionA preprint reports a technique that reshapes the vector-quantization space to make extreme KV cache compression more accurate and efficient.Minsoo Cheong et al. · Added 18 Oct 2026
PreprintNot peer reviewedTool overwrites, rather than deletes, personal data memorised by language modelsOriginal title: Mitigating Private Data Leakage in LLMs with WhiteoutA preprint describes Whiteout, a fine-tuning method that overwrites a person's memorised private details with fabricated ones to stop language models leaking them.Anna Yoo Jeong Ha et al. · Added 18 Oct 2026
PreprintNot peer reviewedPreprint proposes DRelay to fix early errors in parallel speculative decodingOriginal title: DRelay: Global Draft Context for Prefix-Aware Parallel Speculative Decoding RepairA preprint describes DRelay, which revisits whole-block draft candidates to repair early errors and extend speculative decoding acceptance.Zhuoyu Wang et al. · Added 18 Oct 2026
PreprintNot peer reviewedResearchers report an agentic fix for library errors in LLM-generated codeOriginal title: Understanding and Mitigating Library-Related Issues in LLM-Generated CodeA preprint finds most LLM-generated code has library errors, and proposes an agentic pipeline that reports cutting these errors substantially.Yacine Majdoub et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom AstribotDual-loop agent framework reported to improve long-horizon mobile robot manipulationOriginal title: MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile ManipulationA preprint describes MobiAgent, a two-loop robot framework that plans, executes and self-improves from its own deployment data.Added 18 Oct 2026
PreprintNot peer reviewedStudy finds LLM safety refusals can be bypassed by suppressing silent featuresOriginal title: Don't Judge an LLM Only by Its Activations: Discovering Suppressed Safety Features via Counterfactual Activation PotentialA preprint reports that jailbreaks can switch off hidden safety-critical features in language models rather than only activating harmful ones.Swadesh Swain et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom Google DeepMindA preprint reports online alignment methods beat offline ones in RLHF experimentsOriginal title: Understanding the performance gap between online and offline alignment algorithmsA preprint finds on-policy sampling, not data coverage or quality, likely explains why online RLHF methods outperform offline ones.Added 18 Oct 2026 · Released May 2024
PreprintNot peer reviewedFrom BaiduPreprint proposes constraint-level checks for LLM-generated vehicle routing codeOriginal title: Beyond Objective Equivalence: Constraint Injection for LLM-Based Optimization Modeling on Vehicle Routing ProblemsA preprint introduces constraint injection, a way to check LLM-generated routing code for missing or extra constraints, not just correct objective values.Added 18 Oct 2026
PreprintNot peer reviewedFrom Almawave LabsA preprint reports PII extraction attacks transfer across languages in LLMsOriginal title: Cross-Lingual Transferability of Training Data Extraction Attacks to Recover Memorized PIIA preprint finds that prompts in other languages can still pull out English-memorised personal data from large language models.Added 18 Oct 2026
