PreprintNot peer reviewedA preprint reports synthetic scientific coding problems that boost an LLM’s SciCode accuracyOriginal title: SciWalker: Synthesizing Scientific Coding Problems with Operator Graphs and Execution FeedbackResearchers describe SciWalker, a system that auto-generates scientific coding problems and reports large training gains on a 9B model.Chenxi Li et al. · Added 28 Sep 2026
PreprintNot peer reviewedA preprint reports a graph-based method for fairer step-level credit in agentic reinforcement learningOriginal title: Back to the Definition: Estimating Step-Level Advantages via Trajectory Graphs for Agentic Reinforcement LearningResearchers report a graph-based credit-assignment method, GRAFT, that improves multi-turn agentic RL over GRPO and similar baselines.Xincheng Yao et al. · Added 28 Sep 2026
PreprintNot peer reviewedResearchers report LLM-based red teaming framework severely degrades DRL cyber defendersOriginal title: Trident : How to Break Deep Reinforcement Learning Cyber Defenses (Agentic)A preprint introduces Trident, an LLM agent trained via reinforcement learning that the authors report sharply weakens autonomous cyber defenders.Ryozo Masukawa et al. · Added 28 Sep 2026
PreprintNot peer reviewedA preprint introduces Benchmark Radar, a searchable database of AI evaluation benchmarksOriginal title: Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and EvaluationThe authors describe a living catalog and search tool for finding AI benchmarks, their datasets, code and reported scores.Koutian Wu et al. · Added 28 Sep 2026
PreprintNot peer reviewedResearchers report a calibrated-answer model detects many AI alignment failures without fine-tuningOriginal title: Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment FailuresA preprint finds that a single generic question to the Jev model ranks most alignment failures well, at much lower cost than judge models.Ruoqi Guo et al. · Added 28 Sep 2026
Conference paperReview status unknownResearchers report robot backdoors triggered by the robot’s own actionsOriginal title: Silent Sabotage: Internal State Triggered Backdoor Attacks on LLM-Powered Robotic SystemsResearchers demonstrate a stealthy backdoor in LLM-controlled robots that activates only after a rare sequence of the robot's own prior actions.Doniyorkhon Obidov et al. · Added 28 Sep 2026
PreprintNot peer reviewedA preprint reports a scanner that spots poisoned code-generation models with 97%+ accuracyOriginal title: Detecting Data Poisoning in Code Generation LLMs via Black-Box, Vulnerability-Oriented ScanningResearchers describe CodeScan, a black-box method to detect poisoned or backdoored code-generation models by scanning for known vulnerability patterns.Shenao Yan et al. · Added 28 Sep 2026
Journal article✓ Peer reviewedDynamic defence uses a model’s own layers to resist LLM jailbreaksOriginal title: Dynamic Deep Prompt Optimization for Defending Against Jailbreak Attacks on LLMsAn AAAI conference paper introduces a method that generates defensive embeddings from a model's own layers to resist jailbreak attacks.Doniyorkhon Obidov et al. · Added 27 Sep 2026 · Released Mar 2026
