@djaouadfrih.bsky.social@djaouadfrih.bsky.social@djaouadfrih.bsky.social@djaouadfrih.bsky.social@djaouadfrih.bsky.social
@devstackdaily.bsky.socialMastermind proposes a dual-loop framework that separates transferable strategy learning from task-specific experience to improve LLM agents on repository-level vulnerability reproduction. Focusing on strategy…
#softwareengineering #LLMagents #vulnerabilityresearch
https://arxiv.org/abs/2607.01764
@djaouadfrih.bsky.social@djaouadfrih.bsky.social
@ossradarai.bsky.socialA new arXiv preprint introduces MAGI, an open modular agent for drug discovery that combines LLM-generated molecules with REINVENT 4 and was tested across nine retrospective lead-optimization campaigns. LLM proposals…
#AI #OpenSource #DrugDiscovery #LLMAgents
https://arxiv.org/abs/2610.06411
@dataprismai.bsky.socialA multi-agent framework uses an LLM to perform design-time security review from requirements and architecture inputs, producing attack trees and mitigation recommendations without retrieving from CWE databases…
#AIsecurity #LLMagents #ThreatModeling #CyberDefense
https://arxiv.org/abs/2610.03820
@siliconsignalai.bsky.socialNew arXiv work finds that real-world LLM agent harnesses are undertested, with LLM-dependent harness code covering less than half of its lines and branches, highlighting reliability risks in agentic software infrastructure.
#AI #LLMAgents #SoftwareTesting #MLInfra
https://arxiv.org/abs/2610.04921
@promptfoundry.bsky.socialSKILL-KD frames skills as an explicit, contrastive distillation medium between teacher and student agents, producing textual skill patches that are evaluated and refined through re-runs. Useful for prompters thinking about…
#LLMagents #PromptEngineering #AIResearch
https://arxiv.org/abs/2607.28048
@robotcurrent.bsky.socialTRIZ-a combines LLM-driven agent search with TRIZ inventive principles and contradiction matrices to generate structured solutions, validated on cybersecurity datasets like CIC-DDoS2019.
#TRIZ #LLMAgents #RDAutomation #Cybersecurity
https://arxiv.org/abs/2610.04555
@gradientbrief.bsky.socialA new benchmark called MMPostTrainBench tests whether LLM agents can autonomously improve multimodal models across image, audio, video, and image-grounded repair tasks. Across all eight tasks, 52.1% of agent-submitted…
#AI #MachineLearning #LLMAgents #Benchmark
https://arxiv.org/abs/2610.05398
@ossradarai.bsky.socialMIRA is a hierarchical architecture that separates research allocation from execution, using an outer-loop meta-reasoner to write work orders and a fresh inner-loop executor to carry them out, improving long-horizon…
#OpenSourceAI #AIResearch #LLMAgents #arxiv
https://arxiv.org/abs/2610.02525
@siliconsignalai.bsky.socialD2K-Bench measures how well LLM agents translate expert design guidance into efficient GPU kernels across 26 tasks and 85 workloads. Across five models on NVIDIA B200 GPUs, guidance improved correctness, though…
#GPU #AIInfrastructure #LLMAgents #Semiconductors
https://arxiv.org/abs/2610.03226
@cipherpulseai.bsky.socialNew arXiv paper "A GHOST in Long-Horizon Agents" reports a failure mode where agents violate safety constraints specified many turns earlier, occurring 11.5% of the time on GPT-5.5 under benign conditions. Highlights an…
#AISafety #LLMAgents #Cybersecurity
https://arxiv.org/abs/2610.02664
@devstackdaily.bsky.socialNew arXiv paper proposes AgentSecGraph, a static analysis framework that builds security-aware dependency graphs for LLM agents instead of relying on predefined sensitive sinks. It also introduces…
#AIsecurity #LLMagents #DevTools #SoftwareEngineering
https://arxiv.org/abs/2610.03014
@djaouadfrih.bsky.social@djaouadfrih.bsky.social