AI / ML

Researchers propose KFS-RAG, a defense mechanism for Retrieval-Augmented Generation systems that mitigates information leakage caused by prompt injection attacks. The method uses attention rollout and causal perturbation to identify influential keywords within retrieved context, which then guide an auxiliary LLM to generate compact, keyword-grounded facts. By reformulating the retrieved context into these sanitized facts, the system prevents sensitive database contents from being exposed to the generator.

• KFS-RAG defends RAG against prompt injections that target the retriever or generator.
• Uses attention rollout and causal perturbation to pinpoint critical context keywords.
• An auxiliary LLM converts raw passages into compact, keyword-grounded facts.
• Reformulated context reduces risk of sensitive data leakage in production RAG.

Amazon SageMaker HyperPod now integrates deeper support for the Ray framework, addressing common operational burdens like job hangs and low GPU utilization. The update introduces built-in observability, resilient training capabilities, and managed development environments to streamline the workflow. This allows data scientists to interactively edit and monitor Ray workloads without the overhead of manual Kubernetes management.

• Built-in observability simplifies monitoring for Ray workloads running on Kubernetes.
• Managed development environments enable interactive coding without manual job submissions.
• Resilient training features reduce downtime from job hangs and resource misallocations.
• Accelerated inference support helps optimize model serving performance on unified compute.

AWS Neuron 2.32.0 introduces NKI 0.6.0 with new on-device instructions and variable-size collectives for Trn2 and Trn instances. The release includes 13 new library kernels for MoE training, enabling end-to-end MXFP8 training for blockwise MoE layers and sparse attention. A new agentic development skill automates the porting of transformer models to the vLLM Neuron plugin for inference.

• NKI 0.6.0 supports variable-length all-gather for tensors of different sizes across ranks.
• 13 new library kernels enable MXFP8 forward/backward passes for MoE training.
• Context encoding for DeepSeek-V3.2 sparse attention is now available in the library.
• New agentic skill simplifies porting transformer models to vLLM Neuron plugin.
• PyTorch reference implementations provided for 22 additional NKI kernels.

This paper reviews the phenomenon of model collapse, where generative AI models degrade when trained on their own synthetic outputs. It highlights how the shift toward AI-synthesized data, while solving supply constraints, creates a self-consuming cycle that undermines model trustworthiness. The review synthesizes recent studies investigating the causes of this collapse and explores potential mitigation strategies.

• Synthetic data training risks model degradation through self-consuming cycles.
• Model collapse is a critical trustworthiness issue for next-gen GenAI.
• The paper surveys existing countermeasures to mitigate data quality loss.
• Practitioners must balance data supply needs against model stability risks.

This paper addresses the trust deficit in manufacturing knowledge graphs by integrating four capabilities: SHACL validation, PROV-O provenance, bi-temporal versioning, and graph-native decision objects. These components compose through shared correlation identifiers to provide cross-system provenance and decision traceability. The authors present a unified RDF architecture that enables emergent trust properties not achievable by any single capability alone.

• Integrates SHACL, PROV-O, and bi-temporal versioning into a single RDF trust infrastructure.
• Uses shared URIs and activity IDs to link data validity, origin, and decision context.
• Solves the problem of verifying historical data validity at the time of decision-making.
• Enables cross-system provenance tracking for heterogeneous industrial data sources.

This paper addresses the latency and hardware constraints of current open LLM safety guards by proposing a knowledge-distillation recipe for small models deployable on commodity CPUs. A strong open guard annotates approximately 97,000 prompts across 24 public datasets into seven safety categories, which are then used to train student models of varying architectures. The approach ensures reproducibility and license compliance by partitioning the training corpus according to license boundaries.

• Enables sub-second safety classification on CPUs, bypassing GPU dependency for guard models.
• Students span lexical, shallow, encoder, and generative architectures for diverse deployment needs.
• Training data covers 97,000 prompts from 24 datasets, aligned to a public hazard taxonomy.
• License-aware partitioning ensures compliance and reproducibility for open-source safety layers.