Saturday, July 25, 2026

Postgres 19 refines the autovacuum process by abandoning the strict catalog-order execution model. The new logic introduces urgency-based prioritization, ensuring tables near transaction ID wraparound thresholds are processed before those merely needing routine statistical updates. This change addresses the previous egalitarian approach where critical maintenance tasks were delayed by less urgent ones.

• Autovacuum workers now sort tables by urgency rather than catalog appearance.
• Tables approaching transaction ID wraparound receive immediate, high-priority attention.
• Routine statistical updates are deprioritized to prevent system shutdown risks.
• Postgres 19 enhances reliability by reducing latency in critical maintenance tasks.

Percona now includes CycloneDX 1.6 Software Bill of Materials in JSON format for all PBM 2.15.0 and PCSM 0.9.0+ release artifacts. This covers binary tarballs, RPM, DEB packages, and Docker images, enabling immediate vulnerability scanning upon deployment. Practitioners can use tools like Trivy or Grype to parse these SBOMs and identify CVEs without delay.

• PBM 2.15.0+ and PCSM 0.9.0+ include CycloneDX 1.6 JSON SBOMs in all artifacts
• Scan tarballs, RPMs, DEBs, and Docker images with Trivy or Grype immediately
• Use 'trivy image --sbom-sources oci' for rapid Docker image vulnerability checks
• Enable day-one CVE visibility without waiting for external security advisories

AWS has launched Claude Opus 5, its most advanced model, available via Amazon Bedrock and Claude API. The release emphasizes production-grade coding capabilities, resilient long-running agents, and enhanced reasoning for complex enterprise documents. Zero data retention (ZDR) compliance is built in to support sensitive workloads.

• Opus 5 handles long-running agents that recover from errors autonomously.
• Coding features adapt strategies dynamically during complex tasks.
• ZDR support enables secure deployment for regulated enterprise data.
• Access is available through Amazon Bedrock and the Claude API.

This study investigates whether layer-wise differences, previously leveraged in Transformer-based models for contrastive decoding, exist within Mixture-of-Experts (MoE) architectures to mitigate hallucinations. The authors find that while shared-expert MoEs lack these internal variance signals, distinct expert activation patterns emerge across different MoE configurations in higher layers. These findings suggest that expert-aware decoding can effectively address hallucinations in MoE models where traditional transformer-based methods fail.

• Contrastive decoding works in MoEs only when distinct expert activation patterns exist, not in shared-expert variants.
• Higher layers in MoEs show significant variance in expert usage, enabling hallucination mitigation strategies.
• Prior transformer-focused hallucination methods may not generalize to MoE architectures without adaptation.
• Expert-aware decoding offers a viable path for reducing hallucinations in large-scale MoE models.

Domyn-Small is a 10-billion-parameter open-weight model released under the MIT license, trained on 9 trillion multilingual tokens. Its post-training pipeline includes continued pre-training to extend the native context to 32K tokens, scaling to 128K at inference. The reasoning capabilities are enhanced through a multi-stage RL phase featuring GRPO with verifiable rewards and DPO across five domains including math, code, and tool calling.

• Open-weight MIT license allows unrestricted commercial and internal usage.
• Native 32K context extends to 128K during inference for long-document tasks.
• Reasoning optimized via GRPO with verifiable rewards and DPO alignment.
• Trained on 9 trillion multilingual tokens for strong cross-lingual performance.
• Specialized fine-tuning covers math, code, QA, and tool calling domains.

Production RAG systems face split-knowledge attacks where individually benign documents create false associations when combined. Existing per-document filters like LlamaGuard cannot detect these structurally invisible threats. TopoGuard addresses this by building a semantic similarity graph of retrieved documents to identify and mitigate these cross-document attacks.

• Split-knowledge attacks bypass per-document safety filters by combining benign inputs.
• TopoGuard uses graph theory to detect false associations across retrieved documents.
• Semantic similarity graphs are built to expose structural attack patterns.
• Current LLM safety tools are insufficient for multi-document context threats.

This paper analyzes preference-based post-training by decomposing parameter updates into spectral components, revealing a consistent head-tail organization. The dominant endpoint shift is carried by a compact spectral head that emerges early, while a heterogeneous residual tail remains. This framework allows preference-induced updates to be isolated, recomposed, and intervened on directly.

• Decomposes LoRA updates spectrally to reveal structural patterns in preference tuning.
• A compact spectral head drives the main behavioral shift and appears early in training.
• Heterogeneous residual tails remain distinct from the dominant head components.
• Updates become modular objects that can be isolated, recomposed, or intervened on.

This paper addresses the performance degradation that occurs when increasing compression ratios in Large Language Models. It proposes combining low-rank decomposition and quantization, challenging the assumption that these methods are orthogonal. The study investigates their interaction to achieve higher compression without significant accuracy loss.

• Existing compression methods hit a bottleneck where higher ratios cause sharp accuracy drops.
• Low-rank decomposition and quantization can be combined to overcome this limit.
• Researchers must understand the interaction between these two techniques, not treat them as separate.
• This approach aims to reduce memory and compute overhead while maintaining model fidelity.

AWS has introduced aws-bench as a research preview to objectively measure how accurately and efficiently AI agents perform real-world tasks on AWS. The benchmark provides a public suite of test cases derived from actual AWS usage patterns, covering investigation, troubleshooting, and infrastructure creation. Each test case pairs natural language queries with specific cloud resource states and ground-truth answers, enabling consistent scoring and failure diagnosis for model providers and researchers.

• Use aws-bench to objectively score AI agent performance on AWS-specific tasks like troubleshooting and infrastructure setup.
• Access public test cases derived from real AWS usage to ensure reproducible and verifiable benchmarking results.
• Leverage defined resource states and ground-truth answers to diagnose agent failures and improve foundation models.
• Evaluate model providers and internal agents using a consistent framework tailored for AWS infrastructure operations.

The paper introduces AlphaAgent, a framework that separates evidence retrieval from final report generation using explicit skill contracts. A dedicated retrieval skill translates user queries into material-specific intents, searches a curated index of over 300,000 papers, and reformulates queries if initial evidence is insufficient. This approach addresses the difficulty of handling heterogeneous tasks within single retrieve-then-generate architectures.

• Decouples retrieval and generation via skill contracts for better task isolation
• Retrieval skill reformulates queries to improve evidence quality in metallurgy
• Targets heterogeneous materials science tasks beyond standard RAG pipelines
• Leverages a curated index of 300,000+ papers for domain-specific accuracy

ComposioHQ has released a curated repository of over 1,000 Claude Skills and plugins designed to enhance AI workflows across platforms like Claude.ai, Cursor, and Gemini CLI. The collection emphasizes real-world actions by leveraging their MCP Gateway, which provides secure access to 1,000+ app integrations with built-in authentication and audit logs. This resource aims to simplify connecting AI agents to external tools for production-grade reliability.

• Curated list of 1000+ production-ready Claude Skills for diverse use cases.
• Supports multiple agents including Claude Code, Codex, Cursor, and Gemini CLI.
• MCP Gateway centralizes access with auth, access controls, and audit logs.
• Enables real-world actions by connecting agents to 1000+ external apps.

A retrieval-augmented, multi-agent LLM framework with human oversight was tested for detecting cutaneous immune-related adverse events in clinical notes. The system improved F1 score to 0.88 and Cohen's kappa to 0.82 compared to manual review, while halving the average review time. This approach demonstrates a scalable method for extracting accurate adverse event data across organ systems.

• Human-in-the-loop LLMs boost accuracy (F1 0.88) over manual review (0.77).
• Inter-rater agreement improves significantly (kappa 0.82 vs 0.50) with AI assistance.
• Average review time is reduced by approximately 50% compared to unassisted process.
• Retrieval-augmented multi-agent architecture enables transparent and scalable data extraction.
• Framework is designed for identifying immune-related toxicities across various organ systems.

Expedia Group launched STAR, an internal observability platform that leverages LLMs to accelerate production incident investigations. The system analyzes service telemetry via structured workflows to generate root cause assessments while maintaining human oversight. Built on FastAPI, Datadog, Celery, Redis, and Langfuse, STAR aims to streamline response efforts without removing engineers from the loop.

• STAR combines service telemetry with LLMs to automate root cause assessment generation.
• The platform uses structured workflows to keep engineers in control of incident response.
• Core stack includes FastAPI, Datadog, Celery, Redis, and Langfuse for observability.
• Aim is to reduce investigation time while ensuring human validation of AI findings.

AWS has generally available the Durable Execution SDK for .NET, allowing C# developers to build resilient, long-running workflows directly within Lambda. This SDK extends the event-driven model by automatically checkpointing progress and pausing execution for up to a year while waiting on external events. Developers can now implement complex scenarios like payment pipelines and AI agent orchestration without custom tracking or external orchestration services.

• Use the idiomatic C# SDK to build durable functions that auto-checkpoint state.
• Pause execution for up to 12 months while waiting on external events or approvals.
• Eliminate the need for custom progress tracking or external orchestration services.
• Ideal for long-running workflows like payment processing and human-in-the-loop tasks.

AWS Lambda durable functions now allow encryption of execution state using a customer managed KMS key, replacing the default AWS-owned key. This update gives users direct control over key rotation and access permissions for durable execution data. The feature addresses strict data governance requirements in regulated industries like finance and healthcare.

• Encrypt durable execution state with your own KMS key instead of the default AWS-owned key.
• Gain granular control over key rotation policies for compliance in regulated sectors.
• Manage access permissions to execution history and state via your managed keys.
• Enables Lambda workflows to meet stricter data governance and audit requirements.

LikeC4 is a modeling language and toolset that generates up-to-date software architecture diagrams directly from code. Inspired by the C4 Model and Structurizr DSL, it allows teams to visualize, collaborate on, and evolve architectural views. Users can customize notation, element types, and nesting levels to fit specific project needs.

• Generates live diagrams from code to keep architecture views current
• Flexible notation allows custom element types and nesting levels
• CLI tool enables quick previewing via npx likec4 start
• Supports collaboration and evolution of software architecture models

AWS Secrets Manager now publishes direct events to Amazon EventBridge whenever a secret value changes. This replaces the previous reliance on parsing multiple CloudTrail API calls like PutSecretValue and rotation events. The new native integration allows for real-time, event-driven workflows to detect active value changes immediately.

• Eliminates need to parse multiple CloudTrail events for secret changes
• Enables direct routing to Lambda, SNS, SQS, or Step Functions
• Supports proactive cache invalidation when credentials rotate or update

Zalando engineered an in-process, client-side load balancer to manage an API processing approximately one million requests per second. This architecture shift delivered more predictable latency and reduced infrastructure costs while improving failure visibility. The design prioritizes direct client-side decision-making over traditional server-side proxying.

• In-process client-side balancing reduces dependency on external proxy layers.
• Handling 1M RPS requires robust internal state management within the client.
• Predictable latency improvements justify the architectural complexity.
• Better failure visibility stems from centralized client-side metrics.
• Infrastructure cost drops result from optimized request routing efficiency.

A security camera from Hanwha shipped with a hardcoded GitHub admin token visible on its web login interface. The token was exposed directly in the HTML source of the login page, granting unauthorized access to the associated repository. This represents a critical failure in credential management for embedded IoT devices.

• Audit all IoT devices for hardcoded credentials in web interfaces or source code.
• Scan repositories for accidental commits of tokens, especially from vendor firmware.
• Treat vendor-supplied hardware as untrusted until credentials are rotated.
• Implement network segmentation to limit blast radius of compromised IoT devices.

Amazon SageMaker AI inference now supports G7e instances in Asia Pacific (Seoul and Tokyo) and Europe (London). These instances utilize up to eight NVIDIA RTX PRO 6000 Blackwell GPUs, delivering 2.3x performance gains over G6e. The expansion allows for lower-latency deployment of large language models up to 70B parameters closer to end users in these regions.

• G7e instances are now live in Seoul, London, and Tokyo for lower-latency AI inference.
• Performance increases by up to 2.3x compared to previous-generation G6e instances.
• Supports models up to 70B parameters with up to 768 GB total GPU memory per instance.
• Features 5th Gen Intel Xeon CPUs and 1,600 Gbps EFA networking for high throughput.

AWS Network Load Balancer now supports listener rules that route connections to different target groups based on the source IP address type. A single dual-stack NLB can direct IPv6 client traffic to IPv6 targets and IPv4 traffic to IPv4 targets, preserving the original client IP end-to-end. This eliminates the previous trade-off of maintaining separate load balancers or losing client IP information through protocol translation.

• Route IPv6 and IPv4 traffic to same-family target groups on a single dual-stack NLB
• Preserve original client IP address end-to-end without protocol translation
• Remove need for DNS-based client splitting across separate load balancers
• Simplify architecture by handling conditional routing at Layer 3

Amazon ECS Service Connect now prioritizes service-to-service traffic within the same Availability Zone, reducing both data transfer fees and latency. The feature dynamically adjusts traffic weights as endpoints scale, maintaining balanced loads while keeping local requests local. This eliminates the previous trade-off between application resilience and cross-zone networking costs.

• Reduces cross-AZ data transfer costs by keeping intra-service traffic local to the originating AZ.
• Lowers latency for service-to-service calls through same-AZ routing prioritization.
• Dynamically adjusts traffic weights during scaling events to maintain load balance.
• Removes the need to choose between high availability and network cost efficiency.

Amazon EC2 I8ge instances are now generally available in Europe (London) and Canada (Central). Powered by AWS Graviton4 processors, these storage-optimized instances deliver up to 60% better compute performance than Graviton2-based Im4gn instances. They also provide up to 55% better real-time storage performance per TB and significantly lower I/O latency.

• I8ge instances launch in EU (London) and Canada (Central) regions today.
• Graviton4 processors offer 60% compute performance gains over Graviton2.
• Third-gen Nitro SSDs improve storage performance by 55% per TB vs Im4gn.
• Local NVMe storage scales up to 120TB with reduced I/O latency variability.
• Ideal for latency-sensitive workloads requiring rapid local storage access.

Amazon EC2 C7a instances are now available in the US West (N. California) region, offering compute-optimized performance powered by 4th Gen AMD EPYC processors. These instances deliver up to 50% higher performance than C6a, featuring AVX-512, VNNI, and bfloat16 capabilities alongside DDR5 memory for increased bandwidth. With support for up to 128 EBS volumes per instance, C7a is designed for latency-sensitive workloads requiring high-speed data access.

• C7a instances launch in US West (N. California), expanding regional availability.
• Performance up 50% over C6a using 4th Gen AMD EPYC Genoa processors.
• DDR5 memory provides 2.25x bandwidth for latency-sensitive workloads.
• Supports AVX-512, VNNI, and bfloat16 for advanced compute tasks.
• Capacity to attach up to 128 EBS volumes per instance increases flexibility.

Amazon EC2 M8a instances, powered by 5th Gen AMD EPYC processors, are now available in the Asia Pacific (Hyderabad) region. These instances offer up to 30% higher performance and 19% better price-performance than M7a, with 45% more memory bandwidth for latency-sensitive workloads. Specific benchmarks show up to 60% faster GroovyJVM and 39% faster Cassandra performance compared to previous generations.

• M8a instances use 5th Gen AMD EPYC (Turin) CPUs with 4.5 GHz max frequency.
• Up to 30% higher performance and 19% better price-performance vs M7a.
• 45% more memory bandwidth improves latency-sensitive workload handling.
• SAP-certified with 12 sizes, including 2 bare metal options.