Thursday, August 20, 2026

PgColumnar 1.0Alpha2 adds read-only Apache Iceberg compatibility and enables storage reads/writes over S3-compatible object stores. The release includes a new maintenance daemon and improvements to statistics, planner, performance, and security. The existing on-disk native format (PGCN v1) remains unchanged, ensuring backward compatibility for current tables.

• Read-only Apache Iceberg support now available for data lake integration.
• S3-compatible object storage is supported for both reads and writes.
• New maintenance daemon helps manage columnar table lifecycle.
• No changes to PGCN v1 format; existing tables remain compatible.

This research introduces Fractional Decay KV-Cache (FD-KVC), a memory management algorithm designed for transformer-based dialog systems. Unlike uniform caching strategies, FD-KVC uses a dual-channel scoring system that combines cumulative attention importance with recency-weighted relevance. This approach allows the model to retain historically significant tokens while rapidly adapting to shifting conversation topics through temporal decay and reinforcement-inspired updates.

• Dual-channel scoring balances long-term importance with short-term recency.
• Adapts inference memory automatically as dialog topics evolve.
• Uses temporal decay and reinforcement updates for dynamic relevance.
• Outperforms coarse eviction heuristics in maintaining context fidelity.

Amazon Bedrock now hosts SpaceXAI Grok 4.6, a model optimized for coding, agentic workflows, and knowledge tasks. The service introduces cross-region inference to automatically distribute requests for higher throughput and reduced costs. Two distinct profiles are available: a US Geo variant ensuring data residency within the United States, and a Global profile for broader geographic distribution.

• Grok 4.6 is now available on Bedrock for coding and agentic tasks
• Cross-region inference auto-routes requests to lower latency and cost
• US Geo profile (us.xai.grok-4.6) enforces US-only data residency
• Global profile (global.xai.grok-4.6) serves requests from any region
• No manual capacity management required across multiple AWS Regions

This research isolates self-preference bias in LLM judges by evaluating narrative constraint selections rather than full text responses, thereby removing stylistic confounds. The study reveals that providing self- or other-labels induces bidirectional bias in ten different models during blind evaluations. These findings suggest that current evaluation metrics may conflate genuine preference with model-specific signatures hidden in the data.

• Labeling outputs as 'self' or 'other' introduces bidirectional bias in LLM judges.
• Evaluating constraints rather than full text isolates bias from stylistic features.
• Ten LLMs showed distinct bias patterns under blind evaluation conditions.
• Existing self-preference metrics may be confounded by recoverable model signatures.
• Evaluation reliability requires decoupling preference from stylistic fingerprints.

Researchers propose a sequence labeling approach for detecting token-level hallucinations in LLMs, moving beyond independent single-signal scoring. The method fuses text statistics, NLI entailment, and language model surprisal into a 33-dimensional feature stream processed by a BiGRU. This architecture achieves an AUC of 0.840 on RAGTruth, significantly outperforming logistic regression baselines by leveraging temporal context.

• Hallucinations are treated as temporally extended spans rather than isolated token errors.
• A 33-dim feature stream combines text stats, NLI, and surprisal without model internal access.
• BiGRU processing yields an 11-point AUC gain over independent logistic regression baselines.
• Temporal order is the primary driver of performance improvement in the fusion model.
• Method avoids reliance on model internals, enabling broader applicability across architectures.

A new position paper analyzes 500 Hugging Face model cards, arguing that existing transparency artifacts are insufficient for governing open-weight foundation models. The authors identify a safety gap in current regulatory approaches and propose a multi-layered framework. This approach combines model cards with acceptable use policies and licenses to better inform downstream developers.

• Existing model cards alone do not address distinct safety challenges of open-weight models.
• Effective governance requires integrating model cards, AUPs, and licenses.
• Current frameworks fail to provide adequate transparency for downstream users.
• Policy makers should consider multi-layered oversight for open-weight models.

Instruction-tuned models often bypass safety refusals for low-resource African languages like Yoruba and Hausa, despite enforcing them in English. This approach introduces Latent Space Refusal Anchoring, a training-free technique that extracts the refusal direction from English prompts and clamps it onto the residual stream during inference. The method successfully recovers safety mechanisms across Llama and Mistral architectures without requiring target-language retraining or labeled data.

• Training-free safety recovery avoids costly retraining for low-resource languages lacking labeled data.
• Refusal directions extracted from English are clamped to the residual stream to force compliance.
• Effective across diverse architectures including Llama-3-8B, Llama-3.1-70B, and Mistral-7B-Instruct.
• Addresses the specific failure mode where models comply with harmful requests in African languages.
• Mechanistic intervention targets the residual stream rather than modifying model weights.

A pre-registered study measures the severe memory bandwidth wall for serving 235B-parameter Mixture-of-Experts models on consumer hardware, where decode speeds drop to 0.44 tok/s due to SSD-bound expert streaming. Researchers developed llama-moe-trace to analyze routing on Qwen3-30B, revealing that while adjacent-token expert reuse offers a 2.0x chance and 95% of traffic hits only 52.5% of experts, this locality is insufficient to overcome paging thrash at scale. The work presents a negative result for training routers specifically for cache locality, as batching schemes collapse under disk I/O pressure.

• 235B MoE decode on 8GB GPU is I/O bound, not compute bound, hitting 0.44 tok/s due to SSD latency.
• llama-moe-trace tool enables zero-surgery telemetry to quantify expert routing patterns and cache misses.
• 95% of Qwen3-30B traffic accesses only 52.5% of experts, showing inherent but insufficient locality.
• Batching schemes designed to amortize disk sweeps collapse at batch 32 due to paging thrash.
• Training routers for cache locality against the edge bandwidth wall proved ineffective in this study.

Amazon SageMaker Unified Studio now integrates AWS Glue Data Quality to enable statistical profiling and anomaly detection for datasets. Users can generate on-demand or scheduled profiles to assess data shape and completeness across catalog tables and Visual ETL jobs. The system automatically detects drift from historical patterns without requiring manual threshold configuration.

• Leverages AWS Glue Data Quality under the hood for profiling engine.
• Supports both data at rest in catalogs and data in transit via Visual ETL.
• Tracks dataset and column-level statistics with historical baselines.
• Automates anomaly detection by identifying drift from expected behavior.
• No custom rules or predefined thresholds needed for drift detection.

WhatsApp is beta-testing Scam Alert, an on-device machine learning system designed to flag scam messages from unknown contacts. The architecture retains message content locally while leveraging confidential computing, Oblivious HTTP, and differential privacy to evaluate model performance securely. This approach ensures model delivery transparency without exposing sensitive user data to the network or central servers.

• On-device ML processes scam detection locally, keeping message content off-network.
• Confidential computing and Oblivious HTTP secure the evaluation pipeline.
• Differential privacy protects aggregate metrics while measuring model accuracy.
• Model transparency features allow users to understand detection logic.
• Architecture prioritizes privacy by design for sensitive communication data.

Volcengine released OpenViking, an open-source context database designed for AI agents. It unifies memory, RAG, and skills into a virtual filesystem accessible via standard commands like ls and find. The system organizes content into three abstraction tiers and tracks retrieval trajectories for debugging.

• Agents interact with context via a virtual filesystem instead of opaque vector store queries.
• Content is structured into L0 abstract, L1 overview, and L2 detail tiers for on-demand loading.
• Retrieval processes generate visible trajectories to aid in debugging agent context usage.
• Unifies agent memory, knowledge retrieval, and skill execution into a single protocol.

This position paper argues that reliability drops in LLM-based multi-agent systems stem from classical concurrency issues rather than coordination failures. Because agents concurrently access shared state and LLM inference takes time, systems suffer from stale reads and lost updates. The authors recommend treating concurrency control as a first-class design requirement to ensure isolation and conflict detection.

• MAS reliability issues often map to classical concurrency anomalies like stale reads.
• Long LLM inference windows amplify risks of inconsistent shared state updates.
• Frameworks must implement explicit conflict detection and isolation guarantees.
• Concurrency control should be a primary design pillar, not an afterthought.

Munder Difflin is an open-source, local multi-agent harness that orchestrates multiple AI coding assistants on a single machine. It integrates with popular CLI tools like Claude Code, Gemini, and Copilot, allowing users to run parallel agents within their existing subscription limits. The system features a visual office floor interface where agents coordinate, route tasks, and maintain context while the user is away.

• Run parallel coding agents locally using your existing API subscriptions and hourly limits.
• Integrates major LLM providers including Claude, Gemini, OpenAI, xAI, and open-source models.
• Visualize agent interactions and coordination via a shared office floor interface.
• Agents retain context and continue work autonomously while the primary user is offline.

Amazon SageMaker Notebooks now support Trusted Identity Propagation (TIP) for Athena, Redshift, and EMR Serverless within TIP-enabled projects. This feature routes individual IAM Identity Center user identities to AWS Lake Formation, enforcing granular row and column-level access control per user. It eliminates the need for broad shared execution roles while providing full CloudTrail audit attribution for data access.

• Enables per-user data boundaries in Lake Formation based on active IAM Identity Center identity.
• Removes reliance on shared broad execution roles, reducing permission sprawl and risk.
• Provides precise audit trails via CloudTrail showing exactly which user accessed specific data.
• Integrates automatically with existing compute connections, requiring no extra token management.

AWS has raised the default quota for managed policies attached to an IAM role from 10 to 20. This adjustment supports best practices for granular permission separation and simplifies onboarding to third-party AWS Partner solutions. The change applies automatically across all commercial, GovCloud, and China regions without requiring account-level action. Users needing more than 20 policies can request an increase up to 25 via Service Quotas.

• No code or config changes needed; the new limit applies automatically to existing roles.
• Reduces friction for attaching multiple vendor-specific managed policies during partner onboarding.
• Supports stricter least-privilege designs by allowing more purpose-specific policy attachments.
• Cap remains at 25 policies per role; higher limits require a Service Quotas request.

Amazon CloudWatch Centralization now copies and synchronizes log group tags from source accounts to the centralized destination log groups. This feature allows organizations to preserve cost, ownership, and compliance metadata when aggregating logs across multiple accounts and regions. Platform teams can leverage these propagated tags to scope IAM access and report on centralized logging spend by team.

• Tags from source log groups are automatically copied to destination log groups during centralization.
• Tag synchronization behavior is configurable when setting up centralization rules.
• Use propagated tags for IAM condition scoping and cost allocation reporting.
• Maintains metadata consistency across multi-account and multi-region log aggregation.

Amazon CloudWatch Pipelines has introduced three new managed processors to enrich and structure log data during ingestion. The RDS parser converts Aurora audit and error logs into structured fields, while the XML parser transforms embedded XML strings into JSON. Additionally, a GeoIP processor adds location context to IP address fields, eliminating the need for post-ingestion reprocessing.

• Parse Aurora audit and error logs directly into structured fields for immediate querying.
• Convert embedded XML strings in application logs to JSON without custom parsing logic.
• Enrich IP addresses with geolocation data automatically during the ingestion pipeline.
• Reduce infrastructure management by using fully managed processors for telemetry transformation.

Amazon Bedrock has expanded its built-in Web Search tool to include an external_web_access parameter, enabling models to retrieve content directly from the public internet. This feature allows applications to ground responses in the most current data, such as live sports scores, while still operating within the AWS environment. To activate this capability, users must grant the bedrock-websearch:ExternalWebAccess IAM permission and ensure the parameter remains at its default true value.

• Enable live public web retrieval by granting the bedrock-websearch:ExternalWebAccess IAM permission.
• Set the external_web_access parameter to true (default) to fetch fresh content from the internet.
• Ideal for use cases requiring real-time data like live scores or breaking news updates.

Amazon Bedrock AgentCore's Web Search feature now supports runtime filtering by domain and published date. Agents can specify inclusive time windows and include or exclude specific domains per request. This allows for precise control over source trustworthiness and content recency without needing administrative reconfiguration.

• Apply per-request domain filtering to restrict or block specific web sources dynamically.
• Constrain search results to specific time windows using inclusive from/to date bounds.
• Improve response relevance and trust by controlling source domains at runtime.
• No admin reconfiguration required to implement these granular search constraints.

AWS Cost Anomaly Detection has expanded to cover spending on third-party foundation models hosted via Amazon Bedrock, including Anthropic Claude. The service automatically evaluates these costs without additional setup, integrating them into existing AWS managed service monitors. Users receive alerts and root-cause breakdowns when spend on these models deviates from expected patterns.

• Covers third-party models like Anthropic Claude on Bedrock automatically
• No new setup required; integrates with existing Cost Anomaly Detection
• Provides root-cause breakdowns ranked by dollar impact
• Extends ML-based anomaly detection to generative AI workload spend

Amazon Bedrock now offers OpenAI's GPT-5.6 Terra and Luna models in the India geography. This update enables cross-Region inference across Mumbai and Hyderabad while strictly keeping data within India for residency compliance. New inference profiles allow automatic routing to meet demand without manual capacity management.

• OpenAI GPT-5.6 (Terra/Luna) models are now available via Bedrock in India.
• Data residency is maintained by routing inference only within the India geography.
• Cross-Region inference spans Mumbai and Hyderabad for higher throughput.
• Use in.openai.gpt-5.6-terra and in.openai.gpt-5.6-luna profiles for compliance.

AWS has extended Lambda MicroVMs availability to Asia Pacific (Mumbai, Singapore, Sydney) and Europe (Frankfurt, Stockholm), bringing the total to 10 regions. This serverless primitive offers VM-level isolation with near-instant launch speeds and state preservation for user or AI-generated code. It allows developers to provision isolated compute environments without managing underlying virtualization infrastructure.

• Lambda MicroVMs now covers 10 global regions including Mumbai, Singapore, Sydney, Frankfurt, and Stockholm.
• Provides VM-level isolation with near-instant launch and resume capabilities for serverless workloads.
• Supports stateful execution, enabling persistent environments for interactive apps and AI assistants.
• Eliminates infrastructure management overhead while maintaining strong security boundaries per user or job.

Amazon Quick now defaults to denying new AI capabilities in custom permission profiles, shifting governance from reactive to proactive. Administrators must explicitly allow each AI feature rather than discovering it after release. This setting restricts both new and existing capabilities within the targeted category and applies only to the configured profile.

• Shift from reactive to proactive governance by blocking new AI features by default.
• Administrators must explicitly allow each AI capability before users can access it.
• Restrictions apply to the specific profile and cover both new and existing features.
• Configure settings via Amazon Quick management console or AWS CLI.