Thursday, July 16, 2026

This research shifts focus from full query generation to the prerequisite step of retrieving relevant tables and columns from massive schemas that exceed model context windows. The authors introduce a corpus-adaptive embedding approach and recast five major text-to-SQL datasets into retrieval tasks across diverse enterprise-scale structures. Initial results indicate that standard embedding models struggle with this granularity, highlighting a critical gap in current LLM pipelines for handling large database schemas.

• Schema retrieval is a distinct bottleneck before SQL generation in large-scale DBs.
• Corpus-adaptive embeddings outperform generic off-the-shelf models for this task.
• Evaluation now includes table and column granularity across five standard datasets.
• Existing text-to-SQL benchmarks reveal significant gaps in context management.
• Practitioners should prioritize schema indexing strategies over raw query generation.

Article URL: https://mort.coffee/home/sqlite-editions/ Comments URL: https://news.ycombinator.com/item?id=48928135 Points: 266 # Comments: 106

• Read the full item at Hacker News (100+ points).

Researchers introduce interventional grounding audits to test whether LLM chain-of-thought steps genuinely depend on stated premises. The black-box technique substitutes a target predicate in a single premise and checks if the model's normalized conclusion changes. Evaluated on ProntoQA with GPT-4o, the method achieved an F1 score of 0.806 in detecting proof-tree dependencies.

• Proves CoT reasoning may appear logical without true premise dependency.
• Uses black-box predicate substitution to isolate step-level causal links.
• Validated on ProntoQA with GPT-4o showing 0.806 F1 for dependency detection.
• Offers a concrete metric for auditing LLM logical grounding capabilities.

Research into attention-based KV cache eviction reveals that non-stationary filtering disproportionately retains structural noise over content. In schema-dense inputs like nested JSON, delimiters and whitespace carry significantly more attention energy than actual data values. This structural-role bias causes exact-match accuracy to collapse from 88% to 0% at a 5% compression budget due to signal degradation.

• Attention mass is a flawed metric for content retention in structured data contexts.
• Structural tokens like delimiters dominate KV cache, starving content tokens of space.
• Aggressive KV eviction (5% budget) can cause total accuracy failure on JSON schemas.
• Current eviction algorithms need structural-aware weighting to preserve signal integrity.

OriginBlame introduces a system that tracks author identity through data processing pipelines at the record and token level, addressing the gap in locating specific training data for removal requests. Unlike existing tools that operate at the file or dataset level and force over-deletion, this approach resolves revocation requests into precise forget sets via deterministic queries. Evaluation on nearly 220,000 Wikipedia pages shows it reduces dataset-level over-deletion from 101x down to 1.3x, with integration overheads ranging from 1.3% to 19% depending on the pipeline.

• Eliminates catastrophic over-deletion by pinpointing exact records for unlearning requests.
• Propagates author identity through pipelines to create deterministic forget sets.
• Adds 1.3-4.0% throughput overhead on HuggingFace and up to 19% on Datatrove.
• Addresses practical gaps in data contributor removal compliance for model trainers.

This paper addresses the engineering instability of scaling Generative Flow Networks to large language models, where learned prompt-conditional partition functions often cause gradient issues. The authors demonstrate that this partition function, previously considered essential for normalization, can be safely removed. This simplification reduces overhead and improves stability in post-training pipelines for reasoning models.

• Learned partition functions in GFlowNets cause gradient instability at scale, adding unnecessary engineering overhead.
• Replacing the partition function with a simpler alternative stabilizes training for large reasoning models.
• Distribution-matching RL remains viable for LLMs if the normalization component is decoupled from prompt conditioning.
• Simplifying the GFlowNet architecture enables better scaling for math and code generation tasks.

This paper analyzes on-policy distillation (OPD) in LLM post-training, identifying it primarily as an exploration catalyst that steers students toward correct reasoning paths via dense token-level guidance. The study reveals that prompt diversity outweighs per-problem sampling counts and that OPD effectiveness depends entirely on guiding signal quality. A critical Student-Teacher Mismatch pathology emerges when large distributional gaps between teacher and student models derail this exploration process.

• OPD acts as an exploration catalyst, not a capability ceiling expander.
• Prompt diversity is more critical than high per-problem sampling rates.
• Guiding signal quality is the single most important factor for success.
• Large teacher-student distribution gaps cause Student-Teacher Mismatch pathologies.
• Monitor distributional gaps to prevent exploration derailment during training.

Amazon GuardDuty now includes AI Protection to monitor AWS AI services like Bedrock and SageMaker for targeted threats. The feature analyzes CloudTrail events to identify risks such as prompt injection, anomalous model invocations, and cost harvesting attacks. This provides automatic visibility into AI workload security without requiring custom tooling or manual configuration.

• GuardDuty now covers Bedrock and SageMaker for AI-specific threat detection.
• Automated monitoring detects prompt injection and anomalous invocation patterns.
• Cost harvesting attacks via excessive GPU and token consumption are flagged.
• No custom tooling or manual configuration is required for setup.

AWS Security Hub now automatically discovers and catalogs AI workloads across the organization to address visibility gaps. It leverages AWS Config resources to inventory assets from managed services like Amazon Bedrock, Bedrock AgentCore, and Amazon SageMaker without requiring extra configuration. This central view helps security teams connect AI assets to active threats and misconfigurations.

• Security Hub now provides a continuously updated, organization-wide view of AI assets.
• Automatically discovers AI workloads using AWS Config resources from managed services.
• Covers Amazon Bedrock, Bedrock AgentCore, and Amazon SageMaker out of the box.
• Requires no additional configuration for inventorying these specific managed AI services.
• Enables security teams to identify and secure AI assets they previously could not see.

Google DeepMind's Gemma-4-E2B-it is available in Amazon SageMaker JumpStart, offering a multimodal model optimized for local execution. It processes text, image, and audio inputs to generate text, featuring a built-in step-by-step reasoning mode. Key capabilities include object detection, document parsing, video understanding, native function calling, and code generation across multiple languages.

• Gemma-4-E2B-it is now accessible via SageMaker JumpStart for AWS customers.
• Model supports multimodal inputs: text, image, and audio with text output.
• Includes native function calling for agentic workflows and step-by-step reasoning.
• Optimized for efficient local execution on AWS infrastructure.
• Handles OCR, chart comprehension, and code correction with multilingual support.

AWS has added OpenAI's privacy-filter to SageMaker JumpStart, introducing a bidirectional token-classification model for PII detection and masking. This tool enables high-throughput data sanitization by labeling input sequences in a single forward pass to identify entities like emails, names, and account numbers. The model is designed for context-aware processing and can be deployed with minimal effort to support secure data workflows on AWS infrastructure.

• OpenAI's privacy-filter is now available in SageMaker JumpStart for PII detection and masking.
• Model uses bidirectional token classification to label PII spans in a single forward pass.
• Supports detection of emails, names, account numbers, dates, URLs, and secrets.
• Designed for high-throughput sanitization workflows that can run on-premises or on AWS.
• Deployable with few clicks via SageMaker JumpStart for rapid integration.

AWS has added Qwen3-VL-Embedding-2B and Qwen3-Reranker-4B to SageMaker JumpStart for building search pipelines. The embedding model handles initial recall across text, images, and video, while the reranker refines results. These models work in tandem to support cross-modal retrieval tasks on AWS infrastructure.

• Qwen3-VL-Embedding-2B supports mixed modalities including text, images, screenshots, and video.
• Qwen3-Reranker-4B provides precise re-ranking after the initial embedding-based recall stage.
• Both models are now available via SageMaker JumpStart for immediate deployment.
• Designed for comprehensive search pipelines leveraging AWS foundation model infrastructure.

Small language models struggle with structural blindness when predicting molecular properties from SMILES strings. This research introduces a Context-Augmented Prompting framework that integrates graph neural networks at inference time to provide predictive hints and explanatory subgraphs. Evaluations on MUTAG and Tox21 datasets show that enriching prompts with these graph-based tools significantly improves performance across multiple prompting configurations.

• SLMs often miss graph-topological cues when relying solely on SMILES sequences.
• The new framework enables agentic tool use by invoking a GNN expert at inference.
• GNNs provide both predictive hints with confidence and instance-specific subgraphs.
• Performance gains were confirmed on MUTAG and Tox21 across five prompting setups.

This arXiv paper introduces Oracle Agent Memory, a system designed to handle the complexity of retaining and managing state for long-horizon AI agents. Rather than relying solely on document retrieval, the architecture treats memory as a lifecycle involving ingestion, extraction, consolidation, and revision. It leverages Oracle Database to provide durable, scoped, and retrievable state that persists across sessions and conversations.

• Moves beyond simple RAG by treating agent memory as a structured, durable lifecycle problem.
• Uses Oracle Database as the native substrate for state retention and retrieval.
• Addresses key challenges in scoping, latency, and the revision of accumulated knowledge.
• Focuses on practical deployment needs for long-horizon agent interactions.

Amazon OpenSearch Service now supports the Agent Toolkit for AWS, allowing AI coding agents like Claude Code and Cursor to manage domains and Serverless collections via natural language. This integration leverages the AWS Model Context Protocol (MCP) server to execute API calls automatically. A curated skill routes requests across migration, operations, and search capabilities.

• AI agents can now provision and manage OpenSearch domains and Serverless collections directly.
• Integration uses AWS MCP server to handle API calls based on natural language prompts.
• Supported agents include Claude Code, Kiro, and Cursor for development workflows.
• Automates migration from self-managed to fully managed OpenSearch environments.
• Enables semantic and vector search operations through natural language instructions.

This survey examines how autonomous agents evolve capabilities through experience with minimal human intervention. It introduces a system-level framework viewing agents as configurations of foundation models paired with operational scaffolds like prompts, memory, and tools. Self-improvement is formalized as a self-induced update operator that commits changes to either model parameters or these scaffold components. The work categorizes existing research by what is updated and the signals driving those updates.

• Self-improving agents are transitioning from research to production environments.
• Framework treats agents as config of foundation model plus operational scaffold.
• Improvement is formalized as self-induced updates to parameters or scaffolds.
• Prior work organized by update target and driving signals.

This paper addresses the challenge of modifying AI agent harnesses, which coordinate prompts, state, and tools. As systems evolve, developers struggle to map behavioral requirements to distributed, tightly coupled code. The work proposes methods to make these harnesses readable and editable for both human engineers and coding agents.

• AI agent harnesses are critical infrastructure often overlooked compared to the foundation model.
• Behavioral requirements are hard to map to code in large, distributed, and coupled repositories.
• Standard code search and indexing are insufficient for understanding behavioral intent.
• Improved navigability is essential for both human developers and autonomous coding agents.
• The work focuses on making harnesses editable as models and APIs continuously change.

This paper introduces Safety Sentry, a guard model that replaces binary safe/unsafe classification with a three-way EXECUTE-ASK-REFUSE routing decision. By evaluating actions at the instance level rather than the category level, it distinguishes between inherent harm and contextual appropriateness. The approach aims to reduce unnecessary interruptions while ensuring critical safety checks are enforced where needed.

• Replaces binary guardrails with EXECUTE, ASK, or REFUSE routing for finer control.
• Evaluates safety at the individual action instance level, not just the category.
• Separates inherent harm from contextual appropriateness to reduce false positives.
• Uses a lightweight guard model to minimize inference overhead for real-time agents.

MyAG introduces a graph-based architecture that decouples LLM agent construction into component, workflow, and search graphs. This separation enables flexible reuse of modules across different execution strategies and supports hierarchical composition via recursive nodes. The framework includes monitoring and visualization tools to help practitioners analyze performance and efficiency tradeoffs in agent deployments.

• Decouples agent logic, execution flow, and search strategies into distinct graph layers.
• Enables modular reuse of components across varying execution contexts.
• Supports hierarchical system design through recursive node definitions.
• Provides built-in monitoring and visualization for runtime analysis.

FixItFlow uses LLMs to automate the creation of troubleshooting guides by extracting diagnostic patterns from historical cloud incidents. The system synthesizes structured guides with verified commands and applies strict validation to prevent hallucinations. Evaluation with 26 engineers showed 61.5% clarity ratings and a 2.3x reduction in mitigation time.

• Automates guide creation from historical incident data, reducing manual effort.
• Extracts diagnostic patterns from engineer actions to build structured content.
• Enforces strict validation to ensure commands are verified and accurate.
• Reduces incident mitigation time by 2.3x in practitioner evaluations.

Existing classical and LLM-based methods struggle to identify root causes in the large-scale, multimodal OpenRCA dataset due to its lack of detailed domain knowledge. The authors introduce a Structured Multi-Agent RCA pipeline that significantly improves accuracy over prior baselines. This new approach is designed to function effectively in both domain-knowledge and knowledge-free operating modes.

• Classical causal discovery and existing LLM multi-agents fail reliably on the OpenRCA benchmark.
• OpenRCA's lack of domain knowledge makes root cause identification particularly challenging.
• New Structured Multi-Agent RCA pipeline shows substantial performance gains.
• Solution supports both domain-knowledge and knowledge-free operational contexts.

AWS Lambda now offers a one-click prompt in the console to configure coding agents with AWS Serverless skills and the Serverless Model Context Protocol. This automation embeds serverless best practices directly into the agent's workflow, replacing the previous need to manually search documentation for configuration details. The feature is accessible to developers at any stage of their Lambda journey, from initial exploration to function creation.

• Eliminates manual configuration steps for coding agents in the Lambda console
• Automatically installs AWS Serverless skills and Serverless MCP server
• Embeds serverless best practices into agent workflows from the start
• Available across all Lambda entry points for consistent developer experience

AWS Lambda now allows referencing source code directly from customer-managed S3 buckets, eliminating the need for intermediate copies in Lambda-managed storage. This change removes the previous 75GB per-region storage limit and reduces function activation time by skipping the copy step during creation and updates. The feature specifically benefits workloads with numerous functions and layers that previously required quota increase tickets.

• Bypass the 75GB per-region storage limit by storing code in your own S3 buckets
• Reduce cold start and update latency by eliminating the intermediate copy step
• No longer need to submit support tickets for storage quota increases
• Simplifies deployment pipelines by removing Lambda-managed storage management

Amazon Managed Service for Apache Flink now integrates AI Agent Skills to guide developers through common tasks like creation, troubleshooting, and scaling. These skills provide expert, up-to-date guidance to help teams accelerate development and maintain application health without deep specialized knowledge. The feature supports integration with existing AI coding agents such as Kiro to assist with version upgrades and cost optimization.

• AI Agent Skills automate guidance for Flink app creation, scaling, and monitoring.
• Reduces dependency on specialized Flink expertise for routine operational tasks.
• Integrates with existing tools like Kiro to streamline the developer workflow.
• Simplifies upgrades to new Flink versions, such as 2.2, with guided assistance.

AWS Database Migration Service now allows Schema Conversion tasks to be provisioned via AWS CloudFormation templates. This update integrates generative AI capabilities to enhance the code conversion experience for heterogeneous migrations. The feature supports converting database objects between systems such as Oracle, SQL Server, PostgreSQL, and MySQL.

• Automate DMS Schema Conversion setup using Infrastructure as Code workflows.
• Leverage built-in generative AI to improve heterogeneous code conversion quality.
• Simplify migration pipelines for Oracle/SQL Server to PostgreSQL/MySQL targets.
• Eliminate manual console steps for schema conversion task provisioning.

AWS has extended support for Graviton4-based R8g and M8g database instances to numerous new regions, including Europe, Asia Pacific, and South America. R8g is now generally available for Aurora and RDS engines, delivering up to 40% performance gains and 29% better price-performance compared to Graviton3. M8g support for RDS engines has also widened to additional locations across North America, Europe, and Asia.

• R8g Graviton4 instances are now GA in 10+ new regions for Aurora and RDS.
• M8g instances see expanded regional availability for RDS PostgreSQL, MySQL, and MariaDB.
• Expect up to 40% performance improvement and 29% cost efficiency over Graviton3.
• Review regional quotas and capacity plans for new Graviton4 instance types.

Amazon RDS and Aurora now offer R8gd and M8gd instances in additional regions, bringing Optimized Reads capabilities to PostgreSQL, MySQL, and MariaDB. These instances leverage local NVMe storage for ephemeral data like temporary tables, significantly reducing network I/O. This architecture delivers up to 165% better throughput and 120% improved price-performance compared to R6g instances for Aurora PostgreSQL.

• R8gd and M8gd instances are now available in 12 and 6 additional regions respectively.
• Local NVMe storage handles ephemeral data, lowering query latency for complex operations.
• Aurora PostgreSQL sees up to 165% throughput gains and 120% price-performance improvements.
• Optimized Reads support extends to RDS for MySQL and MariaDB alongside PostgreSQL.

Amazon EC2 G7e instances, powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, are now live in Frankfurt, Stockholm, and Mumbai. These instances deliver up to 2.3x inference performance over G6e and support up to 8 GPUs with 96 GB memory each. They are optimized for LLMs, agentic AI, spatial computing, and hybrid graphics-AI workloads.

• G7e instances launched in EU (Frankfurt, Stockholm) and APAC (Mumbai) regions
• Up to 2.3x inference performance improvement compared to G6e instances
• Features 8x NVIDIA RTX PRO 6000 Blackwell GPUs with 96 GB memory each
• Optimized for LLMs, multimodal AI, spatial computing, and physical AI workloads
• Powered by 5th Gen Intel Xeon processors supporting up to 192 vCPUs

Amazon RDS now permits up to four storage modifications per database instance within a rolling 24-hour window. This change eliminates the previous six-hour cool-off period, allowing operators to initiate new adjustments immediately after the prior optimization completes. The update supports increasing volume size, changing storage types, and adjusting performance without downtime across PostgreSQL, MariaDB, and other supported engines.

• Removes the six-hour cool-off period between storage modifications for faster operational agility.
• Allows up to four storage changes (size, type, performance) per instance in a 24-hour window.
• Modifications are non-disruptive, keeping applications running with minimal performance impact.
• Improves responsiveness to sudden data growth or unexpected workload spikes.
• Available on Amazon RDS for PostgreSQL, MariaDB, and other supported engines.

AWS Elastic Disaster Recovery now accelerates failover for workloads already hosted on Amazon EC2 by bypassing unnecessary preparation steps. This optimization reduces recovery time by up to 65% for Windows instances and 40% for Linux. The service automatically handles networking, drivers, and licensing while allowing administrators to toggle this faster mode per account or server.

• Skip prep steps for AWS-hosted source servers to cut RTO significantly.
• Up to 65% faster recovery for Windows and 40% for Linux workloads.
• Networking, drivers, and licensing are applied automatically during failover.
• Control granularity allows enabling faster recovery per account or individual server.

Amazon CloudWatch now supports a lookup processor that enriches log events by matching fields against uploaded CSV reference tables within the pipeline. This feature allows engineers to append metadata, such as mapping IP addresses to application teams or user IDs to details, directly during ingestion. The processor executes these matches in real-time, eliminating the need for external services to perform data enrichment.

• Upload CSV lookup tables directly in the CloudWatch Pipeline config for immediate use.
• Enrich logs with external context like team ownership or user details at ingestion time.
• Matches log fields against table keys to append specified metadata columns to events.
• Simplifies workflows by removing the need for separate enrichment microservices.