#llm
Every summary, chronological. Filter by category, tag, or source from the rail.
Reducing LLM Hallucinations with Governed Semantic Definitions
The GROUND framework mitigates LLM hallucinations in enterprise analytics by enforcing a layer of governed semantic definitions, ensuring models query data based on verified business logic rather than raw natural language interpretation.
Refusal Is Not Robustness: LLMs Fabricate on Uninformative Data
Large Language Models often fail to identify uninformative input, choosing to confidently fabricate clinical assessments rather than admitting a lack of sufficient data.
Explaining ICU Mortality Predictions with LLM Agentic Pipelines
This study demonstrates the feasibility of using standalone LLMs and pre-specified agentic pipelines to interpret complex ICU mortality risk models, providing a path toward more transparent clinical decision support.
CIFQA: Deterministic Multi-Agent Framework for Financial Analysis
CIFQA is a multi-agent framework designed to improve financial query accuracy by replacing non-deterministic LLM reasoning with a structured, tool-grounded execution pipeline.
Anthropic's Automated Researcher: A Leap in Self-Improving AI
Anthropic researchers have developed an Automated Alignment Researcher (AAR) that outperforms human researchers at improving model alignment, doing so at a fraction of the cost and time.
Architecting Production-Grade LLM Gateways
LLM gateways require a shift from standard API engineering: prioritize per-request fallbacks over circuit breakers, track latency per-route rather than globally, and treat guardrails as unreliable services that require explicit fail-open/closed policies.
7 Modular Design Patterns for AI Coding Agents
Improve AI coding agent performance by replacing long, confusing prompts with modular 'skills'—specialized text files that the agent loads dynamically only when needed.
Google Cloud TechBuilding Context Engines for AI Agents
AI agents fail at complex tasks because they lack organizational context, leading to 'satisfaction of search' errors. A context engine provides intent, conventions, and historical data, reducing token waste and preventing compounding logic errors.
Audit Your Coding Agent Configuration
Agent configuration files like CLAUDE.md and custom skills suffer from 'rot' and bloat. Regular audits using tools like /doctor are essential to remove stale instructions, as modern models often perform better with leaner, more focused context.
Optimizing Agentic Inference: KV Cache Routing and P/D Disaggregation
Agentic workloads require moving beyond steady-state benchmarks. By implementing KV cache-aware routing and decoupling prefill from decode compute, teams can achieve 4x faster time-to-first-token and significantly smoother inter-token latency.
Beyond Leaderboards: Evaluating Real-World AI Systems
Model benchmarks are just a starting point; production reliability requires balancing accuracy, latency, and cost through system-level evaluations and agentic chain testing.
Optimizing Code Models with Function-Level Execution Feedback
Improving code generation models by using granular, function-level execution feedback rather than binary pass/fail signals to guide preference optimization.
Reducing Medical AI Sycophancy via Gated Activation Steering
Gated Activation Steering (GAS) improves medical LLM reliability by dynamically suppressing internal representations associated with sycophancy and hallucinations during inference, without requiring model retraining.
Optimizing Masked Diffusion LLMs for Real-World Hardware
This paper provides a characterization of Masked Diffusion LLMs, identifying unique computational bottlenecks and proposing hardware-aware design principles to improve inference efficiency.
RENDER: A Framework for Controlling Evidence in LLM Memory Evaluation
RENDER is a new evaluation framework designed to isolate and measure how LLMs process and recall specific evidence within their context windows, addressing the limitations of existing memory benchmarks.
Building Real-Time Voice AI Agents with Gemini Live
Gemini Live enables bidirectional, audio-native conversations by using WebSockets for streaming and built-in voice activity detection to handle interruptions and tool execution.
Google Cloud TechBuilding and Scaling Multi-Agent AI Systems on GKE
A practical guide to deploying AI agents on GKE, using the Model Context Protocol for infrastructure troubleshooting, and implementing secure sandboxing for AI-generated code.
Radar: Making Podcast Audio Discoverable for AI Agents
Radar is a podcast search engine and API that transcribes and indexes audio, enabling AI agents to process spoken content, track entity mentions, and analyze advertising trends.
Optimizing Documentation for AI Agents
To drive AI-agent adoption of your library, stop relying on web search. Instead, ship bundled markdown files directly within your package and provide hand-curated llms.txt files to ensure agents have accurate, token-efficient context.
Scaling AI Agents: Lessons from Snowflake's GTM Assistant
Successfully deploying AI agents at scale requires prioritizing quality over coverage, aggressive change management, and a willingness to rearchitect as user expectations evolve.
Architecture-Aware Credit Transport for LLM Reinforcement Learning
The paper introduces a method to improve LLM reinforcement learning by aligning credit assignment with the underlying computational architecture, ensuring rewards are distributed based on actual processing paths.
SchemaRouter: Field-Aware Tool Routing for Agentic RAG
SchemaRouter improves agentic RAG efficiency by using field-aware routing, which maps user queries to specific tool schemas rather than relying on generic semantic similarity.
Hate Speech Classification in Roman Urdu: PEFT vs. Prompt Engineering
A comparative study evaluating Parameter-Efficient Fine-Tuning (PEFT) against prompt engineering for detecting hate speech in Roman Urdu, highlighting the trade-offs between computational efficiency and classification accuracy in low-resource linguistic contexts.
KVBoost: Accelerating LLM Inference via Chunk-Level Cache Reuse
KVBoost improves LLM inference latency by 4.49x by enabling chunk-level KV cache reuse regardless of position, using a dual-hash keying scheme and deviation-guided recomputation to maintain accuracy.
4 Common Loop Engineering Failures and How to Fix Them
Loop engineering automates repetitive tasks by setting goals and retrying, but it often fails due to runaway costs, confirmation bias, vague objectives, or excessive complexity. Success requires strict stop rules, external evaluation, concrete metrics, and transitioning to graph-based architectures for complex workflows.
Google Cloud TechAnthropic Unifies Claude Memory Across Chat and Cowork
Anthropic has merged the memory systems for Claude chat and Claude Cowork, allowing the AI to retain context across different workflows and giving users manual control to edit or delete stored information.
AI Agents: Why the Harness Matters More Than the Model
AI system performance is driven by the 'agentic harness'—the tools, memory, and loops surrounding the model—rather than just the model itself. Distinguishing between the 'brain' (model) and the 'jar' (harness) is essential for building effective AI agents.
Structurally Indirect Prerequisite Eviction in Agentic Memory
Agentic memory systems often fail not due to retrieval errors, but because 'prerequisite' information is evicted from context before it can be used, creating a structural failure in long-term reasoning.
OpenAI's Shift to Agentic Workflows for Non-Engineers
OpenAI is expanding beyond coding tools with 'ChatGPT Work,' an agentic platform designed to automate complex, multi-step tasks across common business software, aiming to move AI from simple Q&A to autonomous project execution.
Bridging SQL and Vector Data with Agentic Workflows
Digital librarian AI agents solve the 'what vs. why' data gap by orchestrating queries across structured SQL databases and unstructured vector databases to provide grounded, context-aware answers.
IBM TechnologyShowing 30 of 1272