Brief

A new prototype called LLM-Context-Optimization-Engine has been developed to address failures in long-running Large Language Model applications. These failures often stem from selecting the wrong context, rather than pure reasoning errors. The engine benchmarks various context selection policies, including sliding windows, full history, and retrieval methods, to determine which pieces of prior state are most relevant for the next model call. An importance-based selection policy demonstrated a high retention rate of critical facts within a limited budget, highlighting the need for memory policies over simple memory storage in persistent LLM applications. AI

This article outlines a practical, multi-layered framework for programmatically evaluating the quality of Large Language Model (LLM) outputs. It emphasizes defining specific quality dimensions such as correctness, format compliance, safety, and consistency based on the use case. The framework includes deterministic checks for immediate failure detection and semantic similarity measures using sentence embeddings for free-form text evaluation. AI

This article details how to build a cross-cloud Retrieval-Augmented Generation (RAG) workflow using ChromaDB, a vector database, across Azure and AWS. It focuses on enhancing Large Language Model (LLM) capabilities by integrating external data sources. The guide aims to provide practical steps for developers looking to implement such a system in a multi-cloud environment. AI

A new architectural approach has been developed to integrate generative AI with monolithic enterprise systems like Oracle E-Business Suite (EBS) without altering the core legacy code. This method involves creating a lightweight semantic layer that acts as a plugin, translating complex technical data structures into understandable business terms for AI models. This abstraction layer prevents AI hallucinations and ensures accurate data interpretation, even in heavily customized environments, by operating on virtual data marts instead of direct database access. AI

Researchers have developed LEAP, a closed-loop framework that uses a domain-specific large language model combined with active learning to discover additives for perovskite solar cells. This LLM is trained to extract knowledge from scientific literature and represent molecules, which then informs a Bayesian optimization process for prioritizing additives. Experimental validation showed improved additive prioritization, leading to higher power conversion efficiencies in perovskite devices. AI

Researchers have developed a new entity linking agent designed to improve question answering systems by more effectively connecting natural language mentions to knowledge base entries. This agent, built upon a large language model, mimics human cognitive processes to identify entity mentions, retrieve candidates, and make linking decisions. Experiments demonstrated the agent's robustness and effectiveness in both tool-based entity linking and overall question answering tasks. AI

Researchers have introduced TPMM-DPO, a novel method for aligning large language models that addresses issues of error accumulation in iterative Direct Preference Optimization. This new approach treats the sequence of policy models as an optimization trajectory, adaptively merging them with learned weights to create a more stable and robust reference model. Experiments demonstrate that TPMM-DPO significantly improves generation quality and performance, outperforming standard iterative DPO by mitigating degradation in later training stages. AI

Researchers have introduced the Geno-Synthetic Algorithm (GSA), a novel coevolutionary framework designed to optimize complex design objects with heterogeneous parameters. Unlike traditional methods that flatten diverse data types into a single format, GSA partitions gene families by type and evolves them using type-native operators before assembling them into executable phenotypes. An open-source implementation is available, and empirical studies show GSA's unique ability to handle complex-valued descriptors and embedding vectors, making it applicable to areas like large language model prompt and embedding optimization. AI

This cluster contains two identical Mastodon posts linking to a KDnuggets article. The article lists ten Python libraries useful for developing applications that utilize Large Language Models. AI

Researchers at Together AI have developed AutoJudge, a novel method to accelerate large language model inference. This technique automates the curation of task-specific datasets, enabling lossy speculative decoding without manual annotation. AutoJudge identifies critical tokens that impact downstream quality, achieving up to a 2x speedup over standard speculative decoding with minimal accuracy loss. AI