Researchers propose a user-product relevance metric for offline evaluation of Dynamic Product Advertising (DPA) retrieval systems using large language models (LLMs) as semantic relevance judges, enabling faster and more reliable iteration before A/B testing.

This paper proposes a new approach to selecting large language models (LLMs) for strategic interactions, where the opponent's behavior is unknown and can adapt over time. The authors define a meta game on top of the base game and evaluate various selector methods, finding that a lightweight multi-armed bandit algorithm performs more reliably than others.

A paper introduces a timestamp theory of decision validity and a decision ledger for LLMs, which helps maintain decision accuracy by routing new information to decisions already made. This approach factors accuracy into two capabilities: routing and temporal judgment, and is proven to work across various LLM scales.

Researchers propose a new approach to AI safety by inferring interaction safety from partial model output, allowing for safer and more efficient generation of responses in LLM-based information systems.

This paper proposes a budget-aware optimizer for LLM ORDER BY queries, leveraging heuristic rules and LLM-as-Judge evaluation to dynamically select the near-optimal access path. The optimizer achieves ranking accuracy on par with or superior to static methods across various benchmarks.

Researchers introduce PRIME, a framework for optimizing index-side representation of documents for LLM agents, leading to improved retrieval quality and reduced computational costs.

Researchers tested language models on the Wikipedia Game, a task requiring information retrieval from large document collections. They compared sentence embedding models and LLMs across languages and datasets, providing code for building and navigating local graphs of Wikipedia pages and paper citations.

A new AI agent tool retrieval framework called RATED is proposed, which combines semantic matching with dynamic reliability adaptation to improve tool availability during execution. RATED achieves a 52.2% relative MRR improvement over existing retrievers and reduces fault selection rate under API outage scenarios.

Web$hell is an interface that unifies web access under a single access model, enabling large language model agents to conduct broad information gathering, perform browser actions, and revisit acquired evidence within a single trajectory. It achieves strong task performance at lower cost and opens new possibilities for more capable and efficient LLM agents on the web.

Researchers introduce KnowForge, a pipeline for open-domain multimodal document QA, consisting of a hybrid-retrieval knowledge-point index, a Locator agent, a bridge judge, an evidence pool, and a Deep Reader agent. KnowForge achieves state-of-the-art results on various benchmarks, outperforming existing models and systems.

A new system, called HybridSourceTracker, uses a combination of vector search and fingerprinting to efficiently track the provenance of code snippets generated by large language models, addressing plagiarism and license compliance concerns.

A new method called SkillContrast is introduced for agent skill reranking, using a training-free selector that compares retrieved skills and retains their differing text with local context. This approach yields 54-72 more clean hits than TF-IDF query selection and uses 51.1-58.8% fewer model-input tokens.

Gated Memory proposes a new formation framework for conversational AI memory systems, introducing an admission gate and conditional enrichment stage to improve memory quality. This framework evaluates candidate facts against the full utterance context before extraction and grounds admitted facts in an entity scope taxonomy with privacy constraints.

RIT-RAG (Retrieval-Induced Tree RAG) is a new method that combines content retrieval with structural navigation to improve the accuracy of language models in answering questions. It builds a tree for each document and uses its structure to guide the model in retrieving relevant information. RIT-RAG outperforms other methods on several benchmarks, including a new benchmark of 2.84 million technical-documentation webpages.