Researchers introduced JOINGR, a join-aware table retrieval method for Text-to-SQL, treating the database join graph as the retrieval space. It selects semantically similar anchor tables, traverses join edges, and aggregates scores. JOINGR improves recall over baselines on the BEAVER benchmark and transfers across benchmarks.

Researchers introduce Galahad, a memory layer for LLMs that enables stateful inference by storing and reusing model key-value state, reducing computation costs and energy consumption.

Researchers present SkillSeek, an open-source two-stage skill retriever for LLM agents, achieving parity with LLM-mediated retrieval loops at a significantly reduced cost. The SkillSeek system uses a standard IR recipe and is exposed over MCP, making it a strong default for agent-skill retrieval.

Researchers propose component-aware feedback for self-evolving programs, which improves the efficiency of LLM-guided evolutionary search by identifying the components that contribute to fitness metric changes. This can lead to faster and more stable program search, especially for multi-component systems. The method is demonstrated on LLM reranking, a multi-objective optimization problem.

Researchers have discovered a vulnerability in agentic search systems that use retrieval-augmented generation (RAG). An attacker can compromise the retriever and inject a backdoor that can manipulate the search results, suppress useful evidence, or steer the agent towards prolonged search, without modifying the underlying corpus. The backdoor can be concealed by purifying it and leaving behind a weakened signature that fools detectors.

Researchers propose a hybrid defense, VaccineBooster, to improve alignment in language models under harmful fine-tuning attacks. It combines embedding perturbation and weight-level gradient attenuation, achieving better results than previous methods.

Researchers evaluated the effectiveness of gradient-based jailbreak detection methods in multi-turn dialogue settings. They found that existing methods can detect jailbreaks in synthetic conversations but struggle with realistic conversations. The results suggest that reliable deployment requires calibration on realistic benign conversations and consideration of various attack types and model architectures.

Mnemon is a memory agent that uses a two-system approach to LLM memory, dividing work into fast System 1 judgments and slow System 2 planning. It achieves high scores on LoCoMo and LongMemEval-S benchmarks with low context and achieves parity with published results using a reasoning model.

ThuRunel is a dynamic decoupling approach for structured advisory dialogue, combining a finite-state belief management framework, a chain-of-thought teacher synthesis protocol, and learned generation adapters. It has been deployed as a bilingual web application and achieved consistent improvements in elicitation completeness and specialist brief quality against eleven baselines.

GeoOutageBench is a benchmark for assessing LLM-based geospatiotemporal KGQA for multimodal outage and resilience analysis. It considers a spatiotemporal KG that integrates visual, textual, and structured data from various sources. The benchmark provides a competency query taxonomy at different difficulty levels and evaluates three tasks: LLMs' understanding of ambiguous geospatiotemporal questions, ontology utility, and answer accuracy of multimodal KGQA retrieval.