FanarGuard: A Culturally-Aware Moderation Filter for Arabic Language Models

SwiftMem has been introduced as a query-aware agentic memory system designed to enhance the efficiency of large language model (LLM) agents by enabling sub-linear retrieval through specialized indexing techniques. This system addresses the limitations of existing memory frameworks that rely on exhaustive retrieval methods, which can lead to significant latency issues as memory storage expands.

PrivGemo has been introduced as a privacy-preserving framework designed for knowledge graph (KG)-grounded reasoning, addressing the risks associated with using private KGs in large language models (LLMs). This dual-tower architecture maintains local knowledge while allowing remote reasoning through an anonymized interface, effectively mitigating semantic and structural exposure.

A new offline reinforcement learning (RL) framework named STO-RL has been proposed to enhance policy learning from pre-collected datasets, particularly in long-horizon tasks with sparse rewards. By utilizing large language models (LLMs) to generate temporally ordered subgoal sequences, STO-RL aims to improve the efficiency of reward shaping and policy optimization.

The introduction of STAGE (Screenplay Text, Agents, Graphs and Evaluation) marks a significant advancement in the field of narrative understanding, providing a comprehensive benchmark for evaluating knowledge graph construction, scene-level event summarization, long-context screenplay question answering, and in-script character role-playing across 150 films in English and Chinese.

A new approach called MHEL-LLaMo has been introduced for multilingual historical entity linking, utilizing a combination of a Small Language Model (SLM) and a Large Language Model (LLM). This unsupervised ensemble method addresses challenges in processing historical texts, such as linguistic variation and noisy inputs, by leveraging a multilingual bi-encoder for candidate retrieval and an instruction-tuned LLM for predictions.

A recent study emphasizes the importance of data curation in machine translation, particularly for low-resource languages. The research introduces LALITA, a framework designed to optimize the selection of source sentences for creating parallel corpora, focusing on English-Hindi bi-text to enhance machine translation performance.

A recent study analyzed the false refusal behavior of large language models (LLMs) in the context of hate speech detoxification, revealing that these models disproportionately refuse tasks involving higher semantic toxicity and specific target groups, particularly in English datasets.

Recent research highlights that while KV cache reuse can enhance efficiency in multi-agent large language model (LLM) systems, it can negatively impact the performance of LLM judges, leading to inconsistent selection behaviors despite stable end-task accuracy.