The stateless architecture of Large Language Models inherently lacks the mechanism to preserve dynamic context, compelling agents to redundantly reprocess history to maintain long-horizon autonomy. While latent memory offers a solution, current approaches are hindered by architectural segregation, relying on auxiliary encoders that decouple memory from the reasoning backbone. We propose FlashMem, a framework that distills intrinsic memory directly from transient reasoning states via computation reuse. Leveraging the property that internal representations uniquely encode input trajectories, FlashMem identifies the last hidden state as a sufficient statistic for the interaction history. This enables a Shared-KV Consolidator to synthesize memory by attending directly to the backbone’s frozen cache, eliminating redundant re-parameterization. Furthermore, a parameter-free Cognitive Monitor leverages attention entropy to adaptively trigger consolidation only when high epistemic uncertainty is detected. Experiments demonstrate that FlashMem matches the performance of heavy baselines while reducing inference latency by 5 times, effectively bridging the gap between efficiency and persistent cognition.

Large language models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks. However, as the model size and the input sequence’s length increase, the linearly increasing key-value (KV) cache significantly degrades inference throughput. Therefore, grouped-query attention (GQA), as an alternative to multi-head attention (MHA), has been widely introduced into LLMs. In this work, we propose a cost-effective method for converting MHA into GQA with any compression ratio of KV heads. The key point of our method lies in the application of Procrustes analysis to the attention heads, which enhances the similarity among attention heads while preserving computational invariance, thereby improving the model’s post-training performance. Subsequently, we employ L₀ regularization to prune redundant parameters. The model after pruning can be adapted to the standard GQA framework. Experimental results show that our strategy can compress up to 87.5% KV heads of LLaMA2-7B model and 75% KV heads of Sheared-LLaMA-1.3B with acceptable performance degradation. Our code is released at https://github.com/fpcsong/mha2gqa.

Integrating Large Language Models (LLMs) with existing Knowledge Graph (KG) databases presents a promising avenue for enhancing LLMs’ efficacy and mitigating their “hallucinations”. Given that most KGs reside in graph databases accessible solely through specialized query languages (e.g., Cypher), it is critical to connect LLMs with KG databases by automating the translation of natural language into Cypher queries (termed as “Text2Cypher” task). Prior efforts tried to bolster LLMs’ proficiency in Cypher generation through Supervised Fine-Tuning (SFT). However, these explorations are hindered by the lack of annotated datasets of Query-Cypher pairs, resulting from the labor-intensive and domain-specific nature of such annotation. In this study, we propose SyntheT2C, a methodology for constructing a synthetic Query-Cypher pair dataset, comprising two distinct pipelines: (1) LLM-based prompting and (2) template-filling. SyntheT2C is applied to two medical KG databases, culminating in the creation of a synthetic dataset, MedT2C. Comprehensive experiments demonstrate that the MedT2C dataset effectively enhances the performance of backbone LLMs on Text2Cypher task via SFT. Both the SyntheT2C codebase and the MedT2C dataset will be released.

Integrating information from various reference databases is a major challenge for Retrieval-Augmented Generation (RAG) systems because each knowledge source adopts a unique data structure and follows different conventions. Retrieving from multiple knowledge sources with one fixed strategy usually leads to under-exploitation of information. To mitigate this drawback, inspired by Mix-of-Expert, we introduce Mix-of-Granularity (MoG), a method that dynamically determines the optimal granularity of a knowledge source based on input queries using a router. The router is efficiently trained with a newly proposed loss function employing soft labels. We further extend MoG to MoG-Graph (MoGG), where reference documents are pre-processed as graphs, enabling the retrieval of distantly situated snippets. Experiments demonstrate that MoG and MoGG effectively predict optimal granularity levels, significantly enhancing the performance of the RAG system in downstream tasks. The code of both MoG and MoGG will be made public.