Tool-augmented LLMs invoke external functions to extend their capabilities, but errors in the invocation decision, such as calling a tool when none is needed or omitting a needed call, can produce unreliable outputs and unnecessary cost. A lightweight remedy is to prepend retrieved examples so LLMs decide tool use in context. However, existing retrievers rank examples by semantic similarity alone. Lexically close or semantically close queries can require opposite behavior, so the retrieved examples may be behaviorally inconsistent and silently mislead the model. We propose Behavior-Aligned Retrieval (BAR), a backbone-agnostic training recipe that teaches a dense retriever a behavior-aware similarity, keeping semantically related candidates close only when their tool-use behavior is compatible. BAR does not predict invocation labels; instead, it ranks demonstrations while leaving the final tool-use decision to the LLM. Applied to multiple retrieval backbones, including BERT, Contriever, and Qwen3-Reranker, BAR consistently improves invocation reliability and reduces unnecessary API calls across 14 LLMs and 3 benchmarks.

Automatic extraction of Medical Decision Trees (MDTs) is an important task for Clinical Decision Support Systems, but remains challenging due to the complex logical structures involved. Recent approaches based on Large Language Models (LLMs) commonly adopt generate-then-revise pipelines with self-correction to improve reasoning accuracy. However, we find that applying revision uniformly across all samples can be harmful, as it often alters already correct and stable predictions, leading to over-correction on simple cases. To alleviate this issue, we propose R³-MDT, an adaptive inference framework that dynamically selects reasoning strategies based on sample difficulty. We use model uncertainty and output consistency as internal signals to estimate difficulty and route samples into Easy, Medium, and Hard categories. For Easy samples with confident and consistent predictions, we directly retain the initial outputs. Medium-difficulty samples, which exhibit ambiguity among multiple plausible candidates, are handled via reranking to select the optimal decision tree. Hard samples, characterized by high uncertainty and structural inconsistency, are revised to rectify intractable generation failures. By adaptively selecting inference strategies based on difficulty levels, R³-MDT avoids unnecessary revisions while effectively correcting underlying reasoning failures. Experiments on the Text2DT benchmark show that R³-MDT alleviates the over-correction problem and achieves state-of-the-art performance, while auxiliary language- and task-shift evaluations indicate the broader transferability of its routing principle.

Retrieval-Augmented Generation (RAG) has been widely adopted to enhance large language models (LLMs) by incorporating external knowledge. However, the two main existing paradigms struggle with multi-hop reasoning: aggregate-first approaches suffer from high construction costs and limited adaptability to dynamic knowledge, while dynamic-first approaches rely heavily on LLM reasoning and are prone to error propagation across reasoning steps. To address these limitations, we propose SR-RAG, a symbolic reasoning framework for multi-hop question answering. SR-RAG integrates the advantages of both paradigms by dynamically generating sub-questions, performing information retrieval and symbolic encoding based on an on-the-fly graph, and using a symbolic verifier to formally validate intermediate reasoning steps to ensure the correctness of intermediate answers and the completeness of the reasoning chain . We evaluate SR-RAG on multiple multi-hop benchmarks and a medical dataset. Experimental results demonstrate that it significantly improves both accuracy and robustness.

Named Entity Recognition (NER) and Relation Extraction (RE) are two fundamental and interdependent tasks in information extraction (IE), aiming to identify entities and relations from unstructured text. Recently, generative methods have become mainstream instead of discriminative methods for IE, especially joint multi-task IE, due to their promising performance and flexibility. For joint NER and RE, existing methods suffer from misalignment between entities and relations, as well as misalignment among relations. To address these issues, we propose AnchorAlign, a novel generative method enhanced by anchor alignment. Specifically, we first introduce an anchor entity selection mechanism to identify key entities in the text as anchor points, which serve as semantic pivots to bridge the two tasks. Then, we design a dual-level anchor alignment module: at the semantic level, we construct a cross-task semantic alignment space to align the semantic representations of anchor entities and their associated relations; at the generation level, we introduce an anchor-guided generation constraint to guide the model to generate entities and relations with strict alignment based on the anchor points. Extensive experiments on five benchmark datasets show that AnchorAlign outperforms state-of-the-art baselines, demonstrating its effectiveness. Our work provides a new perspective for optimizing the joint modeling of NER and RE, and has potential to be extended to more complex multi-task IE such as NER and Event Extraction (EE).

Retrieval-Augmented Generation (RAG) grounds large language models (LLMs) in external evidence, but existing RAG systems remain poorly matched to visually rich documents. OCR-based methods discard layout and non-textual visual cues, and although vision-language models (VLMs) can better understand native visual information, existing visual retrieval approaches often compress entire pages into global embedding vectors, obscuring fine-grained spatial information such as table cells and local annotations. Meanwhile, discrete Top-K selection in cascaded architectures blocks gradient flow between retrieval and generation, preventing joint optimization.In this work, we propose E2E-VisRAG, an end-to-end visual RAG framework that operates directly on native visual patches. Building upon patch-level visual encoding, we introduce a differentiable Top-K selection mechanism to enable joint training of retrieval and generation, and employ a gated residual fusion module to inject retrieval-enhanced visual features while suppressing noise. This design abandons the traditional paradigms of relying on OCR, global embeddings, and retrieval–generation separation, maximizing the preservation of original document visual semantics. Experiments on Vidore and MMLongBench-Doc show that, under matched 3B generator capacity, E2E-VisRAG outperforms non-end-to-end visual retrieval baselines, improving both Recall@1 and Accuracy@1 over OCR-based, global-embedding, and patch-level retrieval methods.

Measuring Semantic Textual Similarity (STS) is a fundamental task in biomedical text processing, which aims at quantifying the similarity between two input biomedical sentences. Unfortunately, the STS datasets in the biomedical domain are relatively smaller but more complex in semantics than common domain, often leading to overfitting issues and insufficient text representation even based on Pre-trained Language Models (PLMs) due to too many biomedical entities. In this paper, we propose EARA, an entity-aligned, attention-based and retrieval-augmented PLMs. Our proposed EARA first aligns the same type of fine-grained entity information in each sentence pair with an entity alignment matrix. Then, EARA regularizes the attention mechanism with an entity alignment matrix with an auxiliary loss. Finally, we add a retrieval module that retrieves similar instances to expand the scope of entity pairs and improve the model’s generalization. The comprehensive experiments reflect that EARA can achieve state-of-the-art performance on both in-domain and out-of-domain datasets. Source code is available.

Transformer and its variants have achieved great success in natural language processing. Since Transformer models are huge in size, serving these models is a challenge for real industrial applications. In this paper, we propose , a highly efficient inference library for models in the Transformer family. includes a series of GPU optimization techniques to both streamline the computation of Transformer layers and reduce memory footprint. supports models trained using PyTorch and Tensorflow. Experimental results on standard machine translation benchmarks show that achieves up to 14x speedup compared with TensorFlow and 1.4x speedup compared with , a concurrent CUDA implementation. The code will be released publicly after the review.

This is the system description of the Harbin Institute of Technology Shenzhen (HITSZ) team for the first and second subtasks of the fifth Social Media Mining for Health Applications (SMM4H) shared task in 2020. The first task is automatic classification of tweets that mention medications and the second task is automatic classification of tweets in English that report adverse effects. The system we propose for these tasks is based on bidirectional encoder representations from transformers (BERT) incorporating with knowledge graph and retrieving evidence from online information. Our system achieves an F1 of 0.7553 in task 1 and an F1 of 0.5455 in task 2.

The prediction of the relationship between the disease with genes and its mutations is a very important knowledge extraction task that can potentially help drug discovery. In this paper, we present our approaches for trigger word detection (task 1) and the identification of its thematic role (task 2) in AGAC track of BioNLP Open Shared Task 2019. Task 1 can be regarded as the traditional name entity recognition (NER), which cultivates molecular phenomena related to gene mutation. Task 2 can be regarded as relation extraction which captures the thematic roles between entities. For two tasks, we exploit the pre-trained biomedical language representation model (i.e., BERT) in the pipe of information extraction for the collection of mutation-disease knowledge from PubMed. And also, we design a fine-tuning technique and extra features by using multi-task learning. The experiment results show that our proposed approaches achieve 0.60 (ranks 1) and 0.25 (ranks 2) on task 1 and task 2 respectively in terms of F₁ metric.

The Biological Text Mining Unit at BSC and CNIO organized the first shared task on chemical & drug mention recognition from Spanish medical texts called PharmaCoNER (Pharmacological Substances, Compounds and proteins and Named Entity Recognition track) in 2019, which includes two tracks: one for NER offset and entity classification (track 1) and the other one for concept indexing (track 2). We developed a pipeline system based on deep learning methods for this shared task, specifically, a subsystem based on BERT (Bidirectional Encoder Representations from Transformers) for NER offset and entity classification and a subsystem based on Bpool (Bi-LSTM with max/mean pooling) for concept indexing. Evaluation conducted on the shared task data showed that our system achieves a micro-average F1-score of 0.9105 on track 1 and a micro-average F1-score of 0.8391 on track 2.