Large Language Models (LLMs) have shown improved generation performance through retrieval-augmented generation (RAG) following the retriever-reader paradigm, which supplements model inputs with externally retrieved knowledge. However, prior work often evaluates RAG holistically, assessing the retriever and reader jointly, making it difficult to isolate the true contribution of retrieval, particularly given the prompt sensitivity of LLMs used as readers. We move beyond perplexity and introduce Spectrum Projection Score (SPS), a lightweight and supervision-free metric that allows the reader to gauge the semantic alignment of a retrieved summary with its hidden representation by comparing the area formed by generated tokens from the summary, and the principal directions of subspace in the reader and to measure the relevance. Building on SPS we present xCompress, an inference‑time controller framework that dynamically samples, ranks, and compresses retrieval summary candidates. Extensive experiments on five QA benchmarks with four open-sourced LLMs show that SPS not only enhances performance across a range of tasks but also provides a principled perspective on the interaction between retrieval and generation.

Large language models (LLMs) have recently pushed open-domain question answering (ODQA) to new frontiers.However, prevailing retriever-reader pipelines often depend on multiple rounds of prompt-level instructions, leading to high computational overhead, instability, and suboptimal retrieval coverage.In this paper, we propose EmbQA, an embedding-level framework that alleviates these shortcomings by enhancing both the retriever and the reader.Specifically, we refine query representations via lightweight linear layers under an unsupervised contrastive learning objective, thereby reordering retrieved passages to highlight those most likely to contain correct answers.Additionally, we introduce an exploratory embedding that broadens the model's latent semantic space to diversify candidate generation and employs an entropy-based selection mechanism to choose the most confident answer automatically.Extensive experiments across three opensource LLMs, three retrieval methods, and four ODQA benchmarks demonstrate that EmbQA substantially outperforms recent baselines in both accuracy and efficiency.

Complex narrative contexts often challenge language models' ability to follow instructions, and existing benchmarks fail to capture these difficulties.To address this, we propose Concise-SAE, a training-free framework that improves instruction following by identifying and editing instruction-relevant neurons using only natural language instructions, without requiring labelled data.To thoroughly evaluate our method, we introduce FREEIN-STRUCT, a diverse and realistic benchmark of 1,212 examples that highlights the challenges of instruction following in narrative-rich settings.While initially motivated by complex narratives, Concise-SAE demonstrates stateof-the-art instruction adherence across varied tasks without compromising generation quality.

Large Language Models (LLMs) struggle with complex reasoning due to limited diversity and inefficient search. We propose Soft Reasoning, an embedding-based search framework that optimises the embedding of the first token to guide generation. It combines (1) embedding perturbation for controlled exploration and (2) Bayesian optimisation to refine embeddings via a verifier-guided objective, balancing exploration and exploitation. This approach improves reasoning accuracy and coherence while avoiding reliance on heuristic search. Experiments demonstrate superior correctness with minimal computation, making it a scalable, model-agnostic solution.

In recent years, the growing complexity of database management systems (DBMSs) and the proliferation of SQL dialects have created significant challenges for database migration, federation, and integration. These challenges arise from the disparities between SQL dialects across different DBMSs, hindering seamless communication and system interoperability. SQL translation, the process of converting SQL queries from a source dialect DBMS to a target dialect DBMS, plays a crucial role in addressing these challenges. To facilitate this process, we introduce DLBench, the first comprehensive benchmark designed to evaluate the SQL translation capabilities of Large Language Models (LLMs). The benchmark includes two datasets: BIRDTrans, which covers real-world database query scenarios across seven DBMSs, and ButterTrans, which spans a broader spectrum of SQL types and encompasses extensive DBMS dialect features. We collect high-quality databases and SQL statements, applying a rigorous multi-step cleaning process that ensures data quality through SQL92-based checks and dialect-specific parser validation. Additionally, both LLM-based and human annotations are used to guarantee the correctness and completeness of the dataset. We demonstrate the utility of DLBench through extensive experiments, which show that the benchmark effectively evaluates the SQL translation ability of LLMs. The results highlight the potential of LLMs for SQL translation tasks and provide insights into areas for further improvement.

While Large language models (LLMs) have the capability to iteratively reflect on their own outputs, recent studies have observed their struggles with knowledge-rich problems without access to external resources.In addition to the inefficiency of LLMs in self-assessment, we also observe that LLMs struggle to revisit their predictions despite receiving explicit negative feedback.Therefore, We propose Mirror, a Multiple-perspective self-reflection method for knowledge-rich reasoning, to avoid getting stuck at a particular reflection iteration.Mirror enables LLMs to reflect from multipleperspective clues, achieved through a heuristic interaction between a Navigator and a Reasoner.It guides agents toward diverse yet plausibly reliable reasoning trajectory without access to ground truth by encouraging (1) diversity of directions generated by Navigator and (2) agreement among strategically induced perturbations in responses generated by the Reasoner.The experiments on five reasoning datasets demonstrate that Mirror's superiority over several contemporary self-reflection approaches.Additionally, the ablation study studies clearly indicate that our strategies alleviate the aforementioned challenges.The code is released at https://github.com/hanqi-qi/Mirror.git.

Zero-shot stance detection (ZSSD) aims to detect the stance for an unseen target during the inference stage.In this paper, we propose a joint contrastive learning (JointCL) framework, which consists of stance contrastive learning and target-aware prototypical graph contrastive learning.Specifically, a stance contrastive learning strategy is employed to better generalize stance features for unseen targets.Further, we build a prototypical graph for each instance to learn the target-based representation, in which the prototypes are deployed as a bridge to share the graph structures between the known targets and the unseen ones.Then a novel target-aware prototypical graph contrastive learning strategy is devised to generalize the reasoning ability of target-based stance representations to the unseen targets.Extensive experiments on three benchmark datasets show that the proposed approach achieves state-ofthe-art performance in the ZSSD task 1 .

Stance detection aims to identify the stance of the text towards a target. Different from conventional stance detection, Zero-Shot Stance Detection (ZSSD) needs to predict the stances of the unseen targets during the inference stage. For human beings, we generally tend to reason the stance of a new target by linking it with the related knowledge learned from the known ones. Therefore, in this paper, to better generalize the target-related stance features learned from the known targets to the unseen ones, we incorporate the targeted background knowledge from Wikipedia into the model. The background knowledge can be considered as a bridge for connecting the meanings between known targets and the unseen ones, which enables the generalization and reasoning ability of the model to be improved in dealing with ZSSD. Extensive experimental results demonstrate that our model outperforms the state-of-the-art methods on the ZSSD task.

Aiming at improving the performance of relay transmission in multi-flow streaming overlay networks, a performance evaluation model to optimize the global weighted average latency was proposed. This model supports multiple senders and multiple receivers. A distributed heuristic allocation algorithm was proposed to optimize the transmission performance. In this algorithm, the bottleneck bandwidth shared by multiple flows is allocated based on the path weighted average latency which is computed between any pair of nodes in the network. Source node uses a two stages feedback pattern to allocate the outgoing flow to multiple available paths. Simulation results show that there is linear dependence between the weighted average latency and the data slip hit ratio which is a metric of streaming network performance, the heuristic allocation algorithm can effectively lower the network overall transmission latency and can effectively adjust the traffic allocation. It gets 4%-17% lower in the network overall weighted average latency comparing with that of the pattern of average allocation shared bottle bandwidth.

The current peer-to-peer (P2P) live streaming networks can be generally classified into two categories: tree-based and data-driven. The tree-based approach suffers from three limitations: interruptive delivery due to failures of high level nodes, unfair uploading (out-going) bandwidth utilization in leaf nodes and bandwidth bottleneck in nodes near the root. The data driven approach has been widely studied recently to tackle the defects of the tree-based approach mentioned above. However the tree-based approach still has its advantages: deterministic delivery path length and predictable delay, and natural support to PUSH mode content delivery. Because of these advantages the tree-based approach will not be simply replaced by the data-driven approach. Based on this consideration, we propose a cluster-based approach to remedy the disadvantages of the normal tree-based approach and meanwhile retain its advantages as much as possible. By grouping peers into clusters, the content delivery tree constructed by clusters can maintain a stable overlay structure and transmission direction in a dynamic network environment. Simulation results show that our approach can effectively overcome the shortages of the single tree-based approach and outperform the data-driven approach in terms of deterministic content delivery path and predictable path length.