In this work, we conduct a systematic analysis of Native Sparse Attention (NSA) and propose targeted improvements that enhance long-context modeling. A key insight is that alternating between local (sliding-window) and global (compression/selective) attention across layers, rather than using fixed patterns, enables more effective propagation of long-range dependencies and substantially boosts performance on long-sequence tasks. Meanwhile, we further refine NSA’s branches with Latent Attention that the sliding-window branch is enhanced with Multi-head Latent Attention (MLA) while compression and selective branches adopt Group-head Latent Attention (GLA). These changes reduce KV-cache memory by 50% versus NSA while improving the model’s common-sense reasoning and long-text understanding capabilities. Experiments on models from 340M to 1.3B parameters (trained on 15B and 100B tokens) show our method matches or exceeds full attention and native sparse attention in both common-sense reasoning and long-context understanding tasks.

Reinforcement learning (RL) is widely applied to boost the performance of pretrained models, yet its training efficiency is severely constrained by rollout generation. While speculative decoding based on multi-token prediction (MTP) offers a potential acceleration pathway, its widespread adoption is hindered by the absence of MTP in vanilla pretrained models and the rapid degradation of the MTP acceptance length in RL training. To address these issues, this paper proposes MTP-RL, a two-stage framework that pioneers effective training of MTPs in RL and accelerates the rollout phase for diverse models. It involves a pipeline to equip the multi-layer parameter-sharing MTP for all models and an innovative advantage-aware MTP optimization strategy to facilitate policy-aligned training of MTPs. Experiments demonstrate that our method not only achieves stable growth of acceptance length during RL training, but also accelerates RL rollouts, achieving an average 23.1%–55.3% reduction in rollout time compared to baselines.

In Text-to-SQL tasks, existing LLM-based methods often include extensive database schemas in prompts, leading to long context lengths and increased prefilling latency. While user queries typically focus on recurrent table sets—offering an opportunity for KV cache sharing across queries—current inference engines, such as SGLang and vLLM, generate redundant prefix cache copies when processing user queries with varying table orders. To address this inefficiency, we propose precomputing table representations as KV caches offline and querying the required ones online. A key aspect of our approach is the computation of table caches while preserving primary foreign key relationships between tables. Additionally, we construct a Table Trie structure to facilitate efficient KV cache lookups during inference. To enhance cache performance, we introduce a cache management module with a query-reranking strategy to improve cache hit rates and a computation-loading pipeline for parallelizing model inference and cache loading. Experimental results show that our proposed TableCache achieves up to a 3.62× speedup in Time to First Token (TTFT) with negligible performance degradation.

Reinforcement Learning (RL) has become essential for eliciting complex reasoning capabilities in Large Language Models (LLMs). However, the substantial memory overhead of storing Key-Value (KV) caches during long-horizon rollouts acts as a critical bottleneck, often prohibiting efficient training on limited hardware. While existing KV compression techniques offer a remedy for inference, directly applying them to RL training induces a severe policy mismatch, leading to catastrophic performance collapse. To address this, we introduce Sparse-RL, which empowers stable RL training under sparse rollouts. We show that instability arises from a fundamental policy mismatch among the dense old policy, the sparse sampler policy, and the learner policy. To mitigate this issue, Sparse-RL incorporates Sparsity-Aware Rejection Sampling and Importance-based Reweighting to correct the off-policy bias introduced by compression-induced information loss. Experimental results show that Sparse-RL reduces rollout overhead compared to dense baselines while preserving the performance. Furthermore, Sparse-RL inherently implements sparsity-aware training, significantly enhancing model robustness during sparse inference deployment.

We introduce TableLLM, a robust large language model (LLM) with 8 billion parameters, purpose-built for proficiently handling tabular data manipulation tasks, whether they are embedded within documents or spreadsheets, catering to real-world office scenarios. We propose a distant supervision method for training, which comprises a reasoning process extension strategy, aiding in training LLMs to understand reasoning patterns more effectively as well as a cross-way validation strategy, ensuring the quality of the automatically generated data. To evaluate the performance of TableLLM, we have crafted benchmarks tailored to address both document and spreadsheet formats as well as constructed a well-organized evaluation pipeline capable of handling both scenarios. Thorough evaluations underscore the advantages of TableLLM when compared to various existing general-purpose and tabular data-focused LLMs. We have publicly released the model checkpoint, source code, benchmarks, and a web application for user interaction on this anonymized repository.

Despite advancements in large language models (LLMs), non-factual responses still persist in fact-seeking question answering. Unlike extensive studies on post-hoc detection of these responses, this work studies non-factuality prediction (NFP), predicting whether an LLM will generate a non-factual response prior to the response generation. Previous NFP methods have shown LLMs’ awareness of their knowledge, but they face challenges in terms of efficiency and transferability. In this work, we propose a lightweight model named Factuality Lens (FacLens), which effectively probes hidden representations of fact-seeking questions for the NFP task. Moreover, we discover that hidden question representations sourced from different LLMs exhibit similar NFP patterns, enabling the transferability of FacLens across different LLMs to reduce development costs. Extensive experiments highlight FacLens’s superiority in both effectiveness and efficiency.

Speculative decoding (SD) has been demonstrated as an effective technique for lossless LLM inference acceleration. Retrieval-based SD methods, one kind of model-free method, have yielded promising speedup, but they often rely on single retrieval resources, inefficient retrieval methods, and are constrained to certain tasks. This paper presents a novel retrieval-based speculative decoding method that adapts the suffix automaton (SAM) for efficient and accurate draft generation by utilizing the generating text sequence and static text corpus. Unlike existing n-gram matching methods, SAM-Decoding finds the exact longest suffix match, achieving an average time complexity of O(1) per generation step of SAM update and suffix retrieval.It can also integrate with existing methods, adaptively selecting a draft generation strategy based on match length to generalize to broader domains. Extensive experiments on Spec-Bench show that our method is 18% faster than other retrieval-based SD methods. Additionally, when combined with advanced EAGLE-2, it provides an additional speedup of 3.28% – 11.13% across various-sized LLM backbones.

Current large language models (LLMs) often struggle to produce accurate responses on the first attempt for complex reasoning tasks like code generation. Prior research tackles this challenge by generating multiple candidate solutions and validating them with LLM-generated unit tests. The execution results of unit tests serve as reward signals to identify correct solutions. As LLMs always confidently make mistakes, these unit tests are not reliable, thereby diminishing the quality of reward signals. Motivated by the observation that scaling the number of solutions improves LLM performance, we explore the impact of scaling unit tests to enhance reward signal quality. Our pioneer experiment reveals a positive correlation between the number of unit tests and reward signal quality, with greater benefits observed in more challenging problems. Based on these insights, we propose CodeRM-8B, a lightweight yet effective unit test generator that enables efficient and high-quality unit test scaling. Additionally, we implement a dynamic scaling mechanism that adapts the number of unit tests based on problem difficulty, further improving efficiency. Experimental results show that our approach significantly improves performance across various models on three benchmarks (e.g., with gains of 18.43 for Llama3-8B and 3.42 for GPT-4o-mini on HumanEval Plus). The parameters of CodeRM-8B and corresponding training data will be available upon publication.

Current inference scaling methods, such as Self-consistency and Best-of-N, have proven effective in improving the accuracy of LLMs on complex reasoning tasks. However, these methods rely heavily on the quality of candidate responses and are unable to produce correct answers when all candidates are incorrect. In this paper, we propose a novel inference scaling strategy, CoT-based Synthesizer, which leverages CoT reasoning to synthesize superior answers by analyzing complementary information from multiple candidate responses, even when all candidates are flawed. To support a lightweight and cost-effective implementation, we introduce an automated data generation pipeline that creates diverse training data. This enables smaller LLMs trained on this data to improve the inference accuracy of larger models, including API-based LLMs. Experimental results across four benchmark datasets with seven policy models demonstrate that our method significantly enhances performance, with gains of 11.8% for Llama3-8B and 10.3% for GPT-4o on the MATH dataset. The corresponding training data and code are publicly available on the repository.

Pruning has become a widely adopted technique for reducing the hardware requirements of large language models (LLMs). To recover model performance after pruning, post-training is commonly employed to mitigate the resulting performance degradation. While post-training benefits from larger datasets, once the dataset size is already substantial, increasing the training data provides only limited performance gains. To balance post-training cost and model performance, it is necessary to explore the optimal amount of post-training data. Through extensive experiments on the Llama-3 and Qwen-2.5 series models, pruned using various common pruning methods, we uncover the scaling Law for Post-training after model Pruning, referred to as the P² Law. This law identifies four key factors for predicting the pruned model’s post-training loss: the model size before pruning, the number of post-training tokens, the pruning rate, and the model’s loss before pruning. Moreover, P² Law can generalize to larger dataset sizes, larger model sizes, and higher pruning rates, offering valuable insights for the post-training of pruned LLMs.

Direct Preference Optimization (DPO) has proven effective in complex reasoning tasks like math word problems and code generation. However, when applied to Text-to-SQL datasets, it often fails to improve performance and can even degrade it. Our investigation reveals the root cause: unlike math and code tasks, which naturally integrate Chain-of-Thought (CoT) reasoning with DPO, Text-to-SQL datasets typically include only final answers (gold SQL queries) without detailed CoT solutions. By augmenting Text-to-SQL datasets with synthetic CoT solutions, we achieve, for the first time, consistent and significant performance improvements using DPO.Our analysis shows that CoT reasoning is crucial for unlocking DPO’s potential, as it mitigates reward hacking, strengthens discriminative capabilities, and improves scalability. These findings offer valuable insights for building more robust Text-to-SQL models. To support further research, we publicly release the code and CoT-enhanced datasets: https://github.com/RUCKBReasoning/DPO_Text2SQL.

Conversational Question Generation (CQG) enhances the interactivity of conversational question-answering systems in fields such as education, customer service, and entertainment. However, traditional CQG, focusing primarily on the immediate context, lacks the conversational foresight necessary to guide conversations toward specified conclusions. This limitation significantly restricts their ability to achieve conclusion-oriented conversational outcomes. In this work, we redefine the CQG task as Conclusion-driven Conversational Question Generation (CCQG) by focusing on proactivity, not merely reacting to the unfolding conversation but actively steering it towards a conclusion-oriented question-answer pair. To address this, we propose a novel approach, called Proactive Conversational Question Planning with self-Refining (PCQPR). Concretely, by integrating a planning algorithm inspired by Monte Carlo Tree Search (MCTS) with the analytical capabilities of large language models (LLMs), PCQPR predicts future conversation turns and continuously refines its questioning strategies. This iterative self-refining mechanism ensures the generation of contextually relevant questions strategically devised to reach a specified outcome. Our extensive evaluations demonstrate that PCQPR significantly surpasses existing CQG methods, marking a paradigm shift towards conclusion-oriented conversational question-answering systems.

Recent works on knowledge base question answering (KBQA) retrieve subgraphs for easier reasoning. The desired subgraph is crucial as a small one may exclude the answer but a large one might introduce more noises. However, the existing retrieval is either heuristic or interwoven with the reasoning, causing reasoning on the partial subgraphs, which increases the reasoning bias when the intermediate supervision is missing. This paper proposes a trainable subgraph retriever (SR) decoupled from the subsequent reasoning process, which enables a plug-and-play framework to enhance any subgraph-oriented KBQA model. Extensive experiments demonstrate SR achieves significantly better retrieval and QA performance than existing retrieval methods. Via weakly supervised pre-training as well as the end-to-end fine-tuning, SR achieves new state-of-the-art performance when combined with NSM (He et al., 2021), a subgraph-oriented reasoner, for embedding-based KBQA methods. Codes and datasets are available online (https://github.com/RUCKBReasoning/SubgraphRetrievalKBQA)