Long-context capability has become increasingly critical for next-generation Large Language Models (LLMs). However, the quadratic time complexity of standard softmax attention introduces substantial computation and memory overhead as sequence length grows, limiting the scalability of LLMs in long-context scenarios. Beyond existing approaches such as linear recurrent architectures and KV-cache compression methods, we propose Kwai Summary Attention (KSA), a novel attention mechanism that reduces sequence modeling cost by compressing historical contexts into learnable summary tokens. Specifically, summary tokens are periodically inserted into the input sequence and interleaved with text tokens, partitioning the sequence into distinct chunks while enabling text tokens and summary tokens to operate under different attention scopes. This design preserves dense local expressiveness through full attention among neighboring text tokens, while modeling long-range dependencies efficiently via summary tokens. Empirical results demonstrate that hybrid-KSA consistently outperforms existing hybrid attention variants in long-context modeling. Moreover, KSA is fully orthogonal to GQA and MLA during inference, and combining KSA with these methods achieves up to 8x KV-cache compression while preserving model performance.

Large language models (LLMs) have demonstrated strong capabilities in processing long contexts, enabling them to tackle tasks involving long textual inputs such as multi-turn conversations, legal documents, or retrieved documents in Retrieval-Augmented Generation (RAG) systems. However, despite their ability to handle long sequences, the resulting decoding latency and memory overhead remain substantial, posing challenges for real-world deployment. Recent advances in KV Cache reuse have shown potential to mitigate these costs, but these methods still suffer from notable performance degradation. To address this issue, we conduct an in-depth investigation of recomputation-based reuse methods and observe that the recomputed tokens often fail to align with the context segments most relevant to the question. This misalignment hinders proper updates to the critical contextual representations. Therefore, we propose the Attention-Aware Accurate KV Cache Fusion algorithm (A³), which precomputes and selectively fuses the KV Cache of text chunks based on their relevance to the question, achieving accurate integration with minimal computational overhead. Extensive experiments on various benchmarks and LLMs demonstrate that A³ achieves competitive task performance compared to five baselines while reducing the time-to-first-token (TTFT) by 2×.

Recent reasoning models such as OpenAI-o1 and DeepSeek-R1 have shown strong performance on complex tasks including mathematical reasoning and code generation. However, this performance gain comes with substantially longer output sequences, leading to significantly increased deployment costs. In particular, long outputs require large KV caches, resulting in high memory consumption and severe bandwidth pressure during attention computation. Most existing KV cache optimization methods are designed for long-input, short output scenarios and are ineffective for the long-output setting of reasoning models. Moreover, importance estimation in prior work is computationally expensive and becomes prohibitive when continuous re-evaluation is required during long generation. To address these challenges, we propose LongFlow, a KV cache compression method with an efficient importance estimation metric derived from an intermediate result of attention computation using only the current query. This design introduces negligible computational overhead and requires no auxiliary storage. We further develop a custom kernel that fuses FlashAttention, importance estimation, and token eviction into a single optimized operator, improving system-level efficiency. Experiments show that LongFlow achieves up to an 11.8x throughput improvement with 80% KV cache compression with minimal impact on model accuracy.

Reinforcement learning with verifiable rewards (RLVR) has been effective on tasks with structured solutions like math and coding, but its reliance on simple, rule-based verifiers creates a fundamental bottleneck. We find their applicability is surprisingly narrow even in structured domains, a limitation that is compounded at scale: rule-based systems can paradoxically degrade in performance as multi-domain, free-form training data increases. To overcome these challenges, we propose a new RLVR framework that uses a generative verifier to provide soft, probabilistic rewards. Our key insight is that powerful LLMs show high agreement with human evaluators when judging answer correctness given a ground-truth reference, allowing us to automate reward generation without costly human annotation. Our experiments demonstrate the effectiveness of this approach. We show that a compact 7B generative reward model can guide a 7B policy model to decisively outperform models up to 10x its size, including the 72B Qwen2.5-Instruct (by a margin of +8.6%). This effectiveness is robust, holding true across diverse training datasets with answers sourced from experts, web users, and other LLMs, and generalizes strongly to seven out-of-distribution benchmarks. Our work provides a scalable and effective framework for extending RLVR beyond the limitations of pattern-based verification to complex, noisy, real-world domains.

The powerful generative capacity of Large Language Models (LLMs) has instigated a paradigm shift in recommendation. However, existing generative models (e.g., OneRec) operate as implicit predictors, critically lacking the capacity for explicit and controllable reasoning—a key advantage of LLMs. To bridge this gap, we propose OneRec-Think, a unified framework that seamlessly integrates dialogue, reasoning, and personalized recommendation. OneRec-Think incorporates: (1) Itemic Alignment: cross-modal Item-Textual Alignment for semantic grounding; (2) Reasoning Activation: Reasoning Scaffolding to activate LLM reasoning within the recommendation context; and (3) Reasoning Enhancement, where we design a recommendation-specific reward function that accounts for the multi-validity nature of user preferences. Experiments across public benchmarks show state-of-the-art performance. Moreover, our proposed “Think-Ahead” architecture enables effective industrial deployment, achieving a 0.159% gain in APP Stay Time and validating the practical efficacy of the model’s explicit reasoning capability.

Autoregressive language models demonstrate excellent performance in various scenarios. However, the inference efficiency is limited by its one-step-one-word generation mode, which has become a pressing problem recently as the models become increasingly larger. Speculative decoding employs a “draft and then verify” mechanism to allow multiple tokens to be generated in one step, realizing lossless acceleration. Existing methods mainly adopt fixed heuristic draft structures, which do not adapt to different situations to maximize the acceptance length during verification. To alleviate this dilemma, we propose OPT-Tree, an algorithm to construct adaptive and scalable draft trees, which can be applied to any autoregressive draft model. It searches the optimal tree structure that maximizes the mathematical expectation of the acceptance length in each decoding step. Experimental results reveal that OPT-Tree outperforms the existing draft structures and achieves a speed-up ratio of up to 3.2 compared with autoregressive decoding. If the draft model is powerful enough and the node budget is sufficient, it can generate more than ten tokens in a single step. Our code is available at https://github.com/Jikai0Wang/OPT-Tree.

Rapid integration of large language models (LLMs) into societal applications has intensified concerns about their alignment with universal ethical principles, as their internal value representations remain opaque despite behavioral alignment advancements. Current approaches struggle to systematically interpret how values are encoded in neural architectures, limited by datasets that prioritize superficial judgments over mechanistic analysis. We introduce ValueLocate, a mechanistic interpretability framework grounded in the Schwartz Values Survey, to address this gap. Our method first constructs ValueInsight, a dataset that operationalizes four dimensions of universal value through behavioral contexts in the real world. Leveraging this dataset, we develop a neuron identification method that calculates activation differences between opposing value aspects, enabling precise localization of value-critical neurons without relying on computationally intensive attribution methods. Our proposed validation method demonstrates that targeted manipulation of these neurons effectively alters model value orientations, establishing causal relationships between neurons and value representations. This work advances the foundation for value alignment by bridging psychological value frameworks with neuron analysis in LLMs.

The impressive capabilities of Large Language Models (LLMs) come at the cost of substantial computational resources during deployment. While KV Cache can significantly reduce recomputation during inference, it also introduces additional memory overhead. KV Cache quantization presents a promising solution, striking a good balance between memory usage and accuracy. Previous research has shown that the Keys are distributed by channel, while the Values are distributed by token. Consequently, the common practice is to apply channel-wise quantization to the Keys and token-wise quantization to the Values. However, our further investigation reveals that a small subset of unusual tokens exhibit unique characteristics that deviate from this pattern, which can substantially impact quantization accuracy. To address this, we develop a simple yet effective method to identify these tokens accurately during the decoding process and exclude them from quantization as outlier tokens, significantly improving overall accuracy. Extensive experiments show that our method achieves significant accuracy improvements under 2-bit quantization and can deliver a 6.4 times reduction in memory usage and a 2.3 times increase in throughput.

Large Language Models (LLMs) have demonstrated an impressive capability known as In-context Learning (ICL), which enables them to acquire knowledge from textual demonstrations without the need for parameter updates.However, many studies have highlighted that the model’s performance is sensitive to the choice of demonstrations, presenting a significant challenge for practical applications where we lack prior knowledge of user queries.Consequently, we need to construct an extensive demonstration pool and incorporate external databases to assist the model, leading to considerable time and financial costs.In light of this, some recent research has shifted focus towards zero-shot ICL, aiming to reduce the model’s reliance on external information by leveraging their inherent generative capabilities. Despite the effectiveness of these approaches, the content generated by the model may be unreliable, and the generation process is time-consuming.To address these issues, we propose Demonstration Augmentation for In-context Learning (DAIL), which employs the model’s previously predicted historical samples as demonstrations for subsequent ones.DAIL brings no additional inference cost and does not rely on the model’s generative capabilities.Our experiments reveal that DAIL can significantly improve the model’s performance over direct zero-shot inference and can even outperform few-shot ICL without any external information.

Although pre-trained language models (PLM) have achieved great success in question answering (QA), their robustness is still insufficient to support their practical applications, especially in the face of distribution shifts. Recently, test-time adaptation (TTA) has shown great potential for solving this problem, which adapts the model to fit the test samples at test time. However, TTA sometimes causes model collapse, making almost all the model outputs incorrect, which has raised concerns about its stability and reliability. In this paper, we delve into why TTA causes model collapse and find that the imbalanced label distribution inherent in QA is the reason for it. To address this problem, we propose Anti-Collapse Fast test-time adaptation (Anti-CF), which utilizes the source model‘s output to regularize the update of the adapted model during test time. We further design an efficient side block to reduce its inference time. Extensive experiments on various distribution shift scenarios and pre-trained language models (e.g., XLM-RoBERTa, BLOOM) demonstrate that our method can achieve comparable or better results than previous TTA methods at a speed close to vanilla forward propagation, which is 1.8× to 4.4× speedup compared to previous TTA methods.

Goal-oriented dialogue systems face a trade-off between fluent language generation and task-specific control. While supervised learning with large language models is capable of producing realistic text, how to steer such responses towards completing a specific task without sacrificing language quality remains an open question. In this work, we formulate goal-oriented dialogue as a partially observed Markov decision process, interpreting the language model as a representation of both the dynamics and the policy. This view allows us to extend techniques from learning-based control, such as task relabeling, to derive a simple and effective method to finetune language models in a goal-aware way, leading to significantly improved task performance. We additionally introduce a number of training strategies that serve to better focus the model on the task at hand. We evaluate our method, Context-Aware Language Models (CALM), on a practical flight-booking task using AirDialogue. Empirically, CALM outperforms the state-of-the-art method by 7% in terms of task success, matching human-level task performance.

Human language understanding operates at multiple levels of granularity (e.g., words, phrases, and sentences) with increasing levels of abstraction that can be hierarchically combined. However, existing deep models with stacked layers do not explicitly model any sort of hierarchical process. In this paper, we propose a recursive Transformer model based on differentiable CKY style binary trees to emulate this composition process, and we extend the bidirectional language model pre-training objective to this architecture, attempting to predict each word given its left and right abstraction nodes. To scale up our approach, we also introduce an efficient pruning and growing algorithm to reduce the time complexity and enable encoding in linear time. Experimental results on language modeling and unsupervised parsing show the effectiveness of our approach.