Recurrent neural networks (RNNs), such as linear attention and state-space models, have gained popularity due to their constant per-token complexity when processing long contexts. However, these recurrent models struggle with tasks that require accurate recall of contextual information from long contexts, because all contextual information is compressed into a fixed-size recurrent state. Previous studies have shown that recall ability is positively correlated with the recurrent state size, yet directly training RNNs with large recurrent states results in high training costs. In this paper, we introduce StateX, a post-training framework that efficiently expands the states of pre-trained RNNs. For two popular classes of RNNs, linear attention and state-space models, we design post-training architectural modifications in StateX, to scale up the state size with no or negligible increase in model parameters. Experiments on models with up to 1.3B parameters demonstrate that StateX efficiently enhances the recall and in-context learning performance of RNNs without incurring high post-training costs or compromising other capabilities.

Chain-of-thought (CoT) supervised fine-tuning (SFT) is widely adopted to improve reasoning ability, yet we find that it systematically degrades long-context recall in hybrid linear-attention models. Across architectures including HypeNet and Jet-Nemotron, retrieval performance on Needle-In-A-Haystack (NIAH) deteriorates substantially after CoT-SFT, and the degradation becomes more severe under harder retrieval settings and longer context windows. For example, HypeNet-9B on NIAH-S2@256K decreases from 67.2% to 9.4%. We attribute this to CoT-SFT biasing attention gradients toward short-range patterns, disrupting query-key projections (W_Q,W_K) that are responsible for long-range routing. Motivated by this observation, we propose **QK-Restore**, a training-free method that restores only W_Q and W_K from the pre-SFT checkpoint while preserving all other post-SFT parameters. We further introduce a Procrustes variant to balance routing preservation and reasoning adaptation. Across architectures, **QK-Restore** consistently restores long-context capability at zero training cost while preserving reasoning performance; for instance, on HypeNet-5B it improves S3@256K from 65.4% to 76.4% while maintaining strong reasoning performance.

Modern language models increasingly adopt hybrid architectures that combine full attention with efficient attention modules, such as sliding-window attention (SWA) and recurrent sequence mixers. However, how these efficient modules shape model capabilities remains poorly understood. To address this gap, we conduct a systematic analysis across hybrid architectures from three perspectives: scaling behavior, mechanism analysis, and architecture design. First, from a scaling perspective, we find that efficient-attention design primarily affects how fast long-context capability emerges, while different hybrids eventually converge to comparable long-context performance under sufficient training. Second, mechanistically, we show that long-range retrieval is mainly carried by full attention, whereas efficient attention shapes its optimization trajectory. This explains a counter-intuitive phenomenon we call Large-Window Laziness: larger SWA windows can delay the formation of retrieval heads in full-attention layers. Third, guided by this mechanism, we show that applying No Positional Encoding (NoPE) to only the full-attention layers of a small-window SWA hybrid substantially improves long-context performance with negligible impact on short-context performance.

Hybrid architectures that combine softmax attention blocks and recurrent neural networks (RNNs) have shown a desirable performance-throughput tradeoff for long-context modeling, but their adoption and studies are hindered by the prohibitive cost of large-scale pre-training from scratch. Some recent studies have shown that pre-trained softmax attention blocks can be converted into RNN blocks through parameter transfer and knowledge distillation. However, these transfer methods require substantial amounts of training data (more than 10B tokens), and the resulting hybrid models also exhibit poor long-context performance, which is the scenario where hybrid models enjoy significant inference speedups over Transformer-based models. In this paper, we present HALO (Hybrid Attention via Layer Optimization), a pipeline for distilling Transformer models into RNN-attention hybrid models. We then present HypeNet, a hybrid architecture with superior length generalization enabled by a novel position encoding scheme (named HyPE) and various architectural modifications. We convert the Qwen3 series into HypeNet using HALO, achieving performance comparable to the original Transformer models while enjoying superior long-context performance and efficiency. The conversion requires just 2.3B tokens, less than 0.01% of their pre-training data.

Grouped-Query Attention (GQA) is a widely adopted strategy for reducing the computational cost of attention layers in large language models (LLMs). However, current GQA configurations are often suboptimal because they overlook how context length influences inference cost. Since inference cost grows with context length, the most cost-efficient GQA configuration should vary accordingly. In this work, we analyze the relationship among context length, model size, GQA configuration, and model loss, and introduce two innovations: (1) we decouple the total head size from the hidden size, enabling more flexible control over attention FLOPs; and (2) we jointly optimize the model size and the GQA configuration to arrive at a better allocation of inference resources between attention layers and other components. Our analysis reveals that commonly used GQA configurations are highly suboptimal for long-context scenarios. Moreover, we propose a recipe for deriving cost-optimal GQA configurations. Our results show that for long-context scenarios, one should use fewer attention heads while scaling up the model size. Configurations selected by our recipe can reduce both memory usage and FLOPs by more than 50% compared to Llama-3’s GQA, with no degradation in model capabilities. Our findings offer valuable insights for designing efficient long-context LLMs.

This study presents a multi-modal multi-granularity tokenizer specifically designed for analyzing ancient Chinese scripts, focusing on the Chu bamboo slip (CBS) script used during the Spring and Autumn and Warring States period (771-256 BCE) in Ancient China. Considering the complex hierarchical structure of ancient Chinese scripts, where a single character may be a combination of multiple sub-characters, our tokenizer first adopts character detection to locate character boundaries. Then it conducts character recognition at both the character and sub-character levels. Moreover, to support the academic community, we assembled the first large-scale dataset of CBSs with over 100K annotated character image scans. On the part-of-speech tagging task built on our dataset, using our tokenizer gives a 5.5% relative improvement in F1-score compared to mainstream sub-word tokenizers. Our work not only aids in further investigations of the specific script but also has the potential to advance research on other forms of ancient Chinese scripts.

Large language models (LLMs) can make predictions using parametric knowledge – knowledge encoded in the model weights – or contextual knowledge – knowledge presented in the context. In many scenarios, a desirable behavior is that LLMs give precedence to contextual knowledge when it conflicts with the parametric knowledge, and fall back to using their parametric knowledge when the context is irrelevant. This enables updating and correcting the model’s knowledge by in-context editing instead of retraining. Previous works have shown that LLMs are inclined to ignore contextual knowledge and fail to reliably fall back to parametric knowledge when presented with irrelevant context. In this work, we discover that, with proper prompting methods, instruction-finetuned LLMs can be highly controllable by contextual knowledge and robust to irrelevant context. Utilizing this feature, we propose EREN (Edit models by REading Notes) to improve the scalability and robustness of LLM editing. To better evaluate the robustness of model editors, we collect a new dataset, that contains irrelevant questions that are more challenging than the ones in existing datasets. Empirical results show that our method outperforms current state-of-the-art methods by a large margin. Unlike existing techniques, it can integrate knowledge from multiple edits, and correctly respond to syntactically similar but semantically unrelated inputs (and vice versa). The source code can be found at https://github.com/thunlp/EREN.

As large language models (LLMs) increasingly permeate daily lives, there is a growing demand for real-time interactions that mirror human conversations. Traditional turn-based chat systems driven by LLMs prevent users from verbally interacting with the system while generating responses.To overcome these limitations, we adapt existing LLMs to duplex models so that they can listen to users while generating output and dynamically adjust themselves to provide instant feedback.Specifically, we divide the queries and responses of conversations into several time slices and then adopt a time-division-multiplexing (TDM) encoding-decoding strategy to process these slices pseudo-simultaneously.Furthermore, to make LLMs proficient enough to handle real-time conversations, we build a fine-tuning dataset consisting of alternating time slices of queries and responses and covering typical feedback types in instantaneous interactions.Our experiments show that although the queries and responses of conversations are segmented into incomplete slices for processing, LLMs can preserve their original performance on standard benchmarks with a few fine-tuning steps on our dataset. Automatic and human evaluation indicate that duplex models make user-AI interactions more natural and human-like, and greatly improve user satisfaction compared to vanilla LLMs. Our duplex model and dataset will be released soon.

Processing and reasoning over long contexts is crucial for many practical applications of Large Language Models (LLMs), such as document comprehension and agent construction. Despite recent strides in making LLMs process contexts with more than 100K tokens, there is currently a lack of a standardized benchmark to evaluate this long-context capability. Existing public benchmarks typically focus on contexts around 10K tokens, limiting the assessment and comparison of LLMs in processing longer contexts. In this paper, we propose , the first LLM benchmark featuring an average data length surpassing 100K tokens. comprises synthetic and realistic tasks spanning diverse domains in English and Chinese. The tasks in are designed to require an understanding of long dependencies in contexts and make simply retrieving a limited number of passages from contexts not sufficient for these tasks. Based on , we evaluate several state-of-the-art LLMs tailored for processing long contexts. The experimental results indicate that existing long-context LLMs still require significant advancements to process 100K+ contexts effectively. Furthermore, we present three intriguing analyses regarding the behavior of LLMs processing long context. Our code and data is released.

Tokenization is fundamental to pretrained language models (PLMs). Existing tokenization methods for Chinese PLMs typically treat each character as an indivisible token. However, they ignore the unique feature of the Chinese writing system where additional linguistic information exists below the character level, i.e., at the sub-character level. To utilize such information, we propose sub-character (SubChar for short) tokenization. Specifically, we first encode the input text by converting each Chinese character into a short sequence based on its glyph or pronunciation, and then construct the vocabulary based on the encoded text with sub-word segmentation. Experimental results show that SubChar tokenizers have two main advantages over existing tokenizers: 1) They can tokenize inputs into much shorter sequences, thus improving the computational efficiency. 2) Pronunciation-based SubChar tokenizers can encode Chinese homophones into the same transliteration sequences and produce the same tokenization output, hence being robust to homophone typos. At the same time, models trained with SubChar tokenizers perform competitively on downstream tasks. We release our code and models at https://github.com/thunlp/SubCharTokenization to facilitate future work.

For many real-world applications, the user-generated inputs usually contain various noises due to speech recognition errors caused by linguistic variations or typographical errors (typos). Thus, it is crucial to test model performance on data with realistic input noises to ensure robustness and fairness. However, little study has been done to construct such benchmarks for Chinese, where various language-specific input noises happen in the real world. In order to fill this important gap, we construct READIN: a Chinese multi-task benchmark with REalistic And Diverse Input Noises. READIN contains four diverse tasks and requests annotators to re-enter the original test data with two commonly used Chinese input methods: Pinyin input and speech input. We designed our annotation pipeline to maximize diversity, for example by instructing the annotators to use diverse input method editors (IMEs) for keyboard noises and recruiting speakers from diverse dialectical groups for speech noises. We experiment with a series of strong pretrained language models as well as robust training methods, we find that these models often suffer significant performance drops on READIN even with robustness methods like data augmentation. As the first large-scale attempt in creating a benchmark with noises geared towards user-generated inputs, we believe that READIN serves as an important complement to existing Chinese NLP benchmarks. The source code and dataset can be obtained from https://github.com/thunlp/READIN.

Recently, pre-trained language models (PLMs) have achieved great success on various NLP tasks and have shown a trend of exponential growth in model size. To alleviate the unaffordable computational costs brought by the size growth, model compression has been widely explored. Existing efforts have achieved promising results in compressing medium-sized models for specific tasks, while task-agnostic compression for big models with over billions of parameters is rarely studied. Task-agnostic compression can provide an efficient and versatile big model for both prompting and delta tuning, leading to a more general impact than task-specific compression. Hence, we introduce a task-agnostic compression toolkit BMCook for big models. In BMCook, we implement four representative compression methods, including quantization, pruning, distillation, and MoEfication. Developers can easily combine these methods towards better efficiency. To evaluate BMCook, we apply it to compress T5-3B (a PLM with 3 billion parameters). We achieve nearly 12x efficiency improvement while maintaining over 97% of the original T5-3B performance on three typical NLP benchmarks. Moreover, the final compressed model also significantly outperforms T5-base (a PLM with 220 million parameters), which has a similar computational cost. BMCook is publicly available at https://github.com/OpenBMB/BMCook.