Humans need to study only a handful of well-written textbooks to master a discipline and attempt its hardest problems. We argue that an ideal self-evolution method should share the same property, that is autonomously learning from raw training material for transferable problem-solving capability. However, we still lack a direct measurement for it. We introduce StudyBench, a controlled physics benchmark that directly measures how efficiently a self-evolution method converts training material into capability. We organise the test set into an Application Set, consisting of difficult textbook problems and evaluating absorption ability, and a Transfer Set, consisting of olympiad-level problems and evaluating transfer ability. Benchmarking representative self-evolution methods across three base models, we find that improvements on the Application Set rarely translate to the harder Transfer Set. A guidance ablation exposes a Guidance Gap: even the strongest method closes only a small fraction of what the same material unlocks when supplied as in-context guidance. Besides, every method hits a Compute Plateau, saturating well before exhausting its compute budget. The remaining gap is therefore a method problem rather than a data or compute problem. By offering a clean and controlled benchmark, StudyBench turns self-evolution progress from an open-ended pursuit into a measurable target for future research. Our code is released at https://github.com/thunlp/StudyBench.

Self-evolving agents improve future behavior from interaction experience, yet existing evaluations typically optimize under fixed execution conditions and do not test recovery after those conditions change. To address this gap, we introduce PACE-Bench (**P**hysics **A**daptation via **C**ode **E**volution), a simulator-grounded benchmark of 144 source-to-target adaptation pairs across six physics domains. Each pair links a source environment to a mutated target environment with the same goal and interface. A code-driven design that succeeds in the source fails in the target, where agents must iteratively adapt it into a working target design using diagnostic sandbox feedback within a limited attempt budget. We compare ten self-evolving methods from four paradigms. The benchmark remains far from saturated: Reflexion + Qwen3-14B succeeds on only 35.9% of full-benchmark pairs, while GPT-5.5 solves 66.7% of the Statics subset under the full budget. Together, these results show that simulator-grounded reflection is more reliable than unverified self-revision, while memory anchors agents to early designs and broad tree search explores without converging. Even revealing exact physical changes does not raise the performance ceiling, pointing to mechanism redesign rather than parameter inference as the central bottleneck. Data and code are available at https://github.com/thunlp/PACE-Bench.

Large Language Models (LLMs) have demonstrated strong performance across a wide range of NLP tasks. However, they often exhibit suboptimal behaviors and inconsistencies when exposed to unfamiliar external information, underscoring their limitations in effectively leveraging such knowledge. Inspired by constructivist learning theory, we propose ThinkNote, a novel framework that enhances the external knowledge utilization of LLMs through a two-stage constructivist cognitive modeling process. Specifically, ThinkNote performs knowledge assimilation to align new information with the model’s parametric memory, forming a coherent internal representation. It then applies thought accommodation to adapt internal reasoning, thereby promoting more consistent and reliable outputs. Extensive experimental results demonstrate that ThinkNote achieves a 10% improvement over strong baseline methods on various question-answering benchmarks. Further analysis indicates that ThinkNote effectively integrates and utilizes external knowledge to help LLMs generate accurate responses and improves their self-consistency. All data and code will be publicly available at https://github.com/OpenMatch/ThinkNote.

Masked diffusion models (MDMs), which leverage bidirectional attention and a denoising process, are narrowing the performance gap with autoregressive models (ARMs). However, their internal attention mechanisms remain under-explored. This paper investigates the attention behaviors in MDMs, revealing the phenomenon of Attention Floating. Unlike ARMs, where attention converges to a fixed sink, MDMs exhibit dynamic, dispersed attention anchors that shift across denoising steps and layers. Further analysis reveals its Shallow Structure-Aware, Deep Content-Focused attention mechanism: shallow layers utilize floating tokens to build a global structural framework, while deeper layers allocate more capability toward capturing semantic content. Empirically, this distinctive attention pattern provides a mechanistic explanation for the strong in-context learning capabilities of MDMs, allowing them to double the performance compared to ARMs in knowledge-intensive tasks. All codes and datasets are available at https://github.com/NEUIR/Attention-Floating.

In-context learning (ICL) tackles downstream tasks conditioned on demonstrations. This paradigm relies heavily on neighbor-based matching between queries and demonstrations, limiting out-of-distribution (OOD) performance. For better generalization, rule-based inductive reasoning derives task-level regularities, which we term task-level inductive representations (TIR), before directly answering. However, existing methods lack generalizable inductive ability across diverse tasks. Inspired by meta-learning, we propose Cognitive-Decoupling meta-Induction (CoDI) for TIR-based inductive reasoning on general tasks. CoDI comprises a Cognitive-Decoupling Framework for query-blind TIR induction, and a Dual-Loop Meta-Inductive Learning method for cross-task inducer optimization. Empirically, CoDI achieves substantial performance gains on five ID and OOD benchmarks, and preserves OOD gains without explicit induction at inference. Code is available at https://github.com/Yuqian-Wang2277/codi-meta-induction.

Decoding throughput gains from larger inference batches are often limited by GPU memory, mainly due to the large size of KV cache. Training-free KV cache offloading mitigates this issue by storing redundant context on the CPU and fetching sparse KV subsets, but often harms long-generation quality due to training-inference mismatch of sparse patterns. Trainable sparse attention can improve quality, yet its unconstrained KV accesses are inefficient for offloading and may incur excessive CPU–GPU transfers. To this end, we propose NOSA, a trainable sparse attention mechanism designed for KV cache offloading. NOSA explicitly bounds CPU–GPU KV transfers, enabling low communication overhead and high decoding throughput. We further develop NOSI, an offloading inference system that realizes these efficiency gains. Experiments on 1,3,8B LLMs show that NOSA outperforms KV cache offloading baselines on general, long-input, and long-generation tasks, while improving decoding throughput by up to 5.04×, 1.92×, and 1.83× over FullAttn, InfLLMv2, and ShadowKV, respectively.

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.

The efficiency of long-video inference remains a critical bottleneck, mainly due to the dense computation in the prefill stage of Large Multimodal Models (LMMs). Existing methods either compress visual embeddings or apply sparse attention on a single GPU, yielding limited acceleration or degraded performance and restricting LMMs from handling longer, more complex videos. To overcome these issues, we propose APB-V, a sequence-parallel framework with optimized attention that accelerates long-video inference across multiple GPUs. By distributing approximate attention, APB-V reduces computation and increases parallelism, enabling efficient processing of more visual embeddings without compression and thereby improving task performance. System-level optimizations, such as load balancing and fused forward passes, further unleash the potential of APB-V, delivering speedups of 12.72x, 1.70x, and 1.18x over FlashAttn, ZigZagRing, and APB, without notable performance loss.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by integrating external knowledge. A critical yet underexplored challenge in RAG is document segmentation, also known as document chunking. Existing widely-used rule-based chunking methods usually lead to suboptimal splits, where overly large chunks introduce irrelevant information and small chunks lack semantic coherence. Existing semantic-based approaches either require costly LLM calls or fail to adaptively group contextually related sentences. To address these limitations, we propose PIC, Pseudo-Instruction for document Chunking), a simple yet effective method that leverages document summaries as pseudo-instructions to guide chunking. By computing semantic similarity between sentences and the summary, PIC dynamically groups sentences into chunks that align with the document’s key themes, ensuring semantic completeness and relevance to potential user instructions. Experiments on multiple open-domain question-answering benchmarks demonstrate that PIC can significantly improve retrieval accuracy (Hits@k) and end-to-end QA performance (Exact Match) without any additional training.

While long-context inference is crucial for advancing large language model (LLM) applications, its prefill speed remains a significant bottleneck. Current approaches, including sequence parallelism strategies and compute reduction through approximate attention mechanisms, still fall short of delivering optimal inference efficiency. This hinders scaling the inputs to longer sequences and processing long-context queries in a timely manner. To address this, we introduce APB, an efficient long-context inference framework that leverages multi-host approximate attention to enhance prefill speed by reducing compute and enhancing parallelism simultaneously. APB introduces a communication mechanism for essential key-value pairs within a sequence parallelism framework, enabling a faster inference speed while maintaining task performance. We implement APB by incorporating a tailored FlashAttn kernel alongside optimized distribution strategies, supporting diverse models and parallelism configurations. APB achieves speedups of up to 9.2×, 4.2×, and 1.6× compared with FlashAttn, RingAttn, and StarAttn, respectively, without any observable task performance degradation.

Legal Argument-Pair Extraction (LAE) is dedicated to the identification of interactive arguments targeting the same subject matter within legal complaints and corresponding defenses. This process serves as a foundation for automatically recognizing the focal points of disputes. Current methodologies predominantly conceptualize LAE as a supervised sentence-pair classification problem and usually necessitate extensive manual annotations, thereby constraining their scalability and general applicability. To this end, we present an innovative approach to LAE that focuses on fine-grained alignment of argument pairs, building upon coarse-grained complaint-defense pairs. This strategy stems from two key observations: 1) In general, every argument presented in a legal complaint is likely to be addressed by at least one corresponding argument in the defense. 2) It’s rare for multiple complaint arguments to be addressed by a single defense argument; rather, each complaint argument usually corresponds to a unique defense argument. Motivated by these insights, we develop a specialized pre-training framework. Our model employs pre-training objectives designed to exploit the coarse-grained supervision signals. This enables expressive representations of legal arguments for LAE, even when working with a limited amount of labeled data. To verify the effectiveness of our model, we construct the largest LAE datasets from two representative causes, private lending, and contract dispute. The experimental results demonstrate that our model can effectively capture informative argument knowledge from unlabeled complaint-defense pairs and outperform the unsupervised and supervised baselines by 3.7 and 2.4 points on average respectively. Besides, our model can reach superior accuracy with only half manually annotated data. The datasets and code can be found in https://github.com/thunlp/LAE.

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.

Speculative decoding is a widely used method that accelerates the generation process of large language models (LLMs) with no compromise in model performance. It achieves this goal by using an existing smaller model for drafting and then employing the target LLM to verify the draft in a low-cost parallel manner. Under such a drafting-verification framework, drafting efficiency has become a bottleneck in the final speedup of speculative decoding. Therefore, generating longer drafts at less cost can lead to better decoding speedup. To achieve this, we introduce Ouroboros, which can generate draft phrases to parallelize the drafting process and meanwhile lengthen drafts in a training-free manner. The experimental results on various typical text generation tasks show that Ouroboros can achieve speedups of up to 2.4× over speculative decoding and 3.9× over vanilla decoding, without fine-tuning draft and target models. Code available at https://github.com/thunlp/Ouroboros.

Legal case retrieval (LCR) aims to provide similar cases as references for a given fact description. This task is crucial for promoting consistent judgments in similar cases, effectively enhancing judicial fairness and improving work efficiency for judges. However, existing works face two main challenges for real-world applications: existing works mainly focus on case-to-case retrieval using lengthy queries, which does not match real-world scenarios; and the limited data scale, with current datasets containing only hundreds of queries, is insufficient to satisfy the training requirements of existing data-hungry neural models. To address these issues, we introduce an automated method to construct synthetic query-candidate pairs and build the largest LCR dataset to date, LEAD, which is hundreds of times larger than existing datasets. This data construction method can provide ample training signals for LCR models. Experimental results demonstrate that model training with our constructed data can achieve state-of-the-art results on two widely-used LCR benchmarks. Besides, the construction method can also be applied to civil cases and achieve promising results. The data and codes can be found in https://github.com/thunlp/LEAD.

Large language models (LLMs) have achieved remarkable results on NLP tasks but at the expense of huge parameter sizes and the consequent computational costs. In this paper, we propose Variator, a parameter-efficient acceleration method that enhances computational efficiency through plug-and-play compression plugins. Compression plugins are designed to reduce the sequence length via compressing multiple hidden vectors into one and trained with original LLMs frozen. Different from traditional model acceleration methods, which compress LLMs to smaller sizes, Variator offers two distinct advantages: (1) In real-world applications, the plug-and-play nature of our compression plugins enables dynamic selection of different compression plugins with varying acceleration ratios based on the current workload. (2) The compression plugin comprises a few compact neural network layers with minimal parameters, significantly saving storage and memory overhead, particularly in scenarios with a growing number of tasks. We validate the effectiveness of Variator on seven datasets. Experimental results show that Variator can save 53% computational costs using only 0.9% additional parameters with a performance drop of less than 2%. Moreover, when the model scales to billions of parameters, Variator matches the strong performance of uncompressed LLMs. Our code and checkpoints will be released to facilitate future work.

This work examines the presence of modularity in pre-trained Transformers, a feature commonly found in human brains and thought to be vital for general intelligence. In analogy to human brains, we consider two main characteristics of modularity: (1) functional specialization of neurons: we evaluate whether each neuron is mainly specialized in a certain function, and find that the answer is yes. (2) function-based neuron grouping: we explore to find a structure that groups neurons into modules by function, and each module works for its corresponding function. Given the enormous amount of possible structures, we focus on Mixture-of-Experts as a promising candidate, which partitions neurons into experts and usually activates different experts for different inputs. Experimental results show that there are functional experts, where clustered are the neurons specialized in a certain function. Moreover, perturbing the activations of functional experts significantly affects the corresponding function. Finally, we study how modularity emerges during pre-training, and find that the modular structure is stabilized at the early stage, which is faster than neuron stabilization. It suggests that Transformer first constructs the modular structure and then learns fine-grained neuron functions. Our code and data are available at https://github.com/THUNLP/modularity-analysis.

Large-scale pre-trained models (PTMs) have been widely used in document-oriented NLP tasks, such as question answering. However, the encoding-task coupling requirement results in the repeated encoding of the same documents for different tasks and queries, which is highly computationally inefficient. To this end, we target to decouple document encoding from downstream tasks, and propose to represent each document as a plug-and-play document module, i.e., a document plugin, for PTMs (PlugD). By inserting document plugins into the backbone PTM for downstream tasks, we can encode a document one time to handle multiple tasks, which is more efficient than conventional encoding-task coupling methods that simultaneously encode documents and input queries using task-specific encoders. Extensive experiments on 8 datasets of 4 typical NLP tasks show that PlugD enables models to encode documents once and for all across different scenarios. Especially, PlugD can save 69% computational costs while achieving comparable performance to state-of-the-art encoding-task coupling methods. Additionally, we show that PlugD can serve as an effective post-processing way to inject knowledge into task-specific models, improving model performance without any additional model training. Our code and checkpoints can be found in https://github.com/thunlp/Document-Plugin.

Injecting external knowledge can improve the performance of pre-trained language models (PLMs) on various downstream NLP tasks. However, massive retraining is required to deploy new knowledge injection methods or knowledge bases for downstream tasks. In this work, we are the first to study how to improve the flexibility and efficiency of knowledge injection by reusing existing downstream models. To this end, we explore a new paradigm plug-and-play knowledge injection, where knowledge bases are injected into frozen existing downstream models by a knowledge plugin. Correspondingly, we propose a plug-and-play injection method map-tuning, which trains a mapping of knowledge embeddings to enrich model inputs with mapped embeddings while keeping model parameters frozen. Experimental results on three knowledge-driven NLP tasks show that existing injection methods are not suitable for the new paradigm, while map-tuning effectively improves the performance of downstream models. Moreover, we show that a frozen downstream model can be well adapted to different domains with different mapping networks of domain knowledge. Our code and models are available at https://github.com/THUNLP/Knowledge-Plugin.

Recognizing facts is the most fundamental step in making judgments, hence detecting events in the legal documents is important to legal case analysis tasks. However, existing Legal Event Detection (LED) datasets only concern incomprehensive event types and have limited annotated data, which restricts the development of LED methods and their downstream applications. To alleviate these issues, we present LEVEN a large-scale Chinese LEgal eVENt detection dataset, with 8,116 legal documents and 150,977 human-annotated event mentions in 108 event types. Not only charge-related events, LEVEN also covers general events, which are critical for legal case understanding but neglected in existing LED datasets. To our knowledge, LEVEN is the largest LED dataset and has dozens of times the data scale of others, which shall significantly promote the training and evaluation of LED methods. The results of extensive experiments indicate that LED is challenging and needs further effort. Moreover, we simply utilize legal events as side information to promote downstream applications. The method achieves improvements of average 2.2 points precision in low-resource judgment prediction, and 1.5 points mean average precision in unsupervised case retrieval, which suggests the fundamentality of LED. The source code and dataset can be obtained from https://github.com/thunlp/LEVEN.

Distant supervision (DS) has been widely adopted to generate auto-labeled data for sentence-level relation extraction (RE) and achieved great results. However, the existing success of DS cannot be directly transferred to more challenging document-level relation extraction (DocRE), as the inevitable noise caused by DS may be even multiplied in documents and significantly harm the performance of RE. To alleviate this issue, we propose a novel pre-trained model for DocRE, which de-emphasize noisy DS data via multiple pre-training tasks. The experimental results on the large-scale DocRE benchmark show that our model can capture useful information from noisy data and achieve promising results.

Legal Artificial Intelligence (LegalAI) focuses on applying the technology of artificial intelligence, especially natural language processing, to benefit tasks in the legal domain. In recent years, LegalAI has drawn increasing attention rapidly from both AI researchers and legal professionals, as LegalAI is beneficial to the legal system for liberating legal professionals from a maze of paperwork. Legal professionals often think about how to solve tasks from rule-based and symbol-based methods, while NLP researchers concentrate more on data-driven and embedding methods. In this paper, we introduce the history, the current state, and the future directions of research in LegalAI. We illustrate the tasks from the perspectives of legal professionals and NLP researchers and show several representative applications in LegalAI. We conduct experiments and provide an in-depth analysis of the advantages and disadvantages of existing works to explore possible future directions. You can find the implementation of our work from https://github.com/thunlp/CLAIM.

Relational facts are an important component of human knowledge, which are hidden in vast amounts of text. In order to extract these facts from text, people have been working on relation extraction (RE) for years. From early pattern matching to current neural networks, existing RE methods have achieved significant progress. Yet with explosion of Web text and emergence of new relations, human knowledge is increasing drastically, and we thus require “more” from RE: a more powerful RE system that can robustly utilize more data, efficiently learn more relations, easily handle more complicated context, and flexibly generalize to more open domains. In this paper, we look back at existing RE methods, analyze key challenges we are facing nowadays, and show promising directions towards more powerful RE. We hope our view can advance this field and inspire more efforts in the community.

Legal Judgment Prediction (LJP) aims to predict the judgment result based on the facts of a case and becomes a promising application of artificial intelligence techniques in the legal field. In real-world scenarios, legal judgment usually consists of multiple subtasks, such as the decisions of applicable law articles, charges, fines, and the term of penalty. Moreover, there exist topological dependencies among these subtasks. While most existing works only focus on a specific subtask of judgment prediction and ignore the dependencies among subtasks, we formalize the dependencies among subtasks as a Directed Acyclic Graph (DAG) and propose a topological multi-task learning framework, TopJudge, which incorporates multiple subtasks and DAG dependencies into judgment prediction. We conduct experiments on several real-world large-scale datasets of criminal cases in the civil law system. Experimental results show that our model achieves consistent and significant improvements over baselines on all judgment prediction tasks. The source code can be obtained from https://github.com/thunlp/TopJudge.