Despite recent advances in safety alignment, large language models (LLMs) remain highly susceptible to adversarial attacks, while the internal mechanisms behind such vulnerabilities are still poorly understood. Existing gradient-based attribution methods offer valuable interpretability for analyzing information storage and processing in LLMs. However, they are inapplicable to adversarial attacks, which typically occur in open-ended generation settings without fixed ground-truth outputs. To address these challenges, we propose a novel similarity-based gradient attribution method to identify key neurons sensitive to adversarial behaviors in open-ended generation tasks. The detected neurons, termed targeted neurons, play a critical role in safety training. Building on this neuron-level perspective, we uncover two key neuronal patterns: (i) universal neurons that are consistently exploited across multiple attack strategies, and (ii) interference neurons that hinder safety improvements when fine-tuned indiscriminately, providing mechanistic insights into the interpretability of adversarial vulnerabilities. Inspired by these findings, we propose a neuron-level defense strategy, Targeted Neuron Tuning (TNT), which selectively fine-tunes the identified targeted neurons for specific attacks. Experimental evaluations across multiple LLM architectures and scales demonstrate that TNT substantially improves model robustness against a wide range of jailbreak attacks, achieving safe rates exceeding 90% and even approaching 100%, while preserving general task performance, enabling precise and robust safety interventions. Warning: This paper contains example data that may be harmful.

Recently, large reasoning models have achieved impressive performance, but their lengthy reasoning processes incur substantial inference overhead. To mitigate this issue, we propose the concept of reasoning vectors, representations extracted from the model’s hidden states, which can guide the model towards generating more concise and accurate responses. Building upon this, we present ERRV, a training framework that elicits efficient reasoning through reasoning vectors, which enables the model to generate high-quality responses during reinforcement learning. By performing targeted policy optimization on both accuracy and length objectives, ERRV effectively activates the model’s latent capability for efficient reasoning. Our experiments demonstrate that after training with ERRV, the model achieves approximately 30% reduction in reasoning length while maintaining stable accuracy, without guidance from the reasoning vector during inference. This establishes a trade-off between efficiency and performance. Furthermore, we identify key properties of reasoning vectors: robustness, characterized by high similarity before and after training, and generalizability, demonstrating applicability across base models, distilled models, RL-trained models, parameter-merged models, and mixed-thought models. These properties collectively guarantee the reliability and broad applicability of our approach.

Reinforcement learning with verifiable rewards (RLVR) has emerged as a prominent paradigm for enhancing the reasoning capabilities of large language models (LLMs). However, the entropy of LLMs usually collapses during RLVR training, leading to premature convergence to suboptimal local minima and hindering further performance improvement. Although various approaches have been proposed to mitigate entropy collapse, a comprehensive study of entropy in RLVR remains lacking. To bridge this gap, we conduct extensive experiments to investigate the entropy dynamics of LLMs trained with RLVR and analyze how model entropy correlates with response diversity, calibration, and performance across various benchmarks. Our results identify three key factors that influence entropy: the clipping thresholds in the optimization objective, the number of off-policy updates, and the diversity of the training data. Furthermore, through both theoretical analysis and empirical validation, we demonstrate that tokens with positive advantages are the primary drivers of entropy collapse. Motivated by this insight, we propose Positive-Advantage Reweighting, a simple yet effective approach that regulates model entropy by adjusting the loss weights assigned to tokens with positive advantages during RLVR training, while maintaining competitive performance.

Large language models (LLMs) have achieved remarkable success across a wide range of natural language processing tasks, yet their performance remains heavily biased toward high-resource languages. Tibetan, despite its cultural significance and large speaker population, is still substantially underrepresented. In this work, we present a comprehensive pipeline for advancing Tibetan language modeling through large-scale data curation and continual pre-training. We construct a 72 GB high-quality Tibetan corpus, the largest to date, and adapt Qwen2.5-7B through balanced multilingual continual pre-training with Tibetan, Chinese, and English, followed by multilingual instruction tuning. To further scale capacity efficiently, we extend the dense model to a 50B-A10B Mixture-of-Experts architecture. Due to the absence of standardized Tibetan benchmarks, we build multiple evaluation datasets via high-quality translation and human verification. Experimental results show that both dense and MoE models consistently outperform existing open-source and Tibetan-focused models of similar scale across diverse tasks. Our work advances Tibetan-centric LLM research and provides transferable insights for extending LLMs to other low-resource languages. We will release the model weights, evaluation benchmarks, and detailed data processing documentation in the follow-up.

Reinforcement learning (RL)-based post-training often improves the reasoning performance of large language models (LLMs) beyond the training domain, while supervised fine-tuning (SFT) frequently leads to general capabilities forgetting. However, the mechanisms underlying this contrast remain unclear.To bridge this gap, we present a feature-level mechanistic analysis methodology to probe RL generalization using a controlled experimental setup, where RL- and SFT-tuned models are trained from the same base model on identical data. Leveraging our interpretability framework, we align internal activations across models within a shared feature space and analyze how features evolve during post-training.We find that SFT rapidly introduces many highly specialized features that stabilize early in training, whereas RL induces more restrained and continually evolving feature changes that largely preserve base models’ representations. Focusing on samples where RL succeeds but the base model fails, we identify a compact, task-agnostic set of features that directly mediate generalization across diverse tasks. Feature-level interventions confirm their causal role: disabling these features significantly degrades RL models’ generalization performance, while amplifying them improves base models’ performance. The code is available at https://github.com/danshi777/RL-generalization.

Concept editing aims to control specific concepts in large language models (LLMs) and is an emerging subfield of model editing. Despite the emergence of various editing methods in recent years, there remains a lack of rigorous theoretical analysis and a unified perspective to systematically understand and compare these methods. To address this gap, we propose a unified paradigm for concept editing methods, in which all forms of conceptual injection are aligned at the neuron level. We study four representative concept editing methods: Neuron Editing (NE), Supervised Fine-tuning (SFT), Sparse Autoencoder (SAE), and Steering Vector (SV). Then we categorize them into two classes based on their mode of conceptual information injection: indirect (NE, SFT) and direct (SAE, SV). We evaluate above methods along four dimensions: editing reliability, output generalization, neuron level consistency, and mathematical formalization. Experiments show that SAE achieves the best editing reliability. In output generalization, SAE captures features closer to human-understood concepts, while NE tends to locate text patterns rather than true semantics. Neuron-level analysis reveals that direct methods share high neuron overlap, as do indirect methods, indicating methodological commonality within each category. Our unified paradigm offers a clear framework and valuable insights for advancing interpretability and controlled generation in LLMs.

Previous benchmarks for evaluating large language models (LLMs) have primarily emphasized quantitative metrics, such as data volume. However, this focus may neglect key qualitative data attributes that can significantly impact the final rankings of LLMs, resulting in unreliable leaderboards. In this paper, we investigate whether current LLM benchmarks adequately consider these data attributes. We specifically examine three attributes: diversity, redundancy, and difficulty. To explore these attributes, we propose a framework with three separate modules, each designed to assess one of the attributes. Using a method that progressively incorporates these attributes, we analyze their influence on the benchmark. Our experimental results reveal a meaningful correlation between LLM rankings on the revised benchmark and the original benchmark when these attributes are accounted for. These findings indicate that existing benchmarks often fail to meet all three criteria, highlighting a lack of consideration for multifaceted data attributes in current evaluation datasets.

Ensuring large language models (LLM) behave consistently with human goals, values, and intentions is crucial for their safety but yet computationally expensive. To reduce the computational cost of alignment training of LLMs, especially for those with a huge number of parameters, and to reutilize learned value alignment, we propose ConTrans, a novel framework that enables weak-to-strong alignment transfer via concept transplantation. From the perspective of representation engineering, ConTrans refines concept vectors in value alignment from a source LLM (usually a weak yet aligned LLM). The refined concept vectors are then reformulated to adapt to the target LLM (usually a strong yet unaligned base LLM) via affine transformation. In the third step, ConTrans transplants the reformulated concept vectors into the residual stream of the target LLM. Experiments demonstrate the successful transplantation of a wide range of aligned concepts from 7B models to 13B and 70B models across multiple LLMs and LLM families. Remarkably, ConTrans even surpasses instruction-tuned models in terms of truthfulness. Experiment results validate the effectiveness of both inter-LLM-family and intra-LLM-family concept transplantation. Our work successfully demonstrates an alternative way to achieve weak-to-strong alignment generalization and control.

Large Language Models (LLMs) have demonstrated remarkable abilities in text comprehension and logical reasoning, indicating that the text representations learned by LLMs can facilitate their language processing capabilities. In neuroscience, brain cognitive processing signals are typically utilized to study human language processing. Therefore, it is natural to ask how well the text embeddings from LLMs align with the brain cognitive processing signals, and how training strategies affect the LLM-brain alignment? In this paper, we employ Representational Similarity Analysis (RSA) to measure the alignment between 23 mainstream LLMs and fMRI signals of the brain to evaluate how effectively LLMs simulate cognitive language processing. We empirically investigate the impact of various factors (e.g., pre-training data size, model scaling, alignment training, and prompts) on such LLM-brain alignment. Experimental results indicate that pre-training data size and model scaling are positively correlated with LLM-brain similarity, and alignment training can significantly improve LLM-brain similarity. Explicit prompts contribute to the consistency of LLMs with brain cognitive language processing, while nonsensical noisy prompts may attenuate such alignment. Additionally, the performance of a wide range of LLM evaluations (e.g., MMLU, Chatbot Arena) is highly correlated with the LLM-brain similarity.

With the increasing capability of large language models (LLMs), LLM-as-a-judge has emerged as a new evaluation paradigm. Compared with traditional automatic and manual evaluation, LLM evaluators exhibit better interpretability and efficiency. Despite this, existing LLM evaluators suffer from limited use scenarios and poor flexibility. To mitigate these issues, we propose Praetor, a fine-grained generative LLM evaluator with instance-level customazable evaluation criteria. To train Praetor, we curate a large-scale dataset guided with a hierarchical guideline covering a wide range of tasks and instance-level evaluation criteria. We train Praetor on this dataset in a multi-task learning fashion, which enables to evaluate LLMs in either pointwise grading or pairwise comparison way and support two languages simultaneously with a high flexibility of setting evaluation criteria. Extensive experiments demonstrate that Praetor outperforms previous LLM evaluators and instruction-tuned LLMs on multiple benchmarks, setting new SOTA results. It also exhibits the potential for generating critiques as scalable feedback to further improve LLMs. Our model and related resources are released at https://github.com/tjunlp-lab/Praetor.

Chinese Large Language Models (LLMs) have recently demonstrated impressive capabilities across various NLP benchmarks and real-world applications. However, the existing benchmarks for comprehensively evaluating these LLMs are still insufficient, particularly in terms of measuring knowledge that LLMs capture. Current datasets collect questions from Chinese examinations across different subjects and educational levels to address this issue. Yet, these benchmarks primarily focus on objective questions such as multiple-choice questions, leading to a lack of diversity in question types. To tackle this problem, we propose LHMKE, a Large-scale, Holistic, and Multi-subject Knowledge Evaluation benchmark in this paper. LHMKE is designed to provide a comprehensive evaluation of the knowledge acquisition capabilities of Chinese LLMs. It encompasses 10,465 questions across 75 tasks covering 30 subjects, ranging from primary school to professional certification exams. Notably, LHMKE includes both objective and subjective questions, offering a more holistic evaluation of the knowledge level of LLMs. We have assessed 11 Chinese LLMs under the zero-shot setting, which aligns with real examinations, and compared their performance across different subjects. We also conduct an in-depth analysis to check whether GPT-4 can automatically score subjective predictions. Our findings suggest that LHMKE is a challenging and advanced testbed for Chinese LLMs.

Increasing the number of parameters in large language models (LLMs) usually improves performance in downstream tasks but raises compute and memory costs, making deployment difficult in resource-limited settings. Quantization techniques, which reduce the bits needed for model weights or activations with minimal performance loss, have become popular due to the rise of LLMs. However, most quantization studies use pre-trained LLMs, and the impact of quantization on instruction-tuned LLMs and the relationship between perplexity and benchmark performance of quantized LLMs are not well understood. Evaluation of quantized LLMs is often limited to language modeling and a few classification tasks, leaving their performance on other benchmarks unclear. To address these gaps, we propose a structured evaluation framework consisting of three critical dimensions: (1) knowledge & capacity, (2) alignment, and (3) efficiency, and conduct extensive experiments across ten diverse benchmarks. Our experimental results indicate that LLMs with 4-bit quantization can retain performance comparable to their non-quantized counterparts, and perplexity can serve as a proxy metric for quantized LLMs on most benchmarks. Furthermore, quantized LLMs with larger parameter scales can outperform smaller LLMs. Despite the memory savings achieved through quantization, it can also slow down the inference speed of LLMs. Consequently, substantial engineering efforts and hardware support are imperative to achieve a balanced optimization of decoding speed and memory consumption in the context of quantized LLMs.

Large language models (LLMs) have demonstrated prowess in a wide range of tasks. However, many LLMs exhibit significant performance discrepancies between high- and low-resource languages. To mitigate this challenge, we present FuxiTranyu, an open-source multilingual LLM, which is designed to satisfy the need of the research community for balanced and high-performing multilingual capabilities. The base model, FuxiTranyu-8B, features 8 billion parameters and is trained from scratch on meticulously balanced multilingual data that contains 600 billion tokens covering 43 natural languages and 16 programming languages. We also develop two instruction-tuned models: FuxiTranyu-8B-SFT which is fine-tuned on a diverse multilingual instruction dataset, and FuxiTranyu-8B-DPO which is further refined with DPO on a preference dataset for enhanced alignment ability. Extensive experiments on a wide range of multilingual benchmarks demonstrate the competitive performance of FuxiTranyu against existing multilingual LLMs, e.g., BLOOM-7B, PolyLM-13B, and Mistral-7B-Instruct. Both neuron and representation interpretability analyses reveal that FuxiTranyu achieves consistent multilingual representations across languages. To promote further research into multilingual LLMs, we release both the base and instruction-tuned FuxiTranyu models together with 58 pre-training checkpoints at HuggingFace and Github.

Chinese LLMs demonstrate impressive performance on NLP tasks, particularly on discipline knowledge benchmarks, with some results approaching those of GPT-4. Previous research has viewed these advancements as potential outcomes of data contamination or leakage, prompting efforts to create new detection methods and address evaluation issues in LLM benchmarks. However, there has been a lack of comprehensive assessment of the evolution of Chinese LLMs. To address this gap, this paper offers a thorough investigation of Chinese LLMs on discipline knowledge evaluation, delving into the advancements of various LLMs, including a group of related models and others. Specifically, we have conducted six assessments ranging from knowledge memorization to comprehension for robustness, encompassing tasks like predicting incomplete questions and options, identifying behaviors by the contaminational fine-tuning, and answering rephrased questions. Experimental findings indicate a positive correlation between the release time of LLMs and their memorization capabilities, but they struggle with variations in original question-options pairs. Additionally, our findings suggest that question descriptions have a more significant impact on LLMs’ performance.

The rapid development of Chinese large language models (LLMs) poses big challenges for efficient LLM evaluation. While current initiatives have introduced new benchmarks or evaluation platforms for assessing Chinese LLMs, many of these focus primarily on capabilities, usually overlooking potential alignment and safety issues. To address this gap, we introduce OpenEval, an evaluation testbed that benchmarks Chinese LLMs across capability, alignment and safety. For capability assessment, we include 12 benchmark datasets to evaluate Chinese LLMs from 4 sub-dimensions: NLP tasks, disciplinary knowledge, commonsense reasoning and mathematical reasoning. For alignment assessment, OpenEval contains 7 datasets that examines the bias, offensiveness and illegalness in the outputs yielded by Chinese LLMs. To evaluate safety, especially anticipated risks (e.g., power-seeking, self-awareness) of advanced LLMs, we include 6 datasets. In addition to these benchmarks, we have implemented a phased public evaluation and benchmark update strategy to ensure that OpenEval is in line with the development of Chinese LLMs or even able to provide cutting-edge benchmark datasets to guide the development of Chinese LLMs. In our first public evaluation, we have tested a range of Chinese LLMs, spanning from 7B to 72B parameters, including both open-source and proprietary models. Evaluation results indicate that while Chinese LLMs have shown impressive performance in certain tasks, more attention should be directed towards broader aspects such as commonsense reasoning, alignment, and safety.

Spoken texts (either manual or automatic transcriptions from automatic speech recognition (ASR)) often contain disfluencies and grammatical errors, which pose tremendous challenges to downstream tasks. Converting spoken into written language is hence desirable. Unfortunately, the availability of datasets for this is limited. To address this issue, we present CS2W, a Chinese Spoken-to-Written style conversion dataset comprising 7,237 spoken sentences extracted from transcribed conversational texts. Four types of conversion problems are covered in CS2W: disfluencies, grammatical errors, ASR transcription errors, and colloquial words. Our annotation convention, data, and code are publicly available at https://github.com/guozishan/CS2W.

In a multilingual neural machine translation model that fully shares parameters across all languages, an artificial language token is usually used to guide translation into the desired target language. However, recent studies show that prepending language tokens sometimes fails to navigate the multilingual neural machine translation models into right translation directions, especially on zero-shot translation. To mitigate this issue, we propose two methods, language embedding embodiment and language-aware multi-head attention, to learn informative language representations to channel translation into right directions. The former embodies language embeddings into different critical switching points along the information flow from the source to the target, aiming at amplifying translation direction guiding signals. The latter exploits a matrix, instead of a vector, to represent a language in the continuous space. The matrix is chunked into multiple heads so as to learn language representations in multiple subspaces. Experiment results on two datasets for massively multilingual neural machine translation demonstrate that language-aware multi-head attention benefits both supervised and zero-shot translation and significantly alleviates the off-target translation issue. Further linguistic typology prediction experiments show that matrix-based language representations learned by our methods are capable of capturing rich linguistic typology features.