Linqi Song

Ke Wang, Houxing Ren, Aojun Zhou et al.

The recently released GPT-4 Code Interpreter has demonstrated remarkable proficiency in solving challenging math problems, primarily attributed to its ability to seamlessly reason with natural language, generate code, execute code, and continue reasoning based on the execution output. In this paper, we present a method to fine-tune open-source language models, enabling them to use code for modeling and deriving math equations and, consequently, enhancing their mathematical reasoning abilities. We propose a method of generating novel and high-quality datasets with math problems and their code-based solutions, referred to as MathCodeInstruct. Each solution interleaves natural language, code, and execution results. We also introduce a customized supervised fine-tuning and inference approach. This approach yields the MathCoder models, a family of models capable of generating code-based solutions for solving challenging math problems. Impressively, the MathCoder models achieve state-of-the-art scores among open-source LLMs on the MATH (45.2%) and GSM8K (83.9%) datasets, substantially outperforming other open-source alternatives. Notably, the MathCoder model not only surpasses ChatGPT-3.5 and PaLM-2 on GSM8K and MATH but also outperforms GPT-4 on the competition-level MATH dataset. The dataset and models will be released at https://github.com/mathllm/MathCoder.

Hanlin Zhang, Daxin Tan, Dehua Tao et al.

Instruction-guided speech editing requires a model to modify specified speech attributes while preserving unrelated characteristics. Despite rapid progress in Speech Large Language Models (Speech LLMs), systematic evaluation of this capability remains challenging, as existing benchmarks are fragmented across isolated editing tasks. To bridge this gap, we introduce \textbf{SpeechEditBench}, a bilingual multi-attribute benchmark for instruction-guided speech editing. SpeechEditBench encompasses seven atomic editing tasks, as well as compositional editing tasks that integrate multiple operations within a single instruction. We propose an anchor-based evaluation protocol that separately assesses the edit success of target attributes and the preservation of untargeted attributes, leading to three metrics: target success, preservation success, and joint success. Using this benchmark, we evaluate mainstream Speech LLMs and specialized speech editing systems. The results reveal three key findings: (1) no single model performs well across all editing dimensions; (2) closed-source Speech LLMs generally outperform open-source models; (3) compositional editing remains highly challenging, with even the most advanced models struggling to achieve high joint success. SpeechEditBench provides a rigorous diagnostic framework to identify bottlenecks in Speech LLMs, thereby facilitating the development of next-generation Speech LLMs with more robust and precise instruction-guided editing capabilities. Data and code will be released upon acceptance.

Baoliang Chen, Lingyu Zhu, Hanwei Zhu et al.

There is a growing consensus in the research community that the optimization of low-light image enhancement approaches should be guided by the visual quality perceived by end users. Despite the substantial efforts invested in the design of low-light enhancement algorithms, there has been comparatively limited focus on assessing subjective and objective quality systematically. To mitigate this gap and provide a clear path towards optimizing low-light image enhancement for better visual quality, we propose a gap-closing framework. In particular, our gap-closing framework starts with the creation of a large-scale dataset for Subjective QUality Assessment of REconstructed LOw-Light Images (SQUARE-LOL). This database serves as the foundation for studying the quality of enhanced images and conducting a comprehensive subjective user study. Subsequently, we propose an objective quality assessment measure that plays a critical role in bridging the gap between visual quality and enhancement. Finally, we demonstrate that our proposed objective quality measure can be incorporated into the process of optimizing the learning of the enhancement model toward perceptual optimality. We validate the effectiveness of our proposed framework through both the accuracy of quality prediction and the perceptual quality of image enhancement. Our database and codes are publicly available at https://github.com/Baoliang93/IACA_For_Lowlight_IQA.

Zhicheng Yang, Yiwei Wang, Yinya Huang et al.

Large language models (LLMs) have exhibited their problem-solving abilities in mathematical reasoning. Solving realistic optimization (OPT) problems in application scenarios requires advanced and applied mathematics ability. However, current OPT benchmarks that merely solve linear programming are far from complex realistic situations. In this work, we propose OptiBench, a benchmark for End-to-end optimization problem-solving with human-readable inputs and outputs. OptiBench contains rich optimization problems, including linear and nonlinear programming with or without tabular data, which can comprehensively evaluate LLMs' solving ability. In our benchmark, LLMs are required to call a code solver to provide precise numerical answers. Furthermore, to alleviate the data scarcity for optimization problems, and to bridge the gap between open-source LLMs on a small scale (e.g., Llama-3-8b) and closed-source LLMs (e.g., GPT-4), we further propose a data synthesis method namely ReSocratic. Unlike general data synthesis methods that proceed from questions to answers, \ReSocratic first incrementally synthesizes formatted optimization demonstration with mathematical formulations step by step and then back-translates the generated demonstrations into questions. Based on this, we synthesize the ReSocratic-29k dataset. We further conduct supervised fine-tuning with ReSocratic-29k on multiple open-source models. Experimental results show that ReSocratic-29k significantly improves the performance of open-source models.

Shing Yin Wong, Shaocheng Liu, Linqi Song et al.

Proving Shannon-type entropy inequalities is a fundamental task in information theory that often requires constructing non-trivial linear combinations of known constraints, which is a combinatorial search problem that scales poorly with the number of random variables. We investigate whether small-scale large language models (0.6B--1.7B parameters), fine-tuned on atomic proof steps and combined with guided beam search, can automate this process. On a held-out test set of 60 inequalities spanning n=10 to 15 variables, our 0.6B fine-tuned model achieves an 85\% proof success rate with tree search. GPT-5.5 solves 1.7\% samples under zero-shot prompting while Psitip solves 33.3\% samples. A systematic ablation study across training context length (4096 vs.\ 8192 tokens) and data distribution (n=9-skewed vs not skewed) reveals that a 4096-token not skewed training distribution yields the best performance, with extended context and skewed data providing no marginal benefit. We further identify two dominant failure modes -- format failures and step quality degradation -- and verify that the beam-scoring heuristic is essential via a controlled ablation (random scoring reduces success from 83\% to 23\%).

Xiaoming Xue, Cuie Yang, Liang Feng et al.

Sequential transfer optimization (STO), which aims to improve the optimization performance on a task of interest by exploiting the knowledge captured from several previously-solved optimization tasks stored in a database, has been gaining increasing research attention over the years. However, despite the remarkable advances in algorithm design, the development of a systematic benchmark suite for comprehensive comparisons of STO algorithms received far less attention. Existing test problems are either simply generated by assembling other benchmark functions or extended from specific practical problems with limited scalability. The relationships between the optimal solutions of the source and target tasks in these problems are also often manually configured, limiting their ability to model different similarity relationships presented in real-world problems. Consequently, the good performance achieved by an algorithm on these problems might be biased and hard to be generalized to other problems. In light of the above, in this study, we first introduce four concepts for characterizing STO problems and present an important problem feature, namely similarity distribution, which quantitatively delineates the relationship between the optima of the source and target tasks. Then, we present the general design guidelines of STO problems and a particular STO problem generator with good scalability. Specifically, the similarity distribution of a problem can be easily customized, enabling a continuous spectrum of representation of the diverse similarity relationships of real-world problems. Lastly, a benchmark suite with 12 STO problems featured by a variety of customized similarity relationships is developed using the proposed generator. The source code of the problem generator is available at https://github.com/XmingHsueh/STOP-G.

Hanlin Zhang, Daxin Tan, Dehua Tao et al.

Speech tokenizers are a key building block of fully discrete Speech LLMs. Existing tokenizers either prioritize semantic encoding, fuse semantic content with acoustic style inseparably, or achieve incomplete semantic-acoustic disentanglement. To achieve better disentanglement, we propose \textbf{DSA-Tokenizer}, which explicitly disentangles speech into discrete semantic and acoustic tokens via distinct optimization constraints. Specifically, semantic tokens are supervised by ASR to capture linguistic content, while acoustic tokens focus on mel-spectrograms restoration to encode style. We further introduce a hierarchical Flow Matching decoder and a joint reconstruction-context inpainting training strategy, allowing the model to support both high-fidelity reconstruction and cross-utterance voice clone. To speed up inference, we distill the DiT decoder to reduce sampling steps of inference to 4 and improve synthesis quality with GAN fine-tuning. Experiments demonstrate that DSA-Tokenizer provides strong semantic-acoustic disentanglement, reliable controllable voice cloning, and efficient high-fidelity generation with low WER/CER. Moreover, our results suggest that disentangled tokenization provides a more effective interface for downstream large-model speech generation. Audio samples are avaialble at https://anonymous.4open.science/w/DSA_Tokenizer_demo/.

Yinya Huang, Ruixin Hong, Hongming Zhang et al.

In this study, we delve into the realm of counterfactual reasoning capabilities of large language models (LLMs). Our primary objective is to cultivate the counterfactual thought processes within LLMs and rigorously assess these processes for their validity. Specifically, we introduce a novel task, Counterfactual Logical Modification (CLOMO), and a high-quality human-annotated benchmark. In this task, LLMs must adeptly alter a given argumentative text to uphold a predetermined logical relationship. To effectively evaluate a generation model's counterfactual capabilities, we propose an innovative evaluation metric, the decomposed Self-Evaluation Score (SES) to directly evaluate the natural language output of LLMs instead of modeling the task as a multiple-choice problem. Analysis shows that the proposed automatic metric aligns well with human preference. Our experimental results show that while LLMs demonstrate a notable capacity for logical counterfactual thinking, there remains a discernible gap between their current abilities and human performance. Code and data are available at https://github.com/Eleanor-H/CLOMO.

Mingyang Liu, Qingcan Kang, Yuke Wang et al.

Incorporating news into time series forecasting is appealing because news can reveal abrupt exogenous events that historical values alone cannot recover. However, existing LLM-based news-forecasting pipelines face two practical limitations: relevant news articles often exceed the model's context window, and iterative retrieval of supplementary news is typically unguided, leading to redundant updates and slow convergence. We address these issues with a novel framework that combines importance-aware news compression and process-level retrieval supervision. First, we train an importance reward model that estimates the forecasting utility of each article and uses this signal to allocate compression budgets during sequential pairwise fusion, preserving informative content within a fixed context limit. Second, we introduce a process reward model (PRM) that ranks multiple supplementary-news candidates conditioned on the current error profile and the history of previously selected articles, replacing one-shot blind retrieval with quality-controlled selection. Both components are trained offline using historical data with ground truth; inference uses the frozen filtering logic and compression modules without any reflection loop. Experiments on finance, energy, traffic, and bitcoin forecasting benchmarks show that our method improves prediction accuracy over strong baselines, significantly reduces the number of refinement iterations compared to the iterative baseline, and remains effective when relevant articles span thousands of tokens.

Haoyu Chen, Linqi Song, Zhenxing Qian et al.

Data hiding with deep neural networks (DNNs) has experienced impressive successes in recent years. A prevailing scheme is to train an autoencoder, consisting of an encoding network to embed (or transform) secret messages in (or into) a carrier, and a decoding network to extract the hidden messages. This scheme may suffer from several limitations regarding practicability, security, and embedding capacity. In this work, we describe a different computational framework to hide images in deep probabilistic models. Specifically, we use a DNN to model the probability density of cover images, and hide a secret image in one particular location of the learned distribution. As an instantiation, we adopt a SinGAN, a pyramid of generative adversarial networks (GANs), to learn the patch distribution of one cover image. We hide the secret image by fitting a deterministic mapping from a fixed set of noise maps (generated by an embedding key) to the secret image during patch distribution learning. The stego SinGAN, behaving as the original SinGAN, is publicly communicated; only the receiver with the embedding key is able to extract the secret image. We demonstrate the feasibility of our SinGAN approach in terms of extraction accuracy and model security. Moreover, we show the flexibility of the proposed method in terms of hiding multiple images for different receivers and obfuscating the secret image.

Aojun Zhou, Ke Wang, Zimu Lu et al.

Recent progress in large language models (LLMs) like GPT-4 and PaLM-2 has brought significant advancements in addressing math reasoning problems. In particular, OpenAI's latest version of GPT-4, known as GPT-4 Code Interpreter, shows remarkable performance on challenging math datasets. In this paper, we explore the effect of code on enhancing LLMs' reasoning capability by introducing different constraints on the \textit{Code Usage Frequency} of GPT-4 Code Interpreter. We found that its success can be largely attributed to its powerful skills in generating and executing code, evaluating the output of code execution, and rectifying its solution when receiving unreasonable outputs. Based on this insight, we propose a novel and effective prompting method, explicit \uline{c}ode-based \uline{s}elf-\uline{v}erification~(CSV), to further boost the mathematical reasoning potential of GPT-4 Code Interpreter. This method employs a zero-shot prompt on GPT-4 Code Interpreter to encourage it to use code to self-verify its answers. In instances where the verification state registers as ``False'', the model shall automatically amend its solution, analogous to our approach of rectifying errors during a mathematics examination. Furthermore, we recognize that the states of the verification result indicate the confidence of a solution, which can improve the effectiveness of majority voting. With GPT-4 Code Interpreter and CSV, we achieve an impressive zero-shot accuracy on MATH dataset \textbf{(53.9\% $\to$ 84.3\%)}.

Han Wu, Haochen Tan, Kun Xu et al. · tencent-ai

While conversational semantic role labeling (CSRL) has shown its usefulness on Chinese conversational tasks, it is still under-explored in non-Chinese languages due to the lack of multilingual CSRL annotations for the parser training. To avoid expensive data collection and error-propagation of translation-based methods, we present a simple but effective approach to perform zero-shot cross-lingual CSRL. Our model implicitly learns language-agnostic, conversational structure-aware and semantically rich representations with the hierarchical encoders and elaborately designed pre-training objectives. Experimental results show that our model outperforms all baselines by large margins on two newly collected English CSRL test sets. More importantly, we confirm the usefulness of CSRL to non-Chinese conversational tasks such as the question-in-context rewriting task in English and the multi-turn dialogue response generation tasks in English, German and Japanese by incorporating the CSRL information into the downstream conversation-based models. We believe this finding is significant and will facilitate the research of non-Chinese dialogue tasks which suffer the problems of ellipsis and anaphora.

Haochen Tan, Wei Shao, Han Wu et al.

Contrastive learning has shown great potential in unsupervised sentence embedding tasks, e.g., SimCSE. However, We find that these existing solutions are heavily affected by superficial features like the length of sentences or syntactic structures. In this paper, we propose a semantics-aware contrastive learning framework for sentence embeddings, termed Pseudo-Token BERT (PT-BERT), which is able to exploit the pseudo-token space (i.e., latent semantic space) representation of a sentence while eliminating the impact of superficial features such as sentence length and syntax. Specifically, we introduce an additional pseudo token embedding layer independent of the BERT encoder to map each sentence into a sequence of pseudo tokens in a fixed length. Leveraging these pseudo sequences, we are able to construct same-length positive and negative pairs based on the attention mechanism to perform contrastive learning. In addition, we utilize both the gradient-updating and momentum-updating encoders to encode instances while dynamically maintaining an additional queue to store the representation of sentence embeddings, enhancing the encoder's learning performance for negative examples. Experiments show that our model outperforms the state-of-the-art baselines on six standard semantic textual similarity (STS) tasks. Furthermore, experiments on alignments and uniformity losses, as well as hard examples with different sentence lengths and syntax, consistently verify the effectiveness of our method.

Guangfeng Yan, Tan Li, Kui Wu et al.

Communication efficiency and privacy protection are two critical issues in distributed machine learning. Existing methods tackle these two issues separately and may have a high implementation complexity that constrains their application in a resource-limited environment. We propose a comprehensive quantization-based solution that could simultaneously achieve communication efficiency and privacy protection, providing new insights into the correlated nature of communication and privacy. Specifically, we demonstrate the effectiveness of our proposed solutions in the distributed stochastic gradient descent (SGD) framework by adding binomial noise to the uniformly quantized gradients to reach the desired differential privacy level but with a minor sacrifice in communication efficiency. We theoretically capture the new trade-offs between communication, privacy, and learning performance.

Guanzhi Deng, Kuan Wu, Haibo Wang et al.

Mixture-of-Experts (MoE) models have emerged as a dominant paradigm for efficient LLM scaling, yet adapting them to non-English downstream tasks remains challenging. Existing fine-tuning approaches treat MoE models as monolithic learners, ignoring the heterogeneous routing structure that develops during pretraining. We validate across multiple MoE models and downstream tasks that middle layers form a language-universal alignment zone where routing divergence strongly predicts per-language task performance gaps. Building on this observation, we propose RA-MoE (Routing-Aligned MoE Fine-Tuning), a three-stage framework that categorizes parallel task examples into a four-way taxonomy (cc/ci/ic/ii) based on correctness in English and the target language, identifies task-relevant experts in the middle layers, and augments standard SFT with a routing alignment loss that encourages target-language routing on ci-type examples to follow the English task-expert activation pattern. Experiments across three MoE models, three tasks, and six target languages demonstrate that RA-MoE consistently outperforms standard SFT and strong baselines including Routing Steering and RISE, with the ci proportion of a task-language pair serving as a reliable predictor of alignment benefit.

Linquan Wu, Tianxiang Jiang, Yifei Dong et al.

Current multimodal latent reasoning often relies on external supervision (e.g., auxiliary images), ignoring intrinsic visual attention dynamics. In this work, we identify a critical Perception Gap in distillation: student models frequently mimic a teacher's textual output while attending to fundamentally divergent visual regions, effectively relying on language priors rather than grounded perception. To bridge this, we propose LaViT, a framework that aligns latent visual thoughts rather than static embeddings. LaViT compels the student to autoregressively reconstruct the teacher's visual semantics and attention trajectories prior to text generation, employing a curriculum sensory gating mechanism to prevent shortcut learning. Extensive experiments show that LaViT significantly enhances visual grounding, achieving up to +16.9% gains on complex reasoning tasks and enabling a compact 3B model to outperform larger open-source variants and proprietary models like GPT-4o.

Han Wu, Haochen Tan, Mingjie Zhan et al.

Existing dialogue modeling methods have achieved promising performance on various dialogue tasks with the aid of Transformer and the large-scale pre-trained language models. However, some recent studies revealed that the context representations produced by these methods suffer the problem of anisotropy. In this paper, we find that the generated representations are also not conversational, losing the conversation structure information during the context modeling stage. To this end, we identify two properties in dialogue modeling, i.e., locality and isotropy, and present a simple method for dialogue representation calibration, namely SimDRC, to build isotropic and conversational feature spaces. Experimental results show that our approach significantly outperforms the current state-of-the-art models on three dialogue tasks across the automatic and human evaluation metrics. More in-depth analyses further confirm the effectiveness of our proposed approach.

Mengzhe Ruan, Guangfeng Yan, Yuanzhang Xiao et al.

Distributed stochastic gradient descent (SGD) with gradient compression has become a popular communication-efficient solution for accelerating distributed learning. One commonly used method for gradient compression is Top-K sparsification, which sparsifies the gradients by a fixed degree during model training. However, there has been a lack of an adaptive approach to adjust the sparsification degree to maximize the potential of the model's performance or training speed. This paper proposes a novel adaptive Top-K in SGD framework that enables an adaptive degree of sparsification for each gradient descent step to optimize the convergence performance by balancing the trade-off between communication cost and convergence error. Firstly, an upper bound of convergence error is derived for the adaptive sparsification scheme and the loss function. Secondly, an algorithm is designed to minimize the convergence error under the communication cost constraints. Finally, numerical results on the MNIST and CIFAR-10 datasets demonstrate that the proposed adaptive Top-K algorithm in SGD achieves a significantly better convergence rate compared to state-of-the-art methods, even after considering error compensation.

Yuxuan Yao, Haonan Sheng, Qingsong Lv et al.

The escalating scale of Large Language Models (LLMs) necessitates efficient adaptation techniques. Model merging has gained prominence for its efficiency and controllability. However, existing merging techniques typically serve as post-hoc refinements or focus on mitigating task interference, often failing to capture the dynamic optimization benefits of supervised fine-tuning (SFT). In this work, we propose Streaming Merging, an innovative model updating paradigm that conceptualizes merging as an iterative optimization process. Central to this paradigm is \textbf{ARM} (\textbf{A}ctivation-guided \textbf{R}otation-aware \textbf{M}erging), a strategy designed to approximate gradient descent dynamics. By treating merging coefficients as learning rates and deriving rotation vectors from activation subspaces, ARM effectively steers parameter updates along data-driven trajectories. Unlike conventional linear interpolation, ARM aligns semantic subspaces to preserve the geometric structure of high-dimensional parameter evolution. Remarkably, ARM requires only early SFT checkpoints and, through iterative merging, surpasses the fully converged SFT model. Experimental results across model scales (1.7B to 14B) and diverse domains (e.g., math, code) demonstrate that ARM can transcend converged checkpoints. Extensive experiments show that ARM provides a scalable and lightweight framework for efficient model adaptation.

Yinya Huang, Xiaohan Lin, Zhengying Liu et al.

Recent large language models (LLMs) have witnessed significant advancement in various tasks, including mathematical reasoning and theorem proving. As these two tasks require strict and formal multi-step inference, they are appealing domains for exploring the reasoning ability of LLMs but still face important challenges. Previous studies such as Chain-of-Thought (CoT) have revealed the effectiveness of intermediate steps guidance. However, such step-wise annotation requires heavy labor, leading to insufficient training steps for current benchmarks. To fill this gap, this work introduces MUSTARD, a data generation framework that masters uniform synthesis of theorem and proof data of high quality and diversity. MUSTARD synthesizes data in three stages: (1) It samples a few mathematical concept seeds as the problem category. (2) Then, it prompts a generative language model with the sampled concepts to obtain both the problems and their step-wise formal solutions. (3) Lastly, the framework utilizes a proof assistant (e.g., Lean Prover) to filter the valid proofs. With the proposed MUSTARD, we present a theorem-and-proof benchmark MUSTARDSAUCE with 5,866 valid data points. Each data point contains an informal statement, an informal proof, and a translated formal proof that passes the prover validation. We perform extensive analysis and demonstrate that MUSTARD generates validated high-quality step-by-step data. We further apply the MUSTARDSAUCE for fine-tuning smaller language models. The fine-tuned Llama 2-7B achieves a 15.41% average relative performance gain in automated theorem proving, and 8.18% in math word problems. Codes and data are available at https://github.com/Eleanor-H/MUSTARD.

Xingyu Fan, Wei Shao, Jiacheng Liu et al.

Reward models (RMs) are central to aligning large language models (LLMs) with human preferences, powering RLHF and advanced decoding strategies. While most prior work focuses on single-step generation, real-world applications increasingly adopt multi-stage LLM pipelines, where effective reward guidance remains underexplored. We investigate this through code generation for combinatorial optimization, constructing a pipeline that integrates reward models into both formulation and solution stages. We identify a critical challenge: inconsistency between reward model predictions and actual pipeline execution outcomes. To address this, we propose the Pipeline-Adapted Reward Model (PARM), which leverages pipeline-specific data and direct preference optimization to align rewards with downstream feedback. We instantiate PARM as a two-stage pipeline (formulation -> code generation) and evaluate it on four public optimization benchmarks, measuring execution rate and solving accuracy against baselines and sampling methods. A supplementary cross-domain experiment on GSM8K assesses transferability. Results demonstrate that PARM consistently improves pipeline output quality and stability, providing new insights into reward modeling for multi-stage LLM reasoning.

Zekun Qian, Ruize Han, Wei Feng et al.

We study a novel yet practical problem of open-corpus multi-object tracking (OCMOT), which extends the MOT into localizing, associating, and recognizing generic-category objects of both seen (base) and unseen (novel) classes, but without the category text list as prompt. To study this problem, the top priority is to build a benchmark. In this work, we build OCTrackB, a large-scale and comprehensive benchmark, to provide a standard evaluation platform for the OCMOT problem. Compared to previous datasets, OCTrackB has more abundant and balanced base/novel classes and the corresponding samples for evaluation with less bias. We also propose a new multi-granularity recognition metric to better evaluate the generative object recognition in OCMOT. By conducting the extensive benchmark evaluation, we report and analyze the results of various state-of-the-art methods, which demonstrate the rationale of OCMOT, as well as the usefulness and advantages of OCTrackB.

Wei Shao, Lei Huang, Shuqi Liu et al.

Embedding-based neural topic models could explicitly represent words and topics by embedding them to a homogeneous feature space, which shows higher interpretability. However, there are no explicit constraints for the training of embeddings, leading to a larger optimization space. Also, a clear description of the changes in embeddings and the impact on model performance is still lacking. In this paper, we propose an embedding regularized neural topic model, which applies the specially designed training constraints on word embedding and topic embedding to reduce the optimization space of parameters. To reveal the changes and roles of embeddings, we introduce \textbf{uniformity} into the embedding-based neural topic model as the evaluation metric of embedding space. On this basis, we describe how embeddings tend to change during training via the changes in the uniformity of embeddings. Furthermore, we demonstrate the impact of changes in embeddings in embedding-based neural topic models through ablation studies. The results of experiments on two mainstream datasets indicate that our model significantly outperforms baseline models in terms of the harmony between topic quality and document modeling. This work is the first attempt to exploit uniformity to explore changes in embeddings of embedding-based neural topic models and their impact on model performance to the best of our knowledge.

Yuxuan Yao, Sichun Luo, Haohan Zhao et al.

We present CNER-UAV, a fine-grained \textbf{C}hinese \textbf{N}ame \textbf{E}ntity \textbf{R}ecognition dataset specifically designed for the task of address resolution in \textbf{U}nmanned \textbf{A}erial \textbf{V}ehicle delivery systems. The dataset encompasses a diverse range of five categories, enabling comprehensive training and evaluation of NER models. To construct this dataset, we sourced the data from a real-world UAV delivery system and conducted a rigorous data cleaning and desensitization process to ensure privacy and data integrity. The resulting dataset, consisting of around 12,000 annotated samples, underwent human experts and \textbf{L}arge \textbf{L}anguage \textbf{M}odel annotation. We evaluated classical NER models on our dataset and provided in-depth analysis. The dataset and models are publicly available at \url{https://github.com/zhhvvv/CNER-UAV}.

Gangwei Jiang, Caigao Jiang, Zhaoyi Li et al.

Catastrophic forgetting (CF) poses a significant challenge in machine learning, where a model forgets previously learned information upon learning new tasks. Despite the advanced capabilities of Large Language Models (LLMs), they continue to face challenges with CF during continual learning. The majority of existing research focuses on analyzing forgetting patterns through a singular training sequence, thereby overlooking the intricate effects that diverse tasks have on model behavior. Our study explores CF across various settings, discovering that model forgetting is influenced by both the specific training tasks and the models themselves. To this end, we interpret forgetting by examining the function vector (FV), a compact representation of functions in LLMs, offering a model-dependent indicator for the occurrence of CF. Through theoretical and empirical analyses, we demonstrated that CF in LLMs primarily stems from biases in function activation rather than the overwriting of task processing functions. Leveraging these insights, we propose a novel function vector guided training methodology, incorporating a regularization technique to stabilize the FV and mitigate forgetting. Empirical tests on four benchmarks confirm the effectiveness of our proposed training method, substantiating our theoretical framework concerning CF and model function dynamics. We plan to make our code publicly accessible in the near future.

Zhaoyi Li, Xiangyu Xi, Zhengyu Chen et al.

Supervised Fine-Tuning (SFT) on long Chain-of-Thought (CoT) trajectories has become a pivotal phase in building large reasoning models. However, how CoT trajectories from different sources influence the generalization performance of models remains an open question. In this paper, we conduct a comparative study using two sources of verified CoT trajectories generated by two competing models, \texttt{DeepSeek-R1-0528} and \texttt{gpt-oss-120b}, with their problem sets controlled to be identical. Despite their comparable performance, we uncover a striking paradox: lower training loss does not translate to better generalization. SFT on \texttt{DeepSeek-R1-0528} data achieves remarkably lower training loss, yet exhibits significantly worse generalization performance on reasoning benchmarks compared to those trained on \texttt{gpt-oss-120b}. To understand this paradox, we perform a multi-faceted analysis probing token-level SFT loss and step-level reasoning behaviors. Our analysis reveals a difference in reasoning patterns. \texttt{gpt-oss-120b} exhibits highly convergent and deductive trajectories, whereas \texttt{DeepSeek-R1-0528} favors a divergent and branch-heavy exploration pattern. Consequently, models trained with \texttt{DeepSeek-R1} data inherit inefficient exploration behaviors, often getting trapped in redundant exploratory branches that hinder them from reaching correct solutions. Building upon this insight, we propose a simple yet effective remedy of filtering out frequently branching trajectories to improve the generalization of SFT. Experiments show that training on selected \texttt{DeepSeek-R1-0528} subsets surprisingly improves reasoning performance by up to 5.1% on AIME25, 5.5% on BeyondAIME, and on average 3.6% on five benchmarks.

Yunhe Li, Hao Shi, Bowen Deng et al.

Although most of the automated theorem-proving approaches depend on formal proof systems, informal theorem proving can align better with large language models' (LLMs) strength in natural language processing. In this work, we identify a primary bottleneck in informal theorem proving as a lack of insight, namely the difficulty of recognizing the core techniques required to solve complex problems. To address this, we propose a novel framework designed to cultivate this essential reasoning skill and enable LLMs to perform insightful reasoning. We propose $\mathtt{DeepInsightTheorem}$, a hierarchical dataset that structures informal proofs by explicitly extracting core techniques and proof sketches alongside the final proof. To fully exploit this dataset, we design a Progressive Multi-Stage SFT strategy that mimics the human learning process, guiding the model from basic proof writing to insightful thinking. Our experiments on challenging mathematical benchmarks demonstrate that this insight-aware generation strategy significantly outperforms baselines. These results demonstrate that teaching models to identify and apply core techniques can substantially improve their mathematical reasoning.

Yuxuan Yao, Han Wu, Qiling Xu et al.

General-purpose text decoding approaches are usually adopted for dialogue response generation. Although the quality of the generated responses can be improved with dialogue-specific encoding methods, conversational decoding methods are still under-explored. Inspired by \citet{wu2023learning} that a good dialogue feature space should follow the rules of locality and isotropy, we present a fine-grained conversational decoding method, termed \textit{isotropic and proximal search (IPS)}. Our method is designed to generate the semantic-concentrated response, while still maintaining informativeness and discrimination against the context. Experiments show that our approach outperforms existing decoding strategies in the dialogue field across both automatic and human evaluation metrics. More in-depth analyses further confirm the effectiveness of our approach.

Zerui Yang, Weichuan Wang, Yanwei Xu et al.

Existing NL2SQL systems face two critical limitations: (1) they rely on in-context learning with only correct examples, overlooking the rich signal in historical error-fix pairs that could guide more robust self-correction; and (2) test-time scaling approaches often decompose questions arbitrarily, producing near-identical SQL candidates across runs and diminishing ensemble gains. Moreover, these methods suffer from a stark accuracy-efficiency trade-off: high performance demands excessive computation, while fast variants compromise quality. We present Memo-SQL, a training-free framework that addresses these issues through two simple ideas: structured decomposition and experience-aware self-correction. Instead of leaving decomposition to chance, we apply three clear strategies, entity-wise, hierarchical, and atomic sequential, to encourage diverse reasoning. For correction, we build a dynamic memory of both successful queries and historical error-fix pairs, and use retrieval-augmented prompting to bring relevant examples into context at inference time, no fine-tuning or external APIs required. On BIRD, Memo-SQL achieves 68.5% execution accuracy, setting a new state of the art among open, zero-fine-tuning methods, while using over 10 times fewer resources than prior TTS approaches.

Jian Xu, Sichun Luo, Xiangyu Chen et al.

Large Language Models (LLMs) have been integrated into recommendation systems to enhance user behavior comprehension. The Retrieval Augmented Generation (RAG) technique is further incorporated into these systems to retrieve more relevant items and improve system performance. However, existing RAG methods rely primarily on textual semantics and often fail to incorporate the most relevant items, limiting the effectiveness of the systems. In this paper, we propose Representation learning for retrieval-Augmented Large Language model Recommendation (RALLRec). Specifically, we enhance textual semantics by prompting LLMs to generate more detailed item descriptions, followed by joint representation learning of textual and collaborative semantics, which are extracted by the LLM and recommendation models, respectively. Considering the potential time-varying characteristics of user interest, a simple yet effective reranking method is further introduced to capture the dynamics of user preference. We conducted extensive experiments on three real-world datasets, and the evaluation results validated the effectiveness of our method. Code is made public at https://github.com/JianXu95/RALLRec.

Zhaoyi Li, Jiatong Li, Gangwei Jiang et al.

Chain-of-thought (CoT) reasoning has become the standard paradigm for enabling Large Language Models (LLMs) to solve complex problems. However, recent studies reveal a sharp performance drop in reasoning hop generalization scenarios, where the required number of reasoning steps exceeds training distributions while the underlying algorithm remains unchanged. The internal mechanisms driving this failure remain poorly understood. In this work, we conduct a systematic study on tasks from multiple domains, and find that errors concentrate at token positions of a few critical error types, rather than being uniformly distributed. Closer inspection reveals that these token-level erroneous predictions stem from internal competition mechanisms: certain attention heads, termed erroneous processing heads (ep heads), tip the balance by amplifying incorrect reasoning trajectories while suppressing correct ones. Notably, removing individual ep heads during inference can often restore the correct predictions. Motivated by these insights, we propose test-time correction of reasoning, a lightweight intervention method that dynamically identifies and deactivates ep heads in the reasoning process. Extensive experiments across different tasks and LLMs show that it consistently improves reasoning hop generalization, highlighting both its effectiveness and potential.

Xiangqun Zhang, Wei Feng, Ruize Han et al.

Person re-identification (Re-ID) is an important task and has significant applications for public security and information forensics, which has progressed rapidly with the development of deep learning. In this work, we investigate a novel and challenging setting of Re-ID, i.e., cross-domain video-based person Re-ID. Specifically, we utilize synthetic video datasets as the source domain for training and real-world videos for testing, notably reducing the reliance on expensive real data acquisition and annotation. To harness the potential of synthetic data, we first propose a self-supervised domain-invariant feature learning strategy for both static and dynamic (temporal) features. Additionally, to enhance person identification accuracy in the target domain, we propose a mean-teacher scheme incorporating a self-supervised ID consistency loss. Experimental results across five real datasets validate the rationale behind cross-synthetic-real domain adaptation and demonstrate the efficacy of our method. Notably, the discovery that synthetic data outperforms real data in the cross-domain scenario is a surprising outcome. The code and data are publicly available at https://github.com/XiangqunZhang/UDA_Video_ReID.

Haokun Lin, Haobo Xu, Yichen Wu et al.

Quantization of large language models (LLMs) faces significant challenges, particularly due to the presence of outlier activations that impede efficient low-bit representation. Traditional approaches predominantly address Normal Outliers, which are activations across all tokens with relatively large magnitudes. However, these methods struggle with smoothing Massive Outliers that display significantly larger values, which leads to significant performance degradation in low-bit quantization. In this paper, we introduce DuQuant, a novel approach that utilizes rotation and permutation transformations to more effectively mitigate both massive and normal outliers. First, DuQuant starts by constructing the rotation matrix, using specific outlier dimensions as prior knowledge, to redistribute outliers to adjacent channels by block-wise rotation. Second, We further employ a zigzag permutation to balance the distribution of outliers across blocks, thereby reducing block-wise variance. A subsequent rotation further smooths the activation landscape, enhancing model performance. DuQuant simplifies the quantization process and excels in managing outliers, outperforming the state-of-the-art baselines across various sizes and types of LLMs on multiple tasks, even with 4-bit weight-activation quantization. Our code is available at https://github.com/Hsu1023/DuQuant.

Han Wu, Mingjie Zhan, Haochen Tan et al.

Compared to news and chat summarization, the development of meeting summarization is hugely decelerated by the limited data. To this end, we introduce a versatile Chinese meeting summarization dataset, dubbed VCSum, consisting of 239 real-life meetings, with a total duration of over 230 hours. We claim our dataset is versatile because we provide the annotations of topic segmentation, headlines, segmentation summaries, overall meeting summaries, and salient sentences for each meeting transcript. As such, the dataset can adapt to various summarization tasks or methods, including segmentation-based summarization, multi-granularity summarization and retrieval-then-generate summarization. Our analysis confirms the effectiveness and robustness of VCSum. We also provide a set of benchmark models regarding different downstream summarization tasks on VCSum to facilitate further research. The dataset and code will be released at https://github.com/hahahawu/VCSum.

Jian Xu, Shao-Lun Huang, Linqi Song et al.

Gradient-based training in federated learning is known to be vulnerable to faulty/malicious clients, which are often modeled as Byzantine clients. To this end, previous work either makes use of auxiliary data at parameter server to verify the received gradients (e.g., by computing validation error rate) or leverages statistic-based methods (e.g. median and Krum) to identify and remove malicious gradients from Byzantine clients. In this paper, we remark that auxiliary data may not always be available in practice and focus on the statistic-based approach. However, recent work on model poisoning attacks has shown that well-crafted attacks can circumvent most of median- and distance-based statistical defense methods, making malicious gradients indistinguishable from honest ones. To tackle this challenge, we show that the element-wise sign of gradient vector can provide valuable insight in detecting model poisoning attacks. Based on our theoretical analysis of the \textit{Little is Enough} attack, we propose a novel approach called \textit{SignGuard} to enable Byzantine-robust federated learning through collaborative malicious gradient filtering. More precisely, the received gradients are first processed to generate relevant magnitude, sign, and similarity statistics, which are then collaboratively utilized by multiple filters to eliminate malicious gradients before final aggregation. Finally, extensive experiments of image and text classification tasks are conducted under recently proposed attacks and defense strategies. The numerical results demonstrate the effectiveness and superiority of our proposed approach. The code is available at \textit{\url{https://github.com/JianXu95/SignGuard}}

Wei Shao, Lemao Liu, Yinqiao Li et al.

Current online translation services require sending user text to cloud servers, posing a risk of privacy leakage when the text contains sensitive information. This risk hinders the application of online translation services in privacy-sensitive scenarios. One way to mitigate this risk for online translation services is introducing privacy protection mechanisms targeting the inference stage of translation models. However, compared to subfields of NLP like text classification and summarization, the machine translation research community has limited exploration of privacy protection during the inference stage. There is no clearly defined privacy protection task for the inference stage, dedicated evaluation datasets and metrics, and reference benchmark methods. The absence of these elements has seriously constrained researchers' in-depth exploration of this direction. To bridge this gap, this paper proposes a novel "Privacy-Preserving Machine Translation" (PPMT) task, aiming to protect the private information in text during the model inference stage. For this task, we constructed three benchmark test datasets, designed corresponding evaluation metrics, and proposed a series of benchmark methods as a starting point for this task. The definition of privacy is complex and diverse. Considering that named entities often contain a large amount of personal privacy and commercial secrets, we have focused our research on protecting only the named entity's privacy in the text. We expect this research work will provide a new perspective and a solid foundation for the privacy protection problem in machine translation.

Zhaoyi Li, Gangwei Jiang, Hong Xie et al.

LLMs have marked a revolutonary shift, yet they falter when faced with compositional reasoning tasks. Our research embarks on a quest to uncover the root causes of compositional reasoning failures of LLMs, uncovering that most of them stem from the improperly generated or leveraged implicit reasoning results. Inspired by our empirical findings, we resort to Logit Lens and an intervention experiment to dissect the inner hidden states of LLMs. This deep dive reveals that implicit reasoning results indeed surface within middle layers and play a causative role in shaping the final explicit reasoning results. Our exploration further locates multi-head self-attention (MHSA) modules within these layers, which emerge as the linchpins in accurate generation and leveraing of implicit reasoning results. Grounded on the above findings, we develop CREME, a lightweight method to patch errors in compositional reasoning via editing the located MHSA modules. Our empirical evidence stands testament to CREME's effectiveness, paving the way for autonomously and continuously enhancing compositional reasoning capabilities in language models.

Guanzhi Deng, Bo Li, Ronghao Chen et al.

Mixture-of-Experts (MoE) has become a prominent paradigm for scaling Large Language Models (LLMs). Parameter-efficient fine-tuning (PEFT), such as LoRA, is widely adopted to adapt pretrained MoE LLMs to downstream tasks. However, existing approaches assign identical LoRA ranks to all experts, overlooking the intrinsic functional specialization within MoE LLMs. This uniform allocation leads to resource mismatch, task-relevant experts are under-provisioned while less relevant ones receive redundant parameters. We propose a Dynamic Rank LoRA framework named DR-LoRA, which dynamically grows expert LoRA ranks during fine-tuning based on task-specific demands. DR-LoRA employs an Expert Saliency Scoring mechanism that integrates expert routing frequency and LoRA rank importance to quantify each expert's demand for additional capacity. Experts with higher saliency scores are prioritized for rank expansion, enabling the automatic formation of a heterogeneous rank distribution tailored to the target task. Experiments on multiple benchmarks demonstrate that DR-LoRA consistently outperforms standard LoRA and static allocation strategies under the same parameter budget, achieving superior task performance with more efficient parameter utilization.

Haokun Lin, Haoli Bai, Zhili Liu et al.

Vision-language pre-trained models have achieved impressive performance on various downstream tasks. However, their large model sizes hinder their utilization on platforms with limited computational resources. We find that directly using smaller pre-trained models and applying magnitude-based pruning on CLIP models leads to inflexibility and inferior performance. Recent efforts for VLP compression either adopt uni-modal compression metrics resulting in limited performance or involve costly mask-search processes with learnable masks. In this paper, we first propose the Module-wise Pruning Error (MoPE) metric, accurately assessing CLIP module importance by performance decline on cross-modal tasks. Using the MoPE metric, we introduce a unified pruning framework applicable to both pre-training and task-specific fine-tuning compression stages. For pre-training, MoPE-CLIP effectively leverages knowledge from the teacher model, significantly reducing pre-training costs while maintaining strong zero-shot capabilities. For fine-tuning, consecutive pruning from width to depth yields highly competitive task-specific models. Extensive experiments in two stages demonstrate the effectiveness of the MoPE metric, and MoPE-CLIP outperforms previous state-of-the-art VLP compression methods.

Weichuan Wang, Mingyang Liu, Linqi Song et al.

In recent years, the non-deterministic properties of language models have garnered considerable attention and have shown a significant influence on real-world applications. However, such properties remain under-explored in machine translation (MT), a complex, non-deterministic NLP task. In this study, we systematically evaluate modern MT systems and identify temperature-constrained Non-Deterministic MT (ND-MT) as a distinct phenomenon. Additionally, we demonstrate that ND-MT exhibits significant potential in addressing the multi-modality issue that has long challenged MT research and provides higher-quality candidates than Deterministic MT (D-MT) under temperature constraints. However, ND-MT introduces new challenges in evaluating system performance. Specifically, the evaluation framework designed for D-MT fails to yield consistent evaluation results when applied to ND-MT. We further investigate this emerging challenge by evaluating five state-of-the-art ND-MT systems across three open datasets using both lexical-based and semantic-based metrics at varying sampling sizes. The results reveal a Buckets effect across these systems: the lowest-quality candidate generated by ND-MT consistently determines the overall system ranking across different sampling sizes for all reasonable metrics. Furthermore, we propose the ExpectoSample strategy to automatically assess the reliability of evaluation metrics for selecting robust ND-MT.

Yuxuan Yao, Han Wu, Zhijiang Guo et al.

Large language models (LLMs) have demonstrated outstanding performance across various tasks, yet they still exhibit limitations such as hallucination, unfaithful reasoning, and toxic content. One potential approach to mitigate these issues is learning from human or external feedback (e.g. tools). In this paper, we introduce an intrinsic self-correct reasoning framework for LLMs that eliminates the need for human feedback, external tools, and handcraft prompts. The proposed framework, based on a multi-step reasoning paradigm \textbf{Le}arning from \textbf{Co}rrectness (\textsc{LeCo}), improves reasoning performance without needing to learn from errors. This paradigm prioritizes learning from correct reasoning steps, and a unique method to measure confidence for each reasoning step based on generation logits. Experimental results across various multi-step reasoning tasks demonstrate the effectiveness of the framework in improving reasoning performance with reduced token consumption.

Gangwei Jiang, Yahui Liu, Zhaoyi Li et al.

Recent advances in reasoning with large language models (LLMs) have popularized Long Chain-of-Thought (LCoT), a strategy that encourages deliberate and step-by-step reasoning before producing a final answer. While LCoTs have enabled expert-level performance in complex tasks, how the internal structures of their reasoning chains drive, or even predict, the correctness of final answers remains a critical yet underexplored question. In this work, we present LCoT2Tree, an automated framework that converts sequential LCoTs into hierarchical tree structures and thus enables deeper structural analysis of LLM reasoning. Using graph neural networks (GNNs), we reveal that structural patterns extracted by LCoT2Tree, including exploration, backtracking, and verification, serve as stronger predictors of final performance across a wide range of tasks and models. Leveraging an explainability technique, we further identify critical thought patterns such as over-branching that account for failures. Beyond diagnostic insights, the structural patterns by LCoT2Tree support practical applications, including improving Best-of-N decoding effectiveness. Overall, our results underscore the critical role of internal structures of reasoning chains, positioning LCoT2Tree as a powerful tool for diagnosing, interpreting, and improving reasoning in LLMs.

Tianqi Zhong, Zhaoyi Li, Quan Wang et al.

Compositional generalization, representing the model's ability to generate text with new attribute combinations obtained by recombining single attributes from the training data, is a crucial property for multi-aspect controllable text generation (MCTG) methods. Nonetheless, a comprehensive compositional generalization evaluation benchmark of MCTG is still lacking. We propose CompMCTG, a benchmark encompassing diverse multi-aspect labeled datasets and a crafted three-dimensional evaluation protocol, to holistically evaluate the compositional generalization of MCTG approaches. We observe that existing MCTG works generally confront a noticeable performance drop in compositional testing. To mitigate this issue, we introduce Meta-MCTG, a training framework incorporating meta-learning, where we enable models to learn how to generalize by simulating compositional generalization scenarios in the training phase. We demonstrate the effectiveness of Meta-MCTG through achieving obvious improvement (by at most 3.64%) for compositional testing performance in 94.4% cases.

Shuqi Liu, Han Wu, Bowei He et al.

Recent advances in large language models have led to numerous task-specialized fine-tuned variants, creating a need for efficient model merging techniques that preserve specialized capabilities while avoiding costly retraining. While existing task vector-based merging methods show promise, they typically apply uniform coefficients across all parameters, overlooking varying parameter importance both within and across tasks. We present Sens-Merging, a sensitivity-guided coefficient adjustment method that enhances existing model merging techniques by operating at both task-specific and cross-task levels. Our method analyzes parameter sensitivity within individual tasks and evaluates cross-task transferability to determine optimal merging coefficients. Extensive experiments on Mistral 7B and LLaMA2-7B/13B models demonstrate that Sens-Merging significantly improves performance across general knowledge, mathematical reasoning, and code generation tasks. Notably, when combined with existing merging techniques, our method enables merged models to outperform specialized fine-tuned models, particularly in code generation tasks. Our findings reveal important trade-offs between task-specific and cross-task scalings, providing insights for future model merging strategies.

Zekun Qian, Ruize Han, Junhui Hou et al.

Open-vocabulary multi-object tracking (OVMOT) represents a critical new challenge involving the detection and tracking of diverse object categories in videos, encompassing both seen categories (base classes) and unseen categories (novel classes). This issue amalgamates the complexities of open-vocabulary object detection (OVD) and multi-object tracking (MOT). Existing approaches to OVMOT often merge OVD and MOT methodologies as separate modules, predominantly focusing on the problem through an image-centric lens. In this paper, we propose VOVTrack, a novel method that integrates object states relevant to MOT and video-centric training to address this challenge from a video object tracking standpoint. First, we consider the tracking-related state of the objects during tracking and propose a new prompt-guided attention mechanism for more accurate localization and classification (detection) of the time-varying objects. Subsequently, we leverage raw video data without annotations for training by formulating a self-supervised object similarity learning technique to facilitate temporal object association (tracking). Experimental results underscore that VOVTrack outperforms existing methods, establishing itself as a state-of-the-art solution for open-vocabulary tracking task.

Sichun Luo, Guanzhi Deng, Jian Xu et al.

Personalization is a critical task in modern intelligent systems, with applications spanning diverse domains, including interactions with large language models (LLMs). Recent advances in reasoning capabilities have significantly enhanced LLMs, enabling unprecedented performance in tasks such as mathematics and coding. However, their potential for personalization tasks remains underexplored. In this paper, we present the first systematic evaluation of large reasoning models (LRMs) for personalization tasks. Surprisingly, despite generating more tokens, LRMs do not consistently outperform general-purpose LLMs, especially in retrieval-intensive scenarios where their advantages diminish. Our analysis identifies three key limitations: divergent thinking, misalignment of response formats, and ineffective use of retrieved information. To address these challenges, we propose Reinforced Reasoning for Personalization (\model), a novel framework that incorporates a hierarchical reasoning thought template to guide LRMs in generating structured outputs. Additionally, we introduce a reasoning process intervention method to enforce adherence to designed reasoning patterns, enhancing alignment. We also propose a cross-referencing mechanism to ensure consistency. Extensive experiments demonstrate that our approach significantly outperforms existing techniques.

Shuqi Liu, Yuxuan Yao, Bowei He et al.

Recent advances in large language models have led to specialized models excelling in specific domains, creating a need for efficient model merging techniques. While traditional merging approaches combine parameters into a single static model, they often compromise task-specific performance. However, task-specific routing methods maintain accuracy but introduce substantial storage overhead. We present \texttt{1bit}-Merging, a novel framework that integrates task-specific routing with 1-bit quantized task vectors to balance performance and storage efficiency. Our approach leverages the observation that different task-specific models store knowledge in distinct layers-chat models primarily in attention layers and math/code models in MLP layers, enabling targeted compression strategies. Through extensive experiments with LLaMA2 and Mistral model families across chat, mathematical reasoning, and code generation tasks, we demonstrate that 1bit-Merging achieves comparable or superior performance to existing methods while significantly reducing storage requirements. Our framework offers a practical solution for combining specialized models while maintaining their individual strengths and addressing the storage challenges of current approaches.

Shuqi Liu, Bowei He, Han Wu et al.

Post-training pruning has emerged as a crucial optimization technique as large language models (LLMs) continue to grow rapidly. However, the significant variations in weight distributions across different LLMs make fixed pruning strategies inadequate for multiple models. In this paper, we introduce \textbf{\textsc{OptiShear}}, an efficient evolutionary optimization framework for adaptive LLM pruning. Our framework features two key innovations: an effective search space built on our Meta pruning metric to handle diverse weight distributions, and a model-wise reconstruction error for rapid evaluation during search trials. We employ Non-dominated Sorting Genetic Algorithm III (NSGA-III) to optimize both pruning metrics and layerwise sparsity ratios. Through extensive evaluation on LLaMA-1/2/3 and Mistral models (7B-70B) across multiple benchmarks, we demonstrate that our adaptive pruning metrics consistently outperform existing methods. Additionally, our discovered layerwise sparsity ratios enhance the effectiveness of other pruning metrics. The framework exhibits strong cross-task and cross-model generalizability, providing a cost-effective solution for model compression.

Xiangru Jian, Hao Xu, Wei Pang et al.

The manufacturing sector is increasingly adopting Multimodal Large Language Models (MLLMs) to transition from simple perception to autonomous execution, yet current evaluations fail to reflect the rigorous demands of real-world manufacturing environments. Progress is hindered by data scarcity and a lack of fine-grained domain semantics in existing datasets. To bridge this gap, we introduce FORGE. Wefirst construct a high-quality multimodal dataset that combines real-world 2D images and 3D point clouds, annotated with fine-grained domain semantics (e.g., exact model numbers). We then evaluate 18 state-of-the-art MLLMs across three manufacturing tasks, namely workpiece verification, structural surface inspection, and assembly verification, revealing significant performance gaps. Counter to conventional understanding, the bottleneck analysis shows that visual grounding is not the primary limiting factor. Instead, insufficient domain-specific knowledge is the key bottleneck, setting a clear direction for future research. Beyond evaluation, we show that our structured annotations can serve as an actionable training resource: supervised fine-tuning of a compact 3B-parameter model on our data yields up to 90.8% relative improvement in accuracy on held-out manufacturing scenarios, providing preliminary evidence for a practical pathway toward domain-adapted manufacturing MLLMs. The code and datasets are available at https://ai4manufacturing.github.io/forge-web.

Yuchen Cao, Hanlin Zhang, Jacky Wai Keung et al.

Large language models (LLMs) are increasingly used as quantitative research copilots to translate natural-language strategy specifications into executable trading code. Yet most existing evaluations either focus on static financial knowledge or summarize performance with a single profitability metric, leaving a gap for benchmarking strategy-to-code trading systems as governed, auditable software. We introduce SysTradeBench (SysTB), an iterative build-test-patch benchmark that evaluates LLM-generated trading systems under drift-aware diagnostics. Given a standardized Base Strategy Doc and frozen semantics, each model must produce (i) a strategy card, (ii) executable code, and (iii) mandatory audit logs. A sandboxed harness runs determinism and anti-leakage checks, detects rule drift across iterations, and returns evidence bundles to support constrained patches. SysTradeBench reports multi-dimensional scorecards for spec fidelity, risk discipline, reliability, and out-of-sample robustness indicators, together with cost-effectiveness signals. We evaluate 17 models across 12 strategies. Top models achieve validity above 91.7 percent with strong aggregate scores, but evidence-driven iteration also induces code convergence by Iter2. These findings suggest that LLM iteration complements rather than replaces human quantitative researcher governance: LLMs excel at rapid prototyping and shallow bug fixes, while human oversight remains essential for critical strategies requiring solution diversity and ensemble robustness.