论文检索

Ximing Dong, Shaowei Wang, Dayi Lin, Ahmed Hassan

Optimizing Large Language Model (LLM) performance requires well-crafted prompts, but manual prompt engineering is labor-intensive and often ineffective. Automated prompt optimization techniques address this challenge but the major of them rely on randomly selected evaluation subsets, which fail to represent the full dataset, leading to unreliable evaluations and suboptimal prompts. Existing coreset selection methods, designed for LLM benchmarking, are unsuitable for prompt optimization due to challenges in clustering similar samples, high data collection costs, and the unavailability of performance data for new or private datasets. To overcome these issues, we propose IPOMP, an Iterative evaluation data selection approach for effective Prompt Optimization using real time Model Performance. IPOMP is a two-stage approach that selects representative and diverse samples using semantic clustering and boundary analysis, followed by iterative refinement with real-time model performance data to replace redundant samples. Evaluations on two datasets BIG-bench and LIAR, and two models GPT-3.5 and GPT-4o-mini, show that IPOMP improves effectiveness by at least 1.6% to 3.1%, and stability by at least 50% to 55.5% compared with the best baseline across the studied datasets and models, with minimal computational overhead below 1%. Furthermore, the results demonstrate that our real-time performance-guided refinement approach can be universally applied to enhance existing coreset selection methods.

Tomás Vergara Browne, Alvaro Soto

Motivated by the surge of large language models, there has been a push to formally characterize the symbolic abilities intrinsic to the transformer architecture. A programming language, called RASP, has been proposed, which can be directly compiled into transformer weights to implement these algorithms. However, the tasks that can be implemented in RASP are often uncommon to learn from natural unsupervised data, showing a mismatch between theoretical capabilities of the transformer architecture, and the practical learnability of these capabilities from unsupervised data. We propose tracr-injection, a method that allows us to distill algorithms written in RASP directly into a pre-trained language model. We showcase our method by injecting 3 different algorithms into a language model. We show how our method creates an interpretable subspace within the model’s residual stream, which can be decoded into the variables present in the code of the RASP algorithm. Additionally, we found that the proposed method can improve out-of-distribution performance compared to our baseline, indicating that indeed a more symbolic mechanism is taking place in the inner workings of the model. We release the code used to run our experiments.

Vera Neplenbroek, Arianna Bisazza, Raquel Fernández

Recent generative large language models (LLMs) show remarkable performance in non-English languages, but when prompted in those languages they tend to express higher harmful social biases and toxicity levels. Prior work has shown that finetuning on specialized datasets can mitigate this behavior, and doing so in English can transfer to other languages. In this work, we investigate the impact of different finetuning methods on the model’s bias and toxicity, but also on its ability to produce fluent and diverse text. We reduce biases by finetuning on curated non-harmful text, but find only direct preference optimization to be effective for mitigating toxicity. The mitigation caused by applying these methods in English also transfers to non-English languages. We find evidence that the extent to which transfer takes place can be predicted by the amount of data in a given language present in the model’s pretraining data. However, this transfer of bias and toxicity mitigation often comes at the expense of decreased language generation ability in non-English languages, highlighting the importance of developing language-specific bias and toxicity mitigation methods.

Łukasz Borchmann

ARC Challenge appears more difficult than ARC Easy for modern LLMs primarily due to an evaluation setup that prevents direct comparison of answer choices rather than inherent complexity. Although some researchers have quietly shifted to a more appropriate scheme over the last year, the implications of this change have yet to be widely acknowledged. We highlight this overlooked shift, show how similar evaluation practices falsely imply reasoning deficits in other benchmarks, and demonstrate that fairer methods dramatically reduce performance gaps (e.g. on SIQA) and even yield superhuman results (OpenBookQA). In doing so, we reveal how evaluation shapes perceived difficulty and offer guidelines to ensure that multiple-choice evaluations accurately reflect actual model capabilities.

Yahan Yu, Duzhen Zhang, Yong Ren, Xuanle Zhao, Xiuyi Chen, Chenhui Chu

Multimodal Continual Instruction Tuning (MCIT) empowers Multimodal Large Language Models (MLLMs) to adapt to ever-evolving requirements without continuous costly retraining. However, MCIT faces challenges in mitigating Catastrophic Forgetting (CF) and enhancing Knowledge Transfer (KT). Existing works combine Mixture-of-Expert (MoE) and LoRA to address these. However, using a fixed number of shared LoRA blocks across tasks can lead to the overwriting of acquired knowledge, making MLLMs harder to handle CF and KT. Therefore, we propose the **Prog**ressive **LoRA** framework (ProgLoRA), which contains a progressive LoRA pool and trains a new LoRA block for each incremental task to reduce knowledge interference. Specifically, ProgLoRA has two key mechanisms: task-aware allocation for effectively leveraging acquired knowledge at current task and task recall for realigning the model with learned tasks. Additionally, considering different application scenarios, we design a static ProgLoRA for the more idealized basic setting and a dynamic ProgLoRA for the more realistic challenging setting. Experiments on the latest MCIT benchmark demonstrate that ProgLoRA outperforms existing approaches.

Jilong Li, Zhenxi Song, Jiaqi Wang, Meishan Zhang, Honghai Liu, Min Zhang, Zhiguo Zhang

Current EEG/MEG-to-text decoding systems suffer from three key limitations: (1) reliance on teacher-forcing methods, which compromises robustness during inference, (2) sensitivity to session-specific noise, hindering generalization across subjects, and (3) misalignment between brain signals and linguistic representations due to pre-trained language model over-dominance. To overcome these challenges, we propose BrainECHO (\textbf{B}rain signal decoding via v\textbf{E}ctor-quantized spe\textbf{C}trogram reconstruction for W\textbf{H}isper-enhanced text generati\textbf{O}n), a multi-stage framework that employs decoupled representation learning to achieve state-of-the-art performance on both EEG and MEG datasets. Specifically, BrainECHO consists of three stages: (1) Discrete autoencoding, which transforms continuous Mel spectrograms into a finite set of high-quality discrete representations for subsequent stages. (2) Frozen alignment, where brain signal embeddings are mapped to corresponding Mel spectrogram embeddings in a frozen latent space, effectively filtering session-specific noise through vector-quantized reconstruction, yielding a 3.65% improvement in BLEU-4 score. (3) Constrained decoding fine-tuning, which leverages the pre-trained Whisper model for audio-to-text translation, balancing signal adaptation with knowledge preservation, and achieving 74%-89% decoding BLEU scores without excessive reliance on teacher forcing. BrainECHO demonstrates robustness across sentence, session, and subject-independent conditions, passing Gaussian noise tests and showcasing its potential for enhancing language-based brain-computer interfaces.

V.S.D.S.Mahesh Akavarapu, Hrishikesh Terdalkar, Pramit Bhattacharyya, Shubhangi Agarwal, Dr. Vishakha Deulgaonkar, Chaitali Dangarikar, Pralay Manna, Arnab Bhattacharya

Large Language Models (LLMs) have demonstrated remarkable generalization capabilities across diverse tasks and languages. In this study, we focus on natural language understanding in three classical languages—Sanskrit, Ancient Greek and Latin—to investigate the factors affecting cross-lingual zero-shot generalization. First, we explore named entity recognition and machine translation into English. While LLMs perform equal to or better than fine-tuned baselines on out-of-domain data, smaller models often struggle, especially with niche or abstract entity types. In addition, we concentrate on Sanskrit by presenting a factoid question–answering (QA) dataset and show that incorporating context via retrieval-augmented generation approach significantly boosts performance. In contrast, we observe pronounced performance drops for smaller LLMs across these QA tasks. These results suggest model scale as an important factor influencing cross-lingual generalization. Assuming that models used such as GPT-4o and Llama-3.1 are not instruction fine-tuned on classical languages, our findings provide insights into how LLMs may generalize on these languages and their consequent utility in classical studies.

Keyuan Cheng, Xudong Shen, Yihao Yang, Tengyue Wang, Yang Cao, Muhammad Asif Ali, Hanbin Wang, Lijie Hu, Di Wang

Large language models (LLMs) have demonstrated remarkable proficiency in handling a wide range of tasks within the software engineering domain, but their ability to perform code migration—adapting code to different environments—remains underexplored. In this work, we propose a novel benchmark, : Code Migration Across Environment, designed to evaluate LLMs’ performance in handling code migration tasks. The benchmark comprises 922 data points across 19 Python and Java packages, offering three tasks to systematically evaluate code migration: identifying version-incompatible functions, determining function changes, and adapting code to target environments. Experimental evaluation of across seven LLMs revealed an average pass@1 rate of 26.50%, with GPT-4o performing best at 43.84%. We highlight our key findings as follows: (i) LLMs are more familiar with newer function versions, making them better at migrating legacy code, and (ii) a logical inconsistency where LLMs sometimes identify irrelevant function changes for the target migration environment.

Jiashun Cheng, Aochuan Chen, Nuo Chen, Ziqi Gao, Yuhan Li, Jia Li, Fugee Tsung

Low-Rank Adaptation (LoRA) has emerged as a prominent technique for fine-tuning large foundation models. Despite its successes, the substantial parameter redundancy, which limits the capacity and efficiency of LoRA, has been recognized as a bottleneck. In this work, we systematically investigate the impact of redundancy in fine-tuning LoRA and reveal that reducing density redundancy does not degrade expressiveness. Based on this insight, we introduce Spectral-encoding Low-Rank Adaptation (SeLoRA), which harnesses the robust expressiveness of spectral bases to re-parameterize LoRA from a sparse spectral subspace. Designed with simplicity, SeLoRA enables seamless integration with various LoRA variants for performance boosting, serving as a scalable plug-and-play framework. Extensive experiments substantiate that SeLoRA achieves greater efficiency with fewer parameters, delivering superior performance enhancements over strong baselines on various downstream tasks, including commonsense reasoning, math reasoning, and code generation.

Wenhao Li, Yuxin Zhang, Gen Luo, Daohai Yu, Rongrong Ji

While long-context large language models (LLMs) exhibit remarkable document processing capabilities, their prohibitively high training costs often hinder customized applications. To mitigate this issue, we propose __Sequential Chunk-wise Optimization (SeCO)__, a memory-efficient training paradigm that partitions lengthy inputs into manageable chunks. Each chunk independently constructs its computational graph and performs localized backpropagation, ensuring that only one chunk’s forward activations are stored in memory. Building on SeCO, we further introduce __Sparse Chunk-wise Optimization (SpaCO)__, which reduces computational overhead by selectively propagating gradients to specific chunks and incorporates a carefully designed compensation factor to ensure unbiased gradient estimation. SpaCO decouples the computational cost of backpropagation from the context length, enabling training time to gradually converge to inference time as sequences become longer. Implemented as lightweight training wrappers, both SeCO and SpaCO offer substantial practical benefits. For example, when fine-tuning an 8B model with LoRA on a single RTX 3090 GPU, SeCO expands maximum sequence length from 1K to 16K tokens, while SpaCO demonstrates accelerated training speed—achieving up to 3× faster than SeCO under the same experimental setup. These innovations provide new insights into optimizing long-context models, making them more accessible for practical applications. We have open-sourced the code at https://anonymous.4open.science/r/seco-CCBD.

Yuxia Geng, Runkai Zhu, Jiaoyan Chen, Jintai Chen, Xiang Chen, Zhuo Chen, Shuofei Qiao, Yuxiang Wang, Xiaoliang Xu, Sheng-Jun Huang

Disentanglement of visual features of primitives (i.e., attributes and objects) has shown exceptional results in Compositional Zero-shot Learning (CZSL). However, due to the feature divergence of an attribute (resp. object) when combined with different objects (resp. attributes), it is challenging to learn disentangled primitive features that are general across different compositions. To this end, we propose the solution of cross-composition feature disentanglement, which takes multiple primitive-sharing compositions as inputs and constrains the disentangled primitive features to be general across these compositions. More specifically, we leverage a compositional graph to define the overall primitive-sharing relationships between compositions, and build a task-specific architecture upon the recently successful large pre-trained vision-language model (VLM) CLIP, with dual cross-composition disentangling adapters (called L-Adapter and V-Adapter) inserted into CLIP’s frozen text and image encoders, respectively. Evaluation on three popular CZSL benchmarks shows that our proposed solution significantly improves the performance of CZSL, and its components have been verified by solid ablation studies. Our code and data are available at: https://github.com/zhurunkai/DCDA.

Xiaoyang Hu, Richard Lewis

Cognitive tasks originally developed for humans are now increasingly used to study language models. While applying these tasks is often straightforward, interpreting their results can be challenging. In particular, when a model underperforms, it is often unclear whether this results from a limitation in the cognitive ability being tested or a failure to understand the task itself. A recent study argues that GPT 3.5’s declining performance on 2-back and 3-back tasks reflects a working memory capacity limit similar to humans (Gong et al., 2024). By analyzing a range of open-source language models of varying performance levels on these tasks, we show that the poor performance is due at least in part to a limitation in task comprehension and task set maintenance. We challenge the best-performing model with progressively harder versions of the task (up to 10-back) and experiment with alternative prompting strategies, before analyzing model attentions. Our larger aim is to contribute to the ongoing conversation around refining methodologies for the cognitive evaluation of language models.

Lang Qin, Yao Zhang, Hongru Liang, Adam Jatowt, Zhenglu Yang

Medical Dialogue Systems (MDSs) have emerged as promising tools for automated healthcare support through patient-agent interactions. Previous efforts typically relied on an idealized assumption — patients can accurately report symptoms aligned with their actual health conditions. However, in reality, patients often misreport their symptoms, due to cognitive limitations, emotional factors, etc. Overlooking patient misreports can significantly compromise the diagnostic accuracy of MDSs. To address this critical issue, we emphasize the importance of enabling MDSs to “listen to patients” by tackling two key challenges: how to detect misreport and mitigate misreport effectively. In this work, we propose PaMis, a novel framework that can detect patient misreports based on calculating the structural entropy of the dialogue entity graph, and mitigate them through generating controlled clarifying questions. Our experimental results demonstrate that PaMis effectively enhances MDSs reliability by effectively addressing patient misreports during the medical response generation process.

Ming Zhang, Yuhui Wang, Yujiong Shen, Tingyi Yang, Changhao Jiang, Yilong Wu, Shihan Dou, Qinhao Chen, Zhiheng Xi, Zhihao Zhang 等

Process-driven dialogue systems, which operate under strict predefined process constraints, are essential in customer service and equipment maintenance scenarios. Although Large Language Models (LLMs) have shown remarkable progress in dialogue and reasoning, they still struggle to solve these strictly constrained dialogue tasks. To address this challenge, we construct Process Flow Dialogue (PFDial) dataset, which contains 12,705 high-quality Chinese dialogue instructions derived from 440 flowcharts containing 5,055 process nodes. Based on PlantUML specification, each UML flowchart is converted into atomic dialogue units i.e., structured five-tuples. Experimental results demonstrate that a 7B model trained with merely 800 samples, and a 0.5B model trained on total data both can surpass 90% accuracy. Additionally, the 8B model can surpass GPT-4o up to 43.88% with an average of 11.00%. We further evaluate models’ performance on challenging backward transitions in process flows and conduct an in-depth analysis of various dataset formats to reveal their impact on model performance in handling decision and sequential branches. The data is released in https://github.com/KongLongGeFDU/PFDial.

Shangda Wu, Guo Zhancheng, Ruibin Yuan, Junyan Jiang, SeungHeon Doh, Gus Xia, Juhan Nam, Xiaobing Li, Feng Yu, Maosong Sun

CLaMP 3 is a unified framework developed to address challenges of cross-modal and cross-lingual generalization in music information retrieval. Using contrastive learning, it aligns all major music modalities–including sheet music, performance signals, and audio recordings–with multilingual text in a shared representation space, enabling retrieval across unaligned modalities with text as a bridge. It features a multilingual text encoder adaptable to unseen languages, exhibiting strong cross-lingual generalization. Leveraging retrieval-augmented generation, we curated M4-RAG, a web-scale dataset consisting of 2.31 million music-text pairs. This dataset is enriched with detailed metadata that represents a wide array of global musical traditions. To advance future research, we release WikiMT-X, a benchmark comprising 1,000 triplets of sheet music, audio, and richly varied text descriptions. Experiments show that CLaMP 3 achieves state-of-the-art performance on multiple MIR tasks, significantly surpassing previous strong baselines and demonstrating excellent generalization in multimodal and multilingual music contexts.

Rongguang Ye, Ming Tang

Existing pruning methods for large language models (LLMs) focus on achieving high compression rates while maintaining model performance. Although these methods have demonstrated satisfactory performance in handling a single user’s compression request, their processing time increases linearly with the number of requests, making them inefficient for real-world scenarios with multiple simultaneous requests. To address this limitation, we propose a Univeral Model for Customized Compression (UniCuCo) for LLMs, which introduces a StratNet that learns to map arbitrary requests to their optimal pruning strategy. The challenge in training StratNet lies in the high computational cost of evaluating pruning strategies and the non-differentiable nature of the pruning process, which hinders gradient backpropagation for StratNet updates. To overcome these challenges, we leverage a Gaussian process to approximate the evaluation process. Since the gradient of the Gaussian process is computable, we can use it to approximate the gradient of the non-differentiable pruning process, thereby enabling StratNet updates. Experimental results show that UniCuCo is 28 times faster than baselines in processing 64 requests, while maintaining comparable accuracy to baselines.

Junhao Hu, Wenrui Huang, Weidong Wang, Zhenwen Li, Tiancheng Hu, Zhixia Liu, Xusheng Chen, Tao Xie, Yizhou Shan

Large Language Models (LLMs) have demonstrated strong capabilities across various domains, with recent advancements in challenging reasoning tasks such as mathematics and programming. However, solving reasoning tasks often requires an LLM to generate long sequences, incurring O(N) time and memory complexities per token, where N is the current sequence length. To reduce complexities, existing sparsity-based algorithms propose to retain Key-Value (KV) vectors, the intermediate representations of only the most critical tokens. However, these algorithms struggle with the “impossible trinity” of accuracy, time, and memory. For example, the state-of-the-art algorithm, Quest, achieves high accuracy with O(L) time but O(N) memory (L is the cache budget, L \ll N). To address the “impossible trinity”, in this paper, we identify a new attention pattern during the decode stage of reasoning tasks, where milestone tokens (analogous to lemmas in mathematical proofs) emerge, are utilized, and then become unimportant afterward. Based on this pattern, we propose a new algorithm RaaS that identifies milestone tokens and retains their KV vectors until they are no longer needed, achieving high accuracy with O(L) time and O(L) memory complexities.

Keyuan Cheng, Zijian Kan, Zhuoran Zhang, Muhammad Asif Ali, Lijie Hu, Di Wang

Knowledge Editing-Efficiently modifying the knowledge in large language models has gathered great attention. Current benchmarks primarily use multi-hop question answering to assess and analyze newly injected or updated knowledge. However, we argue that these benchmarks fail to effectively evaluate how well the updated models apply this knowledge in real-life scenarios, particularly when questions require complex reasoning involving one-to-many relationships or multi-step logical intersections. To fill in this gap, we introduce a new benchmark, COMPKE: Complex Question Answering under Knowledge Editing, which includes 11,924 complex questions that reflect real-life situations. We perform a comprehensive evaluation of four different knowledge editing methods in COMPKE, and our results show that the performance of these methods varies between different models. For example, MeLLo achieves an accuracy of 39.47 on GPT-4o-mini but drops significantly to 3.83 on Qwen2.5-3B. We further analyze the reasons behind these results from both methodological and model perspectives. Our dataset will be publicly available on GitHub.

Boyang Xue, Hongru Wang, Rui Wang, Sheng Wang, Zezhong Wang, Yiming Du, Bin Liang, Wenxuan Zhang, Kam-Fai Wong

The tendency of Large Language Models (LLMs) to generate hallucinations raises concerns regarding their reliability. Therefore, confidence estimations indicating the extent of trustworthiness of the generations become essential. However, current LLM confidence estimations in languages other than English remain underexplored. This paper addresses this gap by introducing a comprehensive investigation of Multilingual Confidence estimation (MlingConf) on LLMs, focusing on both language-agnostic (LA) and language-specific (LS) tasks to explore the performance and language dominance effects of multilingual confidence estimations on different tasks. The benchmark comprises four meticulously checked and human-evaluated high-quality multilingual datasets for LA tasks and one for the LS task tailored to specific social, cultural, and geographical contexts of a language. Our experiments reveal that on LA tasks English exhibits notable linguistic dominance in confidence estimations than other languages, while on LS tasks, using question-related language to prompt LLMs demonstrates better linguistic dominance in multilingual confidence estimations. The phenomena inspire a simple yet effective native-tone prompting strategy by employing language-specific prompts for LS tasks, effectively improving LLMs’ reliability and accuracy in LS scenarios.

Ke Wang, Junting Pan, Linda Wei, Aojun Zhou, Weikang Shi, Zimu Lu, Han Xiao, Yunqiao Yang, Houxing Ren, Mingjie Zhan 等

Natural language image-caption datasets, widely used for training Large Multimodal Models, mainly focus on natural scenarios and overlook the intricate details of mathematical figures that are critical for problem-solving, hindering the advancement of current LMMs in multimodal mathematical reasoning. To this end, we propose leveraging code as supervision for cross-modal alignment, since code inherently encodes all information needed to generate corresponding figures, establishing a precise connection between the two modalities. Specifically, we co-develop our image-to-code model and dataset with model-in-the-loop approach, resulting in an image-to-code model, FigCodifier and ImgCode-8.6M dataset, the largest image-code dataset to date. Furthermore, we utilize FigCodifier to synthesize novel mathematical figures and then construct MM-MathInstruct-3M, a high-quality multimodal math instruction fine-tuning dataset. Finally, we present MathCoder-VL, trained with ImgCode-8.6M for cross-modal alignment and subsequently fine-tuned on MM-MathInstruct-3M for multimodal math problem solving. Our model achieves a new open-source SOTA across all six metrics. Notably, it surpasses GPT-4o and Claude 3.5 Sonnet in the geometry problem-solving subset of MathVista, achieving improvements of 8.9% and 9.2%.