论文检索

Luisa Geiger, Mareike Hartmann, Michael Sullivan, Alexander Koller

In this paper, we propose a novel, automatic tree-based evaluation metric for LLM-generated step-by-step assembly instructions, that more accurately reflects spatiotemporal aspects of construction than traditional metrics such as BLEU and BERT similarity scores. We apply our proposed metric to the domain of sewing instructions, and show that our metric better correlates with manually-annotated error counts, demonstrating our metric’s superiority for evaluating the spatiotemporal soundness of sewing instructions. Further experiments show that our metric is more robust than traditional approaches against artificially-constructed counterfactual examples that are specifically constructed to confound metrics that rely on textual similarity.

Zhuozhuo Tu, Cheng Chen, Yuxuan Du

The quality of data plays a vital role in the development of Large-scale Generative Models. Understanding how important a data point is for a generative model is essential for explaining its behavior and improving the performance. The influence function provides a framework for quantifying the impact of individual training data on model predictions. However, the high computational cost has hindered their applicability in large-scale applications. In this work, we present RRInf, a novel and principled method for estimating influence function in large-scale generative AI models. We show that influence function estimation can be transformed into a ridge regression problem. Based on this insight, we develop an algorithm that is efficient and scalable to large models. Experiments on noisy data detection and influential data identification tasks demonstrate that RRInf outperforms existing methods in terms of both efficiency and effectiveness for commonly used large models: RoBERTa-large, Llama-2-13B-chat, Llama-3-8B and stable-diffusion-v1.5.

Junzhuo Li, Bo Wang, Xiuze Zhou, Xuming Hu

Mixture-of-Experts (MoE) models offer immense capacity via sparsely gated expert subnetworks, yet adapting them to multiple domains without catastrophic forgetting remains an open challenge. Existing approaches either incur prohibitive computation, suffer cross-domain interference, or require separate runs per domain. We propose DES-MoE, a dynamic expert specialization framework for multi-domain adaptation of Mixture-of-Experts models. DES-MoE addresses catastrophic forgetting through three innovations: (1) an adaptive router balancing pre-trained knowledge retention and task-specific updates via distillation, (2) real-time expert-domain correlation mapping to isolate domain-specific gradients, and (3) a three-phase adaptive fine-tuning schedule that progressively freezes non-specialized parameters. Evaluated on six domains (math, code, law, etc.), DES-MoE matches single-domain ESFT performance while training one unified model, reduces forgetting by 89% compared to full fine-tuning as domains scale from 2 to 6, and achieves 68% faster convergence than conventional methods. Our work establishes dynamic expert isolation as a scalable paradigm for multi-task MoE adaptation.

Kaiyan Chang, Yonghao Shi, Chenglong Wang, Hang Zhou, Chi Hu, Xiaoqian Liu, Yingfeng Luo, Yuan Ge, Tong Xiao, JingBo Zhu

Test-Time Scaling (TTS) is a promising approach to progressively elicit the model’s intelligence during inference. Recently, training-based TTS methods, such as continued reinforcement learning (RL), have further surged in popularity, while training-free TTS methods are gradually fading from prominence. However, the additional computation overhead of training amplifies the burden on test-time scaling.In this paper, we focus on training-free TTS methods for reasoning. We first design Conditional Step-level Self-refinement, a fine-grained sequential scaling method guided by process verification. On top of its effectiveness, we further combine it with other classical parallel scaling methods at the step level, to introduce a novel inference paradigm called Hybrid Test-Time Scaling. Extensive experiments on five instruction-tuned LLMs across different scales (3B-14B) and families demonstrate that hybrid strategy incorporating various training-free TTS methods at a fine granularity has considerable potential for expanding the reasoning performance boundaries of LLMs.

Zhuowen Han, Xinwei Wu, Dan Shi, Renren Jin, Deyi Xiong

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

Ruobing Wang, Qiaoyu Tan, Yili Wang, Ying Wang, Xin Wang

Designing crystal materials with desired physicochemical properties remains a fundamental challenge in materials science. While large language models (LLMs) have demonstrated strong in-context learning (ICL) capabilities, existing LLM-based crystal generation approaches are limited to zero-shot scenarios and are unable to benefit from few-shot scenarios. In contrast, human experts typically design new materials by modifying relevant known structures which aligns closely with the few-shot ICL paradigm. Motivated by this, we propose CrystalICL, a novel model designed for few-shot crystal generation. Specifically, we introduce a space-group based crystal tokenization method, which effectively reduces the complexity of modeling crystal symmetry in LLMs. We further introduce a condition-structure aware hybrid instruction tuning framework and a multi-task instruction tuning strategy, enabling the model to better exploit ICL by capturing structure-property relationships from limited data. Extensive experiments on four crystal generation benchmarks demonstrate the superiority of CrystalICL over the leading baseline methods on conditional and unconditional generation tasks.

Haijiang Liu, Qiyuan Li, Chao Gao, Yong Cao, Xiangyu Xu, Xun Wu, Daniel Hershcovich, Jinguang Gu

Introducing **MARK**, the **M**ulti-st**A**ge **R**easoning framewor**K** for cultural value survey response simulation, designed to enhance the accuracy, steerability, and interpretability of large language models in this task. The system is inspired by the type dynamics theory in the MBTI psychological framework for personality research. It effectively predicts and utilizes human demographic information for simulation: life-situational stress analysis, group-level personality prediction, and self-weighted cognitive imitation. Experiments on the World Values Survey show that MARK outperforms existing baselines by 10% accuracy and reduces the divergence between model predictions and human preferences. This highlights the potential of our framework to improve zero-shot personalization and help social scientists interpret model predictions.

Yeliang Xiu, Yongmei Liu

Non-monotonic reasoning (NMR) refers to the fact that conclusions may be invalidated by new information. It is widely used in daily life and legal reasoning. An NMR task usually has multiple extensions, which are sets of plausible conclusions. There are two reasoning modes – skeptical and credulous reasoning, depending on whether to believe facts in all extensions or any one extension. Despite some preliminary work exploring the NMR abilities of LLMs, the multi-extension NMR capabilities of LLMs remain underexplored. In this paper, we synthesize a multi-extension NMR dataset MultiLogicNMR, and construct two variants of the dataset with more extensions or text diversity. We propose a neural-symbolic framework MultiLogicNMRer for multi-extension NMR. Experimental evaluation with the datasets shows that LLMs still face significant challenges in NMR abilities, and reveal the effectiveness of our neural-symbolic framework, with an average accuracy gain of about 15% compared to prompt-based methods, and even outperforming some fine-tuning methods. All code and data are publicly available.

Marija Sakota, Robert West

Many recent approaches to structured NLP tasks use an autoregressive language model M to map unstructured input text x to output text y representing structured objects (such as tuples, lists, trees, code, etc.), where the desired output structure is enforced via constrained decoding. During training, these approaches do not require the model to be aware of the constraints, which are merely implicit in the training outputs y. This is advantageous as it allows for dynamic constraints without requiring retraining, but can lead to low-quality output during constrained decoding at test time. We overcome this problem with Boosted Constrained Decoding (BoostCD) which combines constrained and unconstrained decoding in two phases: Phase 1 decodes from the base model M twice, in constrained and unconstrained mode, obtaining two weak predictions. In phase 2, a learned autoregressive boosted model combines the two weak predictions into one final prediction. The mistakes made by the base model with vs. without constraints tend to be complementary, which the boosted model learns to exploit for improved performance. We demonstrate the power of BoostCD by applying it to closed information extraction. Our model, BoostIE, outperforms prior approaches both in and out of distribution, addressing several common errors identified in those approaches.

Kai Chen, Zihao He, Taiwei Shi, Kristina Lerman

Steerability, or the ability of large language models (LLMs) to adapt outputs to align with diverse community-specific norms, perspectives, and communication styles, is critical for real-world applications but remains under-evaluated. We introduce STEER-BENCH, a benchmark for assessing population-specific steering using contrasting Reddit communities. Covering 30 contrasting subreddit pairs across 19 domains, STEER-BENCH includes over 10,000 instruction-response pairs and validated 5,500 multiple-choice questions with corresponding silver labels to test alignment with diverse community norms. It systematically assesses how effectively LLMs understand community-specific instructions, their resilience to adversarial steering attempts, and their ability to accurately represent diverse cultural and ideological perspectives. Our evaluation of 13 popular LLMs using STEER-BENCH reveals that while human experts achieve an accuracy of 81% with silver labels, the best-performing models reach only around 65% accuracy depending on the domain and configuration. Some models lag behind human-level alignment by over 15 percentage points, highlighting significant gaps in community-sensitive steerability.

Gabriele Sarti, Vilém Zouhar, Malvina Nissim, Arianna Bisazza

Word-level quality estimation (WQE) aims to automatically identify fine-grained error spans in machine-translated outputs and has found many uses, including assisting translators during post-editing. Modern WQE techniques are often expensive, involving prompting of large language models or ad-hoc training on large amounts of human-labeled data. In this work, we investigate efficient alternatives exploiting recent advances in language model interpretability and uncertainty quantification to identify translation errors from the inner workings of translation models. In our evaluation spanning 14 metrics across 12 translation directions, we quantify the impact of human label variation on metric performance by using multiple sets of human labels. Our results highlight the untapped potential of unsupervised metrics, the shortcomings of supervised methods when faced with label uncertainty, and the brittleness of single-annotator evaluation practices.

Kaikai An, Fangkai Yang, Liqun Li, Junting Lu, Sitao Cheng, Shuzheng Si, Lu Wang, Pu Zhao, Lele Cao, Qingwei Lin 等

Recent advances in retrieval-augmented generation (RAG) have substantially improved question-answering systems, particularly for factoid ‘5Ws’ questions. However, significant challenges remain when addressing ‘1H’ questions, specifically how-to questions, which are integral for decision-making and require dynamic, step-by-step responses. The key limitation lies in the prevalent data organization paradigm, chunk, which commonly divides documents into fixed-size segments, and disrupts the logical coherence and connections within the context. To address this, we propose THREAD, a novel data organization paradigm enabling systems to handle how-to questions more effectively. Specifically, we introduce a new knowledge granularity, ‘logic unit’ (LU), where large language models transform documents into more structured and loosely interconnected LUs. Extensive experiments across both open-domain and industrial settings show that THREAD outperforms existing paradigms significantly, improving the success rate of handling how-to questions by 21% to 33%. Additionally, THREAD demonstrates high adaptability across diverse document formats, reducing retrieval information by up to 75% compared to chunk, and also shows better generalizability to ‘5Ws’ questions, such as multi-hop questions, outperforming other paradigms.

Ming Li, Nan Zhang, Chenrui Fan, Hong Jiao, Yanbin Fu, Sydney Peters, Qingshu Xu, Robert Lissitz, Tianyi Zhou

While Large Reasoning Models (LRMs) generate extensive chain-of-thought reasoning, we lack a principled framework for understanding how these thoughts are structured. In this paper, we introduce a novel approach by applying Schoenfeld’s Episode Theory, a classic cognitive framework for human mathematical problem-solving, to analyze the reasoning traces of LRMs. We annotated thousands of sentences and paragraphs from model-generated solutions to math problems using seven cognitive labels (e.g., Plan, Implement, Verify). The result is the first publicly available benchmark for the fine-grained analysis of machine reasoning, including a large annotated corpus and detailed annotation guidebooks. Our preliminary analysis reveals distinct patterns in LRM reasoning, such as the transition dynamics between cognitive states. This framework provides a theoretically grounded methodology for interpreting LRM cognition and enables future work on more controllable and transparent reasoning systems.

Cheryl Lee, Chunqiu Steven Xia, Longji Yang, Jen-tse Huang, Zhouruixing Zhu, Lingming Zhang, Michael R. Lyu

Software debugging is a time-consuming endeavor involving a series of steps, such as fault localization and patch generation, each requiring thorough analysis and a deep understanding of the underlying logic. While large language models (LLMs) demonstrate promising potential in coding tasks, their performance in debugging remains limited. Current LLM-based methods often focus on isolated steps and struggle with complex bugs. In this paper, we propose the first end-to-end framework, UniDebugger, for unified debugging through multi-agent synergy. It mimics the entire cognitive processes of developers, with each agent specialized as a particular component of this process rather than mirroring the actions of an independent expert as in previous multi-agent systems. Agents are coordinated through a three-level design, following a cognitive model of debugging, allowing adaptive handling of bugs with varying complexities. Experiments on extensive benchmarks demonstrate that UniDebugger significantly outperforms state-of-the-art repair methods, fixing 1.25x to 2.56x bugs on the repo-level benchmark, Defects4J. This performance is achieved without requiring ground-truth root-cause code statements, unlike the baselines. Our source code is available on an anonymous link: https://github.com/BEbillionaireUSD/UniDebugger.

Jiwen Zhang, Ya-Qi Yu, Minghui Liao, WenTao Li, Jihao Wu, Zhongyu Wei

Graphical User Interface (GUI) agents are expected to precisely operate on the screens of digital devices. Existing GUI agents merely depend on current visual observations and plain-text action history, ignoring the significance of history screens. To mitigate this issue, we propose **UI-Hawk**, a multi-modal GUI agent specially designed to process screen streams encountered during GUI navigation. UI-Hawk incorporates a history-aware visual encoder to handle the screen sequences. To acquire a better understanding of screen streams, we select four fundamental tasks—UI grounding, UI referring, screen question answering, and screen summarization. We further propose a curriculum learning strategy to subsequently guide the model from fundamental tasks to advanced screen-stream comprehension.Along with the efforts above, we have also created a benchmark FunUI to quantitatively evaluate the fundamental screen understanding ability of MLLMs. Extensive experiments on FunUI and GUI navigation benchmarks consistently validate that screen stream understanding is essential for GUI tasks.Our code and data are now available at https://github.com/IMNearth/UIHawk.

Eugene Jang, Kimin Lee, Jin-Woo Chung, Keuntae Park, Seungwon Shin

Tokenization is a crucial step that bridges human-readable text with model-readable discrete tokens. However, recent studies have revealed that tokenizers can be exploited to elicit unwanted model behaviors. In this work, we investigate incomplete tokens, i.e., undecodable tokens with stray bytes resulting from byte-level byte-pair encoding (BPE) tokenization. We hypothesize that such tokens are heavily reliant on their adjacent tokens and are fragile when paired with unfamiliar tokens. To demonstrate this vulnerability, we introduce improbable bigrams: out-of-distribution combinations of incomplete tokens designed to exploit their dependency. Our experiments show that improbable bigrams are significantly prone to hallucinatory behaviors. Surprisingly, the same phrases have drastically lower rates of hallucination (90% reduction in Llama3.1) when an alternative tokenization is used. We caution against the potential vulnerabilities introduced by byte-level BPE tokenizers, which may introduce blind spots to language models.

Mingxuan Xia, Zhijie Jiang, Haobo Wang, Junbo Zhao, Tianlei Hu, Gang Chen

Hierarchical text classification aims to classify documents into multiple labels within a hierarchical taxonomy, making it an essential yet challenging task in natural language processing. Recently, using Large Language Models (LLM) to tackle hierarchical text classification in a zero-shot manner has attracted increasing attention due to their cost-efficiency and flexibility. Given the challenges of understanding the hierarchy, various HTC prompting strategies have been explored to elicit the best performance from LLMs.However, our empirical study reveals that LLMs are highly sensitive to these prompting strategies—(i) within a task, different strategies yield substantially different results, and (ii) across various tasks, the relative effectiveness of a given strategy varies significantly. To address this, we propose a novel ensemble method, HiEPS, which integrates the results of diverse prompting strategies to promote LLMs’ reliability. We also introduce a path-valid voting mechanism for ensembling, which selects a valid result with the highest path frequency score. Extensive experiments on three benchmark datasets show that HiEPS boosts the performance of single prompting strategies and achieves SOTA results. The source code is available at https://github.com/MingxuanXia/HiEPS.

Jiaxin Liu, Yixuan Tang, Yi Yang, Kar Yan Tam

Large Language Models (LLMs) are increasingly used in decision-making scenarios that involve risk assessment, yet their alignment with human economic rationality remains unclear. In this study, we investigate whether LLMs exhibit risk preferences consistent with human expectations across different personas. Specifically, we propose an evaluation metric called Risk Disparity Score (RDS) and assess whether LLM-generated responses reflect appropriate levels of risk aversion or risk-seeking behavior based on individual’s persona. Our results reveal that while LLMs make reasonable decisions in simplified, personalized risk contexts, their performance declines in more complex economic decision-making tasks. To address this, we test whether current state-of-art alignment methods such as Direct Preference Optimization(DPO) and In Context Learning(ICL) can enhance LLM adherence to persona-specific risk preferences. We find DPO can improve the economic rationality of LLMs in loss-related parameters, offering a step toward more human-aligned AI decision-making.

Yunkai Dang, Mengxi Gao, Yibo Yan, Xin Zou, Yanggan Gu, Jungang Li, Jingyu Wang, Peijie Jiang, Aiwei Liu, Jia Liu 等

Multimodal large language models (MLLMs) have recently achieved state-of-the-art performance on tasks ranging from visual question answering to video understanding. However, existing studies have concentrated mainly on visual–textual misalignment, leaving largely unexplored the MLLMs’ ability to preserve an originally correct answer when confronted with misleading information. We reveal a response uncertainty phenomenon: across nine standard datasets, twelve state-of-the-art open-source MLLMs overturn a previously correct answer in 65% of cases after receiving a single deceptive cue. To systematically quantify this vulnerability, we propose a two-stage evaluation pipeline: (1) elicit each model’s original response on unperturbed inputs; (2) inject explicit (false-answer hints) and implicit (contextual contradictions) misleading instructions, and compute the misleading rate—the fraction of correct-to-incorrect flips. Leveraging the most susceptible examples, we curate the Multimodal Uncertainty Benchmark (MUB), a collection of image–question pairs stratified into low, medium, and high difficulty based on how many of twelve state-of-the-art MLLMs they mislead. Extensive evaluation on twelve open-source and five closed-source models reveals a high uncertainty: average misleading rates exceed 86%, with explicit cues over 67.19% and implicit cues over 80.67%. To reduce the misleading rate, we then fine-tune all open-source MLLMs on a compact 2,000-sample mixed-instruction dataset, reducing misleading rates to 6.97% (explicit) and 32.77% (implicit), boosting consistency by nearly 29.37% on highly deceptive inputs, and slightly improving accuracy on standard benchmarks.

Hua Cai, Shuang Zhao, Liang Zhang, Xuli Shen, Qing Xu, Weilin Shen, Zihao Wen, Tianke Ban

Reasoning-focused large language models (LLMs) are rapidly evolving across various domains, yet their capabilities in handling complex legal problems remains underexplored. In this paper, we introduce Unilaw-R1, a large language model tailored for legal reasoning. With a lightweight 7-billion parameter scale, Unilaw-R1 significantly reduces deployment cost while effectively tackling three core challenges in the legal domain: insufficient legal knowledge, unreliable reasoning logic, and weak business generalization. To address these issues, we first construct Unilaw-R1-Data, a high-quality dataset containing ~17K distilled and screened chain-of-thought (CoT) samples. Based on this, we adopt a two-stage training strategy combining Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), which significantly boosts the model’s performance on complex legal reasoning tasks and supports interpretable decision-making in legal AI applications. To assess legal reasoning ability, we also introduce Unilaw-R1-Eval, a dedicated benchmark designed to evaluate models across single- and multi-choice legal tasks. Unilaw-R1 demonstrates strong results on authoritative benchmarks, outperforming all models of similar scale and achieving performance on par with the much larger DeepSeek-R1-Distill-Qwen-32B (54.9%). Following domain-specific training, it also showed significant gains on LawBench and LexEval, exceeding Qwen-2.5-7B-Instruct (46.6%) by an average margin of 6.6%. Code is available at: https://github.com/Hanscal/Unilaw-R1.