The application of visual-to-music generation (VTM) is rapidly growing. However, current VTM methods struggle with capturing the relationship between visuals and music in open-domain settings, mainly due to two challenges: the lack of large-scale, high-quality visual-music paired datasets and the absence of direct semantic correspondence between visuals and music. In this work, we propose CoT-VTM, a framework that distills Chain-of-Thought (CoT) reasoning to enable visual-to-music generation without paired data, while efficiently producing music aligned with visual content in open-domain settings. We first bridge the gap between visual, music, and text data using appropriate foundation models. Next, we identify key elements of the visual-music relationship and design a CoT prompt for visual-to-music mapping. To fully distill the reasoning of CoT, we incorporate latent information from intermediate reasoning steps as supervisory signals alongside visual and music supervision. Finally, we design a two-stage mapping distillation training process: the first stage uses discriminative MLP modules, while the second uses a generative embedding diffusion model (EDM). Our model achieves optimal performance on both image-to-music and video-to-music tasks. Project page: https://xxkkxxx.github.io/cot-vtm/
论文检索
输入标题、作者或关键词，从 3,086 篇学术成果中精准定位
Real-world instructions with multiple constraints pose a significant challenge to existing large language models (LLMs). An observation is that the LLMs exhibit dramatic performance fluctuation when disturbing the order of the incorporated constraints. Yet, none of the existing works has systematically investigated this position bias problem in the field of multi-constraint instruction following. To bridge this gap, we design a probing task where we quantitatively measure the difficulty distribution of the constraints by a novel Difficulty Distribution Index (CDDI). Through the experimental results, we find that LLMs are more performant when presented with the constraints in a “hard-to-easy” order. This preference can be generalized to LLMs with different architecture or different sizes of parameters. Additionally, we conduct an explanation study, providing an intuitive insight into the correlation between the LLM’s attention and constraint orders. Our code and dataset are publicly available at https://github.com/meowpass/PBIF.
TransBench: Breaking Barriers for Transferable Graphical User Interface Agents in Dynamic Digital Environments
PDF ↗Graphical User Interface (GUI) agents, which autonomously operate on digital interfaces through natural language instructions, hold transformative potential for accessibility, automation, and user experience. A critical aspect of their functionality is grounding — the ability to map linguistic intents to visual and structural interface elements. However, existing GUI agents often struggle to adapt to the dynamic and interconnected nature of real-world digital environments, where tasks frequently span multiple platforms and applications while also being impacted by version updates. To address this, we introduce TransBench, the first benchmark designed to systematically evaluate and enhance the transferability of GUI agents across three key dimensions: cross-version transferability (adapting to version updates), cross-platform transferability (generalizing across platforms like iOS, Android, and Web), and cross-application transferability (handling tasks spanning functionally distinct apps). TransBench includes 15 app categories with diverse functionalities, capturing essential pages across versions and platforms to enable robust evaluation. Our experiments demonstrate significant improvements in grounding accuracy, showcasing the practical utility of GUI agents in dynamic, real-world environments. Our code and data will be publicly available at GitHub.
A^2ATS: Retrieval-Based KV Cache Reduction via Windowed Rotary Position Embedding and Query-Aware Vector Quantization
PDF ↗Long context large language models (LLMs) pose significant challenges for efficient serving due to the large memory footprint and high access overhead of KV cache.Retrieval-based KV cache reduction methods can mitigate these challenges, typically by offloading the complete KV cache to CPU and retrieving necessary tokens on demand during inference.However, these methods still suffer from unsatisfactory accuracy degradation and extra retrieval overhead.To address these limitations, this paper proposes A^2ATS, a novel retrieval-based KV cache reduction method.A^2ATS aims to obtain an accurate approximation of attention scores by applying the vector quantization technique to key states, thereby enabling efficient and precise retrieval of the top-K tokens.First, we propose Windowed Rotary Position Embedding, which decouples the positional dependency from query and key states after position embedding.Then, we propose query-aware vector quantization that optimizes the objective of attention score approximation directly.Finally, we design the heterogeneous inference architecture for KV cache offloading, enabling long context serving with larger batch sizes.Experimental results demonstrate that A^2ATS can achieve a lower performance degradation with similar or lower overhead compared to existing methods, thereby increasing long context serving throughput by up to 2.7 \times.
Recent advancements in AI-generated content (AIGC) have heightened concerns about harmful outputs, such as misinformation and malicious misuse.Existing detection methods face two key limitations:(1) lacking real-world AIGC scenarios and corresponding risk datasets, and(2) both traditional and multimodal large language models (MLLMs) struggle to detect risks in AIGC.Towards this end, we introduce **AIGuard**, the first benchmark for AIGC risk detection in real-world e-commerce. It includes 253,420 image-text pairs (i.e., the risk content and risk description) across four critical categories: *abnormal body*, *violating physical laws*, *misleading or illogical context*, and *harmful or problematic message*.To effectively detect these risks, we propose distilling text annotations into dense soft prompts and identifying risk content through image soft prompt matching during inference.Experiments on the benchmark show that this method achieves a 9.68% higher recall than leading multimodal models while using only 25% of the training resources and improving inference speed by 37.8 times.For further research, our benchmark and code are available at [https://github.com/wenh-zhang/aiguard-dataset](https://github.com/wenh-zhang/aiguard-dataset).
Training Turn-by-Turn Verifiers for Dialogue Tutoring Agents: The Curious Case of LLMs as Your Coding Tutors
PDF ↗Intelligent tutoring agents powered by large language models (LLMs) have been increasingly explored to deliver personalized knowledge in areas such as language learning and science education. However, their capabilities in guiding users to solve complex real-world tasks remain underexplored. To address this limitation, in this work, we focus on coding tutoring, a challenging problem that requires tutors to proactively guide students towards completing predefined coding tasks. We propose a novel agent workflow, Trace-and-Verify (TRAVER), which combines knowledge tracing to estimate a student’s knowledge state and turn-by-turn verification to ensure effective guidance toward task completion. We introduce DICT, an automatic evaluation protocol that assesses tutor agents using controlled student simulation and code generation tests. Extensive experiments reveal the challenges of coding tutoring and demonstrate that TRAVER achieves a significantly higher success rate. Although we use code tutoring as an example in this paper, our approach can be extended beyond coding, providing valuable insights into advancing tutoring agents for human task learning.
Large Language Models (LLMs) have shown promise in structured prediction tasks, including regression, but existing approaches primarily focus on point estimates and lack systematic comparison across different methods.We investigate probabilistic regression using LLMs for unstructured inputs, addressing challenging text-to-distribution prediction tasks such as price estimation where both nuanced text understanding and uncertainty quantification are critical.We propose a novel quantile regression approach that enables LLMs to produce full predictive distributions, improving upon traditional point estimates. Through extensive experiments across three diverse price prediction datasets, we demonstrate that a Mistral-7B model fine-tuned with quantile heads significantly outperforms traditional approaches for both point and distributional estimations, as measured by three established metrics each for prediction accuracy and distributional calibration.Our systematic comparison of LLM approaches, model architectures, training approaches, and data scaling reveals that Mistral-7B consistently outperforms encoder architectures, embedding-based methods, and few-shot learning methods.Our experiments also reveal the effectiveness of LLM-assisted label correction in achieving human-level accuracy without systematic bias. Our curated datasets are made available at https://github.com/vnik18/llm-price-quantile-reg/ to support future research.
Synthetic data generation has emerged as a promising approach to enhance the reasoning capabilities of large language models. However, existing methods remain hindered by high costs—either through expensive API access or additional intermediate training—and are limited in their ability to generalize across different domains. To address these challenges, we propose a multi-agent debate framework based on the Socratic questioning strategy, abbreviated as SoDa. Distinguished from previous methods that prioritize data quantity, we highlight the wisdom of Socratic questioning in augmenting reasoning quality by deepening the thinking process to encourage exploration and broadening it to motivate self-reflection on each question. Combined with our efficient production pipeline, SoDa enables scaling while maintaining affordable costs. We use SoDa to generate diverse datasets for mathematics and code generation tasks with the Qwen2.5-7B-Instruct model, successfully fine-tuning a range of foundation models, from general-purpose ones to OpenAI o1-like ones. For mathematics, the experimental results show that SoDa outperforms the performance of existing datasets at the same scale, achieving improvements ranging from 1.3% to 13.5%. Remarkably, SoDa with 30K examples even surpasses the ScaleQuest dataset with 1000K samples, demonstrating significant efficiency. Our findings highlight the potential of SoDa as a universal, scalable, and cost-effective method for enhancing reasoning capabilities in large models across domains.
Direct Preference Optimization (DPO) is broadly utilized for aligning Large Language Models (LLMs) with human values because of its flexibility. Despite its effectiveness, it has been observed that the capability of DPO to generate human-preferred response is limited and the results of DPO are far from resilient. To address these limitations, in this paper we propose a novel Self-Guided Direct Preference Optimization algorithm, i.e., SGDPO, which incorporates a pilot term to steer the gradient flow during the optimization process, allowing for fine-grained control over the updates of chosen and rejected rewards. We provide a detailed theoretical analysis of our proposed method and elucidate its operational mechanism. Furthermore, we conduct comprehensive experiments on various models and benchmarks. The extensive experimental results demonstrate the consistency between the empirical results and our theoretical analysis and confirm the effectiveness of our proposed approach (up to 9.19% higher score).
A Conformal Risk Control Framework for Granular Word Assessment and Uncertainty Calibration of CLIPScore Quality Estimates
PDF ↗This study explores current limitations of learned image captioning evaluation metrics, specifically the lack of granular assessments for errors within captions, and the reliance on single-point quality estimates without considering uncertainty. To address the limitations, we propose a simple yet effective strategy for generating and calibrating distributions of CLIPScore values. Leveraging a model-agnostic conformal risk control framework, we calibrate CLIPScore values for task-specific control variables, tackling the aforementioned limitations. Experimental results demonstrate that using conformal risk control, over score distributions produced with simple methods such as input masking, can achieve competitive performance compared to more complex approaches. Our method effectively detects erroneous words, while providing formal guarantees aligned with desired risk levels. It also improves the correlation between uncertainty estimations and prediction errors, thus enhancing the overall reliability of caption evaluation metrics.
TRANS-ZERO: Self-Play Incentivizes Large Language Models for Multilingual Translation Without Parallel Data
PDF ↗The rise of Large Language Models (LLMs) has reshaped machine translation (MT), but multilingual MT still relies heavily on parallel data for supervised fine-tuning (SFT), facing challenges like data scarcity for low-resource languages and catastrophic forgetting. To address these issues, we propose TRANS-ZERO, a self-play framework that leverages only monolingual data and the intrinsic multilingual knowledge of LLM. TRANS-ZERO combines Genetic Monte-Carlo Tree Search (G-MCTS) with preference optimization, achieving strong translation performance that rivals supervised methods. Experiments demonstrate that this approach not only matches the performance of models trained on large-scale parallel data but also excels in non-English translation directions. Further analysis reveals that G-MCTS itself significantly enhances translation quality by exploring semantically consistent candidates through iterative translations, providing a robust foundation for the framework’s success.
With the rapid emergence of novel capabilities in Large Language Models (LLMs), the need for rigorous multilingual and multiculturalbenchmarks that are integrated has become more pronounced. Though existing LLM benchmarks are capable of evaluating specificcapabilities of LLMs in English as well as in various mid- to low-resource languages, including those in the Southeast Asian (SEA)region, a comprehensive and culturally representative evaluation suite for the SEA languages has not been developed thus far.Here, we present SEA-HELM, a holistic linguistic and cultural LLM evaluation suite that emphasises SEA languages, comprisingfive core pillars: (1) NLP CLASSICS, (2) LLM-SPECIFICS, (3) SEA LINGUISTICS, (4) SEA CULTURE, (5) SAFETY. SEA-HELMcurrently supports Filipino, Indonesian, Tamil, Thai, and Vietnamese. We also introduce the SEA-HELM leaderboard, which allows users to understand models’ multilingual and multicultural performance in a systematic and user-friendly manner. We make the SEA-HELM evaluation code publicly available.
We make the case for language models over logical forms (LFLMs), arguing that such models are more data-efficient than their textual counterparts. To that end, we introduce the \underline{G}\textit{raph-based }\underline{Fo}\textit{rmal-}\underline{L}\textit{ogical }\underline{D}\textit{istributional }\underline{S}\textit{emantics} (GFoLDS) prototype, a pretrained LM over graph representations of logical forms, as a proof-of-concept of LFLMs. Using GFoLDS, we present strong experimental evidence that LFLMs can leverage the built-in, basic linguistic knowledge inherent in such models to immediately begin learning more complex patterns. On downstream tasks, we show that GFoLDS vastly outperforms textual, transformer LMs (BERT) pretrained on the same data, indicating that LFLMs can learn with substantially less data than models over plain text. Furthermore, we show that the performance of this model is likely to scale with additional parameters and pretraining data, suggesting the viability of LFLMs in real-world applications.
As large language models (LLM) become more and more capable in languages other than English, it is important to collect benchmark datasets in order to evaluate their multilingual performance, including on tasks like machine translation (MT). In this work, we extend the WMT24 dataset to cover 55 languages by collecting new human-written references and post-edits for 46 new languages/dialects in addition to post-edits of the references in 8 out of 9 languages in the original WMT24 dataset. We benchmark a variety of MT providers and LLMs on the collected dataset using automatic metrics and find that LLMs are the best-performing MT systems in all 55 languages. However, we caution against using our results to reach strong conclusions about MT quality without a human-based evaluation due to limitations of automatic evaluation metrics, which we leave for future work.
Discourse relations contribute to the structure of a text and can optionally be realized through explicit connectives such as “but” and “while”. But when are these connectives necessary to avoid possible misunderstandings? We investigate this question by first building a corpus of 4,274 text revisions in each of which a connective was explicitly inserted. For a subset of 250 cases, we collect plausibility annotations on other connectives to check whether they would represent suitable alternative relations. The results of this annotation show that several relations are often perceived as plausible in our data. Furthermore, we analyze the extent to which large language models can identify instances with multiple plausible relations as a possible source of misunderstandings. We find that the models predict plausibility of individual connectives with up to 66% accuracy, but they are not reliable in estimating when multiple relations are plausible.
Large Language Models (LLMs) have achieved impressive results across numerous NLP tasks, and fine-tuning them for Machine Translation (MT) has improved their performance. However, vanilla fine-tuning often leads to catastrophic forgetting, compromising the broad general abilities of LLMs and introducing potential security risks. These abilities, which are developed using proprietary and unavailable training data, make simple data replay methods ineffective. To overcome this issue, we propose a novel approach called **Ra**tionale **Dis**tillation. RaDis harnesses the strong generative capabilities of LLMs to create rationales for training data, which are then “replayed” to prevent forgetting. These rationales connect prior knowledge with new tasks, acting as self-distillation targets to regulate the training process. By jointly training on reference translations and self-generated rationales, the model can learn new translation skills while preserving its general abilities across other tasks. Additionally, RaDis provides a fresh perspective on using rationales in the CL field and has the potential to serve as a general continual learning method for a variety of tasks.
We introduce *Slam*, a recipe for training high-quality Speech Language Models (SLMs) on a single academic GPU in 24 hours. We do so through empirical analysis of model initialisation and architecture, synthetic training data, preference optimisation with synthetic data and tweaking all other components. We empirically demonstrate that this training recipe also scales well with more compute getting results on par with leading SLMs in a fraction of the compute cost. We hope these insights will make SLM training and research more accessible. In the context of SLM scaling laws, our results far outperform predicted compute optimal performance, giving an optimistic view to SLM feasibility. See code, data, models, samples - https://pages.cs.huji.ac.il/adiyoss-lab/slamming .
Forget the Token and Pixel: Rethinking Gradient Ascent for Concept Unlearning in Multimodal Generative Models
PDF ↗Gradient Ascent (GA) has emerged as a promising approach for concept unlearning in Multimodal Generative Models (MGMs), such as Multimodal Large Language Models (MLLMs) and Stable Diffusion Models (SDMs). Despite its effectiveness in removing undesired knowledge, GA leads to severe utility degradation in MGMs. In this paper, we explore the mechanism behind this degradation by quantifying two distinct forms of knowledge in MGMs: (i) Conceptual Knowledge, which represents specific information about concepts; (ii) Natural Knowledge, which refers to the ability to produce coherent and logically structured outputs. Our analysis reveals that applying GA globally not only removes the targeted Conceptual Knowledge but also inadvertently diminishes Natural Knowledge, resulting in utility collapse. To address this issue, we propose Forget the Token and Pixel (FTTP), a novel approach that selectively applies GA to targeted Conceptual Knowledge while preserving Natural Knowledge through Gradient Descent (GD). FTTP eliminates the need for additional retain sets and a large number of training steps, thereby reducing computational resource costs. Extensive experiments demonstrate FTTP’s efficiency and superior utility-unlearning tradeoff for both text and image generation tasks. Our source code will be released in the near future.
Large language models (LLMs) have advanced significantly due to the attention mechanism, but their quadratic complexity and linear memory demands limit their performance on long-context tasks. Recently, researchers introduced Mamba, an advanced model built upon State Space Models (SSMs) that offers linear complexity and constant memory. Although Mamba is reported to match or surpass the performance of attention-based models, our analysis reveals a performance gap: Mamba excels in tasks that involve localized key information but faces challenges with tasks that require handling distributed key information. Our controlled experiments suggest that the inconsistency arises from Mamba’s reliance on **local pattern shortcuts** across model scales (10M to 1.4B), which enable Mamba to remember local key information within its limited memory but hinder its ability to retain more dispersed information. Therefore, we introduce a global gate module into the Mamba model to address this issue. Experiments on extensive synthetic tasks, as well as real-world tasks, demonstrate the effectiveness of our method. Notably, with the introduction of only 4M extra parameters, our approach enables the Mamba model (130M) to achieve a significant improvement on tasks with distributed information, increasing its performance from **below 5% to 80%**.
Precise recognition of search intent in Retrieval-Augmented Generation (RAG) systems remains a challenging goal, especially under resource constraints and for complex queries with nested structures and dependencies. This paper presents **QCompiler**, a neuro-symbolic framework inspired by linguistic grammar rules and compiler design, to bridge this gap. It theoretically presents a minimal yet sufficient Backus-Naur Form (BNF) grammar G[q] to formalize complex queries. Unlike previous methods, this grammar maintains completeness while minimizing redundancy. Based on this, QCompiler includes a query expression translator, a Lexical syntax parser, and a Recursive Descent Processor to compile queries into Abstract Syntax Trees (ASTs) for execution. The atomicity of the sub-queries in the leaf nodes ensures more precise document retrieval and response generation, significantly improving the RAG system’s ability to address complex queries.