论文检索

Applications · Language, Speech and Dialog

Siyuan Wang, Yanchen Liu, Xiang Ren

Large Reasoning Models (LRMs) achieve strong reasoning performance by generating long chains of thought (CoTs), yet only a small fraction of these traces meaningfully contributes to answer prediction, while the majority contains repetitive or truncated content. Such output redundancy is further propagated after supervised finetuning (SFT), as models learn to imitate verbose but uninformative patterns, which can degrade performance. To this end, we incorporate integrated gradient attribution to quantify each token's influence on final answers and aggregate them into two segment-level metrics: (1) \textit{attribution strength} measures the overall attribution magnitude; and (2) \textit{direction consistency} captures whether tokens' attributions within a segment are uniformly positive or negative (high consistency), or a mixture of both (moderate consistency). Based on these two metrics, we propose a segment-level selective learning framework to identify important segments with high attribution strength but moderate consistency that indicate reflective rather than shallow reasoning. The framework then applies selective SFT on these important segments while masking loss for unimportant ones. Experiments across multiple models and datasets show that our approach improves accuracy and output efficiency, enabling more effective learning from long reasoning traces.

Applications · Language, Speech and Dialog

Ran Xu, Jingjing Chen, Jiayu Ye, Yu Wu, Jun Yan, Carl Yang, Hongkun Yu

Large Language Models (LLMs) are widely used as judges to evaluate response quality, providing a scalable alternative to human evaluation. However, most LLM judges operate solely on intrinsic text-based reasoning, limiting their ability to verify complex constraints or perform accurate computation. Motivated by the success of tool-integrated reasoning (TIR) in numerous tasks, we propose TIR-Judge, an end-to-end RL framework for training LLM judges that integrates a Python executor for precise evaluation. TIR-Judge is built on three principles: (i) diverse training across verifiable and non-verifiable domains, (ii) flexible judgment formats (pointwise, pairwise, listwise), and (iii) iterative RL that enables bootstrapping directly from a base model without distillation. On six public benchmarks, TIR-Judge surpasses strong reasoning-based judges by up to 6.4% (pointwise) and 7.7% (pairwise), and achieves listwise performance comparable to Claude-Opus-4 despite having only 8B parameters. Remarkably, TIR-Judge-Zero—trained entirely without distillation—matches the performance of the distilled variants, showing that tool-augmented judges can self-improve through iterative reinforcement learning.

Applications · Language, Speech and Dialog

Rocky Klopfenstein, Yang He, Andrew Tremante, Yuepeng Wang, Nina Narodytska, Haoze Wu

Community-driven Text-to-SQL evaluation platforms play a pivotal role in tracking the state of the art of Text-to-SQL performance. The reliability of the evaluation process is critical for driving progress in the field. Current evaluation methods are largely test-based, which involves comparing the execution results of a generated SQL query and a human-labeled ground-truth on a static test database. Such an evaluation is optimistic, as two queries can coincidentally produce the same output on the test database while actually being different. In this work, we propose a new alternative evaluation pipeline, called SpotIt, where a formal bounded equivalence verification engine actively searches for a database that differentiates the generated and ground-truth SQL queries. We develop techniques to extend existing verifiers to support a richer SQL subset relevant to Text-to-SQL. A performance evaluation of ten Text-to-SQL methods on the high-profile BIRD dataset suggests that test-based methods can often overlook differences between the generated query and the ground-truth. Further analysis of the verification results reveals a more complex picture of the current Text-to-SQL evaluation.

Applications · Language, Speech and Dialog

Seungbin Yang, Jihwan Kim, Jaemin Choi, Dongjin Kim, Soyoung Yang, ChaeHun Park, Jaegul Choo

Web information extraction (WIE) is the task of automatically extracting data from web pages, offering high utility for various applications. The evaluation of WIE systems has traditionally relied on benchmarks built from HTML snapshots captured at a single point in time. However, this offline evaluation paradigm fails to account for the temporally evolving nature of the web; consequently, performance on these static benchmarks often fails to generalize to dynamic real-world scenarios. To bridge this gap, we introduce LiveWeb-IE, a new benchmark designed for evaluating WIE systems directly against live websites. Based on trusted and permission-granted websites, we curate natural language queries that require information extraction of various data categories, such as text, images, and hyperlinks. We further design these queries to represent four levels of complexity, based on the number and cardinality of attributes to be extracted, enabling a granular assessment of WIE systems. In addition, we propose Visual Grounding Scraper (VGS), a novel multi-stage agentic framework that mimics human cognitive processes by visually narrowing down web page content to extract desired information. Extensive experiments across diverse backbone models demonstrate the effectiveness and robustness of VGS. We believe that this study lays the foundation for developing practical and robust WIE systems.

Applications · Language, Speech and Dialog

Subrata Biswas, Mohammad Khan, Bashima Islam

Spatial reasoning is fundamental to auditory perception, yet current audio large language models (ALLMs) largely rely on unstructured binaural cues and single- step inference. This limits both perceptual accuracy in direction and distance estimation and the capacity for interpretable reasoning. Recent work such as BAT demonstrates spatial QA with binaural audio, but its reliance on coarse categorical labels (left, right, up, down) and the absence of explicit geometric supervision constrain resolution and robustness. We introduce the $\textbf{Spatial-Acoustic Geometry Encoder (SAGE}$), a geometry-aware audio encoder that aligns binaural acoustic features with 3D spatial structure using panoramic depth images and room-impulse responses at training time, while requiring only audio at inference. Building on this representation, we present $\textbf{OWL}$, an ALLM that integrates $\textbf{SAGE}$ with a spatially grounded chain-of-thought to rationalize over direction-of-arrivals (DoA) and distance estimates. Through curriculum learning from perceptual QA to multi-step reasoning, $\textbf{OWL}$ supports o’clock-level azimuth and DoA estimation. To enable large-scale training and evaluation, we construct and release $\textbf{BiDepth}$, a dataset of over one million QA pairs combining binaural audio with panoramic depth images and room impulse responses across both in-room and out-of-room scenarios. Across two benchmark datasets, our new $\textbf{BiDepth}$ and the public SpatialSoundQA, $\textbf{OWL}$ reduces mean DoA error by $\textbf{11$^{\circ}$}$ through $\textbf{SAGE}$ and improves spatial reasoning QA accuracy by up to $\textbf{25}$% over BAT. Our dataset and code are available at: https://github.com/BASHLab/OWL

Applications · Language, Speech and Dialog

Thinh Pham, Nguyen Nguyen, Pratibha Zunjare, Weiyuan Chen, Yu-Min Tseng, Tu Vu

We introduce SealQA, a challenge benchmark for evaluating SEarch-Augmented Language models on fact-seeking questions where web search yields conflicting, noisy, or unhelpful results. SealQA comes in three flavors: (1) SEAL-0 (main) and (2) SEAL-HARD, both of which assess factual accuracy and reasoning capabilities, where SEAL-0 targets the most challenging questions that frontier non-reasoning models (e.g., GPT-4.1) answer with near-zero accuracy; and (3) LongSeal, which extends SealQA to test long-context, multi-document reasoning in "needle-in-a-haystack" settings. Our evaluation reveals critical limitations in current models. Even frontier reasoning models face significant challenges across SealQA flavors. On SEAL-0, GPT-5 with tools achieves only 43.2% accuracy at its best reasoning effort. We also find that even advanced reasoning models (e.g., DeepSeek-R1) can be vulnerable to noisy search results. Notably, increasing test-time compute does not yield reliable gains across GPT-5 and the o-series of models, with performance often plateauing or even declining early. Finally, while current models are less affected by the "lost-in-the-middle" issue, they still fail to reliably identify relevant documents in LongSeal when faced with numerous distractors. To facilitate future work, we release SealQA at https://huggingface.co/datasets/vtllms/sealqa.

Applications · Language, Speech and Dialog

Hyung-Seok Oh, Deok-Hyeon Cho, Seung-Bin Kim, Seong-Whan Lee

Neural vocoders have recently advanced waveform generation, yielding natural and expressive audio. Among these approaches, iSTFT-based vocoders have recently gained attention. They predict a complex-valued spectrogram and then synthesize the waveform via iSTFT, thereby avoiding learned upsampling stages that can increase computational cost. However, current approaches use real-valued networks that process the real and imaginary parts independently. This separation limits their ability to capture the inherent structure of complex spectrograms. We present ComVo, a Complex-valued neural Vocoder whose generator and discriminator use native complex arithmetic. This enables an adversarial training framework that provides structured feedback in complex-valued representations. To guide phase transformations in a structured manner, we introduce phase quantization, which discretizes phase values and regularizes the training process. Finally, we propose a block-matrix computation scheme to improve training efficiency by reducing redundant operations. Experiments demonstrate that ComVo achieves higher synthesis quality than comparable real-valued baselines, and that its block-matrix scheme reduces training time by 25%. Audio samples and code are available at https://hs-oh-prml.github.io/ComVo/.

Applications · Language, Speech and Dialog

Jakub Krajewski, Amitis Shidani, Dan Busbridge, Sam Wiseman, Jason Ramapuram

While scaling laws for Large Language Models (LLMs) traditionally focus on proxy metrics like pretraining loss, predicting downstream task performance has been considered unreliable. This paper challenges that view by proposing a direct framework to model the scaling of downstream accuracy from the training budget. We demonstrate that for a fixed token-to-parameter ratio, a simple two-parameter scaling law accurately describes this relationship. Our findings are validated by experiments on models with up to 17B parameters trained on up to 350B tokens, showing that downstream capabilities scaling can be described using a scaling law. Furthermore, we extend this framework to extrapolate and predict accuracy of target model with up to 6.7x larger training budget based on a set of smaller experiments. We release a complete list of model losses and downstream evaluation results at various different scales to support reproducibility and encourage future research.

Applications · Language, Speech and Dialog

Xiaoyang Cao, Zelai Xu, Mo Guang, Kaiwen Long, Michiel Bakker, Yu Wang, Chao Yu

Standard human preference-based alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), are a cornerstone technology for aligning Large Language Models (LLMs) with human values. However, these methods are all underpinned by a strong assumption that the collected preference data is clean and that all observed labels are equally reliable. In reality, large-scale preference datasets contain substantial label noise due to annotator errors, inconsistent instructions, varying expertise, and even adversarial or low-effort feedback. This creates a discrepancy between the recorded data and the ground-truth preferences, which can misguide the model and degrade its performance. To address this challenge, we introduce Robust Enhanced Policy Optimization (RE-PO). RE-PO employs an Expectation-Maximization algorithm to infer the posterior probability of each label’s correctness, which is used to adaptively re-weigh each data point in the training loss to mitigate noise. We further generalize this approach by linking a broad class of preference losses to induced probabilistic models. This enables systematic robustification of existing alignment algorithms while preserving exact probabilistic equivalence for likelihood-style losses. Theoretically, under perfect calibration and a population/full-batch setting, we show that RE-PO recovers the true annotator reliability. Our experiments demonstrate RE-PO’s effectiveness as a general framework, generally enhancing four state-of-the-art alignment algorithms (DPO, IPO, SimPO, and CPO) against their corresponding standard versions. When applied to Mistral and Llama 3 models, the RE-PO-enhanced methods improve AlpacaEval 2 win rates by up to 7.0% over their respective baselines.

Applications · Language, Speech and Dialog

Ye Yuan, Amin Shabani, Siqi Liu

Retrieval-Augmented Generation (RAG) systems rely on retrieving relevant evidence from a corpus to support downstream generation. The common practice of splitting a long document into multiple shorter passages enables finer-grained and targeted information retrieval. However, it also introduces challenges when a correct retrieval would require inference across passages, such as resolving coreference, disambiguating entities, and aggregating evidence scattered across multiple sources. Many state-of-the-art (SOTA) reranking methods, despite utilizing powerful large pretrained language models with potentially high inference costs, still neglect the aforementioned challenges. Therefore, we propose Embedding-Based Context-Aware Reranker (EBCAR), a lightweight reranking framework operating directly on embeddings of retrieved passages with enhanced cross-passage understandings through the structural information of the passages and a hybrid attention mechanism, which captures both high-level interactions across documents and low-level relationships within each document. We evaluate EBCAR against SOTA rerankers on the ConTEB benchmark, demonstrating its effectiveness for information retrieval requiring cross-passage inference and its advantages in both accuracy and efficiency. Our source code is available at https://github.com/BorealisAI/EBCAR.

Applications · Language, Speech and Dialog

cheng chen, Haiyan Yin, Ivor Tsang

Large Language Models (LLMs) are increasingly expected to be culturally customisable and personally aligned for natural language understanding (NLU). However, existing methods, from supervised fine-tuning (SFT) to personalised RLHF and prompting, either require costly large-scale annotations or remain constrained by their pretraining distributions. Moreover, acquiring annotations that reflect subjective, diverse, and evolving user preferences is both expensive and labour-intensive. To address these limitations, we propose \textit{\textbf{H}eterogeneous-\textbf{C}onsistency \textbf{C}o-Alignment} (HCC), a training-free annotation paradigm that leverages two heterogeneous models: a knowledge-rich yet potentially overconfident LLM and a task-specialised lightweight model guided by a small user preference set. Together, they verify and co-align misaligned outputs over unlabelled corpora. For verification, HCC introduces the reference-free \textit{\textbf{C}onsistent}-\textit{\textbf{A}nd}-\textit{\textbf{I}nconsistent} (\textbf{CAI}) Ratio, an uncertainty signal derived from inter-model agreements (consistent samples) and disagreements (inconsistent samples) to determine whether refinement is necessary. For co-alignment, HCC employs a non-parametric, embedding-based preference assignment scheme to recalibrate inconsistent samples according to user preferences. Across eight NLU datasets and both open- and closed-source LLMs, HCC consistently improves annotation alignment and, in several tasks, enables \textit{Llama-3-8B} to surpass \textit{GPT-3.5/4o-mini} after co-alignment correction. Moreover, CAI strongly correlates with accuracy and tracks pre- and post-alignment gains, offering a reference-free signal for scaling preference-aligned annotation without ground-truth supervision.

Applications · Language, Speech and Dialog

Hongzhan Lin, Zhiqi Bai, Xinmiao Zhang, Siran Yang, Jiamang Wang, Yunlong Xu, JIAHENG LIU, Yongchi Zhao, Xiang Li, Yuchi Xu 等

Transformer decoders have achieved strong results across tasks, but the memory required for the KV cache becomes prohibitive at long sequence lengths. Although Cross-layer KV Cache sharing (e.g., YOCO, CLA) offers a path to mitigate KV Cache bottleneck, it typically underperforms within-layer methods like GQA. To understand the root cause, we investigate the information flow of keys and values of the top-layers. Our preliminary reveals a clear distribution: values are predominantly derived from the bottom layer, while keys draw more information from both bottom and middle layers. Building upon this, we propose FusedKV, whose top-layer KV caches are a learnable fusion of the most informative ones from the bottom and middle layers. This fusion operates directly on post-RoPE keys, preserving relative positional information without the computational cost of re-applying rotary embeddings. To further improve efficiency, we propose FusedKV-Lite, an cross-layer sharing approach, where top-layer KV caches are directly derived from the bottom-layer values and the middle-layer keys. Compared to FusedKV, FusedKV-Lite reduces I/O overhead at the cost of a slight increase in perplexity. In experiments on LLMs ranging from 332M to 4B parameters, our proposed method reduce 50\% cache memory while achieving lower validation perplexity than the standard Transformer decoder, establishing it as a memory-efficient, high-performance architectural alternative. We have made our Triton implementation available.

Applications · Language, Speech and Dialog

Jasper Dekoninck, Ivo Petrov, Kristian Minchev, Miroslav Marinov, Maria Drencheva, Lyuba Konova, Milen Shumanov, Kaloyan Tsvetkov, Nikolay Drenchev, Lazar Todorov 等

In recent months, large language models (LLMs) have made significant progress in mathematical proof generation, but further advancement is hindered by the lack of a large-scale, high-quality dataset of human-evaluated proofs. While expensive to create, such a dataset is essential for driving improvements in training and addressing key open questions in the field of automated proof generation. Specifically, it remains unknown (1) how large the gap is between natural language and formal proof generation, (2) how final-answer accuracy relates to full proof correctness, and (3) how best-of-n selection strategies can affect proof quality. In this work, we present the Open Proof Corpus (OPC), a dataset comprising over 5,000 human-evaluated proofs produced by state-of-the-art LLMs. The OPC was specifically designed for broad applicability and downstream usage in proof generation research and is the first large dataset of LLM-generated solutions to problems from prestigious mathematics competitions such as the USAMO and IMO. Using the OPC, we address the open questions outlined above and provide new insights into LLMs' strengths and limitations in mathematical reasoning. Finally, to showcase the utility of the OPC, we finetune an 8B-parameter model on the dataset, obtaining a model that matches Gemini-2.5-Pro, and performs close to the best model, GPT-5, on evaluating proof correctness.

Applications · Language, Speech and Dialog

Ethan Hsu, Hong Meng Yam, Ines Bouissou, Aaron John, Raj Thota, Josh Koe, Vivek Putta, G Dharesan, Alexander Spangher, Shikhar Murty 等

Many real-world data science tasks involve complex web-based interactions: finding appropriate data available on the internet, synthesizing multimodal data from different locations, and producing summarized analyses. Existing web benchmarks often focus on simplistic interactions and often do not require diverse tool-using capabilities. Conversely, traditional data science benchmarks typically concentrate on static, highly structured datasets and do not assess end-to-end workflows that encompass data acquisition, cleaning, analysis, and insight generation. In response, we introduce WebDS, the first end-to-end web-based data science benchmark. It comprises 870 web-based data science tasks across 29 diverse websites from structured government data portals to unstructured news media, challenging agents to perform complex, multi-step, tool-based operations, across heterogeneous data formats, to better reflect the realities of modern data analytics. Evaluations of current SOTA LLM agents indicate significant performance gaps in accomplishing these tasks. For instance, Browser Use, which accomplishes 80% of tasks on Web Voyager, completes only 15% of tasks in WebDS, which our analysis suggests is due to new failure modes like poor information grounding, repetitive behavior and shortcut-taking that agents performing WebDS' tasks display. By contrast, humans achieve around 90% accuracy, highlighting a substantial gap between current agents and human performance. By providing a more robust and realistic testing ground, WebDS sets the stage for significant advances in the development of practically useful LLM-based data science.

Applications · Language, Speech and Dialog

Lingjie Jiang, Shaohan Huang, xun wu, Yixia Li, Guanhua Chen, Dongdong Zhang, Furu Wei

Multimodal large language models (MLLMs) have significantly advanced the integration of visual and textual understanding. However, their ability to generate code from multimodal inputs remains limited. In this work, we introduce VisCodex, a unified framework that seamlessly merges vision and coding language models to empower MLLMs with strong multimodal code generation abilities. Leveraging a task vector-based model merging technique, we integrate a state-of-the-art coding LLM into a strong vision-language backbone, while preserving both visual comprehension and advanced coding skills. To support training and evaluation, we introduce the Multimodal Coding Dataset (MCD), a large-scale and diverse collection of 598k samples, including high-quality HTML code, chart image-code pairs, image-augmented StackOverflow QA, and algorithmic problems. Furthermore, we propose InfiBench-V, a novel and challenging benchmark specifically designed to assess models on visually-rich, real-world programming questions that demand a nuanced understanding of both textual and visual contexts. Extensive experiments show that VisCodex achieves state-of-the-art performance among open-source MLLMs and approaches proprietary models like GPT-4o, highlighting the effectiveness of our model merging strategy and new datasets.

Applications · Language, Speech and Dialog

Yen-Ju Lu, Yashesh Gaur, Wei Zhou, Benjamin Muller, Jesus Villalba, Najim Dehak, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Srini Iyer 等

Auto-regressive speech–text models pre-trained on interleaved text tokens and discretized speech tokens demonstrate strong speech understanding and generation, yet remain substantially less compute-efficient than text LLMs, partly due to the much longer sequences of speech tokens relative to text. This modality imbalance disproportionately allocates pre-training and inference compute to speech, potentially hindering effective cross-modal alignment and slowing performance scaling by orders of magnitude. We introduce the Latent Speech-Text Transformer (LST), which aggregates speech tokens into latent speech patches that serve as higher-level autoregressive units. This design aligns the sequence-modeling granularity between speech and text while improving computational efficiency. The resulting patches can align with textual units to facilitate cross-modal knowledge transfer and compactly capture recurring acoustic patterns such as silence. Across story-completion benchmarks under both compute-controlled and data-controlled settings, LST consistently improves speech accuracy while also improving text performance, achieving up to +6.5% absolute gain on speech HellaSwag in compute-controlled training (+5.3% in data-controlled training). Under compute-controlled scaling from 420M to 1.8B parameters in a near compute-optimal regime, gains grow with scale, and improvements persist up to 7B parameters under fixed-token budgets. These benefits extend to downstream tasks: LST stabilizes ASR adaptation and reduces the effective autoregressive sequence length during ASR and TTS inference, lowering computational cost without degrading reconstruction quality. The Code is available at https://github.com/facebookresearch/lst.

Applications · Language, Speech and Dialog

Zhouqi Hua, Wenwei Zhang, Chengqi Lyu, Yuzhe Gu, Songyang Gao, Kuikun Liu, Dahua Lin, Kai Chen

Length generalization, the ability to solve problems of longer sequences than those observed during training, poses a core challenge of Transformer-based large language models (LLMs). Although existing studies have predominantly focused on data-driven approaches for particular arithmetic operations or symbolic manipulation tasks, these approaches tend to be task-specific with limited performance on individual tasks. To pursue a more general solution, this paper focuses on a broader case of reasoning problems that are *computable*, *i.e.*, problems that algorithms can solve, thus can be solved by the Turing machine, which operates over inputs of unbounded length. From this perspective, this paper proposes **T**uring m**A**chine **I**mitation **L**earning (**TAIL**) to improve the length generalization ability of LLMs. TAIL uses computer programs to directly synthesize chain-of-thought (CoT) data that imitate the execution process of a Turing machine, which *linearly* expands the reasoning steps into *atomic* states to alleviate shortcut pattern learning and explicit *memory* fetch mechanism to reduce the difficulties of dynamic and long-range data access. To validate the universality and reliability of TAIL, we construct a challenging synthetic dataset covering 8 classes of algorithms and 18 tasks. Without bells and whistles, TAIL significantly improves the length generalization ability as well as the performance of Qwen2.5-7B in individual tasks using only synthetic data, surpassing previous methods and DeepSeek-R1. The experimental results reveal that the key concepts in the Turing machine, instead of the human-like thinking styles, are indispensable for TAIL for length generalization, through which the model exhibits read-and-write behaviors consistent with the properties of the Turing machine in their attention layers. This work provides a promising direction for future research in the learning of LLM reasoning from synthetic data.

Applications · Language, Speech and Dialog

Kenny Olsen, Mads Østergaard, Karl Ulbæk, Søren Føns Nielsen, Rasmus Malik Høegh Lindrup, Bjørn Jensen, Morten Mørup

In recent years, deep learning-based single-channel speech separation has improved considerably, in large part driven by increasingly compute- and parameter-efficient neural network architectures. Most such architectures are, however, designed with a fixed compute and parameter budget and consequently cannot scale to varying compute demands or resources, which limits their use in embedded and heterogeneous devices such as mobile phones and hearables. To enable such use-cases we design a neural network architecture for speech separation and enhancement capable of early-exit, and we propose an uncertainty-aware probabilistic framework to jointly model the clean speech signal and error variance which we use to derive probabilistic early-exit conditions in terms of desired signal-to-noise ratios. We evaluate our methods on both speech separation and enhancement tasks where we demonstrate that early-exit capabilities can be introduced without compromising reconstruction, and that when trained on variable-length audio our early-exit conditions are well-calibrated and lead to considerable compute savings when used to dynamically scale compute at test time while remaining directly interpretable.

Applications · Language, Speech and Dialog

Tommy Mordo, Sagie Dekel, Tomer Kordonsky, Omer Madmon, Moshe Tennenholtz, Oren Kurland

With the rise of LLMs, publishers now operate in a dual world where traditional search and chat-like systems coexist. We propose a unified, game-theoretic view of this environment and highlight different tools, such as Multi-Agent Reinforcement Learning, that support the development of competitive content-optimization agents.

Applications · Language, Speech and Dialog

Shuoshuo Zhang, Zijian Li, Yizhen Zhang, Jingjing Fu, Lei Song, Jiang Bian, Jun Zhang, Yujiu Yang, Rui Wang

Structured images (e.g., charts and geometric diagrams) remain challenging for multimodal large language models (MLLMs), as perceptual slips can cascade into erroneous conclusions. Intermediate visual cues can steer reasoning; however, existing cue-based methods are constrained with low-fidelity image processing and linear, rigid reasoning patterns, limiting their effectiveness on complex structured-image tasks. In this paper, we propose PixelCraft, a novel multi-agent system for high-fidelity image processing and flexible visual reasoning on structured images. The system comprises a dispatcher, a planner, a reasoner, critics, and a set of visual tool agents. To achieve high-fidelity processing, we construct a high-quality corpus and fine-tune an MLLM into a grounding model, whose pixel-level localizations are integrated with traditional computer vision (CV) algorithms in tool agents. Building on this foundation, PixelCraft facilitates flexible visual reasoning through a dynamic three-stage workflow of tool selection, agent discussion, and self-criticism. Moreover, unlike prior linear reasoning patterns that simply append historical images, PixelCraft maintains an image memory to allow the planner to adaptively revisit earlier visual steps, explore alternative reasoning branches, and dynamically adjust the reasoning trajectory during discussion. Extensive experiments on challenging chart and geometry benchmarks demonstrate that PixelCraft significantly improves visual reasoning performance for advanced MLLMs, setting a new standard for structured image reasoning.