论文检索

Yifan Yang, Juntuo Wang, Yuming Qiao, Xudong Zhang, Chunyang Yu, Yan Li, Xiao Lin, Liang Luo, Dan Meng

With the advancement of Vision-Language Models (VLMs), employing VLM-as-a-Judge for visual evaluation has become a widely adopted metric in vision research. However, existing VLM-as-a-Judge approaches suffer from biased scoring outcomes with low discrimination and lack the capacity for unified multi-attribute compositional assessment. To address these limitations, we propose a novel training paradigm, termed JoPPO (**Jo**int **P**robabilistic **P**olicy **O**ptimization) that enables the VLMs to learn ranking under compositional assessment constraints. We evaluate the JoPPO on image aesthetics as a testbed, a task requiring nuanced understanding of multiple attributes including composition, lighting, color and geometry. Training follows two stages: (1) Supervised Fine-Tuning (SFT) on synthetic composition dataset provided by automated data generation pipeline to instill compositional priors; and (2) Contrastive Joint Conditional Probabilistic Reinforcement Learning: building upon the GRPO algorithm, we introduce JoPPO, which compute reward based on the expected win rate of total scores derived from the conditional distribution of fine-grained attribute scores within batches, effectively enhancing the model's discriminative ability in composite evaluation. Across standard aesthetic benchmarks, our method achieves consistent improvements in ranking consistency, demonstrating strong zero-shot generalization.

Liming Kuang, Yordanka Velikova, Mahdi Saleh, Jan-Nico Zaech, Danda Pani Paudel, Benjamin Busam

Object pose estimation is a fundamental task in computer vision and robotics, yet most methods require extensive, dataset-specific training. Concurrently, large-scale vision language models show remarkable zero-shot capabilities. In this work, we bridge these two worlds by introducing ConceptPose, a framework for object pose estimation that is both training-free and model-free. ConceptPose leverages a vision-language-model (VLM) to create open-vocabulary 3D concept maps, where each point is tagged with a concept vector derived from saliency maps. By establishing robust 3D-3D correspondences across concept maps, our approach allows precise estimation of 6DoF relative pose. Without any object or dataset-specific training, our approach achieves state-of-the-art results on common zero shot relative pose estimation benchmarks, outperforming the strongest baseline by a relative 62% in average ADD(-S) score, including methods that utilize extensive dataset-specific training.

Jiajin Liu, Dongzhe Fan, Chuanhao Ji, Daochen Zha, Qiaoyu Tan

Vision-Language Models (VLMs) have demonstrated remarkable capabilities in aligning and understanding multimodal signals, yet their potential to reason over structured data, where multimodal entities are connected through explicit relational graphs, remains largely underexplored. Unlocking this capability is crucial for real-world applications such as social networks, recommendation systems, and scientific discovery, where multimodal information is inherently structured. To bridge this gap, we present GraphVLM, a systematic benchmark designed to evaluate and harness the capabilities of VLMs for multimodal graph learning (MMGL). GraphVLM investigates three complementary paradigms for integrating VLMs with graph reasoning: (1) VLM-as-Encoder, which enriches graph neural networks through multimodal feature fusion; (2) VLM-as-Aligner, which bridges modalities in latent or linguistic space to facilitate LLM-based structured reasoning; and (3) VLM-as-Predictor, which directly employs VLMs as multimodal backbones for graph learning tasks. Extensive experiments across six datasets from diverse domains demonstrate that VLMs enhance multimodal graph learning via all three roles. Among these paradigms, VLM-as-Predictor achieves the most substantial and consistent performance gains, revealing the untapped potential of vision-language models as a new foundation for multimodal graph learning.

Qiucheng Wu, Jing Shi, Simon Jenni, Kushal Kafle, Tianyu Wang, Shiyu Chang, Handong Zhao

Recent advances in multimodal large language models (MLLMs) have shown great potential for extending vision-language reasoning to professional tool-based image editing, enabling intuitive and creative editing. A promising direction is to use reinforcement learning (RL) to enable MLLMs to reason about and execute optimal tool-use plans within professional image-editing software. However, training remains challenging due to the lack of reliable, verifiable reward signals that can reflect the inherently subjective nature of creative editing. In this work, we introduce RetouchIQ, a framework that performs instruction-based executable image editing through MLLM agents guided by a generalist reward model. RetouchIQ interprets user-specified editing intentions and generates corresponding, executable Lightroom adjustments, bridging high-level aesthetic goals with precise parameter control. To move beyond conventional, rule-based rewards that compute similarity against a fixed reference image using handcrafted metrics, we propose a generalist reward model, an RL fine-tuned MLLM that evaluates edited results through a set of generated metrics on a case-by-case basis. Then, the reward model provides scalar feedback through multimodal reasoning, enabling reinforcement learning with high-quality, instruction-consistent gradients. We curate an extended dataset with 190k instruction-reasoning pairs and establish a new benchmark for instruction-based image editing. Experiments show that RetouchIQ substantially improves both semantic consistency and perceptual quality over previous MLLM-based and diffusion-based editing systems. Our findings demonstrate the potential of generalist reward-driven MLLM agents as flexible, explainable, and executable assistants for professional image editing.

Chuong Huynh, Manh Luong, Abhinav Shrivastava

Multimodal retrieval is the task of aggregating information from queries across heterogeneous modalities to retrieve desired targets. State-of-the-art multimodal retrieval models can understand complex queries, yet they are typically limited to two modalities: text and vision. This limitation impedes the development of universal retrieval systems capable of comprehending queries that combine more than two modalities. To advance toward this goal, we present OmniRet, the first retrieval model capable of handling complex, composed queries spanning three key modalities: text, vision, and audio. Our OmniRet model addresses two critical challenges for universal retrieval: computational efficiency and representation fidelity. First, feeding massive token sequences from modality-specific encoders to Large Language Models (LLMs) is computationally inefficient. We therefore introduce an attention-based resampling mechanism to generate compact, fixed-size representations from these sequences. This shared module is designed to maintain representational diversity and generalization capabilities while remaining sensitive to modality-specific information. Second, compressing rich omni-modal data into a single embedding vector inevitably causes information loss and discards fine-grained details. We propose Attention Sliced Wasserstein Pooling to preserve these fine-grained details, leading to improved omni-modal representations. OmniRet is trained on an aggregation of approximately 6 million query-target pairs spanning 30 datasets. We benchmark our model on 13 retrieval tasks and a MMEBv2 subset. Our model demonstrates significant improvements on composed query, audio and video retrieval tasks, while achieving on-par performance with state-of-the-art models on others. Furthermore, we curate a new Audio-Centric Multimodal Benchmark (ACM). This new benchmark introduces two critical, previously missing tasks--composed audio retrieval and audio-visual retrieval--to more comprehensively evaluate a model's omni-modal embedding capacity. We believe our benchmark will facilitate the development of universal retrieval systems.

Yong Xien Chng, Tao Hu, Wenwen Tong, Xueheng Li, Jiandong Chen, Haojia Yu, Jiefan Lu, Hewei Guo, Hanming Deng, Chengjun Xie 等

Vision-Language Models (VLMs) are limited by static knowledge and insufficient fine-grained visual analysis, hindering their performance on knowledge-intensive and visually complex tasks. While recent research has explored VLMs that employ external tools like search or cropping to enhance model performance, they typically employ tools in isolation and lack the ability to coordinate multiple tools effectively. To address this gap, we propose SenseSearch, the first agentic VLM for search-reasoning that supports adaptive multi-tool coordination via reinforcement learning (RL). Specifically, SenseSearch dynamically integrates the image search, text search, and image crop tools to tackle fine-grained and knowledge-intensive visual understanding challenges. We first construct a high-quality cold-start dataset to instill basic tool-usage behaviors. In the subsequent RL stage, we introduce Batch-Normalized Group Sequence Policy Optimization (BN-GSPO) algorithm to enhance the tool invocation and reasoning ability. To comprehensively evaluate the agentic VLMs on complex visual tasks, we introduce the HR-MMSearch benchmark, the first search-oriented benchmark composed of high-resolution images with knowledge-intensive and search-driven questions. Experiments demonstrate that SenseSearch achieves state-of-the-art performance on open-source search and fine-grained image understanding benchmarks, outperforming baselines by 19.18% on HR-MMSearch. SenseSearch provides a promising path toward agentic VLMs with effective and robust tool invocation capabilities. All code and data are released at https://github.com/OpenSenseNova/SenseNova-MARS.

Ke Niu, Haiyang Yu, Zhuofan Chen, Zhengtao Yao, Weitao Jia, Xiaodong Ge, Jingqun Tang, Benlei Cui, Bin Li, Xiangyang Xue

Computer-Aided Design (CAD) is essential in industrial design, but the complexity of traditional CAD modeling and workflows presents significant challenges for automating the generation of high-precision, editable CAD models. Existing methods, such as 3D reconstruction from sketches, often produce non-editable, approximate models that fall short of meeting the stringent requirements for precision and editability in industrial design. Moreover, the reliance on text or image-based inputs often requires significant manual annotation, limiting their scalability and applicability in industrial settings. To overcome these challenges, we propose the Heterogeneous Collaborative Multi-Expert Reinforcement Learning (CME-CAD) paradigm, a novel training paradigm for CAD code generation. Our approach integrates the complementary strengths of these models, facilitating collaborative learning and improving the model's ability to generate accurate, constraint-compatible, and fully editable CAD models. We introduce a two-stage training process: Multi-Expert Fine-Tuning (MEFT), and Multi-Expert Reinforcement Learning (MERL). Additionally, we present CADExpert, an open-source benchmark consisting of 17,299 instances, including orthographic projections with precise dimension annotations, expert-generated Chain-of-Thought (CoT) processes, executable CADQuery code, and rendered 3D models. The CADExpert dataset is publicly available at https://modelscope.cn/datasets/zhuofanChen/CADExpert.

Nicolas Michel, Maorong Wang, Jiangpeng He, Toshihiko Yamasaki

Deep learning models continue to scale, with some requiring more storage than many large-scale datasets. Thus, we introduce a new paradigm: Continual Distillation (CD), where a student learns sequentially from a stream of teacher models without retaining access to earlier teachers. CD faces two challenges: teacher training data is unavailable, and teachers have varying expertise. We show that external unlabeled data enables Unseen Knowledge Transfer (UKT), allowing the student to acquire information from domains not present in the training data, while known to the teacher. We also show that sequential distillation causes Unseen Knowledge Forgetting (UKF) when transferred knowledge is lost after training on later teachers. To better trade off between UKT and UKF, we propose Self External Data Distillation (SE2D), a method that preserves logits on external data to stabilize learning across heterogeneous teachers. Experiments on multiple benchmarks show that SE2D reduces UKF and improves cross-domain generalization. The code and implementation for this work are publicly available at: https://github.com/Nicolas1203/continual_distillation

Jie Xiao, Yinchao Ma, Yuyang Tang, Dengqing Yang, Jianpeng Yang, Xu Zhou, Qiao Li, Wenfei Yang, Tianzhu Zhang

3D single object tracking (SOT) in point clouds is essential for real-world 3D perception, yet it remains challenging due to data sparsity and large variations in scale and structure across diverse object categories. Most existing methods rely on a category-specific paradigm that trains separate models for each class, severely limiting scalability and generalization in real deployment. Extending these methods to a single model capable of tracking diverse object categories proves inadequate, as the significant variations across categories make it difficult to establish reliable geometric correspondences without category-specific priors. To overcome these limitations, we propose a Unified Structural KeyPoint Tracker (UniKPT), a novel structure-aware and generalizable framework for category-unified 3D point cloud tracking. UniKPT comprises three key modules: (1) an adaptive keypoint extractor that produces scale-aware and semantically meaningful keypoints; (2) a progressive correspondence aligner that establishes robust cross-frame geometric associations; and (3) a confidence-aware structural localization module that suppresses unreliable matches and leverages fine-grained structural relationships for precise 3D localization. Extensive experiments on the nuScenes and KITTI benchmarks show that UniKPT achieves new state-of-the-art performance in category-unified 3D SOT. On the challenging nuScenes dataset, our unified model further surpasses category-specific state-of-the-art trackers by 4.37% in Success and 5.16% in Precision.

Girish, Mohd Mujtaba Akhtar, Orchid Chetia Phukan, Arun Balaji Buduru

The rapid advancement of Audio Large Language Models (ALMs), driven by Neural Audio Codecs (NACs), has led to the emergence of highly realistic speech deepfakes, commonly referred to as CodecFakes (CFs). Consequently, CF detection has attracted increasing attention from the research community. However, existing studies predominantly focus on English or Chinese, leaving the vulnerability of Indic languages largely unexplored. To bridge this gap, we introduce Indic-CodecFake (ICF) dataset, the first large-scale benchmark comprising real and NAC-synthesized speech across multiple Indic languages, diverse speaker profiles, and multiple NAC types. We use IndicSUPERB as the real speech corpus for generation of ICF dataset. Our experiments demonstrate that state-of-the-art (SOTA) CF detectors trained on English-centric datasets fail to generalize to ICF, underscoring the challenges posed by phonetic diversity and prosodic variability in Indic speech. Further, we present systematic evaluation of SOTA ALMs in a zero-shot setting on ICF dataset. We evaluate these ALMs as they have shown effectiveness for different speech tasks. However, our findings reveal that current ALMs exhibit consistently poor performance. To address this, we propose SATYAM, a novel hyperbolic ALM tailored for CF detection in Indic languages. SATYAM integrates semantic representations from Whisper and prosodic representations from TRILLsson using through Bhattacharya distance in hyperbolic space, and subsequently performs the same alignment procedure between the fused speech representation and a input conditioning prompt. This dual-stage fusion framework enables SATYAM to effectively model hierarchical relationships both within speech (semantic–prosodic) and across modalities (speech–text). Extensive evaluations show that SATYAM consistently outperforms competitive end-to-end and ALM-based baselines on the ICF benchmark.

Yu Zou, Yan Chen, Lida He, Qi Zhou, Xiaorui Zhou, Aixi Zhong, Yi Wang, Wei Li, Qingyu Wang, Jiatao Li 等

Understanding financial documents is critical for high-stakes decision-making yet hindered by systemic semantic implicitness: key facts are rarely explicit in surface text and often determined by global structural cues. Missing these cues invites semantic misinterpretations, such as misreading what a number refers to, an outcome unacceptable in high-stakes environments. However, existing Retrieval-Augmented Generation (RAG) systems typically treat structure as a physical navigational skeleton rather than intrinsic semantic knowledge. To address this, we introduce Fin-STAR (Financial STructure-As-Semantics Retrieval), a framework redefining hierarchy as intrinsic semantics. Fin-STAR incorporates a novel Structure-Enriched Semantic Indexing mechanism that augments the hierarchical lineage with snippet-derived virtual nodes, and injects this enriched context via a semantic cross-attention paradigm, rendering implicit cues explicit. By grounding evidence within its structural scope, we preserve factual invariance and ensure contextual integrity. Addressing the lack of granular public datasets, we conduct experiments on FinTierQA Gold, a curated expert benchmark. Results show that Fin-STAR outperforms state-of-the-art hierarchical and graph-based baselines across diverse query complexities, document types, and markets. Notably, ablations confirm that our semantic injection consistently outperforms alternative strategies. Finally, we release FinTierQA, comprising 3.9M pairs automatically constructed from 78k documents via our framework .

Xiao Xia, Dan Zhang, Tianrui Sun

Providing accurate reward signals for code generated by large language models (LLMs) is a significant challenge in applying reinforcement learning (RL) to code generation. Existing methods rely on unit tests to evaluate code correctness and provide rewards, which are hindered by the difficulty of acquiring and verifying reliable unit tests at scale. In this work, we propose CodeRM-NT, a code reward model with no reliance on unit tests. Our method leverages Monte Carlo Tree Search guided by LLMs to generate code snippets and judges execution traces to annotate code with reward signals. We use the rewards to train CodeRM-NT that is capable of providing rewards for code during RL. CodeRM-NT also facilitates curriculum learning by scoring and sorting training samples based on their difficulty. Experimental results demonstrate that training with CodeRM-NT consistently outperforms synthetic unit test-based rewards, yielding superior performance on multiple code generation benchmarks. Additionally, curriculum learning based on CodeRM-NT further enhances model performance. Our code and dataset are available at: https://github.com/THUDM/CodeRM-NT.

Georgii Andriushchenko, Roman Garaev, Lyudmila Rvanova, Artem Shelmanov, Vladimir V. Ivanov

Large language models (LLMs) frequently produce source code that seems correct and well-formed, yet includes hallucinated elements that cause downstream test failures. In this study, we benchmark state-of-the-art uncertainty quantification methods and existing baselines for the task of hallucination detection in source code and introduce a diff-based pipeline to construct a code dataset annotated with line-level hallucinations. Building on this, we train a lightweight Transformer-based detector that uses LLM internal representations to identify hallucinations, substantially outperforming existing methods across several code generation domains. The detector also shows particular promise for enabling self-correction in LLM-based coding agents. We release the first publicly available dataset of line-level code hallucinations, along with the corresponding source code and trained hallucination detectors https://github.com/datapaf/CodeHallucinationDetection

Chenhui Liu, Jianpeng Zhou, Jiahai Wang

Knowledge graph question answering (KGQA) serves as an essential benchmark for KG-enhanced large language models. Among various approaches, agent-based methods have emerged as an effective solution.Existing methods adopt entity-centric exploration that incrementally constructs reasoning paths by selecting and connecting intermediate entities. However, they face two critical limitations. (1) Entity incompleteness vulnerability arises when some intermediate entities lack semantic information beyond opaque IDs, preventing relevance evaluation and leading to discarding valid reasoning paths.(2) Premature entity pruning occurs because beam search retains only top-ranked entities at each step, eliminating candidates before their relevance can be verified.To address these challenges, this paper proposes Chain-of-Relations (CoR) with relation-centric exploration and global entity filtering, reducing dependence on entity completeness and ensuring complete candidate retrieval before constraint validation.Experiments on three benchmark datasets show that CoR consistently outperforms strong baselines in both F1 score and KG-grounded Rate.

Zihan Cheng, Jianxiang Ma, Xiaocui Yang, Peidong Wang, Wen Zhang, Shi Feng, Daling Wang, Yifei Zhang, Mingfu Zhang

Harmful memes convey offensive intent through implicit associations between visual symbols and text, requiring a broad understanding of cultural stereotypes and visual metaphors. Small-scale Multimodal Large Language Models (MLLMs) often lack the knowledge required to identify such implicit hate, whereas Large-scale MLLMs, despite their broader knowledge, exhibit systematic labeling bias. To address these challenges, we propose DR-HM, a Distill-then-Reinforce training framework with cognition-aware data synthesis for harmful meme detection, which aims to transfer knowledge from closed-source models while mitigating their biases. DR-HM introduces a six-step structured data synthesis scheme with self-refinement that decomposes meme analysis into a progressive, human-inspired reasoning process from entity recognition to harmfulness judgment. Based on the synthesized reasoning data, we further adopt a Distill-then-Reinforce training strategy. This approach combines a two-stage Supervised Fine-Tuning (SFT) with an Adaptive Group Relative Policy Optimization (A-GRPO) algorithm, which incorporates class-ratio-aware reward weighting and dynamic KL coefficients. Experiments on three benchmark datasets show that the proposed approach consistently outperforms existing methods and achieves an accuracy of 84.7% on the FHM dataset, approaching the reported performance of human annotators.

Nanxu Gong, Zixin Chen, Haotian Li, Zishu Zhao, Jianxun Lian, Huamin Qu, Yanjie Fu, Xing Xie

Improving the Theory of Mind (ToM) capability of Large Language Models (LLMs) is crucial for effective social interactions between these AI models and humans. However, the existing benchmarks often measure ToM capability improvement through story-reading, multiple-choice questions from a third-person perspective, while ignoring the first-person, dynamic, and open-ended nature of human-AI (HAI) interactions. To directly examine how ToM improvement techniques benefit HAI interactions, we first proposed the new paradigm of interactive ToM evaluation with both perspective and metric shifts. Next, following the paradigm, we conducted a systematic study of four representative ToM enhancement techniques using both four real-world datasets and a user study, covering both goal-oriented tasks (e.g., coding, math) and experience-oriented tasks (e.g., counseling). Our findings reveal that improvements on static benchmarks do not always translate to better performance in dynamic HAI interactions. This paper offers critical insights into ToM evaluation, showing the necessity of interaction-based assessments in developing next-generation, socially aware LLMs for HAI symbiosis.

Chenxi Qing, Junxi Wu, Zheng Liu, Yixiang Qiu, Hongyao Yu, Bin Chen, Hao Wu, Shu-Tao Xia

Recently, large language models (LLMs) are capable of generating highly fluent textual content. While they offer significant convenience to humans, they also introduce various risks, like phishing and academic dishonesty. Numerous research efforts have been dedicated to developing algorithms for detecting AI-generated text and constructing relevant datasets. However, in the domain of Chinese corpora, challenges remain, including limited model diversity and data homogeneity. To address these issues, we propose C-ReD: a comprehensive Chinese Real-prompt AI-generated text Detection benchmark. Experiments demonstrate that C-ReD not only enables reliable in-domain detection but also supports strong generalization to unseen LLMs and external Chinese datasets—addressing critical gaps in model diversity, domain coverage, and prompt realism that have limited prior Chinese detection benchmarks. We release our resources at https://github.com/HeraldofLight/C-ReD.

Guiyang Hou, Xiang Huang, Shangke Lyu, Yuchuan Wu, Weiyao Luo, Xinyu Mei, Yongliang Shen, Weiming Lu, Yongbin Li

Theory of Mind (ToM), the ability to infer others’ mental states from behavior, is pivotal for developing machines with human-level social intelligence. Existing methods endowing LLMs with ToM fall into two paradigms: training-free methods and those repurposing ToM evaluation benchmarks as training data for RL-based fine-tuning. However, training-free methods fail to internalize the augmented ToM into the LLMs. Meanwhile, using evaluation benchmarks as training sources is conceptually problematic and, in practice, results in narrow in-domain overfitting rather than robust ToM. To address the lack of training resources within the ToM community and to empower LLMs with robust ToM, we introduce ToM-Synth, a factorial combinatorial synthesis framework of 6912 social units. This framework enables the systematic synthesis of ToM data, yielding a training dataset of 27,648 instances, termed ToM-Synth-27K. Utilizing ToM-Synth-27K for RL fine-tuning, experimental results demonstrate consistent and significant improvements across models of varying families and scales on ToM, Emotional Intelligence, and Social Commonsense benchmarks. Furthermore, we observe concurrent enhancements in IQ-related tasks (math, science, logic) and effective performance scaling with increasing data scale.

Zijian Wu, Jinjie Ni, Xiangyan Liu, Zichen Liu, Hang Yan, Michael Qizhe Shieh

Vision-language models (VLMs) trained via reinforcement learning with verifiable reward (RLVR) have shown notable progress in scaling test-time compute effectively. In this work, we investigate how synthesized RL data can further improve RLVR. To this end, we propose SynthRL—a scalable and guaranteed pipeline for automatic data scaling in reasoning-oriented RL training. SynthRL comprises three key stages: (1) selecting seed questions with appropriate distribution, (2) augmenting them into more challenging variants while preserving the original answers, and (3) a guaranteed verification stage that ensures near-perfect correctness and difficulty enhancement. Our empirical experiments demonstrate SynthRL’s scalability and effectiveness. When applied to the MMK12 dataset, SynthRL synthesizes over 3.3K additional verifiable, challenging questions from approximately 8K seed samples. Models trained with our synthesized data achieve consistent gains across five out-of-domain visual math reasoning benchmarks, with a significant improvement over baseline models trained on seed data alone. Notably, detailed analysis reveals that the gains are more pronounced on the most challenging evaluation samples, highlighting SynthRL’s effectiveness in eliciting deeper and more complex reasoning patterns.

Sonal Kumar, Sreyan Ghosh, Yueqian Lin, S Sakshi, Ashish Seth, Yiran Chen, Ramani Duraiswami, Dinesh Manocha

Large Audio Language Models have shown impressive performance on single-clip audio language tasks such as automatic speech recognition, captioning, and sound event recognition. Yet, their ability to reason over interleaved multi-audio contexts-where answering a query requires relating information across multiple audio clips-remains limited. We present PolyAudio, a LALM built on Audio Flamingo 3 that targets multi-audio understanding via instruction tuning rather than massive-scale pre-training, and PolyAudio-Instruct, a high-quality instruction-tuning dataset consisting of 1.3M+ QA pairs, spanning over 14 task subsets to empower multi-audio understanding and reasoning. PolyAudio uses an explicit interleaved representation with clip indexing to encourage faithful grounding and reduce ambiguity in multi-clip references. We evaluate PolyAudio on a diverse suite of multi-audio benchmarks alongside standard single-audio tasks. PolyAudio achieves strong performance on multi-audio reasoning, outperforming competitive baselines that are also often limited to reasoning over up-to 2 audio clips, while preserving robust single-clip performance. Overall, our results suggest that precise, academic-scale multi-audio instruction tuning can unlock advanced cross-clip reasoning capabilities, enabling more capable audio-centric assistants.