Synthetic video data with verifiable rewards offers a scalable source of supervision for temporal motion understanding in video-language models, yet it remains unclear which synthetic tasks transfer reliably to real-video benchmarks.We investigate this question through a diagnostic framework that treats each synthetic task as a controlled intervention.Starting from a shared Qwen3-VL-2B checkpoint with frozen vision encoder, we independently train nine programmatic motion tasks with RLVR and evaluate each resulting model across six temporal-motion dimensions, such as direction, rotation, and velocity, yielding an interference matrix of cross-dimensional transfer. The matrix reveals substantial variation in transfer that surface-level label alignment fails to explain: a frame-to-frame velocity task yields a large gain on its target dimension, whereas a nominally aligned trajectory-shape task degrades the dimension it is intended to improve. We find that transfer is better explained by computational alignment, namely whether the training task induces the local, order-sensitive temporal computation that the downstream capability requires.Counterfactual tasks that share semantic targets but differ in computational demands support this account, and the patterns hold on InternVL3-2B. We further introduce a training-free answer-invariance test under frame shuffling that flags shortcut-prone tasks prior to RLVR training. Using the interference matrix to filter tasks yields a task mixture that consistently outperforms uniform mixing on TOMATO, with positive transfer trends on MVBench, TempCompass, and TemporalBench.Our framework thus reframes synthetic video generation as an empirically grounded task-selection problem.

Extending the input modality of Large Language Models (LLMs) to the audio domain is essential for achieving comprehensive multimodal perception. However, it is well-known that acoustic information is intrinsically heterogeneous, entangling attributes such as speech, music, and environmental context. Existing research is limited to a dense, parameter-shared adapter to model these diverse patterns, which induces gradient conflict during optimization, as parameter updates required for distinct attributes contradict each other. To address this limitation, we introduce the MoE-Adapter, a sparse Mixture-of-Experts (MoE) architecture designed to decouple acoustic information. Specifically, it employs a dynamic gating mechanism that routes audio tokens to specialized experts capturing complementary feature subspaces while retaining shared experts for global context, thereby mitigating gradient conflicts and enabling fine-grained feature learning. Comprehensive experiments show that the MoE-Adapter achieves superior performance on both audio semantic and paralinguistic tasks, consistently outperforming dense linear baselines with comparable computational costs. To facilitate future research, our code are publicly available at https://github.com/Alittleegg/Eureka-Audio.

Large Audio Language Models (LALMs) have garnered significant research interest. Despite being built upon text-based large language models (LLMs), LALMs frequently exhibit a degradation in knowledge and reasoning capabilities. We hypothesize that this limitation stems from the failure of current training paradigms to effectively bridge the acoustic-semantic gap within the feature representation space. To address this challenge, we propose CORD, a unified alignment framework that performs online cross-modal self-distillation. Specifically, it aligns audio-conditioned reasoning with its text-conditioned counterpart within a unified model. Leveraging the text modality as an internal teacher, CORD performs multi-granularity alignment throughout the audio rollout process. At the token level, it employs on-policy reverse KL divergence with importance-aware weighting to prioritize early and semantically critical tokens. At the sequence level, CORD introduces a judge-based global reward to optimize complete reasoning trajectories via Group Relative Policy Optimization (GRPO). Empirical results across multiple benchmarks demonstrate that CORD consistently enhances audio-conditioned reasoning and substantially bridges the audio–text performance gap with only 80k synthetic training samples, validating the efficacy and data efficiency of our on-policy, multi-level cross-modal alignment approach.

Dialogue Aspect-based Sentiment Quadruple analysis (DiaASQ) extends ABSA to more complex real-world scenarios (i.e., dialogues), which makes existing generation methods encounter heightened noise and order bias challenges, leading to decreased robustness and accuracy.To address these, we propose the Segmentation-Aided multi-grained Denoising and Debiasing (SADD) method. For noise, we propose the Multi-Granularity Denoising Generation model (MGDG), achieving word-level denoising via sequence labeling and utterance-level denoising via topic-aware dialogue segmentation. Denoised Attention in MGDG integrates multi-grained denoising information to help generate denoised output.For order bias, we first theoretically analyze its direct cause as the gap between ideal and actual training objectives and propose a distribution-based solution. Since this solution introduces a one-to-many learning challenge, our proposed Segmentation-aided Order Bias Mitigation (SOBM) method utilizes dialogue segmentation to supplement order diversity, concurrently mitigating this challenge and order bias.Experiments demonstrate SADD’s effectiveness, achieving state-of-the-art results with a 6.52% F1 improvement.