💡 The paper presents dots.tts, a 2 billion parameter continuous autoregressive text-to-speech model that achieves state-of-the-art performance on multiple benchmarks. The model is trained on a large-scale multilingual corpus and enables efficient low-latency speech generation. The key innovations of the model are threefold. First, the authors train an AudioVAE with multiple objectives to build a semantically structured and prediction-friendly continuous speech space. Second, they use full-history conditioning in the flow-matching head to preserve long-range consistency and reduce drift during generation. Third, they apply reward-free self-corrective post-training to the flow-matching head to further improve robustness and acoustic quality.

The model is evaluated on several benchmarks, including Seed-TTS-Eval, where it achieves the best average performance with word error rates of 0.94, 1.30, and 6.60 percent and similarity scores of 81.0, 77.1, and 79.5 on the Chinese, English, and Chinese hard test sets, respectively. The model also demonstrates strong generation stability, voice cloning ability, and emotional expressiveness on other benchmarks.

To enable efficient inference, the authors apply CFG-aware MeanFlow distillation, which allows for low-latency speech generation with first-packet latencies of 85 and 54 milliseconds in output streaming and dual-streaming modes, respectively. The training and inference code, as well as the pre-trained, post-trained, and MeanFlow-distilled checkpoints, are released under the Apache 2.0 license to facilitate reproducible research and practical deployment.

Overall, the paper presents a significant contribution to the field of text-to-speech synthesis, achieving state-of-the-art performance and enabling efficient low-latency speech generation. The model's ability to generate high-quality speech with strong generation stability, voice cloning ability, and emotional expressiveness makes it a valuable tool for a wide range of applications.

💡 The paper introduces MMAE, a comprehensive benchmark for instruction-based audio editing that evaluates models across multiple modalities and complexity levels. The current evaluation infrastructure for audio editing is fragmented and limited in scope, which motivated the creation of MMAE. This benchmark extends to a broad spectrum of real-world scenarios, covering 7 distinct audio modalities, including sound, speech, music, and their mixtures, and establishes a comprehensive taxonomy spanning 6 levels of task complexity, 2 levels of granularity, and 8 distinct operation types.

The MMAE benchmark was meticulously curated through human-agent collaboration and comprises 2000 high-fidelity samples paired with a pioneering rubric-based evaluation framework. This framework enables a precise, multi-dimensional assessment of both instruction following and context consistency by decomposing free-form tasks into 17741 verifiable criteria.

The paper evaluates leading models using the MMAE benchmark and reveals significant gaps in current model capabilities. The results show that the Exact Match Rate consistently falls below 5% and plummets to 0% in complex, mixed-modality tasks, exposing critical bottlenecks in precise execution and structural robustness. The MMAE benchmark provides a clear diagnostic roadmap and establishes a standardized, long-lasting evaluation paradigm for next-generation audio editing systems, which can serve as a catalyst for future advances in the intelligent creation community.

💡 The paper introduces UniSHARP, a method for universal monocular view synthesis that extends the popular SHARP method to work with various camera systems, including conventional perspective cameras, wide-field-of-view, fisheye, and omnidirectional panoramic settings. The problem with SHARP is that it is designed for pinhole cameras and does not work well with other types of cameras. To overcome this limitation, UniSHARP aligns images in a unified omnidirectional latent space by performing implicit alignment in both feature and Gaussian spaces. This is achieved by arranging Gaussian primitives along rays and radial distances in a ray-based universal representation, and jointly decoding 2D semantic and 3D spatial features extracted from encoders to generate a complete Gaussian cloud. The authors evaluate UniSHARP on a benchmark that covers diverse imaging systems and scenes, and stratify the benchmark by field of view to enable fine-grained assessment of the universal monocular rendering task. The results show that UniSHARP outperforms alternative methods by a large margin, demonstrating its effectiveness in universal monocular rendering across different camera systems. Overall, UniSHARP provides a universal solution for monocular view synthesis that can work with a wide range of camera systems, making it a significant contribution to the field of computer vision.

💡 The paper discusses the challenges of comparing genomic foundation models due to the lack of a unified evaluation protocol and fragmented benchmarks. This makes it difficult to assess the progress of these models and compare their performance. To address this issue, the authors introduce GENEB, a comprehensive benchmark that evaluates 40 genomic foundation models across 100 tasks and 13 functional categories using a unified protocol. GENEB allows for controlled comparison of models based on their scale, architecture, tokenization, and pretraining data, and exposes task-level trade-offs. The analysis of GENEB reveals that model rankings vary significantly across task categories, and that increasing model scale provides only modest and inconsistent gains. The results also show that architectural and pretraining alignment are more important than parameter count. The paper highlights the limitations of current evaluation practices and positions GENEB as a reference framework for comparing and selecting genomic machine learning models in a principled and category-aware manner. Overall, GENEB provides a much-needed benchmark for the genomic machine learning community, enabling more accurate and meaningful comparisons of genomic foundation models.

💡 The paper proposes a self-improving AI framework called SIA that addresses the bottleneck of human involvement in building and improving AI systems. Currently, two separate research approaches exist to tackle this issue: the harness-update school, which updates the task-specific agent's architecture while keeping the model weights fixed, and the test-time training school, which updates the model weights using reinforcement learning pipelines while keeping the harness fixed. However, these two approaches operate in isolation.

The SIA framework combines these two approaches by introducing a language-model feedback agent that simultaneously updates both the model weights and the task-specific agent's architecture. This is achieved through a self-improving loop where the feedback agent provides updates to both the harness and the weights of the task-specific agent.

The authors evaluate the SIA framework across three diverse domains: Chinese legal charge classification, low-level GPU kernel optimization, and single-cell RNA denoising. The results show that combining both harness and weight updates outperforms using only harness updates. The gains are significant, with improvements of 56.6% on the LawBench benchmark, 91.9% runtime reduction on GPU kernels, and 502% improvement on denoising over the initial baseline.

The SIA framework makes the model more agentic by shaping how it searches and acts, while the weight updates build domain-specific intuition that cannot be instilled through prompts or scaffolds alone. Overall, the paper contributes to the development of self-improving AI systems by proposing a novel framework that integrates both harness and weight updates, demonstrating its effectiveness across multiple domains.

💡 The paper introduces OpenSkill, a framework that enables self-evolving agents to develop skills and verification signals from scratch using open-world resources without target-task supervision. The problem addressed is that existing approaches to self-evolving agents require a usable learning loop, such as curated skills or successful trajectories, which may not be available in real-world deployment scenarios. OpenSkill solves this problem by bootstrapping the learning loop, acquiring grounded knowledge and verification anchors from documentation, repositories, and the web, synthesizing them into transferable skills, and refining those skills against self-built virtual tasks.

The method used in OpenSkill involves three main steps. First, it acquires knowledge and verification anchors from open-world resources. Second, it synthesizes this knowledge into transferable skills. Third, it refines these skills against self-built virtual tasks grounded in the anchors, rather than in target answers. This approach allows the agent to develop skills and verification signals without requiring target-task supervision.

The results of OpenSkill are impressive, with the framework attaining the best automated pass rate across three benchmarks and two target agents, while satisfying the no-supervision constraint. Analysis of the results shows that the skills learned by OpenSkill transfer across models without requiring model-specific adaptation, and the self-built verifier aligns with ground-truth outcomes despite never accessing them. Overall, OpenSkill provides a novel approach to open-world self-evolution, enabling self-evolving agents to develop skills and verification signals from scratch using open-world resources, and achieving high automated performance across benchmarks.

💡 The paper introduces ResearchClawBench, a benchmark for evaluating the end-to-end autonomous scientific research capabilities of artificial intelligence agents and large language models. The problem addressed is the difficulty in verifying the autonomous research capabilities of AI agents, which are increasingly used for scientific work. To tackle this, the authors created a benchmark that consists of 40 tasks from 10 scientific domains, each grounded in a real published paper with related literature and raw data. The target paper is hidden during evaluation, and expert-curated criteria are used to assess the performance of the AI agents.

The method used to evaluate the autonomous research agents involves a unified protocol, where seven agents are evaluated, and a lightweight ResearchHarness is used to evaluate seventeen native large language models. The performance of these agents is measured using multimodal rubrics that decompose the target scientific artifacts into weighted criteria, enabling the evaluation of target-paper-level re-discovery while allowing for new discovery.

The results show that current autonomous research agents and large language models are far from reliable re-discovery, with the strongest autonomous agent averaging a score of 21.5 and the strongest large language model averaging a score of 20.7. The mean score of the large language models is only 26.5, indicating significant room for improvement. Error analysis reveals that the failures of these agents concentrate in three areas: experimental protocol mismatch, evidence mismatch, and missing scientific core. The ResearchClawBench benchmark provides a reproducible evaluation frontier for measuring progress toward autonomous scientific research, allowing researchers to track improvements in this area over time.

💡 The paper proposes a novel approach to video world models called latent spatial memory, which stores 3D scene information directly in diffusion latent space. This approach eliminates the need for explicit point cloud memory constructed in RGB space, which is computationally expensive and inherently lossy due to the round trip through pixel space. The authors introduce a framework called Mirage, which constructs the latent spatial memory by lifting latent tokens into 3D via depth-guided back-projection and queries it by synthesizing novel views through direct latent-space warping. This approach avoids pixel-space reconstruction and reduces the computational burden of repeated encoding and rendering. The results show that latent spatial memory achieves significant improvements in video generation speed and memory footprint, with up to 10.57 times faster end-to-end video generation and 55 times reduction in memory footprint compared to explicit 3D baselines. The Mirage framework also attains state-of-the-art performance on WorldScore and strong reconstruction quality on RealEstate10K, demonstrating the effectiveness of the proposed approach. Overall, the paper contributes a new and efficient method for video world models that leverages the geometric prior of the diffusion model to achieve faster and more memory-efficient video generation.

💡 The paper introduces OmniGameArena, a unified benchmark for evaluating vision-language model agents in various game settings. The current benchmarks for these agents have limitations, such as only reporting a single score per agent and game pair, focusing on solo play, and lacking a unified protocol for evaluating different types of agents. To address these gaps, OmniGameArena provides a real-time benchmark of twelve new games built with Unreal Engine 5, covering solo, player versus player, and cooperative play, with a unified action interface.

The benchmark also includes the Improvement Dynamics Curve, a protocol that allows a tool-using reflector language model to refine a skill prompt across multiple rounds, providing additional insights into the agents' performance evolution and skill generalization. This protocol enables the evaluation of how the score evolves across reflection rounds and how the learned skill behaves on held-out task variants.

The paper reports the results of twelve vision-language model agents on the cold-start leaderboard and four top agents under the Improvement Dynamics Curve protocol. The results provide a more comprehensive understanding of the agents' performance and capabilities, beyond just a single score. Overall, OmniGameArena offers a more comprehensive and unified benchmark for evaluating vision-language model agents in game environments, allowing for a more detailed analysis of their strengths and weaknesses.

💡 The paper introduces UniCorn, a self-improvement framework for unified multimodal models that addresses the generation gap in these models. The generation gap refers to the discrepancy between a model's ability to understand multimodal inputs and its ability to generate high-quality outputs. This gap is formalized as Conduction Aphasia, where models can accurately interpret multimodal inputs but struggle to translate that understanding into faithful and controllable synthesis.

To address this issue, UniCorn proposes a simple yet elegant self-improvement framework that eliminates the need for external data or teacher supervision. The framework partitions a single unified multimodal model into three collaborative roles: Proposer, Solver, and Judge. The Proposer generates initial outputs, the Solver refines these outputs, and the Judge evaluates the quality of the refined outputs. Through self-play and cognitive pattern reconstruction, UniCorn generates high-quality interactions and distills latent understanding into explicit generative signals.

The authors introduce UniCycle, a cycle-consistency benchmark based on a Text to Image to Text reconstruction loop, to validate the restoration of multimodal coherence. The results demonstrate that UniCorn achieves comprehensive and substantial improvements over the base model across six general image generation benchmarks. Notably, it achieves state-of-the-art performance on several benchmarks, including TIIF, DPG, CompBench, and UniCycle, and delivers substantial gains on WISE and OneIG.

The contributions of the paper are significant, as UniCorn enhances text-to-image generation while maintaining robust comprehension, demonstrating the scalability of fully self-supervised refinement for unified multimodal intelligence. The results highlight the effectiveness of the self-improvement framework in addressing the generation gap in unified multimodal models, and the potential of UniCorn to improve the performance of these models in various applications. Overall, the paper presents a novel approach to self-improving unified multimodal models, with significant implications for the development of more advanced and effective multimodal models.