Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
50 篇论文 · 按点赞排序
Zhaoyang Wang, Canwen Xu, Boyi Liu +5 authors
Large language model agents trained in synthetic environments with code-driven simulations and database-backed state transitions demonstrate superior out-of-distribution generalization compared to traditional benchmark-specific approaches.
Yitian Gong, Kuangwei Chen, Zhaoye Fei +9 authors
A fully end-to-end Transformer-based audio tokenizer architecture achieves high-fidelity reconstruction across diverse audio domains and enables superior text-to-speech and automatic speech recognition performance.
Huan Zhang, Yizhan Li, Wenhao Huang +18 authors
AI-driven materials science integrates large language models across discovery pipelines from data curation to agent-based experimentation, emphasizing system-level optimization and autonomous goal pursuit.
Yunze Tong, Mushui Liu, Canyu Zhao +9 authors
TP-GRPO addresses reward sparsity in flow matching models by introducing step-level incremental rewards and identifying turning points to capture long-term effects in denoising trajectories.
Yufan Wen, Zhaocheng Liu, YeGuo Hua +4 authors
NarraScore presents a hierarchical framework that uses frozen Vision-Language Models as affective sensors to generate coherent soundtracks for long-form videos by combining global semantic anchors with token-level adaptive modulation.
Zijie Chen, Zhenghao Lin, Xiao Liu +3 authors
A large-scale scientific question dataset and post-training pipeline are developed to improve open-ended science question answering through enhanced data processing and reinforcement learning with rubric-guided evaluation.
Youjin Wang, Run Zhou, Rong Fu +6 authors
A training-free method called Activation Steering Adapter corrects tool calling behavior in language models by using mid-layer activation interventions guided by a probe and router-conditioned steering vectors.
Haodong Li, Jingwei Wu, Quan Sun +14 authors
A new benchmark and evaluation metric are introduced for assessing temporal coherence and dynamic interaction in GUI generation models, revealing significant challenges in maintaining consistency over extended interaction sequences.
Futing Wang, Jianhao Yan, Yun Luo +6 authors
Models require in-context exploration capabilities to scale effectively at test time, but autoregressive generation faces exponential decay in sampling long sequences, which is addressed by a length-incentivized exploration method that improves performance on both in-domain and out-of-domain tasks.
Xinghao Wang, Pengyu Wang, Xiaoran Liu +4 authors
Prism addresses inefficiencies in block-sparse attention for long-context LLM pre-filling by using a spectral-aware approach that improves block selection accuracy through energy-based temperature calibration.
Ivan Bondarenko, Daniil Grebenkin, Oleg Sedukhin +3 authors
A three-component speech-to-text system combines Wav2Vec2, AST, and Whisper models with curriculum learning and uncertainty modeling to improve transcription accuracy and reduce hallucinations in Russian speech recognition.
Shenyuan Gao, William Liang, Kaiyuan Zheng +27 authors
DreamDojo is a foundation world model trained on 44k hours of egocentric human videos that enables efficient simulation of dexterous robotic tasks through continuous latent actions and real-time distillation.
Xinyu Yang, Chenlong Deng, Tongyu Wen +2 authors
LawThinker is an autonomous legal research agent that uses an Explore-Verify-Memorize strategy with a DeepVerifier module to ensure accurate and procedurally compliant legal reasoning through dynamic verification of intermediate steps.
Jingxuan Wei, Honghao He, Caijun Jia +9 authors
Visual reasoning is enhanced by reconstructing logical structures from compressed visual tokens through a DSL-based approach that generates deterministic visual proofs for verification.
Tony Feng, Trieu H. Trinh, Garrett Bingham +26 authors
Aletheia, a math research agent, demonstrates advanced reasoning capabilities by generating and verifying solutions end-to-end in natural language, achieving autonomous research outcomes from Olympiad problems to PhD-level exercises and contributing to AI-assisted mathematical research.
Huai-Hsun Cheng, Siang-Ling Zhang, Yu-Lun Liu
Progressive Semantic Illusions use a generative framework with dual-branch Score Distillation Sampling to create vector sketches that transform semantically through sequential stroke additions, achieving superior recognizability and illusion strength.
Lianhai Ren, Yucheng Ding, Xiao Liu +3 authors
Training instability in large language models is linked to weight matrix stable rank decline and Jacobian alignment, which MSign addresses through matrix sign operations to prevent gradient explosions.
Yifu Qiu, Zheng Zhao, Waylon Li +4 authors
SWIRL is a self-improvement framework that learns world models from state-only sequences by alternating between forward and inverse dynamics modeling with variational information maximization and ELBO maximization, achieving improved performance on various reasoning and planning benchmarks.
Leheng Sheng, Yongtao Zhang, Wenchang Ma +6 authors
GRU-Mem addresses long-context reasoning challenges in LLMs by incorporating text-controlled gates and reinforcement learning rewards to stabilize memory updates and improve computational efficiency.
Jiazhi Yang, Kunyang Lin, Jinwei Li +10 authors
RISE is a robotic reinforcement learning framework that uses a compositional world model to predict multi-view futures and evaluate imagined outcomes, enabling policy improvement through virtual interactions rather than physical trials.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号