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
Yida Cai, Xin Dai, Bingxiang He +5 authors
LexReward introduces a taxonomy-based framework that evaluates legal responses across style, legal elements, and reasoning chains to build interpretable rewards for preference optimization and reinforcement learning.
You-Zhe Xie, Ting-Wei Chou, Yu-Hsuan Li +3 authors
OuroWorld converts static 3D Gaussian Splatting scenes into looping dynamic 3D cinemagraphs with diverse motion via a vision-language guided video synthesis pipeline and an inconsistency-robust periodic 4D deformation model.
Minxing Li, Minghao Han, Weizhi Zhao +10 authors
Robotic in-context learning is formally defined and implemented via a minimal framework that enables robots to learn tasks from visual demonstrations without large-scale pretraining.
Zhentao Tan, Jianrong Zhang, Ruijie Quan +1 authors
PDE-JEPA uses masked-latent prediction, geometry projection, and structured latent dynamics to learn predictive representations for parametric PDEs, improving accuracy and generalization to unseen parameters.
Hanchen Xia, Baoyou Chen, Yutang Ge +5 authors
REMORY enhances long-horizon agent memory by adding learnable soft tokens to textual summaries, improving decision accuracy with minimal context usage.
Chunghyun Park, Beomjun Kim, Seungcheol Park +5 authors
PointWAM is a 3D world action model that forecasts scene and hand point trajectories to guide robot manipulation, improving dexterous task success through pre-training on human videos.
Ruqing Ning, Haibo Meng, Zhishang Xiang +4 authors
MemAdapter adaptively integrates retrieved memories via counterfactual reasoning, context-aware reflection, and evidence-based reasoning to reduce sycophancy in LLM agents.
Caiqi Zhang, Rujun Han, Zifeng Wang +4 authors
VeriHarness strengthens verification of long-horizon LLM agent outputs by using disagreement and consensus analysis with evidence-backed revision, improving selection and final artifact quality without reference answers.
Yerim Oh, Young-Jun Lee, Jaewoo Ahn +2 authors
AI-generated scientific papers exhibit reasoning failures detectable via structural, argument, and artifact measures, and a harness-level revision framework reduces residual slop by enforcing evidentiary grounding.
Yibo Li, Jinhang Qiu, Zhi Zheng +4 authors
Learn2Play Bench evaluates LLM agents on novel text-based games to measure learning from interaction rather than pretrained reasoning, revealing benefits of full experience retention, a human performance gap, and harness design effects.
Seo Hyun Kim, Sunwoo Hong, Younwoo Choi +3 authors
Pivot-SD improves masked diffusion language models by selectively training high-impact token commitments via information-gain-based self-distillation, outperforming full-sequence and reinforcement learning baselines with minimal data.
Xin Wang, Hao Yu, Zhengyang Zhuge +9 authors
TRACE is an FP4 quantization framework for reinforcement learning training of Mixture-of-Experts language models that aligns training and rollout quantization to reduce discrepancy, enabling efficient low-precision rollout with comparable performance to higher-precision baselines.
Quanyu Long, Xiao Chen, Jianda Chen +4 authors
Trace2Env enables realistic environment simulation for LLM agents by reconstructing historical traces into a reusable worldbook, improving fidelity and long-horizon consistency without rebuilding original systems.
Zihan Su, Junhao Zhuang, Yaowei Li +10 authors
Self Gradient Forcing Plus improves autoregressive video generation by separating context-writing and denoising parameters, enhancing visual quality and long-horizon consistency without extra training data.
Yiming Xu, Hongyue Yu, Beihua Yang +8 authors
DecepEval benchmarks LLM agent deception across professional scenarios using a framework of external inducements to measure condition-dependent vulnerability.
Ying Yang, Guiyu Zhang, Lianghua Huang +5 authors
Spatial Memory Intelligence uses multimodal language models to manage long-range spatial memory in video world models through clustering, sparsification, retrieval, and filtering, improving consistency and stability.
Alham Fikri Aji, Faiz Rizki Ramadhan, Zayd M. K. Zuhri +14 authors
We introduce HyperBrowseComp, a multilingual and multimodal browsing benchmark comprising 423 manually authored and human-validated questions across 13 languages, written by native or highly proficient speakers. Questions are designed to be extremely challenging. Each question targets a concise, publicly verifiable answer whose discovery requires locating obscure evidence, following multi-step clue chains, or inspecting heterogeneous sources such as videos, scanned documents, images, or maps. Easier questions are filtered out by evaluating them with models without internet access to reduce the likelihood that they can be answered with parametric knowledge alone. We evaluate several models using provider-native search and a shared external retrieval harness under a common agent protocol. To contextualize model performance and effort, we also conduct a human evaluation on a sample of the questions. HyperBrowseComp provides a challenging testbed for persistent information seeking across languages and evidence modalities, with difficulty arising from discovering and connecting evidence on the open web.
Ruihan Yu, Yu-Ju Tsai, Muyao Niu +9 authors
EditHero benchmarks long-horizon part-level 3D editing, showing that LLM/VLM agents preserve unedited regions better than non-agentic methods but are slower.
Shaokun Zhang, Yifan Zhang, Jian Hu +5 authors
LoGRA reduces reinforcement learning memory demands for large language models by using low-rank gradient sketches and predicted-KL step control, enabling stable training of large models on limited hardware.
Yiqi Liu, Ruifeng Yuan, Yang Wang +7 authors
Researchers benchmark how video embeddings encode physical properties across simulations, finding a trade-off between cross-modal alignment and quantitative recoverability, and show that retrieved physical references improve generated video fidelity.
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