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
Xu Wang, Kaixiang Yao, Miao Pan +4 authors
ProVisE enables standardized evaluation of image-generation models on spatial tasks via visual protocols, revealing complementary strengths between pixel-space expression and text-based reasoning.
Aleksandr Kutsakov, Mariia Sadovina, Georgii Gospodinov +4 authors
A time-aware audio LLM answers questions with explicit timestamps across long recordings by interleaving periodic time markers with continuous audio tokens, achieving strong temporal grounding and supporting time-anchored descriptions.
Paul Furgale, Severin Klingler, James Nolan +12 authors
NOOA treats AI agents as Python objects whose methods and fields define actions, state, and prompts, enabling deterministic testing and LLM-driven runtime completion within a unified programming model.
Junhao Zhuang, Shiyi Zhang, Yuxuan Bian +11 authors
Self Gradient Forcing restores missing supervision for historical memory writing in autoregressive video diffusion by using a two-pass strategy that trains context key-value representations via future-frame losses without full rollout backpropagation.
Junyao Yang, Yucheng Shi, Zongxia Li +6 authors
SAT adapts PPO clipping to staleness in asynchronous RL, tightening updates for high-mismatch rollouts and improving stability on large mixture-of-experts models.
Jin Xu, Kangdi Wang, Ruibin Yuan +24 authors
Qwen-Music is a music generation system that uses tokenized semantic representations, autoregressive language modeling with melody planning, and generative rendering to produce high-fidelity songs from text or reference audio.
Andrei Kuzmenko, Alexandr Maximenko, Aleksandr Kutsakov +5 authors
A Conformer encoder pre-trained with HuBERT-style objectives and cluster-level balancing improves ASR for underrepresented Central Asian languages.
Kailin Jiang, Lei Liu, Jian Xi +8 authors
A framework evaluates document sets for generation quality, diagnoses reranker weaknesses, and optimizes selection via rubric-based signals to improve downstream outputs.
Bowen Zheng, Chao Yi, Dian Chen +21 authors
RecGPT-V3 improves large-scale recommendation by using persistent user memory, hybrid text and semantic-ID reasoning, and compressed latent reasoning to boost engagement and cut serving costs.
Jingyan Shen, Ang Li, Salman Rahman +4 authors
Using chess and math as controlled testbeds, the study shows that pretraining loss predicts reinforcement learning gains and that RL amplifies or surfaces correct reasoning differently by difficulty, linking pretraining scale to post-training reasoning performance.
Guangran Cheng, Chengqi Lyu, Songyang Gao +2 authors
GEPO extends GRPO with group-level entropy-conditioned advantage shaping to balance exploration and exploitation across heterogeneous tasks during LLM reinforcement learning.
Suzhen Zhong, Shayan Noei, Bram Adams +1 authors
The study analyzes over a million pull requests across three code review eras to show that AI agent collaboration accelerates review decisions but does not improve quality, highlighting human-AI interaction patterns as key efficiency drivers.
Tencent WorkBuddy Bench Team, Siqi Cai, Shaopeng Chen +35 authors
We introduce Tencent WorkBuddy Bench, a multi-domain evaluation suite for coding agents; this report documents its construction methodology, scoring protocol, and a cross-model leaderboard. At its core is a unified evaluation framework for constructing and running distribution-informed coding-agent tasks across four work domains - Code, Web, Office, and Security. Rather than adapting public issue text, every task is reverse-engineered from a real commit, pull request, or business scenario and rewritten as a short, colloquial, role-played request, so that a task's prompt is not recoverable by web-searching the underlying issue, pull request, or commit thread. Because the dataset is released openly - task directories, environment images, evaluation harness, tests, and reference solutions - contamination resistance rests on this construction together with dataset versioning rather than on secrecy. The four subsets - repository-level engineering, front-end development, office and business workflows, and red-/blue-team security - probe complementary facets of real work, each with its own verification style. All are packaged in a uniform task-directory format and run, under a uniform and reproducible protocol, on two agent harnesses (CodeBuddy Code and Claude Code); the full open release makes the benchmark reproducible end to end and directly auditable, since any third party can re-run each task and inspect its content. Because each subset uses a different scoring instrument, scores are not comparable across subsets and the suite reports no suite-wide average. We report a cross-model leaderboard across several model families.
Hyunin Lee, Jinglue Xu, Jeffrey Seely +3 authors
Recursive Harness Self-Improvement optimizes user-built agent harnesses via iterative prompt-level refinement to enhance execution traces and agent performance with minimal computation.
Mantha Sai Gopal, Jaison Saji Chacko, Harsh Nandwana +3 authors
REBASE improves training-free in-context segmentation by removing background feature subspaces to suppress spurious correspondences and refine prompt localization.
Jihoon Tack, Philippe Laban, Jennifer Neville
Large language models show significant performance drops when user intent evolves across multi-turn conversations, revealing a critical gap in tracking dynamic goals.
Kunlun Zhu, Xuyan Ye, Zhiguang Han +9 authors
AgentDebugX is an open-source framework that closes the debugging loop through multi-turn root-cause diagnosis and automated recovery for LLM agents.
Jiarui Zhang, Muzi Tao, Shangshang Wang +3 authors
ActiveVision reveals that current multimodal large language models lack robust active visual observation, performing far below humans on tasks requiring repeated perception.
Yuxuan Luo, Peng Zhang, Xinjie Zhang +3 authors
SciForma improves scientific diagram generation by decomposing structural quality into component, arrow, and text axes, using multi-dimensional preference optimization and iterative editing to achieve high structural fidelity.
Ruikang Li, Molin Li, Jiarui Wu +3 authors
A learned end-to-end framework couples smartphone cameras and displays to reduce color reproduction errors across the full capture-to-display pipeline.
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