Curated research — 48 papers
Can AI agents handle real professional work? OccuBench evaluates agents across 100 tasks in 65 specialized domains using language world models, revealing critical gaps in professional task performance.
2026-04-13T00:00:00+00:00
What if reward models could explain their reasoning? RationalRewards teaches reward models to produce explicit critiques before scoring, turning passive evaluators into active optimization tools that improve visual generation at both training and test time.
2026-04-13T00:00:00+00:00
GameWorld introduces a standardized benchmark for evaluating multimodal AI agents in browser-based video games, tackling heterogeneous action interfaces and heuristic verification challenges.
What if the secret to better LLM reasoning is giving hints that are just enough — not too much, not too little? KnowRL breaks problems into atomic Knowledge Points and uses Constrained Subset Search to find the minimal hint that unblocks exploration without leaking answers. On a 1.5B model, it beats GRPO by +9.63 points across 8 benchmarks.
AiScientist treats long-horizon ML research engineering as a systems problem: thin orchestrator control over thick durable state. The File-as-Bus workspace delivers +10.54 pts on PaperBench and 81.82 Any Medal% on MLE-Bench Lite.
Open-source full-stack framework for GUI agents: online RL training, reproducible evaluation across 6 benchmarks and 11+ models, and real-device deployment — +17.1% on ScreenSpot-Pro.
First comprehensive survey on Attention Sink — why Transformers waste attention on uninformative tokens. Covers utilization (StreamingLLM), interpretation, and mitigation across 200+ papers.
A memory-enhanced reward shaping framework that clusters recurring LLM errors and penalizes repetition, boosting pass@1 by up to +4.13 on math benchmarks.
Can LLMs generate correct quantum code across multiple frameworks? QuanBench+ benchmarks code generation for Qiskit, PennyLane, and Cirq, revealing that feedback-based repair boosts Pass@1 from 59.5% to 83.3%.
RefineAnything surgically fixes distorted text, logos, and faces in AI-generated images — a bounding box plus a reference image is all it takes to restore pixel-perfect local details.
FORGE reveals that domain knowledge—not visual grounding—is the bottleneck for manufacturing AI. A 3B model fine-tuned on FORGE matches a model 78× larger.
WildDet3D brings 3D object detection in the wild — any object, any prompt, any image — with a 10× leap over prior state-of-the-art.
What if every agent failure made all agents smarter? SkillClaw shows collective skill evolution is possible.
SFT doesn't just memorize — it generalizes conditionally. Three factors decide when: optimization depth, data quality, and model capability. The cost: reasoning improves, but safety degrades.
The first systematic map of the agentic skill layer — from formal definition to marketplace attacks. This SoK reveals 7 design patterns, introduces trust tiers for skill governance, and documents the ClawHavoc supply-chain attack that compromised 36.8% of marketplace users.
High-frequency text isn't just easier to read — it makes LLMs significantly smarter. Adam's Law proposes TFL, TFD, and CTFT to harness this principle across 4 NLP tasks.
RL-trained LLM agents silently collapse into repetitive templates despite high entropy. Mutual information (+0.39 Spearman) beats entropy (-0.14) as a diagnostic, and SNR-Aware Filtering restores diverse reasoning across 4 environments.
What if AI could paint like a human artist — sketching, inspecting, and refining step by step? This paper introduces process-driven image generation with BAGEL-7B, achieving GenEval 0.83 (+5%) and WISE 0.76 (+6%) through Plan→Sketch→Inspect→Refine cycles.
SkillNet introduces an open infrastructure for creating, evaluating, and connecting AI agent skills at scale, featuring a unified ontology over 200,000+ skills that boosts average rewards by 40% and cuts execution steps by 30% on ALFWorld, WebShop, and ScienceWorld.
What if LLMs had their own operating system for memory? MemOS unifies plaintext, KV cache, and model weights as schedulable resources—achieving state-of-the-art on all major memory benchmarks.
A new training paradigm for IR systems: learning to retrieve from agent trajectories bridges the gap between human-designed search and LLM-powered agent consumption.
Claw-Eval introduces trajectory-aware grading, safety evaluation, and multimodal coverage to build trustworthy benchmarks for autonomous LLM agents.
Video-MME-v2 exposes a 41-point AI-human gap in video understanding, using group-based evaluation to reveal hidden failures in consistency and multimodal reasoning.
VibeVoice synthesizes 90-minute, 4-speaker conversations using next-token diffusion with a 7.5 Hz tokenizer that compresses speech 80× vs Encodec — making long-form multi-speaker TTS feasible in a standard LLM context window.
InCoder-32B-Thinking bridges general and industrial code intelligence through Error-driven Chain-of-Thought synthesis and an Industrial Code World Model, achieving 81.3% on LiveCodeBench and 84.0% on CAD-Coder.
Current top AI models score just 56.3% on tasks humans ace at 93.8% — this benchmark finally exposes exactly where and why multimodal agents fail.
Token warping — rearranging ViT image tokens rather than pixels — enables MLLMs to reason from nearby viewpoints without fine-tuning, consistently outperforming all baselines on the new ViewBench benchmark.
DataFlex unifies dynamic data selection, domain mixture optimization, and sample reweighting into a single LLaMA-Factory-compatible framework, enabling reproducible data-centric LLM training with consistent MMLU gains.
A minimal 4-frame sliding-window baseline beats all published streaming video models with half the GPU memory.
RLSD solves the information leakage problem of on-policy self-distillation by repurposing the teacher as a token-level magnitude evaluator, achieving state-of-the-art on 5 multimodal reasoning benchmarks.
A unified foundation model that jointly generates synchronized audio and video from text prompts, eliminating the need for separate audio and video pipelines.
A simple architectural modification to Transformers that feeds attention outputs back as residuals, improving reasoning and long-context performance without additional parameters.
A three-layer security framework for autonomous AI agents that protects against data leakage, privilege escalation, and malicious tool execution in real-time.
This paper introduces Natural-Language Agent Harnesses (NLAHs), showing that agent control logic can be expressed in editable text rather than code — achieving a 55% performance boost when migrating from code to natural language.
LlamaFactory provides a unified framework for fine-tuning 100+ language models with minimal code, supporting LoRA, QLoRA, RLHF and more out of the box.
PaperBanana automates publication-ready academic illustrations using VLM-powered agents — a potential game-changer for the AI research workflow.
CARLA-Air unifies drone flight and autonomous driving in a single simulation, enabling air-ground cooperative AI research without co-simulation overhead.
Smaller models beat larger ones by auto-synthesizing code harnesses that eliminate illegal moves — a new paradigm for efficient agent design.
A coding agent that automatically discovers better LLM harnesses—achieving rank #1 on TerminalBench-2 and +7.7 points over ACE on text classification, using filesystem access for causal diagnosis.
Could AI achieve superintelligence by building its own tools? This paper argues yes — through the Diligent Learner framework combining test-time search with tool-building.
ACE (Agentic Context Engineering) treats LLM contexts as evolving playbooks, achieving +10.6% on agent benchmarks through systematic context optimization.
RealRestorer leverages large-scale image editing models to achieve generalizable real-world image restoration, tackling complex degradations that previous methods couldn't handle.
Voxtral TTS by Mistral AI generates highly natural multilingual speech from minimal data, setting a new standard for expressive text-to-speech.
Calibri reveals hidden capacity in Diffusion Transformers through lightweight parameter-efficient calibration, achieving significant quality gains with minimal compute overhead.
PixelSmile achieves precise facial expression editing at the pixel level, overcoming semantic overlap issues that plagued prior methods.
Intern-S1-Pro is the world's first trillion-parameter scientific multimodal model, setting new state-of-the-art across 50+ scientific benchmarks.
CUA-Suite unlocks computer-use agent training at scale with 55 hours of expert-annotated desktop video — the largest benchmark of its kind.
Xiaomeng Hu, Yinger Zhang, Fei Huang et al.
Can AI agents handle real professional work? OccuBench evaluates agents across 100 tasks in 65 specialized domains using language world models, revealing critical gaps in professional task performance.
Haozhe Wang, Cong Wei, Weiming Ren et al.
What if reward models could explain their reasoning? RationalRewards teaches reward models to produce explicit critiques before scoring, turning passive evaluators into active optimization tools that improve visual generation at both training and test time.
Mingyu Ouyang, Siyuan Hu, Kevin Qinghong Lin et al.
GameWorld introduces a standardized benchmark for evaluating multimodal AI agents in browser-based video games, tackling heterogeneous action interfaces and heuristic verification challenges.
Linhao Yu, Xingguang Ji, Zhenghan Chen et al.
What if the secret to better LLM reasoning is giving hints that are just enough — not too much, not too little? KnowRL breaks problems into atomic Knowledge Points and uses Constrained Subset Search to find the minimal hint that unblocks exploration without leaking answers. On a 1.5B model, it beats GRPO by +9.63 points across 8 benchmarks.
Guoxin Chen, Jie Chen, Lei Chen et al.
AiScientist treats long-horizon ML research engineering as a systems problem: thin orchestrator control over thick durable state. The File-as-Bus workspace delivers +10.54 pts on PaperBench and 81.82 Any Medal% on MLE-Bench Lite.
Fei Tang, Zhiqiong Lu, Boxuan Zhang et al.
Open-source full-stack framework for GUI agents: online RL training, reproducible evaluation across 6 benchmarks and 11+ models, and real-device deployment — +17.1% on ScreenSpot-Pro.
Zunhai Su, Hengyuan Zhang, Wei Wu et al.
First comprehensive survey on Attention Sink — why Transformers waste attention on uninformative tokens. Covers utilization (StreamingLLM), interpretation, and mitigation across 200+ papers.
Yang Liu, Enxi Wang, Yufei Gao et al.
A memory-enhanced reward shaping framework that clusters recurring LLM errors and penalizes repetition, boosting pass@1 by up to +4.13 on math benchmarks.
Ali Slim, Haydar Hamieh, Jawad Kotaich et al.
Can LLMs generate correct quantum code across multiple frameworks? QuanBench+ benchmarks code generation for Qiskit, PennyLane, and Cirq, revealing that feedback-based repair boosts Pass@1 from 59.5% to 83.3%.
Dewei Zhou, You Li, Zongxin Yang et al.
RefineAnything surgically fixes distorted text, logos, and faces in AI-generated images — a bounding box plus a reference image is all it takes to restore pixel-perfect local details.
Xiangru Jian, Hao Xu, Wei Pang et al.
FORGE reveals that domain knowledge—not visual grounding—is the bottleneck for manufacturing AI. A 3B model fine-tuned on FORGE matches a model 78× larger.
Weikai Huang, Jieyu Zhang, Sijun Li et al.
WildDet3D brings 3D object detection in the wild — any object, any prompt, any image — with a 10× leap over prior state-of-the-art.
Ziyu Ma, Shidong Yang, Yuxiang Ji et al.
What if every agent failure made all agents smarter? SkillClaw shows collective skill evolution is possible.
Qihan Ren, Peng Wang, Ruikun Cai et al.
SFT doesn't just memorize — it generalizes conditionally. Three factors decide when: optimization depth, data quality, and model capability. The cost: reasoning improves, but safety degrades.
Yanna Jiang, Delong Li, Haiyu Deng et al.
The first systematic map of the agentic skill layer — from formal definition to marketplace attacks. This SoK reveals 7 design patterns, introduces trust tiers for skill governance, and documents the ClawHavoc supply-chain attack that compromised 36.8% of marketplace users.
Hongyuan Adam Lu, Z. L., Victor Wei et al.
High-frequency text isn't just easier to read — it makes LLMs significantly smarter. Adam's Law proposes TFL, TFD, and CTFT to harness this principle across 4 NLP tasks.
Zhefei Yu, Sipeng Zheng, Kun Shao et al.
RL-trained LLM agents silently collapse into repetitive templates despite high entropy. Mutual information (+0.39 Spearman) beats entropy (-0.14) as a diagnostic, and SNR-Aware Filtering restores diverse reasoning across 4 environments.
Lei Zhang, Junjiao Tian, Zhipeng Fan et al.
What if AI could paint like a human artist — sketching, inspecting, and refining step by step? This paper introduces process-driven image generation with BAGEL-7B, achieving GenEval 0.83 (+5%) and WISE 0.76 (+6%) through Plan→Sketch→Inspect→Refine cycles.
Yuan Liang, Ruobin Zhong, Haoming Xu et al.
SkillNet introduces an open infrastructure for creating, evaluating, and connecting AI agent skills at scale, featuring a unified ontology over 200,000+ skills that boosts average rewards by 40% and cuts execution steps by 30% on ALFWorld, WebShop, and ScienceWorld.
Zhiyu Li, Shichao Song, Chenyang Xi et al.
What if LLMs had their own operating system for memory? MemOS unifies plaintext, KV cache, and model weights as schedulable resources—achieving state-of-the-art on all major memory benchmarks.
Yuqi Zhou, Sunhao Dai, Changle Qu et al.
A new training paradigm for IR systems: learning to retrieve from agent trajectories bridges the gap between human-designed search and LLM-powered agent consumption.
Bowen Ye, Rang Li, Qibin Yang et al.
Claw-Eval introduces trajectory-aware grading, safety evaluation, and multimodal coverage to build trustworthy benchmarks for autonomous LLM agents.
Chaoyou Fu, Haozhi Yuan, Yuhao Dong et al.
Video-MME-v2 exposes a 41-point AI-human gap in video understanding, using group-based evaluation to reveal hidden failures in consistency and multimodal reasoning.
Zhiliang Peng, Jianwei Yu, Wenhui Wang et al.
VibeVoice synthesizes 90-minute, 4-speaker conversations using next-token diffusion with a 7.5 Hz tokenizer that compresses speech 80× vs Encodec — making long-form multi-speaker TTS feasible in a standard LLM context window.
Jian Yang, Wei Zhang, Jiajun Wu et al.
InCoder-32B-Thinking bridges general and industrial code intelligence through Error-driven Chain-of-Thought synthesis and an Industrial Code World Model, achieving 81.3% on LiveCodeBench and 84.0% on CAD-Coder.
Qianshan Wei, Yishan Yang, Siyi Wang et al.
Current top AI models score just 56.3% on tasks humans ace at 93.8% — this benchmark finally exposes exactly where and why multimodal agents fail.
Phillip Y. Lee, Chanho Park, Mingue Park et al.
Token warping — rearranging ViT image tokens rather than pixels — enables MLLMs to reason from nearby viewpoints without fine-tuning, consistently outperforming all baselines on the new ViewBench benchmark.
Hao Liang, Zhengyang Zhao, Meiyi Qiang et al.
DataFlex unifies dynamic data selection, domain mixture optimization, and sample reweighting into a single LLaMA-Factory-compatible framework, enabling reproducible data-centric LLM training with consistent MMLU gains.
Yujiao Shen, Shulin Tian, Jingkang Yang et al.
A minimal 4-frame sliding-window baseline beats all published streaming video models with half the GPU memory.
Chenxu Yang, Chuanyu Qin, Qingyi Si et al.
RLSD solves the information leakage problem of on-policy self-distillation by repurposing the teacher as a token-level magnitude evaluator, achieving state-of-the-art on 5 multimodal reasoning benchmarks.
Yoav HaCohen, Benny Brazowski, Nisan Chiprut et al.
A unified foundation model that jointly generates synchronized audio and video from text prompts, eliminating the need for separate audio and video pipelines.
Kimi Team, Guangyu Chen, Yu Zhang et al.
A simple architectural modification to Transformers that feeds attention outputs back as residuals, improving reasoning and long-context performance without additional parameters.
Songyang Liu, Chaozhuo Li, Chenxu Wang et al.
A three-layer security framework for autonomous AI agents that protects against data leakage, privilege escalation, and malicious tool execution in real-time.
Linyue Pan, Lexiao Zou, Shuo Guo et al.
This paper introduces Natural-Language Agent Harnesses (NLAHs), showing that agent control logic can be expressed in editable text rather than code — achieving a 55% performance boost when migrating from code to natural language.
LlamaFactory provides a unified framework for fine-tuning 100+ language models with minimal code, supporting LoRA, QLoRA, RLHF and more out of the box.
Dawei Zhu, Rui Meng, Yale Song et al.
PaperBanana automates publication-ready academic illustrations using VLM-powered agents — a potential game-changer for the AI research workflow.
CARLA-Air unifies drone flight and autonomous driving in a single simulation, enabling air-ground cooperative AI research without co-simulation overhead.
Xinghua Lou, Miguel Lázaro-Gredilla, Antoine Dedieu et al.
Smaller models beat larger ones by auto-synthesizing code harnesses that eliminate illegal moves — a new paradigm for efficient agent design.
Aiden Grossman, Sanyam Kapoor, Arjun Desai et al.
A coding agent that automatically discovers better LLM harnesses—achieving rank #1 on TerminalBench-2 and +7.7 points over ACE on text classification, using filesystem access for causal diagnosis.
David Koplow, Tomer Galanti, Tomaso Poggio
Could AI achieve superintelligence by building its own tools? This paper argues yes — through the Diligent Learner framework combining test-time search with tool-building.
Ahanaf Tazwar Shamim, Farhan Sadik, Taiyeong Lee
ACE (Agentic Context Engineering) treats LLM contexts as evolving playbooks, achieving +10.6% on agent benchmarks through systematic context optimization.
Yufeng Yang, Xianfang Zeng, Zhangqi Jiang et al.
RealRestorer leverages large-scale image editing models to achieve generalizable real-world image restoration, tackling complex degradations that previous methods couldn't handle.
Alexander H. Liu, Alexis Tacnet, Andy Ehrenberg et al.
Voxtral TTS by Mistral AI generates highly natural multilingual speech from minimal data, setting a new standard for expressive text-to-speech.
Danil Tokhchukov, Aysel Mirzoeva, Andrey Kuznetsov et al.
Calibri reveals hidden capacity in Diffusion Transformers through lightweight parameter-efficient calibration, achieving significant quality gains with minimal compute overhead.
Jiabin Hua, Hengyuan Xu, Aojie Li et al.
PixelSmile achieves precise facial expression editing at the pixel level, overcoming semantic overlap issues that plagued prior methods.
Yicheng Zou, Dongsheng Zhu, Lin Zhu et al.
Intern-S1-Pro is the world's first trillion-parameter scientific multimodal model, setting new state-of-the-art across 50+ scientific benchmarks.
CUA-Suite Team
CUA-Suite unlocks computer-use agent training at scale with 55 hours of expert-annotated desktop video — the largest benchmark of its kind.