Can vision-language models turn 2D observations into 3D understanding that supports precise action? This requires inferring object geometry, anticipating the effects of rotation, and refining actions through visual feedback. We introduce RoTA-Bench to evaluate these abilities through interactive 3D-to-2D projection alignment. Given a target silhouette, a model observes a rotated object and chooses successive rotations to recover the target projection within a limited action budget. Our automated pipeline constructs silhouette-preserving 3D objects from 2D icons and uses reversible initialization to guarantee solvability, yielding 674 tasks and a stratified 100-task test set. Empirically, we find that (i) a substantial human-model gap persists: GPT-5.6-Sol succeeds on only 17% of tasks, whereas 20 undergraduate participants collectively complete the test set four times with a 100% success rate; (ii) image-space and pose-space progress can diverge: at matched states, the model more often improves silhouette overlap while worsening pose alignment than humans (19.4% vs. 10.9%); and (iii) fine-grained control is sensitive to interaction history: removing it leaves 96.8% of the model's actions at a coarse 30-degree step size and sharply reduces observed task success. These findings highlight the challenge of converting visual evidence into precise 3D control, beyond recognizing or matching individual images.

MAC-AutoML

Vision-language models struggle to convert 2D visual observations into precise 3D rotational control, showing large gaps versus humans in projection alignment tasks.

Qingchuan Ma

QingchuanMa

Yuexiao Ma

Xiang Meng

Yixuan Zou

Tianyu Xie

alexisty

Wang Chen

Yuhui Zeng

Yuhang Wu

Xiawu Zheng

Rongrong Ji

RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment

Can vision-language models turn 2D observations into 3D understanding that supports precise action? This requires inferring object geometry, anticipating the effects of rotation, and refining actions through visual feedback.

Policy-shift distillation reuses the difference between a teacher's log probabilities before and after reinforcement learning (RL) as supervision on student-generated trajectories. However, this endpoint policy shift cannot reveal whether local action preferences evolved consistently during teacher training. The same endpoint shift can result from early and late shifts that reinforce or partially cancel each other. We investigate whether agreement between early and late teacher policy shifts can guide selective transfer. We introduce Waypoint-guided On-Policy Distillation (Waypoint-OPD), which uses an intermediate teacher checkpoint from the same RL run to compare early and late shifts in relative action preferences at student-visited prefixes. We gate the endpoint shift according to the directional agreement between the early and late shifts, suppressing it when they point in opposing directions. Compared with endpoint-only transfer, our method requires only one additional frozen-teacher scoring pass. Using RL-trained Qwen3-4B teachers, we evaluate our method with Qwen3-8B, Qwen3-14B, and Qwen3-30B-A3B as student models in mathematical and code reasoning. In all six domain--student settings, Waypoint-OPD achieves the highest domain-average score among the compared distillation methods. Ablations with the Qwen3-8B student show that removing or shuffling gates reduces mathematical reasoning performance. These results support the use of directional agreement between early and late teacher policy shifts to weight the endpoint shift during student training.

Waypoint-OPD improves policy-shift distillation by gating endpoint shifts using directional agreement between early and late teacher policy changes, boosting reasoning performance with minimal overhead.

Bingxi Hou

Houbingxi

Guochao Jiang

Nothing2Say

Guofeng Quan

Weiqing Li

Wenfeng Feng

wenfengfwf

Guohua Liu

Yuewei Zhang

Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation

Policy-shift distillation reuses the difference between a teacher's log probabilities before and after reinforcement learning (RL) as supervision on student-generated trajectories. However, this endpoint policy shift cannot reveal whether local action preferences evolved consistently during teacher training.

Harness self-evolution revises an agent's software scaffolding, including prompts, memory, workflows, and executable components, while keeping model weights fixed. We formalize a feedback-driven process in which accepted revisions persist across runs. The framework separates update targets from proposal and admission mechanisms. It distinguishes changes to interpreted state from revisions to tools and code, repository-level control logic, and plugin composition. We characterize search through evolutionary optimization and examine its reliability through continual PAC-learning. Under explicit assumptions, the analysis yields an expected generalization bound, a retention bound conditional on historical regression budgets, and a high-probability admission rule using fresh, independent validation data. We examine how persistent harness revisions affect knowledge sharing and organizational workflows, with implications for scientific reliability, physical safety, and institutional accountability. Cases from prior work illustrate model--harness co-design, distributed execution, and plugin reuse. Reliable deployment requires continual validation as task distributions and evaluators change, with independent safeguards for learning, runtime privileges, and consequential actions. We provide a comprehensive 66-page discussion of related work, theory, algorithms, evaluation, applications, and the broader impact of Harness RSI, together with continuously updated code and supplementary materials, at https://github.com/zgcagi/Harness-RSI.

zgcagi

A framework for self-evolving agent harnesses uses feedback-driven revisions and evolutionary optimization with theoretical generalization guarantees while keeping model weights fixed.

Chuyang Wei

weichy2023

Haoxiang Guan

hxkwan

Minghang Zhu

Kexin Zhou

Jiyan He

ustchjy

Xiren Zhou

Huanhuan Chen

Yifei Shen

yshenaw

Self-Evolving Agent Harness

Harness self-evolution revises an agent's software scaffolding, including prompts, memory, workflows, and executable components, while keeping model weights fixed. We formalize a feedback-driven process in which accepted revisions persist across runs.

Online reinforcement learning (RL) with verifiable rewards has driven rapid progress in mathematics, coding, and tool use, but scaling this success to GUI agents remains challenging due to the high cost of interactive environments and the scarcity of realistic training tasks. We introduce MAI-Gym, an open all-in-one platform that unifies GUI agent evaluation, data collection, and online RL training across mobile, computer, and web environments. Through a shared rollout pipeline, MAI-Gym ensures consistent agent behavior across evaluation and training, supports ten established benchmarks, and scales to 20,000 parallel environment instances. Beyond scalable infrastructure, MAI-Gym enables automated synthesis of diverse GUI applications and challenging tasks, including resettable mocked apps and coherent cross-app worlds with execution-validated task verifiers. Using MAI-Gym, we train MAI-Gym-9B, a unified 9B agent for mobile and computer use, achieving state-of-the-art performance among open GUI agents under 10B parameters: 78.4% on MobileWorld and 76.2% on OSWorld-Verified. Importantly, RL in a mocked train-ticketing app transfers to its real-world counterpart, improving task success on a phys- ical phone from 28.8% to 48.1%. We release the framework, synthesis skills, mocked apps, 2,000 verified tasks, and the model checkpoint to support repro- ducible research and further advances in scalable GUI agent training.

Tongyi-MAI

MAI-Gym is an open platform that unifies GUI agent evaluation, data collection, and online reinforcement learning training across mobile, computer, and web environments, enabling scalable synthesis of verified tasks and training of a unified 9B agent with state-of-the-art open performance.

Panrong Tong

Quyu Kong

kongquyu

Pengxiang Li

PengxiangLi

Tianyu Xia

Hanzhang Zhou

Lei Wang

demolei

Chenglin Cai

Long Li

Long Chen

Jianan Zhang

Zhengwen Zeng

Xu Zhang

Xu0307

Gongjie Zhang

Yue Wang

Steven Hoi

MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments

Online reinforcement learning (RL) with verifiable rewards has driven rapid progress in mathematics, coding, and tool use, but scaling this success to GUI agents remains challenging due to the high cost of interactive environments and the scarcity of realistic training tasks. We introduce MAI-Gym, an open all-in-one platform that unifies GUI agent evaluation, data collection, and online RL training across mobile, computer, and web environments.

Spatial intelligence lets an agent perceive and act in the physical world, and many of its judgments must be made at interaction rates: which spatial relation holds, how far an object is, or where to place it. Vision-language models (VLMs) produce these answers by free-form text generation, paying for autoregressive decoding and answer parsing even though the output space is compact: one of a few options, a scalar, or a 2D point. In light of this, we introduce MJev-Spatial, a System 1 multimodal decision model that tackles all three types of spatial understanding tasks within a unified classification framework. Specifically, MJev-Spatial treats each candidate as a class and scores it using the backbone's pretrained query and key projections, without adding parameters. The candidates are the given options for choice tasks, or value bins and image regions refined through coarse-to-fine classification for numerical and point prediction, respectively. At inference, each classification decision takes a single forward pass and generates no answer tokens. To train on heterogeneous spatial data efficiently, we supervise all coarse-to-fine stages jointly with teacher forcing in one forward pass, and combine this with length-aware global batching and cross-GPU load balancing, raising training throughput by 30.8% over the original batching strategy. Built on Qwen3.5-4B and trained on 12.3M question-answer pairs converted from existing spatial datasets, MJev-Spatial achieves the highest overall score (65.64%) among eight tested models, ranks first on seven of twelve benchmarks, and outperforms its unmodified autoregressive base model on all twelve, while running 2.5 times faster than the fastest autoregressive baseline and 5.6 times faster than the base model.

mindlab-research

MJev-Spatial is a fast multimodal decision model that unifies spatial reasoning tasks into a single-pass classification framework, improving speed and accuracy over autoregressive vision-language models.

Luozheng Qin

Fr0zencr4nE

Yuhan Zhu

Changhai Zhou

Qian Qiao

Jun Gao

Yuhua Zhou

Harry Zhang

Xiangyu Zeng

Yuxuan Liu

Kaijie Chen

Andrew Chen

Pony Ma

MJev-Spatial: System One Spatial Intelligence via Native Attention Matching

Spatial intelligence lets an agent perceive and act in the physical world, and many of its judgments must be made at interaction rates: which spatial relation holds, how far an object is, or where to place it. Vision-language models (VLMs) produce these answers by free-form text generation, paying for autoregressive decoding and answer parsing even though the output space is compact: one of a few options, a scalar, or a 2D point.

Audio source separation models must preserve fine spectral structure while keeping the cost of the separator tractable. Band-split architectures reduce this cost by manually partitioning the spectrum into predefined frequency bands and encoding each band as a token, but the same boundaries are used for every mixture. Because informative spectral structure varies with source content, fixed boundaries can merge discriminative components or spend tokens where little resolution is needed. We introduce Dynamic Band Splitting (DBS), an end-to-end framework that adapts boundary locations and band widths to each mixture under a fixed band-token budget. DBS predicts content-conditioned boundary scores within frequency blocks, selects a prescribed number of boundaries using hard Top-K, and applies content-dependent pooling to obtain compact band tokens. After separation, it smooths and expands these tokens to the original STFT grid, refines them along frequency, and decodes complex masks with a frequency-shared head. DBS improved separation over the corresponding static models on DnR v2, MUSDB18-HQ, and EchoSet, with DBS-BSRNN achieving relative gain of 29.0% in average SI-SDR on DnR v2. These results and boundary analyses on DnR v2 show that content-adaptive frequency partitioning benefits multiple separator backbones under a fixed band-token budget. Code and audio demos are available at https://cslikai.cn/Dynamic-Band-Splitting

thu-ml

Dynamic Band Splitting adapts frequency partition boundaries to each audio mixture, improving source separation quality while maintaining a fixed token budget.

Bingzhe Zhang

Kai Li

Xiaolin Hu

JusperLee

Learning Where to Split: Dynamic Band Splitting for Audio Source Separation

Audio source separation models must preserve fine spectral structure while keeping the cost of the separator tractable. Band-split architectures reduce this cost by manually partitioning the spectrum into predefined frequency bands and encoding each band as a token, but the same boundaries are used for every mixture.

Existing agent benchmarks typically evaluate agents on well-specified tasks defined by issue reports, specifications, or task tickets. Real-world customer delivery is different: forward-deployed engineers (FDEs) must clarify ambiguous requirements, work in live systems, and uncover unstated constraints. We introduce FDE-Gym, a benchmark of 95 end-to-end delivery tasks from 20 customer scenarios across 18 sectors, each adapted by industry experts from real-world delivery projects. In simulated customer environments with scripted stakeholders and modifiable customer systems, agents interact with stakeholders and operate on production-like systems to deliver working solutions. FDE-Gym scores a delivery by how much it improves the customer's own success metric, normalized between 0 (unchanged system) and 1 (reference solution), with penalties for stakeholder interactions (messages, trial runs, and waiting time). An online platform lets humans practice FDE skills under the same grading criteria. We evaluate 13 model configurations from seven providers. In the realistic setting, where unstated constraints must be discovered, the best configuration achieves a mean FDE-Gym score of 0.56 and resolves only 14 of the 95 tasks (a score of at least 0.8 with no incident). Discovering unstated constraints is the primary bottleneck: performance drops sharply when models must uncover them without explicit guidance. Yet information discovery alone does not account for the gap: even on the 23 tasks where the best configuration obtains every key fact from stakeholders, it reaches a normalized business uplift of 0.8 or above on only 10. These results reveal a substantial gap between solving well-specified tasks and delivering effective AI systems for real customers.

DeepLethe

FDE-Gym evaluates AI agents on realistic customer delivery tasks requiring clarification of ambiguous requirements and discovery of unstated constraints, revealing a substantial performance gap compared to well-specified benchmarks.

Dongxu Yang

WaylandYang

Hongda Yan

Renxin Wang

Baoli Zhang

Haojia Zuo

Ying Zhao

Youmin Chen

Xiaojian Liao

Congming Gao

Tianyang Jiang

Yixiao Chen

Qing Wang

Zhiguo Yu

Xiaqing Li

Yuhao Zhang

Jiawen Sun

Renzhe Xu

De Mi

Tiancheng Wu

Xianya Fu

Xuanlu Zhao

Bo Feng

Meixuan Ren

Guanming Lin

Ziheng Zhou

Xinhua Wang

FDE-Gym: The Last Mile of AI Delivery

Existing agent benchmarks typically evaluate agents on well-specified tasks defined by issue reports, specifications, or task tickets. Real-world customer delivery is different: forward-deployed engineers (FDEs) must clarify ambiguous requirements, work in live systems, and uncover unstated constraints.

Feed-forward Gaussian reconstruction builds 3D scenes directly from input images for novel view rendering, but learning Gaussian attributes from predicted geometry for high-quality rendering under large viewpoint changes remains challenging. We analyze limitations in Gaussian attribute learning and introduce InfiniSplat v2, which improves rendering quality through architectural and training changes. Our method decouples geometry and attribute prediction with a frozen geometry branch, aligns predicted Gaussians using ground-truth camera poses during training, and adopts large-angle target view supervision. Across six datasets, the feed-forward outputs of the same trained model achieve state-of-the-art mean LPIPS and DISTS among the evaluated methods in both pose-conditioned and pose-free settings. In the pose-conditioned setting, mean LPIPS and DISTS decrease by 25.8% and 32.9%, respectively, compared with the strongest evaluated baselines. We further find that opacity learned with only image reconstruction supervision suppresses geometrically inconsistent points. Project page: https://zju3dv.github.io/InfiniSplat2/

zju3dv

InfiniSplat v2 improves feed-forward 3D Gaussian reconstruction by decoupling geometry and attribute prediction, aligning Gaussians with ground-truth poses, and using large-angle supervision to enhance novel-view rendering quality.

Yongzhen Hu

Hao Yu

Jiawei Wang

Xianming Zeng

Jianyun Xu

Xingliang Liu

Ruizhen Hu

Xiaowei Zhou

Hujun Bao

Sida Peng

Edisoneh

InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views

Feed-forward Gaussian reconstruction builds 3D scenes directly from input images for novel view rendering, but learning Gaussian attributes from predicted geometry for high-quality rendering under large viewpoint changes remains challenging. We analyze limitations in Gaussian attribute learning and introduce InfiniSplat v2, which improves rendering quality through architectural and training changes.

Reinforcement learning (RL) enables language models to improve text-to-image (T2I) generation by rewriting prompts for a frozen image generator. Existing methods typically apply a single signal derived from the image to individual subword tokens, although objects, counts, attributes, and relations form coherent semantic spans. We propose SCOPE, an objective based on Group Relative Policy Optimization (GRPO) that aligns update weighting with this structure. The rewriter jointly generates an enhanced prompt and its semantic partition. SCOPE pools shifts in policy log probabilities within each span to derive a shared coefficient for its token losses while retaining a single advantage for each completion. It requires neither separate span rewards nor additional image renders. Across Qwen-Image, SD3, and FLUX.1-dev, SCOPE achieves average relative gains over vanilla GRPO of 9.48% on GeneVal and 7.62% on T2I-CompBench. Ablation studies show that both semantic boundaries and pooling within spans contribute to these gains, supporting semantic structure as a practical inductive bias for prompt engineering with RL.

university-of-rochester

SCOPE improves text-to-image prompt rewriting via reinforcement learning by aligning policy updates with semantic spans, boosting generation quality without extra rendering.

Yufei Jiao

Jinfa Huang

Zeyue Xue

Jiebo Luo

Jinfa

SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation

Reinforcement learning (RL) enables language models to improve text-to-image (T2I) generation by rewriting prompts for a frozen image generator. Existing methods typically apply a single signal derived from the image to individual subword tokens, although objects, counts, attributes, and relations form coherent semantic spans.

Sparse autoencoders (SAEs) represent model activations with sparse, linearly decoded features, but the downstream effects of those features need not compose additively. We introduce SAE Additivity Residuals (SAR), which measure the behavioral variation left unexplained above a chosen interaction order. Under independent feature subset sampling, population SAR is exactly the normalized energy of interactions above that order and bounds coalition ordering error for the optimal low order approximation. To distinguish compositionality associated with the SAE representation from that induced by downstream computation, we compare learned SAE coordinates with sum preserving random rotations of the same selected subspace and find lower first order residuals for the learned coordinates across all 18 model and behavior settings. The coordinates are nevertheless not fully additive: pairwise terms improve held out coalition ordering in every setting and outperform parameter matched models using randomly selected third order terms. In a combined steering experiment, pairwise modeling raises rank fidelity from 0.5874 to 0.7790 and reduces ordering error from 0.1444 to 0.0606. Together, these results show that SAE coordinates support comparatively compositional behavioral descriptions while faithful joint explanations and interventions still require interactions beyond individual feature effects. Code is available at https://github.com/OpenSAE/SAE-Behavioral-Compositionality.

OpenSAE

Sparse autoencoder features are partially compositional but require interaction terms for accurate behavioral predictions and steering.

Haotian Ma

Hunter26666666

Ziye Mei

Frederic Sala

Behavioral Compositionality of Sparse Autoencoder Representations

Sparse autoencoders (SAEs) represent model activations with sparse, linearly decoded features, but the downstream effects of those features need not compose additively. We introduce SAE Additivity Residuals (SAR), which measure the behavioral variation left unexplained above a chosen interaction order.

Computer use has reached a turning point: the focus is expanding from GUI-only interaction to systems that complete tasks on users' computers through visual and programmatic interfaces. We define a generalist computer-use agent (GCUA) as a system in which the model chooses how to observe, act, and organize its work across visual and programmatic interfaces. Yet frontier implementations, such as GPT-6 Astra's computer-use system, remain largely closed, limiting systematic study of their design. We introduce Open Generalist CUA (OpenGCUA), an ongoing project to develop and open the full stack for GCUAs with the community. This report presents the first release of OpenGCUA. Code execution is a key feature of our agent design: it lets intermediate results determine subsequent operations and enables models to apply their existing programming capabilities directly to computer interaction. We also propose harness-to-model decision transfer, moving selected observations and task-state decisions from fixed harness rules to the model. On OSWorld-v2, our full configuration improves task scores across all eight evaluated models, including GPT-6 Astra and Claude Opus 5, raising mean partial score by +19.6 points and reducing reported steps per task by an average of 53.0% relative to batched GUI-only execution. We release our implementation and prompts, evaluation trajectories from all eight models, and preliminary versions of a real-device harness and a cross-device SDK. The harness addresses real-device challenges not fully captured by benchmarks, while the SDK provides a common programming interface across computers, mobiles and browsers. Across four real macOS tasks under our real-device harness, selected Qwen3.8-Max runs use 19.4 times as many reported reasoning tokens as GPT-6 Astra, highlighting substantial differences between models in real-device execution efficiency.

OpenGCUA introduces an open generalist computer-use agent that integrates visual and programmatic interfaces, harness-to-model decision transfer, and code execution to improve task performance and efficiency across models.

xtyXpastor

Chen Liu

OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use

Computer use has reached a turning point: the focus is expanding from GUI-only interaction to systems that complete tasks on users' computers through visual and programmatic interfaces. We define a generalist computer-use agent (GCUA) as a system in which the model chooses how to observe, act, and organize its work across visual and programmatic interfaces.

With the rapid development of multimodal large language models (MLLMs), evaluating their video understanding capabilities has become increasingly important. Existing video understanding benchmarks cover only part of the evaluation dimensions, limiting systematic diagnosis of omni-modal video understanding. We introduce OmniCompass, a holistic benchmark for omni-modal video grounding and understanding. OmniCompass defines two evaluation tracks: OmniCompass-VL assesses vision-language models on tasks requiring visual evidence, while OmniCompass-AV assesses omni-modal models on tasks requiring joint audio-visual evidence. Each track includes three complementary tasks: temporal grounding (TG), spatial grounding (SG), and question answering (QA), assessing when a referenced target appears, where it is, and what happens. QA covers five perception and three reasoning categories in OmniCompass-VL, and three perception and two reasoning categories in OmniCompass-AV. OmniCompass-VL contains 820 videos and 2,460 instances across domains and three duration types: short, medium, and long. A narrative- and audio-rich subset forms OmniCompass-AV with 240 videos and 720 instances. Each instance is evaluated under two paired settings: text-only and text+image, specifying the same target through text and visual-language guidance. All instances are manually annotated with multi-round quality verification. Evaluations of 50 MLLMs show that open-source models still lag behind proprietary counterparts, particularly in audio-visual understanding. Precise temporal and spatial grounding remains challenging even for leading models, and their predictions still exhibit inconsistencies across the paired target specifications.

VCLab-HKPU

OmniCompass is a holistic benchmark for omni-modal video grounding and understanding that evaluates vision-language and audio-visual reasoning through temporal grounding, spatial grounding, and question answering tasks.

Xuyang Liu

xuyang-liu16

Shihao Wang

Xinyu Wei

Minghan Li

Zeqing Wang

Zuhao Yang

Zichen Wen

Lei Zhang

VCLab-PolyU

OmniCompass: A Holistic Benchmark for Omni-Modal Video Grounding and Understanding

With the rapid development of multimodal large language models (MLLMs), evaluating their video understanding capabilities has become increasingly important. Existing video understanding benchmarks cover only part of the evaluation dimensions, limiting systematic diagnosis of omni-modal video understanding.

AI Research Made Simple