Back to newsletter
·Daily digest

🔬 Research Pulse

Daily Digest

May 26, 2026

🤖 AI

🧠 LLMs

1. MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research

Authors: Dingbang Wu, Rui Hao, Haiyang Wang... Published: 2026-05-25 | Citations: 0 arXiv | PDF

Research Question: How can we build a verifiable, scalable simulation environment for training and evaluating mobile GUI agents on everyday apps, without relying on proprietary backends or unreliable free-text matching for evaluation?

Summary: MobileGym is a browser-hosted, JSON-state mobile GUI simulation platform designed for verifiable evaluation and scalable RL training of mobile agents. It pairs deterministic state-based judges with hundreds of parallel rollouts, and demonstrates that policies trained in simulation (GRPO on Qwen3-VL-4B) transfer to real devices while retaining 95.1% of training gains.

Key Results: MobileGym hosts hundreds of parallel instances per server at ~400 MB memory and ~3s cold start per instance. MobileGym-Bench provides 416 parameterized templates (256 test / 160 train) across 28 apps. GRPO training on Qwen3-VL-4B-Instruct yields +12.8 percentage points on the 256-task test set. Sim-to-real transfer retains 95.1% of simulation training gains on a 59-task real-device subset.

Key Findings:

  • Structured JSON state with deterministic judges eliminates free-text matching failures and provides dense RL reward signal simultaneously
  • Lightweight browser hosting (~400 MB/instance, ~3s cold start) enables hundreds of parallel rollouts on a single server, making online RL practical
  • GRPO training on Qwen3-VL-4B-Instruct improves performance by +12.8 pp on test tasks, and 95.1% of this gain transfers to real-device execution

Technical Novelty: Combines three new pieces: (1) a browser-hosted simulation of everyday mobile apps using structured JSON state that can be forked/compared deterministically, (2) a declarative task-definition framework with parameterized templates, and (3) a single programmatic judging mechanism that serves both as evaluation verdict and dense RL reward — replacing prior free-text matching and emulator-based approaches.

What's New: Prior mobile agent benchmarks either use real emulators (slow, expensive to parallelize) or static traces (no interactive RL). MobileGym is the first to combine high-throughput parallel simulation of everyday apps with deterministic JSON-state judging, enabling both rigorous evaluation and online RL on the same substrate.

Extension Opportunities:

  • Expand the 28-app coverage to include more complex apps (banking, ride-sharing, social media with dynamic content) and test how the layered state model scales to apps with heavier server-side dependencies
  • Use the deterministic JSON state and dense rewards to explore curriculum learning or self-play between agents on multi-step cross-app workflows
  • Build a model-agnostic evaluation harness on top of MobileGym-Bench to systematically compare proprietary (GPT-4V, Claude) vs. open VLMs on identical parameterized tasks with verifiable judges

Replicability: Project page at mobilegym.github.io suggests open release of the platform and MobileGym-Bench. Compute is modest by simulation standards (single server hosts hundreds of instances at 400 MB each); RL fine-tuning used Qwen3-VL-4B-Instruct with GRPO, which is feasible on a small multi-GPU node. Real-device validation requires physical Android devices for the 59-task subset.

Research Gaps:

  • Interaction fidelity is prioritized over backend replication — unclear how well agents trained on simulated apps generalize to apps with heavy server-side logic (payments, real-time data)
  • Sim-to-real validation is on a small 59-task subset; broader real-device benchmarking and failure-mode analysis is needed

2. Claw-Anything: Benchmarking Always-On Personal Assistants with Broader Access to User's Digital World

Authors: Yusong Lin, Xinyuan Liang, Haiyang Wang... Published: 2026-05-25 | Citations: 0 arXiv | PDF

Research Question: How can we benchmark always-on personal assistant agents in realistic settings where they have broad access to a user's digital world (long activity histories, interdependent backends, multi-device GUI/CLI), rather than the narrow slices current benchmarks evaluate?

Summary: Claw-Anything is a benchmark that evaluates LLM agents as always-on personal assistants by simulating months of multi-device user activity, interdependent services, and noisy signals. It shows GPT-5.5 reaches only 34.5% pass@1, and ships a data-generation pipeline (2K environments) that boosts a base model by 23.7%.

Key Results: Introduces Claw-Anything benchmark; GPT-5.5 achieves only 34.5% pass@1, far below prior benchmarks, exposing capability gaps. The automated data-generation pipeline produces 2,000 training environments and improves the base model by 23.7%, demonstrating scalable training utility.

Key Findings:

  • Current frontier agents (GPT-5.5 at 34.5% pass@1) perform substantially worse on broad-context always-on tasks than on prior narrow benchmarks.
  • Agents struggle with noise robustness — irrelevant events and conflicting signals materially degrade performance.
  • Scalable synthetic environment generation (2,000 envs) yields a 23.7% improvement, suggesting data infrastructure is a key lever for always-on agent quality.

Technical Novelty: Three-axis context expansion (long-horizon histories + interdependent backend services + integrated GUI/CLI across devices) combined with multi-round event injection to synthesize months of realistic user state with irrelevant and conflicting signals, plus explicit evaluation of proactive (anticipatory) assistance — prior benchmarks evaluate only reactive, narrow-context tasks.

What's New: First benchmark to jointly stress long-horizon activity history, backend-service interdependence, and unified GUI+CLI multi-device interaction, while explicitly measuring proactive anticipation under realistic noise — a substantially broader evaluation surface than prior assistant benchmarks.

Extension Opportunities:

  • Extend the multi-round event injection simulator to include cross-user/social signals (shared calendars, collaborative docs) to test multi-tenant proactive reasoning.
  • Add adversarial noise generators (deceptive events, prompt-injection in tool outputs) to stress-test robustness beyond benign irrelevant/conflicting signals.
  • Use the 2,000 generated environments to train a smaller open-source agent (e.g., Llama/Qwen) with RL on proactive recommendation rewards and compare against GPT-5.5.

Replicability: Abstract states the benchmark and an automated data-generation pipeline yielding 2,000 training environments will be released. Compute for evaluation is modest (API calls to GPT-5.5-class models); reproducing the 23.7% fine-tuning gain likely requires GPU resources for SFT/RL on a mid-size base model, though exact specs aren't stated.

Research Gaps:

  • No clear evaluation of privacy/safety trade-offs when agents have broad cross-service access.
  • Proactive recommendation quality metrics may not capture user-perceived utility or annoyance — human-in-the-loop evaluation appears absent.

👁️ Vision

1. Rethinking Weak Supervision in Anomaly Detection: A Comprehensive Benchmark

Authors: Xu Yao, Siyuan Zhou, Wu Zhenbo... Published: 2026-05-25 | Citations: 0 arXiv | PDF

Research Question: Are the three branches of weakly supervised anomaly detection (incomplete, inexact, inaccurate supervision) genuinely distinct problems requiring specialized methods, or do they share underlying mechanics that a unified evaluation would expose?

Summary: WSADBench is the first benchmark unifying evaluation of incomplete, inexact, and inaccurate weak supervision for anomaly detection, running 700K+ experiments over 36 algorithms and 4 modalities. It shows the three supervision regimes are deeply correlated rather than independent, and that specialized WSAD methods are quickly outclassed by tabular foundation models and standard classifiers as labels accumulate or distributions shift.

Key Results: Ran over 700K experiments benchmarking 36 algorithms across 4 modalities, systematically varying label quantity, granularity, and quality. Demonstrated (1) strong intrinsic correlations across the three WSAD scenarios, (2) specialized WSAD methods only win in extreme label-scarcity regimes before being overtaken by tabular foundation models and general classifiers as supervision grows or under OOD shift, (3) unlabeled data yields inconsistent and marginal gains versus label refinement, and (4) models show asymmetric sensitivity to different label-noise types.

Key Findings:

  • The three weak-supervision scenarios share fundamental mechanics — treating them as isolated research tracks is unjustified.
  • Specialized WSAD methods only outperform in extreme label scarcity; tabular foundation models dominate otherwise and under OOD.
  • Refining existing labels yields larger gains than leveraging more unlabeled data, and models are asymmetrically sensitive to different label-noise types.

Technical Novelty: The first benchmark to jointly evaluate all three WSAD weak-supervision regimes under standardized protocols, and to include modern tabular foundation models alongside specialized WSAD baselines and generic classifiers in the same comparison.

What's New: Prior WSAD work evaluated each supervision type in isolation with bespoke protocols; this paper imposes a single protocol that varies label quantity, granularity, and quality together, and is the first to pit specialized WSAD methods against tabular foundation models at scale.

Extension Opportunities:

  • Build a unified WSAD algorithm that explicitly exploits the cross-scenario correlations identified, instead of treating incomplete/inexact/inaccurate supervision separately.
  • Develop a label-refinement pipeline (active relabeling, noise correction) and benchmark it against semi-supervised use of unlabeled data on WSADBench to quantify the tradeoff.
  • Extend WSADBench to non-tabular modalities (video, time series, graphs) and probe whether tabular foundation models' dominance persists or breaks.

Replicability: Open-source code and datasets released at github.com/SUFE-AILAB/WSADBench. 700K+ experiments across 36 algorithms and 4 modalities implies substantial compute — likely a multi-GPU cluster over days/weeks for full reproduction, though individual algorithm/dataset slices should be tractable on a single workstation.

Research Gaps:

  • No principled algorithm yet exploits the cross-regime correlations the benchmark uncovers.
  • Limited understanding of why tabular foundation models generalize better under OOD than purpose-built WSAD methods.

🦾 ROBOTICS

1. AnyScene: Towards Highly Controllable Driving Scene Generation at Anywhere and Beyond

Authors: Haiming Zhang, Junfei Zhou, Feng Jiang... Published: 2026-05-25 | Citations: 0 arXiv | PDF

Research Question: How can synthetic driving scene data be generated with fine-grained controllability from arbitrary BEV layouts, without relying on shallow conditioning or reference-frame-dependent video synthesis, to scale simulation of rare safety-critical scenarios for end-to-end autonomous driving?

Summary: AnyScene is a unified occupancy-centric framework that generates controllable driving scenes from arbitrary BEV layouts via a Spatial-Temporal Occupancy Diffusion Transformer, then synthesizes multi-view videos using occupancy as a canonical geometric representation. It decouples scene structure from camera configuration, enabling reference-free, long-horizon, multi-view video generation that generalizes across datasets and supports user-defined layouts.

Key Results: The abstract claims state-of-the-art performance on both occupancy and video generation tasks, demonstrates generalization to unseen/customized BEV layouts (including cross-dataset transfer), and shows measurable downstream benefit for sparse-view 3D reconstruction. Specific numerical benchmarks (e.g., FID, IoU, mIoU on nuScenes/Occ3D) are referenced but not quantified in the provided abstract.

Key Findings:

  • Joint autoregressive tokenization of BEV and occupancy yields finer controllability than shallow conditioning baselines.
  • Treating occupancy as a canonical 3D representation enables reference-frame-free, multi-view video synthesis with flexible camera rigs at inference.
  • Generated data improves downstream sparse-view 3D reconstruction, validating geometric fidelity beyond perceptual metrics.

Technical Novelty: Two specific contributions: (1) a Spatial-Temporal Occupancy Diffusion Transformer that jointly tokenizes BEV and occupancy autoregressively, replacing shallow BEV conditioning with deep cross-modal tokenization; (2) a reference-free, occupancy-as-canonical-geometry view expansion module that supports arbitrary camera rigs at inference, decoupling video synthesis from training-time camera configurations.

What's New: Unlike prior occupancy-guided generators that use shallow conditioning and require reference frames (constraining camera setup and limiting long-horizon coherence), AnyScene unifies BEV→occupancy→video in one autoregressive pipeline with occupancy as the geometry anchor, enabling cross-dataset BEV inputs, user-defined layouts, and flexible cameras at inference.

Extension Opportunities:

  • Couple AnyScene with a reinforcement-learning driving policy loop: generate adversarial BEV layouts that maximize policy failure to systematically mine long-tail scenarios.
  • Extend the Geometry-Grounded View Expansion to support LiDAR/radar modality synthesis from the same occupancy canonical representation for sensor-fusion training data.
  • Integrate text-to-BEV layout generation (LLM-driven) on top of AnyScene to enable natural-language scenario authoring (e.g., 'pedestrian darting from occluded bus at dusk').

Replicability: No code/data release is mentioned in the abstract. Reproduction would likely require nuScenes + Occ3D-style occupancy labels and multi-A100/H100-scale training given diffusion transformer + multi-view video synthesis (typically 64-256 GPU-days for comparable models).

Research Gaps:

  • Quantitative evaluation of safety-critical long-tail scenario coverage (the stated motivation) is not directly addressed by occupancy/video metrics.
  • Lack of closed-loop validation showing that policies trained on AnyScene data transfer to real-world driving performance gains.

2. ParkourFormer: Integrating Predictive Supervision and Sequence Modeling into Parkour Locomotion

Authors: Yanheng Mai, Wenhao Xu, Zirui Huang... Published: 2026-05-25 | Citations: 0 arXiv | PDF

Research Question: How can humanoid parkour locomotion policies move beyond purely reactive observation-to-action mappings to explicitly anticipate future body states and contact transitions across rapidly changing terrains?

Summary: ParkourFormer is a Transformer-based locomotion policy that augments standard sequence modeling with a supervised future-proprioception prediction head, fusing anticipated body dynamics with historical context via cross-attention to produce actions. The single unified policy reaches 93.85% success across stairs, gaps, slopes, rough terrain, and obstacles, beating MLP/MoE/Transformer baselines by up to 42.73% and transferring to a real humanoid.

Key Results: ParkourFormer achieves a 93.85% average traversal success rate on highly challenging multi-terrain benchmarks (stairs, gaps, slopes, rough terrain, obstacles), with improvements of up to 42.73% over MLP, MoE-based MLP, and vanilla Transformer baselines. Validated in both simulation and on a real humanoid robot using a single unified policy across all terrain types.

Key Findings:

  • Explicit supervised prediction of short-horizon future proprioceptive states substantially improves agile locomotion versus implicit temporal modeling
  • A single unified Transformer policy generalizes across five distinct terrain types without per-terrain specialization or mixture-of-experts gating
  • Cross-attention from current state to historical trajectories outperforms causal self-attention baselines for contact-heavy locomotion

Technical Novelty: Reformulating locomotion RL as a future-conditioned decision problem: cross-attention between current state (query) and historical sensorimotor trajectories (keys/values), combined with a supervised prediction head that forecasts short-horizon future proprioceptive states which are then fused with temporal features for action generation. Unlike vanilla Transformers or Decision Transformer-style return-conditioning, supervision comes from explicit future-state prediction.

What's New: Prior humanoid RL policies are reactive (MLP, MoE) or use vanilla Transformers without explicit future modeling. ParkourFormer is novel in coupling sequence modeling with a supervised future-state prediction auxiliary objective whose outputs are fused into the action head — making anticipation a first-class signal rather than emergent.

Extension Opportunities:

  • Extend the prediction head to forecast exteroceptive (vision/depth) future states in addition to proprioception, enabling longer-horizon terrain-aware planning
  • Apply the future-conditioned sequence modeling framework to bimanual manipulation or loco-manipulation tasks where contact anticipation is equally critical
  • Investigate adaptive prediction horizons that lengthen on predictable terrain (flat ground) and shorten during high-frequency contact events (stair descent, landing)

Replicability: Abstract does not mention public code/data release. Reproduction would require a humanoid simulation stack (Isaac Gym/MuJoCo), GPU cluster for RL training (likely 1-8 GPUs for days), procedurally generated parkour terrains, and ideally a humanoid platform (e.g., Unitree H1/G1) for sim-to-real validation.

Research Gaps:

  • No exteroceptive/visual future prediction — anticipation is limited to proprioception, leaving longer-horizon terrain reasoning unaddressed
  • Unclear how performance degrades under sensor noise, latency, or out-of-distribution terrains beyond the benchmark suite

3. LRDDv3: High-Resolution Long-Range Drone Detection Dataset with Range Information and Thermal Data

Authors: Knut Peterson, Zaid Mayers, Azmain Yousuf... Published: 2026-05-25 | Citations: 0 arXiv | PDF

Research Question: How can we address the lack of high-resolution, long-range drone detection datasets that include both RGB and thermal (IR) imagery with explicit range information, which is needed for safe UAV detection in shared airspace?

Summary: LRDDv3 is a large-scale drone detection dataset featuring 102,532 4K RGB images and 29,630 paired thermal IR images of drones at long range, collected over 8 months with explicit per-image range annotations. It targets the gap in high-resolution, multi-modal, long-range UAV detection data needed for safe airspace sharing.

Key Results: The authors compiled LRDDv3, comprising 102,532 long-range RGB images sampled at 5 FPS from 128 distinct video clips across 17 collection days spanning 8 months, plus 29,630 paired IR images. The dataset uses 4K RGB resolution and 640x512 IR resolution, and includes comprehensive per-image drone range annotations — making it one of the first drone detection datasets at this resolution scale with paired RGB-IR modalities.

Key Findings:

  • Existing drone detection datasets lack the combination of 4K resolution, thermal pairing, and range information needed for long-range detection research
  • Collecting across 17 days over 8 months provides meaningful diversity in lighting, location, and background — important for generalization
  • Pairing every IR image with an RGB counterpart enables direct multi-modal fusion research, which prior datasets do not support at this scale

Technical Novelty: Unlike prior drone detection datasets, LRDDv3 combines three rarely co-occurring properties: 4K RGB resolution, paired thermal IR (640x512), and explicit per-image range information — collected across diverse lighting and backgrounds over 8 months to support long-range detection research.

What's New: First drone detection dataset to simultaneously offer 4K RGB resolution, paired 640x512 thermal IR imagery, and per-image range ground truth, sampled densely from real flight video over an extended collection period.

Extension Opportunities:

  • Train and benchmark modern small-object detectors (YOLOv8/v10, DETR variants, RT-DETR) on LRDDv3 with range-conditioned loss functions that weight detection difficulty by distance
  • Develop RGB-IR sensor fusion architectures that exploit the paired modality to improve detection in low-light or visually cluttered conditions, using the range labels for evaluation stratification
  • Build a range-prediction auxiliary head (monocular distance estimation for drones) trained on the explicit range ground truth, enabling collision-avoidance systems beyond pure 2D detection

Replicability: The dataset is publicly hosted at research.coe.drexel.edu/ece/imaple/lrddv3/. No model or training code is mentioned in the abstract. Reproducing data collection would require multiple drones, RGB+thermal cameras, range-measurement instrumentation, and FAA-compliant flight operations; using the dataset to train detectors needs only standard GPU resources (single high-memory GPU sufficient for most detectors at 4K).

Research Gaps:

  • No benchmark detector results or baseline metrics are reported in the abstract, leaving the actual detection difficulty unquantified
  • Single drone-type or fleet diversity, weather variation, and annotation methodology (especially range measurement accuracy) are not described in the abstract

🔥 GitHub Trending

1. wesleysimplicio/simplicio-cli

⭐ 5 stars | Python

Your tasks with 99% accuracy using any LLM (Claude, DeepSeek, Codex, Gemini, Hermes, OpenClaw, Cursor).

agentic-ai ai-agent ai-tools anthropic claude cli

2. Vishwa028/Underwater-Vision-and-Sonar-Object-Detection-System

⭐ 2 stars | Python

Underwater object detection system using Python, OpenCV, PySide6 and Arduino serial communication.

arduino computer-vision desktop-application machine-learning opencv pyside6

3. lambda765/smzdm_bot

⭐ 2 stars | Python

什么值得买智能推送机器人，可根据用户喜好、库存状态等自动筛选推送至飞书

agent feishu-bot llm smzdm

4. Agent-A345/Loan-Approval-ML-Pipeline

⭐ 1 stars | Jupyter Notebook

A machine learning pipeline that predicts loan approval status using applicant financial and demographic data through preprocessing, feature engineering, classification modeling, and predictive analyt

jupyter-notebook loan-approval-prediction machine-learning python

5. ShrekDino/uploaded-consciousness-framework

⭐ 1 stars | Python

Consciousness Simulation Diagnostic Framework. A formal testbed for enumerating, measuring, and classifying failure modes in synthetic consciousness architectures. Implements Szilard thermodynamics, G

active-inference artificial-intelligence consciousness consciousness-simulation cybernetics diagnostic-framework

6. Amirtha655/image-lablels-generator

⭐ 1 stars | Python

Computer vision app that detects objects in images and draws bounding boxes using AWS Rekognition, S3, and Python

amazon-rekognition aws aws-s3 deep-learning machine-learning python

7. nitishpavanreddy/churn-predictor

⭐ 1 stars | Python

Customer Churn Predictor using Machine Learning — Logistic Regression and Random Forest on 7,043 customers. Achieved 79.9% accuracy. Built with Python, pandas, scikit-learn.

churn-prediction classification data-analytics logistic-regression machine-learning python

8. nitishpavanreddy/stock-forecasting

⭐ 1 stars | Python

Stock Price Forecasting using Python, yfinance and Facebook Prophet — pulled 753 days of live Apple (AAPL) stock data, calculated moving averages and volatility, predicted 30-day price direction with

data-analytics machine-learning matplotlib prophet python stock-forecasting

9. HAYDARKILIC/linear_algebra_for_ml

⭐ 1 stars | Jupyter Notebook

The linear algebra underpinning modern machine learning from scratch in pure NumPy then connects every construct back to the model that depends on it

artificial-intelligence deep-learning linear-algebra linear-algebra-for-ml machine-learning python

10. jiachengwang-punch/predictive-analytics-skill

⭐ 1 stars | Unknown

A reusable, multi-model, language-adaptive methodology for end-to-end machine learning analysis of tabular data.

claude-skill codex-skill data-analysis data-science deepseek feature-engineering

11. bradleybeatz1313/eda-ai-training-pipeline

⭐ 1 stars | Python

Automated circuit netlist generation and labeling pipeline for AI/ML training

artificial-intelligence automation circuit-design eda machine-learning ngspice

12. SarathKumarKomathukattil/autonomous-vehicle-perception-quanser-qcar

⭐ 1 stars | Python

End-to-end autonomous vehicle perception system - MSc Mechatronics dissertation deployed on Quanser QCar

autonomous-vehicles computer-vision deep-learning enet lidar machine-learning

13. tej-droid-byte/Bond-Portfolio-Optimization-RL

⭐ 1 stars | Jupyter Notebook

Reinforcement learning–based bond portfolio optimization using DQN agents and interactive Power BI dashboards for trading strategy and portfolio performance analysis.

analytics bond-market deep-learning dqn finance financial-analysis

14. ricky5932TW/End2End-autodrive-image-steer

⭐ 1 stars | Jupyter Notebook

End-to-end self-driving AI in Forza using PyTorch, screen capture, telemetry, Grad-CAM, and virtual controller feedback.

autonomous-driving behavior-cloning cnn computer-vision deep-learning driving-simulator

15. StephenQiu30/rtc-reader

⭐ 1 stars | Swift

基于火山引擎 RTC 的 iOS AI 朗读 Demo — 摄像头实时捕捉，大模型视觉 + TTS，翻页即读

ai computer-vision doubao ios llm rtc

Generated by Research Pulse on 2026-05-26 06:04