Qwen-CUA uses a 397-billion-parameter mixture-of-experts backbone with 17 billion active parameters to operate software while observing screenshots alone. The authors report 86.2 on OSWorld-Verified for the standard model and 87.6 for a larger Max variant; neither receives DOM trees, accessibility metadata, or task-specific APIs.

This edition also covers a world-model reactivity test, coding agents sharing files with users, lightweight failure detection and repair, hidden tool behavior, agent-skill training, runtime-harness editing, faster video generation, robot data converted from human video, humanoid fall recovery, programming practice, and a newly verified 2027 education conference. The research papers were submitted August 3; their results are author-reported and have not been independently reproduced.

What changed overnight

  • Qwen-CUA, WorldExam, SWE-Touch, and eight other newly listed papers provide fresh results on computer use, interactive worlds, coding, agent reliability, video generation, robotics, and programming education.
  • Official National ESEA Conference guides confirm February 23-25, 2027 in Louisville and online. Individual attendance and company exhibit registration are open.

Computer use and interactive worlds

Qwen-CUA reports 86.2 on OSWorld-Verified from screenshot-only control

The Qwen team built Qwen-CUA on a 397B-A17B mixture-of-experts model. It sees only screenshots and sends keyboard and mouse events. Its runtime keeps up to 20 active screenshots, then compresses older visual history in fixed-size blocks. This preserves recent evidence while keeping earlier context available.

Training used nearly 100,000 virtual CPUs, tens of thousands of concurrent environments, and about 40,000 tasks with outcomes that could be checked automatically. The standard model reached 86.2 on OSWorld-Verified and 18.5% binary completion plus 48.4% partial completion on OSWorld 2.0. Qwen-CUA-Max, which has more than one trillion parameters, reached 87.6 and 21.2%/53.3% on the same tests.

The authors also report RedTeamCUA attack success falling from 36.6% with Qwen3.7 to 16.4% with Qwen-CUA. The paper does not show that screenshot-only control is safe across every account, permission, transaction, or destructive action. It also does not isolate how much of the gain comes from the larger model, the rollout fleet, the task set, or the runtime scaffold.

WorldExam finds visual quality does not guarantee a reactive world

Yuxue Yang and collaborators assembled 1,474 cases across eight tasks and four levels: visual quality, control adherence, spatial consistency, and world reactivity. The last level asks whether a generated environment produces plausible scene-dependent consequences that were not explicitly written into the prompt.

Across 20 models, camera-driven systems followed camera controls well but did not support dynamic interaction. Action-driven systems controlled subjects more precisely but often left the surrounding world unchanged. Language-driven systems reacted more often but followed complicated controls less faithfully. No tested model combined broad task coverage with consistently strong results, according to the authors.

Shared coding and agent reliability

SWE-Touch lowers coding-agent resolve rates with conflicting user edits

Yuqiao Tan’s team inserted plausible but task-conflicting code changes into shared workspaces while agents were already solving repository issues. The counter-edits were drawn from task-critical regions and arrived with contextual user messages, testing whether an agent noticed that the code had changed under it.

Across nine coding models on SWE-bench Verified, average resolve rate fell 7.7 percentage points. The drop persisted on longer SWE-Bench Pro and DeepSWE tasks. Failed agents sometimes kept conflicting code or replaced it without re-inspecting the repository and running tests focused on the affected behavior.

Lightweight checks lift repaired-agent task success from 52% to 73% in one experiment

Sunny Dubey evaluated a telemetry monitor on 2,823 completed agent episodes across three frameworks, three local models, and Gemini 2.5 Flash. A one-class echo-state-network ensemble detected 71% of failures at a 5% false-alarm budget and ran in about 200 microseconds per step.

The study also used deterministic checks that recalculated claimed totals from tool results and confirmed required calls. Those checks caught 60% of failures with no false positives in 63 cases, or 96% when call coverage was included. Rolling back flagged runs and executing them again recovered 45% of failures, compared with 16% for a resampling control, and raised task success from 52% to 73% for roughly one extra model call per run. The learned monitor did not transfer cold: AUROC was 0.527 before deployment-specific recalibration and 0.885 after it.

ScrambleToolBench removes meaningful names from terminal tools and requires agents to discover behavior through interaction. It then changes mappings, introduces stochastic failures, and opens actions only during certain time windows.

The authors report that current models often kept obsolete beliefs or searched exhaustively after a mapping changed, even when their earlier observations pointed to a shorter deduction. More test-time reasoning increased the search rather than producing reliable cycle tracing. Persistent memory reduced accumulated mistakes but did not solve structural adaptation.

SKT turns 2,000 public agent skills into 27,164 verified trajectories

Zelin Tan’s team created a pipeline that selects single or combined skills, writes executable tasks, repairs failed generations, and retains only trajectories that complete the task while substantially using every required skill. Starting from 2,000 public skills, it produced 4,000 task packages and 27,164 verified trajectories.

The researchers also built a held-out SkillEval set from a separate skill pool. Supervised fine-tuning on the generated trajectories improved skill use across multiple models and agent runtimes, they report. The abstract does not provide one aggregate score for those gains, so comparisons depend on the paper’s individual test settings.

Harness-R1 learns to patch the runtime around a failing agent

Harness-R1 trains a separate 9-billion-parameter “harness engineer” to edit the software that builds context, exposes tools, validates actions, and handles recovery. Batches of failed target-agent runs become candidate runtime patches, and fresh reruns provide the reward; the target model remains frozen during this stage.

Across WebShop, ALFWorld, and DBBench, the edited harness raised Qwen3.5-9B’s average success from 44.3% to 53.6%. After the target agent itself was fine-tuned, a target-specific harness engineer lifted the average from 59.2% to 64.2%. The gains are tied to those three environments and do not yet show whether learned runtime edits remain maintainable or secure in a production codebase.

Video efficiency and robot control

Token Radius Attention retains 9-19% of video attention interactions

Jiayu Chen’s team observed that different video tokens need different attention budgets. Its training-free Token Radius Attention method estimates a budget from each query’s attention entropy, then builds a local radius that shrinks over time instead of ranking every possible key.

Across seven Wan 2.1, Wan 2.2, and HunyuanVideo text-to-video and image-to-video configurations, the method retained 9-19% of attention interactions and produced 1.56-2.05 times faster generation with competitive quality, according to the authors. Creator-facing latency, memory use, and long-sequence quality still need matched hardware measurements.

Ego2Robot converts human video into 18,561 hours of robot data

Ego2Robot retargets egocentric human manipulation videos into robot actions, synthesizes robot arms into the footage, and filters the results at several quality levels. The resulting dataset contains 18,561 hours across 15 robot body configurations.

The team extended RoboTwin 2.0 so appearance, scene layout, robot body, and task meaning could vary independently. Joint pretraining on synthesized and conventional robot data improved out-of-distribution results across several perturbations, and the authors also report real-robot validation. The paper still depends on the quality of retargeted actions and rendered bodies; those transformations can carry errors that ordinary video scale does not remove.

StableMimic recovers in 100 of 100 push-to-fall trials

StableMimic trains a humanoid beyond ordinary motion tracking by resetting it into prone, supine, off-balance, and intermediate contact states around human get-up references. Separate tracking and recovery experts are blended by a gate that uses the robot’s own body state; deployment does not receive the get-up reference or an external recovery command.

On the retargeted LAFAN1 dance subset, StableMimic recorded the lowest error on all four tracking measures among five methods. In 100 matched push-to-fall trials per method, it recovered in every trial and had the lowest value on six of seven post-fall motion and load measures. Real Unitree G1 demonstrations show recovery and resumed commands, but the reported 100% figure comes from the matched experimental protocol rather than unrestricted real-world falls.

Programming education and events

Arun Balajiee Lekshmi Narayanan and collaborators analyzed logs from 334 students across 11 semesters of introductory and intermediate Java. The practice system included code writing, tracing, completion, visualizations, and explanations.

Active practice was associated with 3.2 times better learning outcomes than passive activity. Among the active formats, code writing had the strongest association with posttest performance; the other formats did not show a comparable link. This is an observational association in one course system, not a randomized comparison proving that code writing alone caused the gain.

National ESEA Conference 2027 opens attendance and company exhibits

The National ESEA Conference will run February 23-25, 2027 at the Kentucky International Convention Center in Louisville and online. The program focuses on U.S. federal education programs for disadvantaged students.

Individual and group registration is open for in-person or virtual attendance. Booth selection for new exhibitors opened July 16. The organizer has not published a final attendee-registration or booth deadline, and its standard presentation proposal window closed June 16.

See the 2027 games and education-technology events calendar.

Watch next

  • Qwen-CUA needs independent testing that separates model scale, training data, runtime memory, and scaffold effects while measuring destructive-action and credential risks.
  • WorldExam and SWE-Touch need public traces that show exactly when models miss environmental reactions or overwrite a user’s changed code.
  • Agent monitors, skill training, and learned harness edits need production tests that track false alarms, maintenance cost, permissions, and regressions over time.
  • Video and robot systems need matched hardware, latency, energy, and physical-failure comparisons beyond author-run tasks.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.