Wonder News Morning: AI agents failed two open-ended research questions after six days

In two open-ended research trials, frontier AI agents had six days and thousands of dollars of compute. They completed the engineering but did not make enough progress to answer either paper’s central question. This morning also brings an agent-first coding workflow, an economically grounded office benchmark, a memory-poisoning test, two-minute autoregressive video, camera reconstruction and planning, articulated 3D objects, embodied-agent evaluation, a compact robot policy, faster on-robot learning, open autonomous-racing hardware, and a new EdTech calendar entry. Unless noted otherwise, the results are author-reported from new preprints and have not been independently replicated.

What Changed Overnight

• ArXiv’s July 30 announcement batch added new work on open-ended research agents, coding and office agents, memory security, video and 3D production, embodied evaluation, robot policies, and educational hardware.
• ASU+GSV opened attendee registration for its April 4–7, 2027 summit, while its GSV Cup interest form is accepting eligible early-stage digital-learning and workforce-skills startups.

Agents, Coding, And Memory

AI agents finish the engineering but miss the research question

Peter Kirgis, Sayash Kapoor, Arvind Narayanan, and collaborators devised “shadow evaluations” around the central questions of two unpublished NeurIPS 2026 submissions. The original paper authors graded the agent work instead of sending generated papers through ordinary peer review.

The study gave the agents six days and thousands of dollars of compute. They completed all engineering without human help, but the authors judged both results too weak to make substantial progress on the research questions and rejected them unambiguously. A second model and scaffold reproduced the pattern. The reported failures included weak judgment about publishable evidence, uncreative responses to design problems, ineffective backtracking, poor resource awareness, and instruction drift. Two cases cannot establish a general capability rate, but the released repositories, logs, reviews, and surveys make the gap inspectable.

SpecFirst makes agents write the behavior down before coding

Yihao Chen and collaborators separate program synthesis into two stages. One agent probes an execute-only reference program and turns documentation plus observations into a structured behavioral specification; a second agent writes code against that specification.

Across all 200 ProgramBench instances and four models, SpecFirst raises test pass rates by 6.9%–21.3% and binary-exploration coverage by 9.4%–18.5% over the single-loop comparison. The benchmark starts from documentation and a behavioral oracle rather than an existing repository, so the result applies to from-scratch construction rather than ordinary issue fixing.

OmegaUse prices long office tasks as well as scoring them

Jingbo Zhou and collaborators built OmegaUse-OfficeVal from 100 privacy-preserving office-suite requests proposed by practitioners. The tasks take people an average of 2.32 hours and carry both a human-labor estimate and a task-price proxy, while code-based verifiers score deliverable quality.

The tested frontier models finish faster and at lower inference cost than people but remain below the human baseline on deliverable quality. The authors have open-sourced the benchmark and code. Full model-by-model quality and cost differences need the paper’s detailed tables rather than the abstract alone.

MemSecBench follows poisoned memory through action and repair

MemSecBench contains 310 cases from 48 code, science, daily-life, and office contexts. Each case runs a controlled Write–Execute–Forget sequence across two agent harnesses, four memory backends, and three language models, producing 24 fixed configurations.

Malicious content persists in memory in 84.2% of cases across those configurations, while 50.3% complete the full write-to-action attack chain. Among successfully poisoned cases, 56.1% achieve selective repair. The largest differences from matched native-memory configurations reach 16.1 percentage points for end-to-end attack success and 41.3 points for repair, showing that the memory stack itself changes the outcome.

Video, 3D, And Camera Control

FreqForcing keeps long video anchored in low frequencies

Jiatong Li and collaborators trace color drift, motion stagnation, and eventual collapse in autoregressive video to accumulating low-frequency energy error. Their training-free Spectral Self-Anchoring method uses low-frequency components from anchor attention for long-term stability while retaining high-frequency local attention for motion.

FreqForcing extends Self-Forcing from the five-second clips used in training to two-minute generation, a 24-fold increase in duration. The authors report improvements over other training-free methods and competitive results against selected trained approaches; creator-facing tests still need to show how subjects, cuts, and scene state survive the full two minutes.

VidMap treats video order as reconstruction evidence

Zador Pataki, Paul-Edouard Sarlin, and Marc Pollefeys combine sequential constraints from SLAM with offline structure-from-motion optimization. VidMap uses temporal order for loop closure, wide-baseline dense matching, and monocular depth priors to recover camera calibration and metric poses from long, uncalibrated video.

The authors report greater robustness and accuracy than the tested classical and learned SLAM and structure-from-motion systems on datasets with extreme motion and visual symmetries. Code is public, but the abstract does not provide one aggregate margin or state the processing cost, both of which matter for asset-ingestion pipelines.

CinemaTraj turns cinematography terms into collision-free 3D paths

Qianru Li and collaborators give a language-model agent RGB-D views and a structured 3D scene graph. The agent decomposes a prompt into moves such as dolly, orbit, crane, pan, tilt, zoom, and arc, then instantiates parametric trajectories that can be optimized to avoid collisions. It also aligns voiceover and subtitles with the resulting camera motion.

Tests use real ScanNet++ environments. The authors report better prompt alignment, trajectory quality, safety, and cinematographic quality than their comparisons, although the abstract does not publish the numerical margins or a creator study.

StructureGS keeps articulated parts compact and connected

Gahye Lee and collaborators add part-oriented bounding boxes to 3D Gaussian-splatting reconstruction. One loss keeps each movable part spatially coherent within its assigned region; another preserves plausible contact between adjacent parts.

The paper reports the best articulated-object reconstruction in its comparisons, with cleaner part geometry and boundaries. The method targets objects with multiple movable components, which is useful for interactive assets, but the abstract does not quantify rig export, animation stability, or performance inside a game engine.

Embodied Systems And Learning Hardware

HumanCLAW finds that models lose track of their own bodies

Siyao Li and collaborators separate high-level action choice from low-level motor execution. A harnessed vision-language model selects atomic skills, while a controller translates each choice into sub-second full-body motion with gravity and collisions; execution disturbances are factored out so the benchmark can score the model’s next-action decisions.

HumanCLAW-Bench contains 1,218 long-horizon, first-person find-navigate-interact episodes across 41 indoor scenes. None of nine tested vision-language models solves the benchmark, and the best success rate is 16.8%. The paper attributes the failures less to recognizing targets than to tracking body position, arrival, and collisions.

TurboVLA removes the language model from the robot policy’s center

Hengyi Xie and collaborators encode vision and language separately, exchange information through a lightweight bidirectional module, and decode continuous action chunks without routing every observation through a large language model.

On LIBERO, the 0.2-billion-parameter policy reports 97.7% average success, 31.2 milliseconds of latency, and 0.9 GB of inference memory on an RTX 4090. That equals about 32 policy updates per second on the stated hardware. The result is from a simulated manipulation suite, not a physical-robot deployment.

SymmGrid multiplies real-robot experience through symmetry

Gabe Everett and collaborators transform complete image-and-action trajectories into geometrically consistent symmetric variants, including homography-based warps for external camera views. The extra trajectories populate the replay buffer during physical-robot reinforcement learning.

Across peg insertion, cable routing, and object relocation, SymmGrid reports wall-clock convergence-speed ratios of 1.37–2.17 and evaluation success-rate ratios of 1.09–1.27 against its comparisons. Training takes 16.6, 10.9, and 79.3 minutes on the three tasks. The tests cover a limited set of contact tasks, but they measure training directly on robots rather than only in simulation.

NeoRacer offers an open autonomous car for labs and classrooms

Koneshka Bandyopadhyay and collaborators describe a pre-assembled, open-source 1:12-scale race car built around a Jetson Orin Nano rated at 67 TOPS, a 270-degree lidar, a 120 FPS global-shutter camera, and a nine-axis inertial sensor.

NeoRacer costs $2,699. The authors say it provides more than three times the compute at less than half the price of the closest pre-assembled comparison. Pilot deployments involved 15 MIT students and 10 Boston University students. Hardware carries CERN-OHL-S v2, software uses GPLv3, and the design files, firmware, and ROS 2 components are public.

Events Calendar

ASU+GSV opens 2027 tickets and startup interest

ASU+GSV Summit 2027 will run April 4–7 at the Manchester Grand Hyatt in San Diego. Individual registration is open while tickets remain; the organizer has not published a final sales date.

The separate GSV Cup interest form is open to eligible pre-seed and seed-stage companies working in digital learning and workforce skills. The organizer has not stated a closing date for that form.

See the 2027 games and education-technology calendar.

Watch Next

• The shadow-evaluation authors released agent logs and reviews, enabling follow-up work to test more research questions, models, and scaffolds under the same expert-grading method.
• SpecFirst can be tested on larger programs where specifications evolve, while OmegaUse can show whether lower inference cost still holds after models are required to meet human deliverable quality.
• Long-video and 3D-camera systems need creator studies that measure continuity, editability, collision repair, and export into production tools.
• HumanCLAW separates action choice from motor control; future evaluations can test whether explicit body-state memory raises the 16.8% ceiling.
• TurboVLA and SymmGrid need broader physical-hardware tests, and NeoRacer’s public design can reveal whether independent labs reproduce its cost and performance claims.

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