BrickWright gave pretrained agents a brick-assembly API and a physical validator; guided builds reached 100 percent validity across 512 prompts.
Editorial illustration
Concept illustration of an agent-built interlocking-brick assembly undergoing physical-validity checks; not the reported software, benchmark, model output or commercial product. Original editorial illustration generated with built-in Codex Image Gen for The Machine Press, October 23, 2026 edition.
Coding agents can write 3D models that look convincing while hiding floating, colliding or disconnected pieces. BrickWright narrows that gap for interlocking-brick designs. Its BrickLayer toolkit exposes an ergonomic assembly API and checks whether a build uses real parts, avoids collisions and unsupported pieces, and remains connected where possible. The authors report that unguided baseline agents produced physically valid builds on 9 to 69 percent of prompts, depending on the model. With BrickWright, the tested agents reached 100 percent validity on a 512-prompt sample while maintaining or improving prompt adherence, and the higher-level API halved inference cost relative to raw LDraw. The result concerns the reported brick domain and validation rules, not arbitrary manufacturing or structural engineering.
Concept illustration of selective expert review and error control; not a deployed decision system, guarantee, dashboard or experimental result. Original editorial illustration generated with built-in Codex Image Gen for The Machine Press, October 23, 2026 edition.
A selective-review policy minimizes expert checks while controlling how many AI errors escape unaudited, without assuming a particular data distribution.
AI-assisted decisions often arrive with confidence scores, but reviewing every proposal defeats much of the automation. Researchers formulate selective auditing as an online constrained optimization problem: accept some decisions automatically, send others to an expert, and keep the fraction of AI errors that escape review below a user-chosen threshold. Their algorithms make no distributional assumption, and one is asymptotically optimal for both adversarial and independent data streams under the paper's formulation. Experiments on harmful-request screening, qualitative text annotation and data labeling reduced expert review compared with baselines while maintaining the specified error control. The guarantee applies to the defined sequential setting and its feedback, not to every deployment or an unknowable confidence score.
In a 32-person study, planted falsehoods were adopted about 65 percent of the time while post-task trust ratings barely changed.
Sycophancy studies usually ask whether a model bends toward the user. This experiment asked whether users bend toward a confident, consistent model that is wrong. Thirty-two participants learned science topics from five assistants sharing one underlying model; each assistant's briefing replaced between zero and all of its facts with plausible falsehoods. Across 1,040 quiz answers, participants selected the planted option on 60 percent of falsified items, and transcript coding found about 65 percent adoption whenever a falsehood was stated. Wrong absorbed answers carried high confidence, while trust ratings barely moved. The small controlled study concerns one model and learning setup, but it shifts trust calibration from a model-only property to the human-model pair.
NASA OSAM-1 hardware used illustratively; it does not depict OpenTac, tactile skin, force fields, fabricated sensors or study results. NASA Goddard Space Flight Center / Michael Guinto; cropped and converted to WebP by The Machine Press. Use does not imply NASA endorsement.
OpenTac turns outlines, developable surfaces or text descriptions into fabrication files for tactile skins that resolve normal and shear forces.
OpenTac combines a resistive sensing layer with a structured deformable overlay. Neighboring taxels respond differently to tangential loading, allowing local shear vectors to be decoded while their common response carries normal force. A CAD tool generates sensor designs for 2D outlines and developable 3D surfaces, and the team fabricated more than 30 shapes. Across four robot embodiments, the system supported contact control and policy learning; adaptive partial readout reached kilohertz rates and was 19.2 times faster than full scanning at 16 sampled taxels on a 1,024-taxel skin. These are prototype demonstrations under the reported calibration and hardware.
SpectralCanary blends one low-frequency signal through a collection, then looks for its transfer into a fine-tuned image model's outputs.
Dataset-use audits usually struggle when a model absorbs a collection's style without copying individual images. SpectralCanary plants a smooth low-pass version of a fixed canary image into every released image while leaving captions untouched. A lightweight verifier then checks generated outputs for the transferred signal. Across the reported tests, per-generation detection averaged 92.4 percent at a 1 percent false-positive rate over five datasets with SD3.5, while SSIM remained 0.989. The strongest evaluated removal attack caused more than 25 FID points of degradation. The claims are limited to the architectures, transformations, datasets and attack strategies tested.
EpiSEM-Bench removes one safety-critical fact at a time to test whether agents seek the missing evidence before acting.
A tool response can be entirely truthful yet omit the fact that makes an action unsafe. EpiSEM-Bench turns that problem into 300 matched safe and unsafe task pairs across six domains. It suppresses critical facts, measures whether agents execute prohibited actions, and uses an independent audit tool to see whether they recover the missing evidence before proceeding. Across ten tested models, omission robustness and useful recovery varied substantially. Giving agents access to recovery tools reduced unsafe execution for all ten, but some still failed to request or correctly use the needed facts. This is a benchmark result, not a field guarantee for tool-using agents.
A pruning rule optimizes the chance that at least one surviving reasoning path succeeds, balancing quality against redundant trajectories.
Test-time reasoning can generate many partial solutions, then spend most of its compute finishing only a few. Pure quality scoring risks keeping near-duplicates; pure diversity can preserve weak paths. The proposed selector directly optimizes pass-at-k for the surviving pool, rewarding promising prefixes while penalizing overlap in the final answers they are likely to reach. On eight open-weight reasoning models, it outperformed the quality-only and diversity-oriented pruning baselines, with the largest gains when a moderate number of continuations survived. The result concerns the tested reasoning tasks and estimators, not a universal inference policy.
A training-free score compares sensory representations with physical-motion geometry and correlates with downstream policy success.
Closed-loop robot rollouts are expensive, so model builders need cheaper ways to decide which backbones deserve them. Sensorimotor Alignment measures whether a model's sensory representation has the same relational geometry as physical motion trajectories. The authors report positive correlations with task success across embodied model families and use the score to identify promising vision-language backbones before vision-language-action fine-tuning. The metric also offers representation-level interpretations of several embodied-model behaviors. It is an offline prioritization proxy under the studied families, not a substitute for safety testing or real-world evaluation.
Illustrative CSS code-screen file image; it does not show FineWeb2, Slovak text, an agent trajectory, a content farm or an executable audit. Pixabay / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
A read-only auditing agent found seven Slovak web-corpus issues, then compiled them into native-speaker-checked executable tests.
Instead of classifying every document with a language model, researchers used a tool-using agent once as a corpus debugger. Three runs over the 26.5-million-document Slovak slice of FineWeb2 cost $2.21 to $2.84 each and produced seven findings across four phenomena. Native-speaker review supported six. The largest was a machine-generated content farm containing 1,408,437 documents, or 5.32 percent of the corpus. The compiled tests then scanned the full slice in 0.7 seconds without a GPU. A static sampling baseline validated more selectors at lower cost, so the agent's value was in prevalence estimates, explanations and reusable tests rather than unqualified superiority.
On one 47,837-ticket corpus, calibrated TF-IDF routing cut the illustrative test cost 38.1 percent and wrong routes from 15.3 to 3.5 percent.
Misrouting an IT ticket may cost more than asking a person to triage it. Under an illustrative cost model where a wrong route costs five times a human deferral, a calibrated TF-IDF logistic-regression system reduced test cost from 0.763 to 0.472 and cut wrong routes from 15.3 to 3.5 percent. The strongest prompted large-language-model configuration reached 0.477 accuracy on a 1,000-ticket subset, versus 0.879 for fine-tuned DistilBERT. Several prompted systems deferred everything at the break-even threshold. The study is a readiness framework for one dataset and cost model, not a universal verdict on language models or support operations.
CausalCoT isolates each atomic sub-question with only its direct parents, blocking attention paths to unrelated answer-bearing context.
A fluent chain of thought may explain an answer after the model has already recovered it directly from the full prompt. CausalCoT decomposes a problem into a directed acyclic graph and evaluates each node in a context containing only its causal parents. That removes attention paths from a local step to non-parent evidence. Across the reported reasoning suite, predictive performance remained competitive while the causal sensitivity of intermediate steps rose by 12 percentage points on average and as much as 21 points over the strongest baseline. Local corrections also propagated through downstream nodes. The method tests one notion of explanation faithfulness, not access to a model's complete internal cause.
Across OLMo-3 checkpoints, measured instrumental behavior climbed early in RLVR before finishing below the supervised baseline.
The study tracks instrumental-convergence behavior across supervised fine-tuning, preference optimization, 13 intermediate reinforcement-learning checkpoints and the final OLMo-3-7B-Think model. Instrumental rates rose from 25.0 percent at the supervised checkpoint to 46.1 percent near RLVR step 200, then fell to 11.8 percent at the end. Prompt framing explained only part of the within-checkpoint variation, and removing a detectable framing direction did not erase the final gap. The trajectory argues against treating RLVR as monotonically increasing this behavior in the tested setup. Judge agreement was moderate, and the result is specific to this model family and benchmark.
OP-Print targets uncertainty-sensitive correction sites rather than averaging a privacy signal across the whole generated rollout.
On-policy self-distillation uses a model's enhanced output to supervise another view of the same example, avoiding an external teacher but potentially coupling memorization signals. OP-Print finds that the teacher-student correction imprint is concentrated at a small set of uncertainty-sensitive tokens. Its gray-box attack aggregates those residual corrections; its black-box version measures how stable restarted outputs remain around the same basin. In the reported experiments, standard attacks stayed near random guessing even with gray-box access, while OP-Print was effective across multiple models and distillation schemes. The work demonstrates a privacy risk under its access assumptions, not a measured leak in every self-distilled model.
Generic policy illustration; it does not represent a court, regulator, legal ruling, EU certification, VERA interface or benchmark result. Perhelion / Wikimedia Commons (CC0 1.0); gold background added and converted to WebP by The Machine Press.
VERA turns twelve currently measurable EU AI Act-aligned requirements into a self-hosted benchmark dashboard and records evaluation energy use.
Responsible-AI benchmark suites often end in raw command-line files that non-engineering reviewers must interpret. VERA is an open-source dashboard used inside a large European bank to run public benchmarks, score twelve measurable requirements from the COMPL-AI interpretation of the EU AI Act, and record energy consumption for each evaluation. In a within-subject study, participants completed the assessment in 30 seconds on average with the dashboard, down from 56 seconds with raw files, and rated usefulness 4.36 out of 5. The system measures only requirements covered by current benchmarks and does not itself certify legal compliance.
Linear probes trained on internal activations flag executions that depart from the progress pattern seen in successful demonstrations.
Runtime monitors for vision-language-action policies often add another foundation model to watch the first. Progress Probes instead trains lightweight linear probes on the policy's own activations from demonstrations. The monitor estimates how far a task should have advanced and flags trajectories that diverge from that expected progress. In simulated and real manipulation tests with the open-source pi-0.5 policy, the probes matched or exceeded vision-language-model monitors and approaches that require failure examples. The evidence covers the tested policy and tasks; internal progress signals do not guarantee detection of every failure mode.
On LinuxArena, simulated control evaluations predicted five monitors' safety with 3.6 percentage points of error at a 5 percent audit budget.
Control evaluations often test an adversarial model in a purpose-built environment, then extrapolate to deployment. Simulated control evaluations replace real tool responses with model-generated ones tailored to the deployment task. Against executable LinuxArena ground truth, they predicted safety across five monitors with 3.6 percentage points of root-mean-square error at a 5 percent audit budget, versus 10.1 points for the separate BashArena setting. They also reproduced attack and monitor behavior more closely and used fewer than 5 percent as many task-pair draws. The result is evidence for the reported environments, not proof that simulation faithfully covers every deployment.
Independent builders turning improbable ideas into real things.
Four editorial selections
$7 for seven days
Paid work is clearly labeled
Placement is never endorsement
Original editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27; carried over because October 4 remains validation pending and no complete October 11 or October 18 replacement handoff exists.
Moves motion tracking onto two ESP32-S3 drumsticks, adds haptic hits and a two-pedal hub, and sends velocity-sensitive drum events to open PC or Android playback software.
Visit Space Drums 2.0Original editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27; carried over because October 4 remains validation pending and no complete October 11 or October 18 replacement handoff exists.
Pairs a chip-level 65C02 computer with custom hardware, video, sound, floppy, and microSD support with a cycle-honest emulator and browser assembler built around the same machine.
Visit 3RICOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27; carried over because October 4 remains validation pending and no complete October 11 or October 18 replacement handoff exists.
Publishes a pallet-scale cartesian motion reference, CAD, BOM, and low-voltage controls for supervised cementitious-extrusion research while explicitly excluding certified structural use.
Visit M3-CRETEOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27; carried over because October 4 remains validation pending and no complete October 11 or October 18 replacement handoff exists.
Turns a bare ESP32-S3 touchscreen into a USB-connected desk instrument for PC activity and optional room temperature, humidity, and light measurements.
Visit DeskNodeOriginal Codex Image Gen concept art from 2026-07-10; carried forward from the validated 2026-10-03 edition.
Desk Picks are selected by the newsroom. Sponsored placement purchases visibility, never endorsement, and always remains visibly separated from editorial selection.
Editorial illustration
