Sixty-three superconducting qubits follow electric flux as charge, matter and sideways deformation compete beyond the reach of short classical simulations.
Editorial illustration
Concept illustration of electric-flux dynamics on a lattice quantum processor; it is not the reported hardware or measurement record. Original editorial illustration generated with built-in Codex Image Gen for The Machine Press.
A new quantum simulation follows electric flux strings in a two-dimensional Z2 lattice gauge theory with dynamical matter and an explicit magnetic plaquette interaction. The researchers used 63 superconducting qubits, including 12 ancillas, to evolve localized strings with static or moving endpoint charges across strongly confined and nonperturbative deconfined regimes. Spatial measurements showed rapid charge and flux spreading when confinement weakened, enhanced matter occupation near a pair-creation resonance and suppressed spreading under strong electric fields. The team used local Gauss-law constraints, ancilla checks and noise rescaling to detect errors in circuits reaching 15 second-order Trotter steps; the result is a laboratory demonstration reported by the authors, not a general solution to gauge dynamics.
Concept illustration of evidence shifting support among competing explanations; it does not reproduce the reported agent or experiments. Original editorial illustration generated with built-in Codex Image Gen for The Machine Press.
An agent that keeps competing explanations in view resolves more scientific questions from the same evidence budget, the authors report.
Many discovery agents test one hypothesis against its negation, so a failed hypothesis says little about which alternative should replace it. Contrastive Discovery instead asks the agent to maintain a mutually exclusive, collectively exhaustive set of explanations and score each experiment by how it moves belief across that set. Across six domains, the authors report that contrast sets resolved 73% of high-shift cases that pointwise evaluation merely rejected; when the sets guided search, the agent resolved 67% more questions under the same experiment budget. In a blinded study, 38 human experts chose the same explanation on 78% of resolved questions, evidence for the method within the evaluated tasks rather than proof of autonomous scientific discovery.
Long-range information stored without entanglement can still demand an entangled measurement to recover it.
A theoretical construction lets local qubit probes record neighboring Ising-spin relationships in a fully separable state. For a range of probe overlaps, every separable readout—including local operations with classical communication—falls toward chance as the distance between target spins grows. A joint readout of the same record stays a fixed amount above chance over an overlapping parameter range. The result separates how quantum information is stored from what kind of measurement is needed to extract it.
Illustrative laboratory-glassware file image; it does not show patients, clinical sites, scans, lesions, models or study results. Rodolfo Clix / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
Adapted DINOv2 models generalized better on every external medical-imaging dataset, while specialist networks still drew sharper boundaries at home.
Researchers compared adaptations of DINOv2 with U-Nets and ResU-Nets trained from scratch across colonoscopy, breast ultrasound and dermoscopy. The adapted foundation model generalized better on every independent external dataset, while task-specific models remained competitive within the training distribution when data were abundant and more precisely delineated boundaries. The foundation model also missed small lesions more often, and the amount of adaptation it needed depended on the imaging domain. The study additionally found duplicate images across two widely used polyp benchmarks, underscoring why cross-site testing matters.
A 1,500-pair benchmark asks models and chemists to rank candidate protein-ligand poses that cofolding systems already know how to produce.
Pyrite turns protein-ligand pose selection into 1,500 pairwise comparisons across diverse complexes. In the reported study, expert chemists using 3D viewers and physics-based scores performed near chance, while an LLM reasoner reached up to 67% accuracy and an agent that wrote features for logistic regression reached 74%. The authors found that LLM-guided feature discovery generally beat direct textual reasoning and that reinforcement learning improved open-weight models. The benchmark measures ranking within its constructed pairs; it does not establish clinical or drug-development performance.
A connectome audit separates structural advantage from the hidden tuning choices that make a circuit model work.
Connectome-constrained models can appear to prove that measured wiring beats shuffled alternatives even when the decisive difference comes from unmeasured physiology. The study compares three protocols across synthetic ground truth and three Drosophila circuit models, including one in which each wiring is tuned only within physiology consistent with independent data. Some claims survived every protocol, but in a mushroom-body model the real wiring's 0.09 advantage became a 0.09 loss under the data-constrained protocol. The authors argue that structural claims need matched optimization, independently constrained physiology and a model that still functions under those constraints.
A memory-augmented vision-language-action policy counted pauses as progress because its demonstrations made time a shortcut for task state.
Researchers paused or lightly moved a robot during counting tasks without changing anything the task depended on. The released memory policy nevertheless stopped early: success fell from 47 to 16 of 50 in one task and from 44 to 10 in another, then recovered when the null-event frames were withheld. Changing timestamps alone made the policy stop early or repeat, while removing explicit time encoding did not eliminate the shortcut because time-sampled memory still filled with pause frames. Indexing memory by events recovered at least half the loss without retraining in the reported tests.
TanCRAFT tracks a partial build, forecasts a piece's motion and intervenes when the likely placement conflicts with the target.
TanCRAFT is a four-stage assistance system for assembly settings where users may choose their own construction order. It tracks scene state, reasons over the partial assembly, forecasts the near-future motion of a manipulated piece and warns when the predicted placement is incompatible with the target. In offline tests and a within-subject tangram study, the authors report fewer mistakes than in unguided or reactive sessions, along with the lowest mean workload and completion time. The evidence comes from a controlled tangram task, not an industrial deployment.
Conceptual Visualising AI file illustration by Tim West; it is not UGG, a scene model, geometry output or benchmark result. Tim West / Google DeepMind / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
UGG treats video, depth and camera pose as partially observed pieces of the same four-dimensional scene.
UGG combines RGB video with depth and camera pose in one generative framework for four-dimensional scene understanding. Different noise levels across modality and time let the model formulate visual synthesis and geometric prediction as conditional completion, so it can train across datasets that expose different subsets of annotation. At inference, it searches multiple plausible scene hypotheses and uses geometric evidence extracted from observed frames to improve predictions. The authors report gains over separate baselines in camera-controlled generation, depth estimation and pose prediction, including sparse-observation settings.
An independent BEHAVIOR Challenge team trained one checkpoint across 211 million demonstration frames, then routed fourteen hard tasks to a coding agent.
An independent team reports a single vision-language-action checkpoint for all 100 long-horizon tasks in the 2026 BEHAVIOR Challenge. Its learned policy scored 0.392 Q across the full set after a roughly one-pass, 200,000-step run over 211 million demonstration frames. Fourteen tasks were separately routed to a coding agent with observation and motion tools; on those tasks the authors report mean Q of 0.639 versus 0.108 for the learned policy. The public submission included only part of those agent results and finished with 0.446 Q and 27.3% full-task success.
A hardware model puts dynamic energy, fidelity and refrigerator attenuation into the same QRAM calculation.
Quantum algorithms often treat coherent access to classical data as free, but a new analysis models a QRAM query as a physical process on superconducting hardware. For a representative bucket-brigade query with a 12-qubit address register, the authors calculate 0.969 fidelity, 3.51 times 10^-4 joules of dynamic energy, 4.38 watts of average dynamic power and 64.4 watts at peak. Changing the operating point and attenuation profile can alter average power by a factor of two or more at fixed target fidelity. These are modeled hardware costs, not measurements from a deployed QRAM.
Earlybird trains on audio cut off mid-word and emits calibrated guesses from a 29-million-parameter streaming model.
Earlybird is a 29-million-parameter, 16.8-megabyte-at-four-bits speech recognizer trained to predict words from audio truncated at random moments, including mid-word. On LibriSpeech test-clean, the authors report that the median word was predicted correctly 90 milliseconds before the speaker finished it, while compared released streaming models emitted words 90 to 400 milliseconds afterward. Sixty-nine percent of words were predicted correctly before completion on that set. Results across six timing sets and eight Open ASR sets position the model as a low-latency research system, not a universal guarantee across speakers or environments.
Resonators coupled to on-chip resistive loads reset multiple transmons while isolating them from warmer external circuitry.
Researchers experimentally demonstrated reset and leakage removal across multiple superconducting transmon qubits using dedicated dissipation units with on-chip resistive loads. The architecture retains Purcell protection while reducing exposure to thermal noise from external wiring. They report a median reset time constant near 48 nanoseconds and, after a 400-nanosecond pulse, residual excited-state populations of 0.29% and 0.67% for two initialization cases. Leakage from the second excited state fell to 0.51% in the reported device.
Illustrative code-screen file image; it does not depict ARE-bench, a research workspace, an agent run or evaluation evidence. Daniil Komov / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
ARE-bench preserves unfinished files, old decisions and failed attempts instead of evaluating agents in a clean room.
ARE-bench contains 75 tasks reconstructed from real researcher sessions and full in-progress workspaces across seven research stages. Its 1,173 binary checklist items include not only the immediate deliverable but also historical decisions and material workspace constraints. Across seven frontier models, the strongest passed every item on 66.2% of tasks. Three models passed about 69% when judged only on delivery contracts and general norms, but 28.4% to 40.0% when historical and workspace requirements were added, showing how clean-task evaluation can overstate reliability.
Two single-flux-quantum protocols address the leakage and gate-time tradeoff that frustrates low-frequency qubit control.
A proposed fluxonium-control strategy uses picosecond single-flux-quantum pulses from Josephson digital circuits. One protocol sends high-frequency pulse packets through the flux channel for a simpler, robust drive; a second combines phase-shifted flux and charge sequences to shorten the gate. Multilevel simulations report single-qubit fidelities above 99.99%, with the simpler protocol tolerating moderate parameter and bias changes. The work is a simulation-based control proposal rather than an experimental gate demonstration.
WikiTrace links distilled lessons back to trajectory summaries and raw evidence so an agent can challenge bad advice before acting.
External agent memory can degrade performance when a concise lesson is wrong, misread or applied outside its original context. WikiTrace stores each lesson with a summary of why its source trajectory succeeded or failed and a link to the raw trajectory. When a lesson is underspecified, conflicts with another or comes from failure, the agent can descend through those links before acting. Across five agent benchmarks, the authors report that evidence-linked memory outperformed lesson-only baselines and reduced memory-induced degradation.
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.
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.
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.
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 2026-09-26 edition.
Desk Picks are selected by the newsroom. Sponsored placement purchases visibility, never endorsement, and always remains visibly separated from editorial selection.
Editorial illustration
