A new dataset keeps 558 complete AI-scientist trajectories, revealing error patterns that nearly identical success rates concealed.
Editorial illustration
Concept illustration: OpenDiscoveryTrace preserves scientific-agent process records; the papers, instruments and marks are symbolic, not a literal trajectory or laboratory. Original editorial illustration generated with built-in Codex Image Gen for The Machine Press, 2026-09-10.
OpenDiscoveryTrace records nine fields for every step—including tool calls, observations, errors, revision triggers and self-reported confidence—across 558 trajectories on 124 scientific tasks. In a pilot analysis of 363 LLM-judged runs, three frontier models finished at similar reported success rates of 84 to 89 percent, yet one produced 2.5 logged errors per trajectory versus 0.08 for another, a thirtyfold difference. Their failure shapes also diverged: tool misuse dominated one model while reasoning errors dominated another. The dataset is designed to make scientific-agent process auditable, though the headline comparisons rely on the paper’s trace schema and judge-based pilot rather than independent replication.
Concept illustration: spectra motivate a supermassive-star interpretation for Little Red Dots; the cutaway, abundance streams and interior are not directly observed. Original editorial illustration generated with built-in Codex Image Gen for The Machine Press, 2026-09-10.
JWST spectra found magnesium-poor, aluminum-rich gas in Little Red Dots—the chemical fingerprint expected from extremely hot hydrogen burning.
Deep SPURS spectroscopy of compact early-universe Little Red Dots found central gas depleted in magnesium and enhanced in aluminum at about one percent of the Sun’s metallicity. The authors argue that ordinary massive stars at those redshifts, along with changes in ionization, geometry or dust, cannot reproduce the reported pattern. Fully convective supermassive-star models can, and the inferred burning conditions imply objects of at least 10,000 solar masses—roughly one hundred times heavier than any star observed today. The result offers one possible bridge between globular-cluster abundance anomalies and seeds of massive black holes, but it remains an interpretation of spectra and stellar models rather than a direct image of such a star.
Population models inferred rockier planets around Sun-like stars and tightly bounded volatile inventories around M dwarfs.
A mixture model linked interior structures to the measured density distributions of small planets in three samples. Under one rocky-population parameterization, planets around FGK stars had characteristic core mass fractions about 16 percent higher than planets around M dwarfs. An alternative model estimated 89.9 to 97.0 percent of FGK-hosted planets as rocky, while the M-dwarf rocky share remained much less certain. Volatile-bearing planets were common in the model but concentrated at small atmospheric and water fractions. These are population-level inferences shaped by sample selection and interior-model assumptions, not direct compositional measurements of individual worlds.
Generic code-screen file image, used illustratively; it does not show the studied skills, subagents, tasks or results. Markus Spiske / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
A study found subagent execution stronger than loading reusable skill packages directly when each skill exposed a clear input-output contract.
The researchers compare two ways of reusing procedural knowledge: putting a skill package into the main agent’s growing context, or invoking that package in a fresh subagent context. On long-horizon tasks, the subagent approach performed better when the package stated a clear input-output contract and encoded actionable procedure. The improvement came with extra coordination tokens, so the result is a trade-off rather than a universal argument for delegation. The paper’s central claim is architectural: how knowledge is invoked can matter as much as what the knowledge contains.
A state-path menu raised ToolBench online success from 0.737 to 0.898 by ordering prerequisite producers before the final action.
Tool retrieval usually ranks interfaces by their relevance to the user’s request, which can surface the final action while omitting the tools that create its inputs. State-Path Tool Menu instead predicts an executable route from the current state to the desired outcome, retrieves missing-input producers and then orders producers before consumers. On ToolBench, the authors report online success rising from 0.737 to 0.898. A 32-tool menu also covered more complete chains than the official 128-tool list, suggesting that executable order can outweigh raw menu size within the evaluated setup.
Decision-focused active learning reached a recycled-magnet enrichment maximum with 16 to 24 experiments instead of 48.
Using records from Pacific Northwest National Laboratory’s CICERO selective-precipitation workflow, the authors retrospectively tested how active learning could choose experiments for critical-material recovery. Adaptive policies found the best recorded neodymium-iron-boron enrichment after 16 to 24 wells, while nonadaptive space filling required 48; two reconstructed policies tied at 16. Results for samarium-cobalt magnets and produced water exposed purity, yield and measurement-assumption trade-offs. The work is conditional and retrospective, and the authors call for a preregistered prospective test before claiming scale-up gains.
Internal-representation probes outperformed surface and sequence baselines across Bash, SQL and Python agent benchmarks.
The study tests whether an agent’s internal residual-stream representations reveal eventual task success before the final answer. Latent Trajectory Dynamics summarizes how representations change across a run, while an Action Representation Probe reads states formed at action decisions. Across Bash, SQL and Python benchmarks and three open model families, both methods consistently beat surface-generation and sequence-based calibration baselines without changing prompts or sampling extra rollouts. The evidence is benchmark-specific and requires access to internal model activations, which limits applicability to closed systems.
ContractEval maps query-active procedural obligations to trace evidence and separates omissions, wrong branches and invariant breaches.
Output-only scoring can approve an answer even when the agent skipped the branch, check or dependency that justified it. ContractEval turns a procedure into obligations activated by the current query, then matches those obligations against a response or execution trace. Under gold expected and observed graphs, it localized every injected structural failure in a controlled audit suite; language-model extraction retained much of the signal but remained sensitive to calibration. The authors explicitly frame the method as an audit aid, not a compliance guarantee.
NASA OSAM-1 robotics file image, used illustratively; it does not depict the studied demonstrations, datasets or world model. Use does not imply NASA endorsement. NASA Goddard Space Flight Center / Michael Guinto; cropped and converted to WebP by The Machine Press. Use does not imply NASA endorsement.
Separating habit, shared dynamics and camera nuisance improved low-shot transfer across three robot datasets.
Teleoperated demonstrations can look multimodal even when the executed action nearly determines the next physical state. The paper models three distinct causes: operator habit in choosing actions, shared physics after the action, and observation nuisance such as camera appearance. Its adaptation rule freezes a shared physics readout and updates only a thin interface. Across StackCube, DROID and RH20T, the split improved low-shot transfer and resisted corrupted adaptation data better than training from scratch, including multi-view pixel tests. The authors do not claim that every latent action is a human habit.
AccelMPC ran onboard a 35-gram drone, cutting solve time up to 15.6-fold and energy-delay product 195.4-fold.
AccelMPC co-designs an alternating-direction solver, numerical representation, FPGA mapping and a custom six-gram circuit board for a 35-gram Crazyflie drone. Hardware experiments ran constrained model-predictive control at one kilohertz with dynamic obstacles and optimization problems exceeding 20,000 variables. Against embedded microcontroller solvers, the authors report up to 15.6 times faster solves and a 195.4-fold improvement in energy-delay product. The open design files support replication, but the measurements describe the team’s hardware and test conditions rather than every tiny-aircraft platform.
A simulated lifting policy reached 93.3 percent task success after reward revision, but only 33.3 percent within its force budget.
A compact visuotactile world model cut simulated endpoint-force prediction error from 1.058 to 0.228 newtons, yet a simple tactile-persistence baseline remained better. After revising the reward on fresh simulated environments, ten-centimeter lifting success rose from 20.0 to 93.3 percent, while only 33.3 percent of runs stayed within an eight-newton per-finger limit, compared with 70.0 percent under direct force feedback. Separate GelSight analysis showed trajectory-level calibration covering 87.36 percent of complete recordings at a nominal 90 percent. The sensing and simulator studies were not transferred into one physical robot demonstration.
Finite-energy GKP states beat squeezed vacuum in the model when their envelope was broad enough, though equal photon budgets reversed the result.
The analysis compares standard coherent-plus-squeezed-vacuum interferometers with a coherent beam paired to a finite-energy Gottesman–Kitaev–Preskill state. With a sufficiently broad envelope, the multi-peak GKP resource produced better modeled phase sensitivity even in the presence of optical loss. At equal or lower mean photon number, however, squeezed vacuum performed better because the GKP envelope narrowed. Loss eventually erased either advantage as both resources approached ordinary vacuum, making energy accounting central to the proposal.
A variational method optimized sparse measurements and nonlinear physics across an entire spacetime grid instead of marching forward step by step.
The proposed variational quantum algorithm encodes a full discrete spacetime solution in one state, then minimizes both mismatch to sparse sensor readings and violation of the governing nonlinear partial differential equation. Numerical demonstrations reconstruct one-dimensional Burgers and Kuramoto–Sivashinsky velocity fields. Joint optimization lets information at all times constrain the result rather than propagating errors through sequential time steps. The paper presents simulations and a compact formulation, not evidence of practical quantum advantage on hardware.
Staged technology file image, used illustratively; it is not a quantum telemetry stream, processor console, syndrome record or leakage measurement. Rafael Minguet Delgado / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
A 156-qubit processor’s execution record identified a distance-one logical input with total variation of at least 0.927.
Quantum error-correction hardware emits syndrome, decoder, reset and timing records separately from its logical answer. The paper proves exponentially small logical-input leakage only under explicit locality, correctability and convergence assumptions, with each logical axis paying its own code distance. On a 156-qubit superconducting processor, the sufficient certificate missed by a factor of 21.5 and could not be invoked. Directly measured records from a distance-one memory identified the input with total variation of at least 0.927; randomized encoding returned the statistic to the floor without extra two-qubit gates. The result warns that fault tolerance does not automatically make operational telemetry private.
JWST spectroscopy measured a highly magnified object at redshift 5.66 with low metallicity and unusually efficient ionizing output.
The Small And Lensed Source Arc, or SALSA, is magnified by more than one hundred times behind the galaxy cluster Abell 2744. JWST NIRSpec observations robustly detected hydrogen-alpha and measured oxygen emission, yielding an oxygen abundance of 7.43 plus or minus 0.09 on the usual logarithmic scale. The source’s reported ionizing efficiency is log xi-ion 25.49 and its Lyman-alpha escape fraction 0.39 plus or minus 0.14. Those properties place SALSA among the most extreme known star-forming sources of its epoch, while the lens model and inferred intrinsic faintness remain essential to the interpretation.
ALMA mapped localized heating in the edge-on Gomez’s Hamburger disk alongside asymmetry, a disk wind and a gas overdensity.
High-resolution ALMA maps of the edge-on Gomez’s Hamburger disk found a narrow one-sided arc of sulfur monoxide near a previously identified gas overdensity. The authors say the localized heating could surround an early giant protoplanet or a disk fragment. Carbon-monoxide observations also show non-Keplerian emission consistent with a disk wind, while the continuum is asymmetric north to south. A revised distance of 139 plus or minus 24 parsecs places the system near Scorpius–Centaurus. The sulfur chemistry is a candidate tracer, not a direct detection of a planet.
BuilderTorbjørn Ludvigsen (tobben) and Hangprinter contributors
Suspends a print head from tensioned lines anchored around a room, replacing a rigid gantry with cable geometry so an open RepRap can work across an unusually large build space.
Visit HangprinterOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-06.
BuilderDave Hakkens and Precious Plastic contributors
Publishes replicable shredders, presses, workspace plans, and shared know-how so small local teams can sort waste plastic and turn it into reusable flakes and sheet material.
Visit Precious PlasticOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-06.
Pairs a square e-paper display with an ESP32-S3 and publishes the hardware, software, documentation, and case files so owners can build and program their own watch faces.
Visit WatchyOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-06.
Turns a desktop trackball into an inspectable kit by publishing its mechanical and electrical design files, assembly documentation, and programmable QMK firmware.
Visit Ploopy Classic 2Original Codex Image Gen concept art from 2026-07-10; carried forward from the validated September 9 edition.
Desk Picks are selected by the newsroom. Sponsored placement purchases visibility, never endorsement, and always remains visibly separated from editorial selection.
Editorial illustration
