Research library
Search papers on equipment connectivity, quality inspection, production analysis, and manufacturing work.
Search and filter
New paper summaries
Summaries are editorial notes, not statements by the original authors. Check the original paper. Updated 2026-11-20.
- 1493Research Agents
DiscoveryAgent: A Framework for Scientist-centric Agentic AI
- Scientists select, refine, or replace hypotheses between research rounds, while audit agents interrogate checkpointed records to trace claims, reproduce analyses, and inform subsequent decisions.
- DISCOVERYAGENT produced the best variable-star classifier among evaluated baselines and, with scientist steering, a protein variant-effect predictor surpassed by only one of 106 ProteinGym baselines.
- Audit agents autonomously identify errors at higher rates than previous AI scientist agents, and scientists collaborating with audit agents identify 14x more valid errors versus scientists performing manual review.
- we developed a multimodal photometric-redshift estimator that resolves complementary failures of images and photometry, achieving 0.971 AUROC for low-redshift classification; documented an 83% rise in Florida homeowners-insurance premiums from 2013 to 2023
- 1492Machine Learning
Heteroscedastic Gaussian Process Q-learning for Sample Efficient Experimentation
- We learn the Q-function directly with a Gaussian process (GP), which avoids learning the dynamics
- If each Bellman target is treated as a noisy measurement of the Q-function, the learning rate that minimises the mean squared error is the Kalman gain, and applying it to every observed transition is equivalent to GP regression.
- Heteroscedastic Gaussian Process Q-learning (HGPQ) sets the noise variance of each target with a heteroscedastic GP, which learns how unreliable the targets are across the states and inputs.
- we prove a high-probability bound on the error of the value estimate, uniform over the explored region.
- Pooled over five simulated benchmarks, which include a chemical reactor and a bioprocess, HGPQ approaches the best return found by any method in fewer episodes than GP, kernel and deep RL baselines.
- 1491LLMs / NLP
K2-Horizon: A Frontier Fleet of Open Source Language Models
- The researchers released six fully open language models from a 0.9B dense model to a 375B-A23B MoE, for general-purpose reasoning and agentic work across math and science, software engineering, and deep research.
- The authors introduced MoVA, applying MoE to attention value projections, and softplus attention gating; the 36B-A4B model nearly matches a 32B dense model trained on the same data with 1/7 the active parameters.
- Uno is a lossless diffusion adapter that speeds up decoding by up to 2.2× without a separate draft model.
- The released artifacts include intermediate checkpoints from every training stage, pre-training data, training logs, and all training, evaluation, and infrastructure code; the pipelines produced roughly 10T synthetic pre-training tokens.
- Four of the models share one 22T-token pre-training corpus and recipe, and the 7B and 3.7B lead SWE-bench Verified by wide margins.
- 1490Research Agents
Read Before You Fit: Text-Derived Constraints for Discovering Differential Equations from Noisy Biological Data
- We introduce P2C, an agentic framework that translates textual descriptions of biological systems into a typed set of formal constraints, and use these constraints to restrict candidate ordinary differential equation (ODE) structures before numerical fitting.
- P2C consists of five stages: Text2Constraint extraction, ODE discovery, symbolic verification, parameter fitting, and model scoring.
- Across the four systems, adding Text2Constraint reduces the worst form error from 1.397 to 0.499 on clean data and from 13.096 to 1.095 for data perturbed with Gaussian noise of σ = 0.1
- P2C also achieves lower form error and trajectory error than several LLM-based baselines over the tested noise range.
- 1489Interpretability
When can an interpretability claim be trusted? Stability and certification of feature dimensionality, ridge leverage and the feature recovery rate
- Small changes to the model weights, a different training seed, or a different threshold can reverse them. In earlier work in this program, a norm threshold alone produced a 24.8% apparent failure of a well-known regularity.
- We characterize exactly where feature dimensionality is discontinuous: where a feature’s weight vector vanishes while some direction is orthogonal to all the other feature vectors.
- We replace feature dimensionality with a regularized (ridge) version and prove sharp stability bounds.
- We also prove certificates that state when a claim “feature i is recovered” must survive a given change of the dictionary.
- Our experiments use toy models, synthetic dictionaries and small dictionaries trained with three seeds; we have not tested real language models.
- 1488Interpretability
A Mechanistic Interpretability Study of an Astronomy Foundation Model
- We present, to our knowledge, the first causal mechanistic interpretability study of an astronomy foundation model, AION-1, focusing on its multimodal advantage in redshift prediction, where combining imaging with photometry outperforms photometry alone.
- We find that this advantage is largest for spatially extended and dust-attenuated galaxies, and that most of it is localized to the first of the encoder’s twelve blocks.
- Spatial extent is causally used, and one attention head dominates linear probing for it, yet ablating that head has negligible effect on redshift prediction: the information is distributed across heads and is decodable again one layer after we erase it.
- Dust attenuation shows the opposite pattern: it is only weakly decodable anywhere in the encoder, yet a single head is necessary for most of the multimodal advantage, particularly for the most dust-attenuated galaxies
- None of the astrophysically motivated operations we tested explains what this head computes.
- 1487AI / General
Living constitutions: Participatory alignment for sovereign foundation models
- We propose treating the development of a constitution as an evolving, participatory process akin to citizen science, iteratively drafted, tested, and refined.
- We therefore formulate five desirable principles: iterative, modular, justifiable, positional, and legitimate.
- propose that provisions be marked as settled, contested, or underdetermined, so that durable disagreement is recorded in the constitution rather than resolved away.
- we show via agent-based modelling that proposed process for maintaining living constitutions is robust against 50% bad actors among citizens – assuming a share of 75% participants with good intent in the deliberative body.
- 1486Research
AI-Driven Discovery of Entanglement Witness in Spin Hamiltonians
- We construct local-unitary-invariant representations of measurable Pauli correlators up to four-body order, sparsely select 16 candidate invariant feature channels, and learn their spatial structure using a Transformer architecture with rotary positional embeddings.
- Using only expectation values of these low-order observables, EntNet achieves > 98% recall on certified entangled states with near-zero false-positive rates on separable states.
- Remarkably, it exhibits transferability across system sizes and distinct Hamiltonian families, including frustrated and disordered spin models, with minimal loss of accuracy.
- decision boundary tracks the excitation gap of Jordan–Wigner quasiparticles. These results establish a transferable and physically interpretable framework
- 1485Research
SciCheck: A Framework for Verifying the Integrity of Deep Learning Training
- The verifying scientist judges the run in two layers: a statistical judge model that compares per-step gradient statistics against the verifier’s own reference run, and a spot-check that re-executes randomly chosen training segments whose checkpoints are committed to the ledger one epoch at a time.
- On an MLP, a CNN, and an autoencoder at 100k–500k parameters, under gradient sign flipping, label poisoning, and fabricated runs, SciCheck reaches F1 of 0.966, 0.906, and 0.976 while rejecting 0–4.4% of honest runs.
- For the attacks we test, the spot-check accounts for most of this accuracy, and we characterize what each layer can detect.
- The effective-signal statistic falls as (1 −2s)2 under sign flipping, so statistics catch strong enough sign flips; fabricated runs are invisible to every snapshot detector we test
- 1484Research Agents
Discovering formally verified quantum programs with AI
- The researchers built a Lean verification library based on quantum Hoare logic (QHL) and a multi-agent research workflow, proving resource reductions on asymptotic algorithms and large-scale scientific simulations.
- Agent teams published independently re-verified solutions, partial results, and dead ends to a shared wiki, and exchanged ideas and critiques in scheduled cross-team conferences.
- Parallel Grover search reduced expected oracle rounds by up to 2.57×, and fixed-point arcsin achieved a 290× spacetime cost reduction.
- On the fusion alpha-particle stopping power problem, the workflow found a solution using 4,338 logical qubits, lowering spacetime cost by 11.8×.
- The authors state that these simulations still demand far more resources than projected for the first fault-tolerant quantum computers, and reaching early hardware requires further orders-of-magnitude improvements.
- 1483Machine Learning
Sage S2.1 Flash Technical Report
- Sage Sci-Perception connects dedicated encoders for molecular graphs and biological sequences and numerical spectrum patches to the shared backbone.
- The team built verifiable environments for tool use, search, and coding where task completion is checked against execution results.
- ToolSpace built 77,760 tool-call pattern nodes from 2,042 distinct tools and organized them into 804 environments.
- Among 33 benchmarks, the model ranked first on OfficeQA Pro (72.2), MLE-bench Lite (90.3), Harvey LAB, and SpreadsheetBench 2.
- It ranked second on OSWorld-Verified (85.2), WebArena-Verified (67.6), and DeepResearch Bench II (54.2).
- Generated SFT trajectories averaged 18.86 tool calls, reaching up to 209.
- 1482Research Agents
Formally verified harness for Shor's order finding algorithm with $O(n^{2+ε})$ gate cost
- The authors built a harness where participants submit only a recursive multiplication design policy and Lean automatically applies correctness and gate-cost proofs.
- They proved as a theorem that the recursive multiplication circuits improve the cost scaling to O(n2+ϵ) for every valid policy.
- Exact optimization is intractable at cryptographic sizes, with an estimated 3.2 × 1051 core-hours to certify the best known tree at n = 2048.
- In Qualtran resource estimation, the best fixed k was 1.820 × 1012 T-gates at k = 4.
- A policy varying k at each level achieved a lower count of 1.685 × 1012 T-gates.
- Without further optimization, the PhaseProduct approach would need a modulus size of about 3.8 million bits to beat the Gidney-Ekerå method.
- 1481AI / Agents
Security of Deep Research Agents: Attack Surface, Defenses, Benchmarks, and Agenda Seeding against Research Planning
- The paper presents a unified threat model splitting deep research agents into retrieval, planning, evidence integration, and report generation.
- Coding 34 primary studies showed that only 26.5% involve the planning stage and none performs end-to-end poisoning evaluation on commercial products.
- The authors proposed the agenda seeding attack, which writes no instructions and only introduces a fictitious evaluation metric.
- Across 3,520 sandboxed research runs under the same budget, GEO and prompt injection were nearly ineffective.
- Agenda seeding made the attacker's target the top pick in 16% to 28% of reports.
- Provenance tagging preserved utility but could be bypassed by multi-site bait.
- 1480Software Engineering
The Hidden Methodological Layer: Uncovering Threats to Validity in LLM-Assisted Empirical Software Engineering
- The authors re-analyzed 69 LLM-assisted research tasks from 50 primary studies.
- Only 30 of the 69 tasks (43.5%) explicitly reported at least one LLM-related threat to validity.
- The automation category had the highest reporting rate, with 22 of 42 tasks (52.4%) reporting threats.
- Threats to empirical evidence construction accounted for 19 occurrences, including 10 LLM classification and labeling errors.
- The authors argue that prompts, model configurations, interaction protocols, and validation procedures must be visible as part of the research method.
- ChatGPT accounted for 38 model usages, the most in the corpus.
- 1479Research Agents
Towards Agentic Scientific Discovery in Distributed Systems
- The authors built Query-ATLAS, a controller that selects tools through an interface exchanging equations and aggregate scores across clients.
- With overlapping support, Query-ATLAS recovered 23 of 36 equations exactly versus 21 for local PySR.
- The difference came from two relativistic mass cases, and the selected expressions had test NRMSE of approximately 4.3 × 10−16.
- With partitioned support, Query-ATLAS recovered 18 versus 19 for local PySR, reversing the ordering.
- Query-ATLAS took 4.70 times the local baseline's mean runtime in the overlapping setting.
- The authors conclude that controlled evaluation of adaptive search is needed rather than a claim of general superiority.
- 1478Research Agents
Eliciting Research Taste in LLMs through Future Research Direction Choice
- The authors built 80 pairwise comparison queries from 19 ML papers published after model training cutoffs.
- Across 8 models, author alignment ranged from 45.0% to 67.5% and stable divergence from 15.0% to 36.2%.
- 45 of 80 pairs elicited stable divergence from at least one model, totaling 169 divergence events.
- Across four prompt variations, models retained divergences with weighted means from 66.9% to 80.5%.
- In the pilot human study, 58 of 95 fixed-choice responses (61.1%) selected the author-recommended direction.
- The most frequent divergence transition was Infrastructure-to-Methodological substitution with 55 events.
- 1477AI / Agents
Omni-Workspace: Benchmarking Omni-Modal AI Assistants on Multimodal Workspace Tasks
- The authors built a benchmark of 380 tasks in workspaces mixing audio, video, image, and document files.
- Evaluating 18 models, the best score was Gemini 3.5 Flash at 65.93 on average, and only two models scored above 60.
- On L2 tasks requiring tool use and code execution, every model scored lower than on L1.
- Error analysis identified multimodal context inflation and task-state drift as recurring problems.
- OmniFocus, which adds working-memory management and explicit task-state tracking, improved both performance and efficiency.
- On Gemini 3.1 Pro Preview, OmniFocus used 271.07k input tokens versus 615.81k for the setting without both mechanisms.
- 1476AI / Agents
The Critic Can Hurt: On the Safety of Critics in LLM-Guided AutoML Search
- The authors built a four-arm audit comparing no critic, random feedback, real critic, and oracle under the same budget and frontier.
- A naive 8B critic scored below the no-critic floor on 7 of 8 binary tabular datasets.
- The worst case on ADULT reached a loss of 0.245 AUC.
- Fault injection showed that random errors are nearly costless while systematically directional errors consume effective search rounds.
- An idealized model yields the necessary condition q(m−1) > β for a critic to help.
- A validation gate and a learned detector reduced the average loss, but neither guaranteed the no-critic baseline on every dataset.
- 1475AI / Agents
Vibe synthesis: hierarchical agentic decision-making for intent-driven multistep molecule synthesis
- The system uses a Manager Agent that keeps the campaign objective and 10 task agents that call computational tools and automated synthesis devices.
- Connecting a Bayesian optimization tool with the reaction device raised the Mitsunobu reaction yield to 81 ± 1% after seven optimization rounds.
- The system reproduced a 10-step total synthesis of catharanthine from a 1987 paper with 722 recorded actions and obtained 160 mg of product.
- In that campaign, active human input was 7.3% and simple confirmation was 4.2%, while agent reasoning accounted for 69.7%.
- In cyclic peptide campaigns with unnatural amino acids, all 15 analogues passed the quality criteria with more than 95% HPLC purity.
- In the final step, the system chose Rh-catalyzed hydrosilylation over the literature multistep reduction and obtained a 95% isolated yield.
- 1474Research Agents
PaperCritic: Training AI Reviewers to Assess the Soundness of Papers
- A human study with seven AI researchers identified soundness and substantiveness as the two quality criteria for review points.
- The authors built PaperSound-23K by labeling 23K review points collected from 1K papers with quality labels.
- They trained PaperCritic with reinforcement learning that rewards the F1 score of precision, from soundness and substantiveness, and recall, from reference coverage.
- PaperCritic-4B achieved an F1 score of 45.1 on the in-domain test set, 21.4 pp higher than the best trained 4B baseline.
- PaperCritic-4B beat the frontier models DeepSeek-V4-Flash-0731 and Claude Opus 5 by 5.8 pp and 6.0 pp in F1, respectively.
- Training on ICLR paper reviewing transferred to Nature-family papers, code review, legal contract review, and story writing feedback.
- 1473Robotics
Toward Resilient Self-Driving Laboratories through Telerobotic Interventions
- The authors built a teleoperation interface that controls a six-degree-of-freedom robotic arm (UR3e) with a 3D mouse for fault recovery in self-driving laboratories.
- Six participants performed a total of 86 trials across three simulated fault scenarios: process interruption, initialization error, and escalated fault.
- The decoupled controller achieved higher success rates than the coupled mode and reduced task completion time by a mean of 25.6 s.
- The decoupled mode produced a mean paired difference of -3.99 rad in rotational motion relative to the coupled mode, with a large effect size (Cohen’s dz = -1.08).
- Multi-waypoint reference-path guidance reduced task duration by a mean of 40.1 s and path length by a mean of 218 mm relative to unguided operation.
- As a limitation, the participant count of 6-8 was small, and guided trials followed unguided trials, so practice effects were not fully separated.
- 1472Research Agents
NeoResearch: Literature- and Experience-Guided Autonomous Time Series Model Discovery
- NeoResearch used a dual-loop workflow coupling an outer architecture-discovery loop with an inner training-optimization loop.
- The agent mined scientific literature from the web, extracting modeling mechanisms, empirical evidence, and applicability conditions into testable hypotheses.
- A hierarchical experience memory linked literature findings with experimental outcomes and consolidated conditional lessons for reuse within and across tasks.
- Across 10 datasets, NeoResearch achieved the lowest error in 19 of 20 dataset-metric comparisons.
- In dual-loop abations, the full system achieved lower errors than both loop-removed variants on all ten tasks.
- As a limitation, on the Windy dataset DLinear and DeepAR achieved lower MSE, so dominance was not universal.
- 1471Interpretability
Interpretable weight-space programming builds transparent transformers and brain encoding models
- iWSP has the agent write weights directly in code and forbids training, forcing iterative revision of human-readable models.
- On 13 algorithmic tasks, all runs reached 100% accuracy on both the visible set and the hidden set.
- The smallest model reaching 100% visible accuracy had a median of 166 parameters in the Opus runs.
- In visual cortex prediction on the NSD dataset, iWSP reached 0.443 in early visual areas, ahead of the best interpretable baseline at 0.400.
- In language fMRI, tracking a jump from 0.067 to 0.104 let iWSP uncover a scan-edge timing confound in the data.
- In auditory ECoG, iWSP alone scored 0.074 below Whisper's 0.082, but combining them raised performance from 0.070 to 0.075.
- 1470Agents
Can You Trust Your Finance Agent? Drift, Diagnosis and Auditable Intervention on One Shipped Financial Research Agent
- Seven identical runs of the agent on 18 questions passed between 105 and 123 of 149 rubric criteria, and the bill varied by up to 17%.
- Classifying 20 failed trajectories under a pre-registered taxonomy assigned 13 to the agent's instructions, 3 to its tools, 4 to the model, and none to its data.
- Four blind readers named the same fault as the analyzer in 75 of 80 cases, supporting the diagnosis's trustworthiness.
- Under a cost objective, the final configuration reached 0.34× the baseline cost on the held-out questions at accuracy within noise.
- The accuracy objective gained 14 points on the development questions but only 1.5 on the held-out ones, with a confidence interval straddling zero.
- As a limitation, this is a case study of one agent and nine test questions, not a general result.
- 1469Research Agents
MineralExp: Benchmarking LLM Agents for Experimental Design in Real-World Mineral Exploration
- The authors organized geochemical experimental design in mineral exploration into 19 experimental types across five mineral-system processes: Source, Transport, Deposition, Modification, and Preservation, turning it into a multi-label problem.
- The benchmark was built from published ore-deposit literature, yielding 277 expert-reviewed cases and 1,572 reference experimental plans.
- Across 28 configurations of 4 agent harnesses and 7 LLMs, overall recall ranged from 70.76% to 91.56%, yet the highest overall F1 was only 63.84%.
- While reference cases involve 3.49 experiments on average, DeepSeek-V4.1-Flash predicts nearly 10 per case, dropping its Precision to approximately 32%.
- In LLM-as-a-Judge evaluation, partial coverage ranged from 83.59% to 93.51%, but strict coverage stayed between 22.65% and 36.83%.
- As a limitation, the domain harness Biomni consumed approximately 67k input tokens per case, more than 12× the input token cost of OpenCode, yet scored lower.
- 1468Machine Learning
Uncertainty-Aware Continual Learning for 2D XRD Interpretation Using Simulated and Experimental Data
- The framework combines MMD-based domain adaptation, CNN-SNGP uncertainty estimation, and replay with knowledge distillation for continual learning.
- All models achieved approximately 94-96% accuracy on simulated and experimental validation subsets, and CNN-SNGP-adapt reached the highest mean experimental validation accuracy of 95.69 ± 4.17%.
- On held-out materials, CNN-SNGP-adapt achieved 50.3 ± 6.9% experimental accuracy, far above the 29.1 ± 10.2% of CNN.
- In paired material-level analysis, the adapted model gained 21.2 percentage points over CNN and 6.0 percentage points over CNN-SNGP.
- The proposed continual-learning strategy achieved higher accuracy on both earlier and newly introduced blocks than naïve fine-tuning, with substantially less forgetting.
- As a limitation, this study is an offline proof-of-concept using predefined data blocks and did not select experiments or control instruments.
- 1467Machine Learning
Diffusion Tree Sampling for Inference Time Adaptation of Crystal Generative Models
- The authors applied DTS to crystal generation by treating the denoising trajectory of diffusion models as a Monte Carlo tree search tree steering toward a reward-tilted distribution.
- When using conditional generative models as search proposals, importance weights correct for double tilting and preserve the target distribution.
- The multi-objective extension MO-DTS keeps per-objective value estimates and samples a preference each traversal, so a single tree samples the whole Pareto front.
- In MatterGen experiments at matched compute, the number of stable structures rose from a baseline of 868 up to 1432, with the stability rate rising from 68.4% to 75.7%.
- In DiffCSP experiments, the stability rate kept increasing as the compute budget grew from 105 to 107 NFEs, showing inference-time scaling.
- As a limitation, the method relies on pretrained predictors for reward evaluation, and experimental synthesis validation remains a separate step.
- 1466Distributed Systems
Scaling Pêlagos: Multi-GPU Orchestration for Relay Belief-Propagation Decoding
- The authors designed a multi-GPU orchestration with one process per GPU, parallel host preprocessing, and shared-memory output collection.
- In the original design, eight-GPU speedups were only 2.20× and 3.22× on the one-gross code, improving to 5.68× and 6.23× after optimization.
- For the two-gross circuit, eight-GPU end-to-end speedups reached 6.92× on MI300X and 7.01× on MI355X.
- Using GPU decoding time, the two-gross speedups were 7.58× and 7.45×, approaching near-linear scaling.
- With 8.0 TB/s memory bandwidth versus 5.3 TB/s on MI300X, MI355X sustained higher throughput at every GPU count.
- As a limitation, batch sizes were bounded by signed 32-bit indexing to 5,114 for one-gross and 1,707 for two-gross.
- 1465Research Agents
Auditable Human-Agent Teams for Autonomous X-ray Experimentation
- VISION-CC ran autonomous synchrotron X-ray experiments with Operator, Analyst, GP, and Writer specialist agents on top of a coding-agent team.
- Across three beamtime sessions spanning more than 50 active hours, it mapped nanoparticle superlattice samples with adaptive sampling.
- Across the three experiments, 61-67% of messages involved interpretation or judgment.
- The share of methodological and scientific contributions increased from 46% in the first experiment to 56-61% in later experiments.
- Provenance showed that 59-68% of error corrections involved multiple participants.
- As a limitation, the audit pipeline's individual labels and links carry uncertainty, and findings rest on aggregate shares.
- 1464Interpretability
Glyph: An Interpretability Engine for AI-Driven Scientific Discovery in Clinical Data
- Glyph is a multistage framework that decomposes foundation model embeddings with sparse autoencoders and interprets them as clinical concepts.
- The researchers embedded abdominal CT volumes, radiology reports, and lab measurements from 41,628 patients using three foundation models.
- Testing 604,185 concept-outcome associations against 141 prognostic outcomes yielded 397 predictive concepts and 4,463 significant associations.
- The discovered associations recovered established prognostic relationships across hepatic, renal, cardiovascular, and oncologic disease and reproduced in held-out external cohorts.
- The top 10 Glyph concepts per axis reproduced more frequently than literature markers, 64.8% versus 41.2%.
- Glyph concepts showed a larger median odds ratio of 1.31 versus 1.16 for literature markers.
From research to product use
Operating capabilities, pilots, and technologies in development are identified separately.
View technology