Our Work

Electrification, not sustainable aviation fuel, delivers the largest modelled local water-quality benefit per tonne of CO₂ avoided-code

The railway intrusion scene video dataset collected by the forward camera of the train and obtained based on the first frame image enhancement

Explanation methods for natural language processing (NLP) come from two literatures that rarely meet: feature attribution and probing for task models, and the interpretability of large language models…

Explanation methods for natural language processing (NLP) come from two literatures that rarely meet: feature attribution and probing for task models, and the interpretability of large language models (LLMs), from chain-of-thought reasoning to sparse autoencoder features and circuit tracing. We place both in a unified taxonomy with four dimensions: scope (local vs. global), mechanism (how the explanation is computed), model access (model-agnostic vs. model-specific), and output form (importance scores, rules, examples, counterfactuals, rationale spans, concepts, natural language). Faithfulness and plausibility come apart, and the erasure tests built for extracted rationales do not apply to a chain of thought, which is an output rather than part of the input, so chains of thought need intervention tests of their own. The access a deployment allows settles which mechanisms are usable before scope or audience matter; our practical decision frameworks start from that question and are traced through three worked deployments. In our synthesis of LLM-era interpretability, a lineage table shows that most current techniques rebuild or extend older ideas under new constraints of scale and access, while two problems are new: keeping reasoning traces monitorable under training pressure, and testing whether models can introspect. Faithfulness metrics disagree with one another, and sparse autoencoders do not yet beat simple baselines on standardized benchmarks, so evaluations should name the tests they ran instead of reporting one score. We also survey applications across healthcare, legal and financial services, social science research, content moderation, and education, and close with six open problems.

BeamZ is an open-source electromagnetic simulation framework for photonic design using the finite-difference time-domain (FDTD) method. It provides CUDA-accelerated simulations, a Python API, and diff…

BeamZ is an open-source electromagnetic simulation framework for photonic design using the finite-difference time-domain (FDTD) method. It provides CUDA-accelerated simulations, a Python API, and differentiable workflows for gradient-based inverse design of photonic devices.

This release fixes several long-standing errors in the dynamics of bioNC. Every dynamic result changes: forward dynamics, inverse dynamics, Lagrange multipliers and energies. Kinematics and inverse ki…

This release fixes several long-standing errors in the dynamics of bioNC. Every dynamic result changes: forward dynamics, inverse dynamics, Lagrange multipliers and energies. Kinematics and inverse kinematics of data-driven models are unchanged. If you have struggled with forward simulations that drifted, it's now fixed. The mass matrix was wrong for every segment The generalized mass matrix is built from the pseudo-inertia of each segment, which had two errors: The Huygens term was wrong: a missing mass factor, and a scalar subtracted from every entry. The pseudo-inertia must be the second moment of mass ∫ n nᵀ dm, not the inertia tensor. G was therefore wrong, and often not positive semi-definite. It now equals the exact rigid-body mass matrix to 1e-16. Across the examples, pendulum periods were off by 20–26% and energy drifted by up to 66 J. Both now match the analytic values, and only RK4 integration error remains. Inverse dynamics, validated against R. Dumas' toolbox Fixed four bugs in the recursion: swapped force/torque outputs, the wrong sign of the child's reaction on its parent, the child's moment that never reached the parent, and the wrong lever arm when transporting it. Ankle, knee and hip loads now match a port of Dumas' Inverse_Dynamics_GC.m to 1e-8 on his gait dataset (see docs/inverse_dynamics_dumas_comparison.md). External forces Every moment transport used the wrong lever-arm sign, which affected any force applied away from the proximal point, in both forward and inverse dynamics. Non-orthogonal segments (#178) compute_transformation_matrix returned B transposed. On non-orthogonal segments, the centre of mass, the inertia, and markers, vectors and muscle via points given in segment coordinates were misplaced (by about 0.1 m in a typical case). Orthogonal segments were not affected. Joints SphereOnPlane: the parent and child Jacobians were swapped. The Feikes parallel knee now simulates cleanly. GROUND_WELD accepts Q_child_ref to fully weld a segment. With only rp/rd, the axial rotation stayed free and forward dynamics was singular. joint_dof_indexes no longer overlaps between joints with several degrees of freedom. New: joint actuation forward_dynamics(..., joint_generalized_forces=...) now works in numpy and casadi. It takes minimal-coordinate torques about each joint's Euler or hinge axes (or forces along its translation axes). Each joint exposes its axes through joint.dof_axes(), and the actuation is power-consistent and checked against the inverse dynamics. Previously the argument was silently ignored. Other fixes potential_energy() now includes g. Examples and documentation README: a new animated section on what the transformation matrix B does (#117). examples/transformation_matrix/: every frame change (segment ↔ natural ↔ global), written in plain numpy and then with bioNC. Feikes knee: a pendulum test with a distal femur view (condyle spheres, tibial planes, ligaments), the knee rhythm and joint loads. actuated_3d_pendulum: a pendulum held still by joint torques, and a constant-torque case. Breaking changes Simulations: forward and inverse dynamics outputs, energies and Lagrange multipliers change. Re-run your simulations; saved snapshots are no longer valid. Saved models: models of non-orthogonal segments whose CoM, inertia or markers were defined in segment coordinates must be rebuilt. Example equilibria: in pendulum_with_force and double_pendulum_with_force, force_equilibrium now applies its force at the centre of mass. It only balanced before because of the lever-arm bug. ExternalForceInLocal: custom transformation matrices must now be B, not its transpose. ##Merge-requests Fix(DYNAMICS): Inverse but consequently forward dynamics too by @Ipuch in https://github.com/Ipuch/bioNC/pull/174 fix(sphereOnPlane) : and feikes by @Ipuch in https://github.com/Ipuch/bioNC/pull/175 Fix compute_transformation_matrix returning B.T by @Ipuch in https://github.com/Ipuch/bioNC/pull/181 ground-weld fix by @Ipuch in https://github.com/Ipuch/bioNC/pull/176 fix(potential energy) by @Ipuch in https://github.com/Ipuch/bioNC/pull/177 feat(actuated pendulum) by @Ipuch in https://github.com/Ipuch/bioNC/pull/179 fix(joint dof index) by @Ipuch in https://github.com/Ipuch/bioNC/pull/180 Full Changelog: https://github.com/Ipuch/bioNC/compare/0.12.1...0.13.0

The ASTERISK 2025 Annual Report provides an overview of the project’s activities and progress during 2025. ASTERISK is a project co-funded by the Clean Hydrogen Partnership and the European Unio…

The ASTERISK 2025 Annual Report provides an overview of the project’s activities and progress during 2025. ASTERISK is a project co-funded by the Clean Hydrogen Partnership and the European Union, dedicated to advancing the development of sustainable seawater electrolysis for renewable hydrogen production. The report summarises the main scientific, technical, communication and dissemination activities carried out during the reporting period. It presents key developments across the project, including progress towards the project’s research and technological objectives, collaboration among consortium partners, stakeholder engagement and activities aimed at increasing the visibility and awareness of seawater electrolysis. The report also provides an overview of ASTERISK’s communication and dissemination efforts during 2025, documenting activities and outputs designed to make the project’s work accessible to scientific, industrial, policy and wider public audiences. The Annual Report offers stakeholders and the wider research community a concise overview of the project’s progress during 2025 and its contribution to the broader objectives of advancing renewable hydrogen technologies in Europe. ASTERISK is co-funded by the Clean Hydrogen Partnership and the European Union.

copy it from the README.

У статті досліджено інституційну архітектуру впровадження та контролю за використанням систем штучного інтелекту у досудовому розслідуванні Великої Британії. Виявлено, що сучасна британська модель поє…

У статті досліджено інституційну архітектуру впровадження та контролю за використанням систем штучного інтелекту у досудовому розслідуванні Великої Британії. Виявлено, що сучасна британська модель поєднує централізовану національну координацію розвитку, тестування та впровадження систем штучного інтелекту в поліції з функціонально спеціалізованими механізмами технічного, етичного, правозахисного, регуляторного та демократичного контролю. Проаналізовано функціональне значення Home Office, DSIT, Government Digital Service, Police Digital Service, National Police Chiefs’ Council, Police.AI, STEAC, локальних етичних комітетів, а також спеціалізованих регуляторних і наглядових інституцій. Особливу увагу приділено превентивній оцінці технологічних ризиків, алгоритмічній прозорості, технічному тестуванню, оцінюванню упередженості та механізмам незалежної етичної експертизи. На прикладі National Data Analytics Solution та практики застосування технології автоматизованого розпізнавання облич досліджено значення превентивного контролю до початку або розширення операційного використання систем ШІ. Встановлено, значення британського досвіду полягає у формуванні функціональних вимог до вітчизняної правової системи: незалежності технічної оцінки від суб’єкта використання технології, окремого етичного та правозахисного контролю, належного документування застосування систем штучного інтелекту та нормативного визначення процесуальних наслідків використання їх результатів. Запропоновано враховувати під час формування української інституційної моделі контролю за використанням систем штучного інтелекту у правоохоронній діяльності принцип поєднання централізованої технічної та організаційної координації з інституційно відокремленими механізмами етичного, правозахисного та правового контролю, а також із чітким визначенням відповідальності за використання технології та процесуальні наслідки її результатів.

Monorepo for the PID Component and related packages

Maqolada oliy ta’lim talabalarida mustaqil ishlash ko‘nikmalarini rivojlantirishning nazariy asoslari va zamonaviy pedagogik yondashuvlari tahlil qilinadi. Mustaqil ta’lim o‘zi…

Maqolada oliy ta’lim talabalarida mustaqil ishlash ko‘nikmalarini rivojlantirishning nazariy asoslari va zamonaviy pedagogik yondashuvlari tahlil qilinadi. Mustaqil ta’lim o‘zini boshqarib o‘qish, maqsad qo‘yish, vaqtni rejalashtirish, axborotni mustaqil izlash va tanqidiy baholash, topshiriq bajarilishini monitoring qilish, feedbackdan foydalanish hamda refleksiya jarayonlarining integratsiyasi sifatida talqin etiladi. Self-regulated learning, flipped learning, loyiha va muammo asosida ta’lim, raqamli ta’lim platformalari, learning analytics, sun’iy intellekt vositalari, e-portfolio va peer learning imkoniyatlari qiyosiy yoritiladi. Xalqaro tadqiqotlar tahlili mustaqil ishlash samaradorligi faqat topshiriq hajmini ko‘paytirish bilan emas, balki talabaning agentligi, metakognitiv nazorati, aniq mezonli feedback, bosqichma-bosqich pedagogik tayanch va raqamli muhitning to‘g‘ri tashkil etilishi bilan bog‘liqligini ko‘rsatadi. Yakunda oliy ta’lim amaliyoti uchun integrativ metodik model va amaliy tavsiyalar taklif qilinadi.