2026-07-26·Daily digest

🔬 Research Pulse

Daily Digest

July 26, 2026

🤖 AI

🧠 LLMs

1. SceneActBench: Can Agents Act on the 3D Scenes They See?

Authors: Yifei Zhao, Xiangxin Zhou, Wenhao Yang... Published: 2026-07-24 | Citations: 0 arXiv | PDF

Research Question: How well can VLM agents actually act on complete multi-object 3D scenes through tool use, rather than merely describe them or manipulate single objects in isolation?

Summary: SceneActBench introduces a unified benchmark for evaluating VLM agents on visually-conditioned action across five 3D scene tasks (520 cases from 210 source instances) using a fixed agent-environment loop and hidden geometric ground truth. Testing 11 proprietary VLM configurations shows overall scores of just 38.6-50.2 with no model dominating across tasks, revealing that current frontier VLMs are far from reliable 3D scene actors.

Key Results: Built SceneActBench with 5 tasks from 210 source instances yielding 520 task cases with paired input conditions (PNG images or sampled video frames plus optional 3D assets). Evaluated 11 proprietary VLM configurations under a unified agent-environment loop with task-specific geometric metrics against hidden ground truth. Overall scores ranged only 38.6-50.2, and no model performed consistently well across all five tasks.

Key Findings:

Frontier proprietary VLMs cluster in a narrow 38.6-50.2 overall score band, indicating no clear leader on 3D scene action
No single configuration wins across all five tasks, suggesting task-specific rather than general 3D-action capability
Multi-object scene action is a distinct and harder regime than the single-object or text-response 3D benchmarks that dominate prior work

Technical Novelty: First benchmark to unify five visually-conditioned 3D action tasks under a single fixed agent-environment loop with hidden geometric ground truth, shifting evaluation from textual description or single-object ops to end-to-end multi-object scene manipulation.

What's New: Moves 3D VLM evaluation from describing scenes or manipulating one object to acting on full multi-object scenes through tools, with a fair fixed-loop protocol and geometric (not textual) scoring across five tasks.

Extension Opportunities:

Add open-source/open-weight VLMs (Qwen-VL, LLaVA, InternVL) to the 11 proprietary configurations to characterize the open vs closed gap on 3D action
Extend the fixed agent loop with iterative self-correction or verification tools (e.g., render-and-critique) to test whether failures are perception vs planning
Generate synthetic training data from the benchmark's geometric metrics as a reward signal for RL fine-tuning of a 3D-acting agent

Replicability: Abstract does not mention released code, data, or asset licenses. Reproduction primarily requires API access to the 11 proprietary VLMs (likely GPT-4o, Claude, Gemini tiers) plus a 3D rendering/simulation stack; compute is API-bound rather than GPU-bound, but total cost scales with 520 cases x 11 configs x agent-loop turns.

Research Gaps:

Failure-mode analysis is promised but no taxonomy of where perception, planning, or tool-invocation breaks is quantified in the abstract
No open-weight models evaluated, leaving the accessible-model frontier uncharacterized

📊 Other AI

1. Learning to Prepare Molecular Ground States with Transformer Models

Authors: Alex Koziell-Pipe, Jasmine Brewer, Jem Guhit... Published: 2026-07-24 | Citations: 0 arXiv | PDF

Research Question: How can we efficiently generate shallow ground-state preparation circuits for large molecules where iterative methods like ADAPT-VQE become computationally prohibitive?

Summary: ADAPT-GQE is a generative transformer framework that learns from ADAPT-VQE reference circuits and is then RL-fine-tuned to synthesize ground-state preparation circuits for electronic structure problems. It cuts circuit-generation time by an order of magnitude versus ADAPT-VQE and produced circuits successfully executed on Quantinuum Helios-1 for imipramine.

Key Results: ADAPT-GQE achieves order-of-magnitude reductions in circuit generation time versus ADAPT-VQE while maintaining comparable or improved state-preparation accuracy. Demonstrated on imipramine (tricyclic antidepressant, a challenging pharmaceutical target) with generated circuits executed on Quantinuum Helios-1 hardware — reportedly the first AI-generated quantum chemistry circuits run on state-of-the-art quantum hardware.

Key Findings:

Transformer models can learn to propose and score high-quality ansatz circuits from ADAPT-VQE demonstrations
RL fine-tuning enables the generator to exceed the accuracy of its supervised training data
The approach scales to pharmaceutically relevant molecules (imipramine) and runs on current trapped-ion hardware

Technical Novelty: A two-stage pipeline: (1) supervised training of a transformer on ADAPT-VQE reference circuits, then (2) RL fine-tuning that lets the generator surpass the accuracy of its supervised training targets. This is distinct from prior GQE work that stopped at imitation of classical solvers.

What's New: First demonstration of an AI-generated quantum chemistry circuit executed on state-of-the-art quantum hardware, plus the use of RL to push beyond the ceiling set by the classical training oracle.

Extension Opportunities:

Extend the transformer to broader molecule classes (e.g., transition-metal catalysts, proteins) by expanding training corpora beyond drug-like molecules
Replace ADAPT-VQE reference generation with cheaper classical surrogates (e.g., CCSD or DMRG-derived targets) to scale training data
Combine ADAPT-GQE with error-mitigation-aware RL rewards so generated circuits are optimized for specific hardware noise profiles (e.g., trapped-ion vs superconducting)

Replicability: Abstract does not mention open-sourced code, weights, or datasets. Reproduction would require an ADAPT-VQE implementation, GPU compute for transformer training/RL, and access to Quantinuum Helios-1 (or a comparable trapped-ion system) for hardware validation — non-trivial but feasible for a well-resourced quantum-ML lab.

Research Gaps:

Scaling to strongly correlated systems and larger active spaces where ADAPT-VQE itself struggles to provide training data
Hardware-aware generation that accounts for connectivity, gate fidelities, and noise across different quantum architectures

2. Hyperball May Not Be a Free Lunch

Authors: Yihao Xiao, Jialong Sun, Zitian Gao... Published: 2026-07-24 | Citations: 0 arXiv | PDF

Research Question: Why do Hyperball-style optimizers (e.g., MuonH) outperform norm-decay variants (MuonWD) in scale-invariant deep networks, and is the advantage due to update direction or effective step size?

Summary: The paper investigates why Hyperball-style optimizers like MuonH outperform weight-decay variants like MuonWD in scale-invariant networks. Through a new angular effective learning rate framework and update decomposition, the authors show the advantage comes from effective step-size evolution rather than an intrinsically better update direction, meaning learning-rate scheduling still matters critically.

Key Results: The authors derive an angular effective learning rate incorporating parameter-update angle, parameter norm, and update norm, showing the conventional norm-based measure is a special case under parameter-update orthogonality. Numerical experiments decompose updates into radial/tangential components, showing radial updates have only a limited direct effect on angular displacement. A heuristic learning-rate-schedule swap reproduces each optimizer's dynamics using the other, demonstrating the difference stems from effective step-size evolution rather than update direction. Pretraining experiments confirm aggressive LR decay accelerates MuonH early but impairs later performance.

Key Findings:

The radial component of optimizer updates has only limited direct effect on the angular effective learning rate, so it does not explain MuonH's late-training advantage over MuonWD
A schedule-swap experiment reproduces each optimizer's dynamics using the other, indicating the primary difference is effective step-size evolution, not update direction
Aggressive LR decay accelerates MuonH early in training but impairs later performance — constant angular velocity does not remove the scheduling problem

Technical Novelty: Introduction of an angular effective learning rate that generalizes prior norm-based measures, plus a radial/tangential decomposition of optimizer updates, combined with a schedule-swap heuristic that isolates 'update direction' vs 'effective step size' as competing explanations for Hyperball's advantage.

What's New: First work to formally decompose Hyperball optimizer advantages into radial vs tangential components and to empirically isolate 'update direction' from 'effective step size' via schedule swapping, challenging the assumption that norm-fixing alone explains Hyperball's success.

Extension Opportunities:

Design adaptive learning-rate schedules specifically tailored to MuonH's angular velocity dynamics to capture early-training speedups without the late-training degradation
Extend the angular effective learning rate framework to other scale-invariant optimizers (Lion, Shampoo variants) and analyze whether the same effective step-size mechanism explains their behavior
Investigate the interaction between Hyperball constraints and layer-wise / parameter-group-specific scheduling in large-scale pretraining runs (>1B params)

Replicability: Code is publicly available at github.com/mangocrazz/hyperball-may-not-be-a-free-lunch. Pretraining experiments implied — likely requires multi-GPU compute (single-node A100/H100 class) for LLM-scale reproduction; the analytical/heuristic experiments are lightweight.

Research Gaps:

No principled LR schedule design that captures MuonH's early-training gains without sacrificing late-training performance
Analysis is limited to specific training configurations; generalization across model scales, architectures, and data regimes remains open

🦾 ROBOTICS

1. A scalable and resource-efficient pipelined p-computer for probabilistic Ising machines

Authors: Deborah Volpe, Eleonora Raimondo, Andrea Grimaldi... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can digital probabilistic Ising machines (PIMs) achieve high-throughput execution on dense, fully-connected combinatorial optimization problems, where memory bandwidth and data movement — not sparse interaction exploitation — become the dominant bottleneck?

Summary: The paper presents a deeply pipelined FPGA architecture for probabilistic Ising machines that overlaps spin evaluation with local-field updates and uses a bandwidth-aware on-chip memory layout, enabling scalable, high-throughput execution on fully-connected (dense) problems. It supports up to 2048 p-bits at 300 MHz and delivers an order-of-magnitude update-rate improvement over a non-pipelined baseline, validated on portfolio optimization and LDPC decoding.

Key Results: The authors built an FPGA architecture with a >20-stage pipelined p-bit update path plus bandwidth-aware on-chip memory for coupling/bias matrices. It supports 512 p-bits at 16-bit fixed-point, 1024 p-bits at 10-bit, and 2048 p-bits at 2-bit coefficients, running up to 300 MHz. At fixed parallelization it delivers an order-of-magnitude higher update rate than an optimized non-pipelined baseline. Validation on portfolio optimization and LDPC decoding shows close agreement with software references and substantial time-to-solution reductions.

Key Findings:

A >20-stage pipeline overlapping spin evaluation and local-field updates yields ~10× higher update rate than an optimized non-pipelined FPGA baseline at equal parallelization
Bandwidth-aware coupling/bias memory organization allows scaling to 512 p-bits (16-bit), 1024 p-bits (10-bit), and 2048 p-bits (2-bit) at up to 300 MHz
On portfolio optimization and LDPC decoding, the pipelined PIM matches software reference results while substantially reducing time-to-solution

Technical Novelty: Prior digital PIMs exploited sparse interactions to hit high throughput, which fails for dense problems. This work introduces a deeply pipelined (>20 stages) update path that overlaps spin evaluation with local-field updates, paired with a bandwidth-aware memory layout for the J-matrix — enabling fully-connected PIM operation without the usual memory-bandwidth stall.

What's New: Unlike prior digital PIMs that scale by exploiting problem sparsity, this design targets dense/fully-connected workloads by treating memory bandwidth as the primary constraint and using deep pipelining plus a co-designed memory layout to sustain throughput.

Extension Opportunities:

Port the pipelined architecture to newer FPGA families (e.g., Versal AI Edge or HBM-enabled Alveo) to push p-bit count beyond 2048 and evaluate scaling on MaxCut/QUBO benchmarks like G-set
Integrate adaptive/annealed pseudo-temperature schedules into the pipeline stages to improve solution quality on hard instances without sacrificing throughput
Extend validation to additional dense workloads — protein folding, traffic assignment, or graph coloring — and benchmark against GPU-based Monte Carlo and D-Wave/CMOS annealers on wall-clock time-to-target-energy

Replicability: The abstract does not mention a public code or HDL release. Reproduction requires a mid-to-high-end FPGA (likely Xilinx UltraScale+ or similar) capable of hosting 512–2048 p-bit datapaths at 300 MHz, plus toolchain (Vivado), and the portfolio optimization / LDPC benchmark datasets. Full replication would need the RTL sources, which appear not to be disclosed in the abstract.

Research Gaps:

Bit-width vs. solution-quality trade-offs are only implicitly explored (2-bit at 2048 p-bits) — systematic accuracy characterization on harder benchmarks is missing
No comparison against GPU, ASIC, or analog/CMOS Ising-machine implementations on standard combinatorial benchmarks (e.g., G-set MaxCut)

2. End-to-End Learning of Safe Optimal Feedback Control in High Dimensions with Control Barrier Function Layers

Authors: Xingjian Li, Kelvin Kan, Deepanshu Verma... Published: 2026-07-22 | Citations: 0 arXiv | PDF

Research Question: How can control barrier function (CBF) safety filters be embedded as differentiable optimization layers in end-to-end feedback controller training so that the approach scales beyond the ~16-state ceiling of prior QP-in-the-loop methods?

Summary: The paper makes end-to-end training of safe feedback controllers scalable to high-dimensional systems by embedding a CBF-based safety filter as a differentiable layer, using operator splitting to solve the safety QP efficiently and Jacobian-Free Backpropagation to differentiate through it without forming Jacobians. It provides nonsmooth-analysis theoretical justification and empirically pushes the tractable regime from ~16 state dimensions to 1200 states / 400 controls on multi-agent nonlinear problems.

Key Results: The authors combine operator splitting with Jacobian-Free Backpropagation (JFB) to differentiate through the CBF-QP safety layer, provide a nonsmooth-analysis justification of the training procedure, and demonstrate scaling to multi-agent nonlinear control problems with state dimensions up to 1200 and control dimensions up to 400 — roughly a 75x jump over the ~16-state prior limit — while preserving hard CBF safety guarantees.

Key Findings:

Operator splitting + JFB removes the differentiation bottleneck that previously capped CBF-in-the-loop policy learning at ~16 state dimensions.
Hard safety guarantees from the CBF filter are preserved during learned-policy execution, not just approximated by a penalty term.
The method trains semi-global feedback controllers on multi-agent nonlinear systems at 1200 states and 400 controls, a scale previously considered infeasible for QP-layer safe learning.

Technical Novelty: Prior CBF-as-layer work (e.g., OptNet-style differentiation of the QP) requires implicit differentiation through KKT conditions, which is O(n^3) and memory-heavy. This paper's novelty is (1) using operator splitting to decompose the safety QP into cheaper subproblems and (2) applying Jacobian-Free Backpropagation so gradients are computed without forming or inverting the Jacobian of the fixed-point map, plus a nonsmooth-analysis proof that the resulting gradients are valid descent directions despite the CBF layer's non-differentiability at active-set boundaries.

What's New: First demonstration that CBF safety layers can be embedded end-to-end at high dimensions by replacing implicit KKT differentiation with Jacobian-free fixed-point gradients over a split QP, with matching nonsmooth-analysis guarantees.

Extension Opportunities:

Swap the hand-designed CBF for a learned neural CBF trained jointly with the policy, using the same JFB + splitting stack to keep gradients tractable at high dimensions.
Port the JFB-differentiated CBF layer into a model-based RL loop (e.g., PPO or SAC with a safety-filtered action head) for multi-robot swarms or high-DOF manipulators where existing safe-RL fails to scale.
Extend the splitting scheme to handle stochastic dynamics or partial observability by combining it with belief-state controllers or diffusion-based policies under CBF constraints.

Replicability: The abstract does not mention a public code release. Reproducing the 1200-state multi-agent experiments would likely need a modern multi-GPU setup (given end-to-end training over long horizons with a QP layer in the loop), but individual smaller-scale runs should fit on a single high-memory GPU. Full replication depends on the authors releasing the CBF specifications, dynamics models, and training scripts.

Research Gaps:

Assumes hand-crafted, valid CBFs are available — learning or synthesizing CBFs for arbitrary high-dimensional systems remains open.
Evaluation is on deterministic nonlinear multi-agent benchmarks; behavior under model mismatch, disturbances, or partial observability is not characterized.

3. Safety and Security: Experimental Validation of Encrypted Model Predictive Control

Authors: Juraj Holaza, Martin Kalúz, Matúš Furka... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can Model Predictive Control (MPC) be executed on untrusted third-party/cloud platforms while preserving closed-loop stability, recursive feasibility, and full privacy of both process data and controller coefficients?

Summary: The paper proposes evaluating an explicit MPC control law — approximated as a polynomial — inside a fully homomorphic encryption scheme so that a cloud-hosted controller sees neither the plant state nor its own coefficients in cleartext. This is the first encrypted-MPC formulation the authors claim preserves stability and recursive feasibility, and they validate it on a real laboratory-scale plant.

Key Results: The authors demonstrate a polynomial approximation of the explicit MPC control law that can be evaluated inside a fully homomorphic encryption (FHE) scheme, and validate it experimentally on a laboratory-scale process. The abstract reports qualitative experimental validation rather than concrete latency/accuracy numbers — specific benchmarks (approximation error, encrypted-evaluation time, sampling period) are not disclosed in the abstract.

Key Findings:

Iterative QP-based encrypted MPC cannot guarantee closed-loop stability or recursive feasibility under FHE constraints
A polynomial approximation of the explicit MPC law is FHE-compatible (add/multiply only) and retains the control-theoretic guarantees of explicit MPC
The scheme was successfully deployed and validated on a laboratory-scale plant with both data and controller parameters encrypted

Technical Novelty: Prior encrypted-MPC work either encrypts only part of the pipeline or loses recursive feasibility/stability guarantees because iterative QP solvers do not compose well with FHE. This paper sidesteps the iterative solver entirely by pre-computing an explicit control law, approximating it with a polynomial (FHE-friendly: only additions and multiplications), and evaluating the polynomial homomorphically — protecting both state measurements and controller coefficients simultaneously.

What's New: Combines explicit MPC (offline QP solved parametrically) with polynomial surrogate modeling to make the control law natively evaluable under FHE — a departure from prior work that tried to encrypt the online optimizer itself and either leaked information or lost stability guarantees.

Extension Opportunities:

Benchmark the polynomial-approximation FHE controller against CKKS-based direct QP solvers on standard process-control testbeds (CSTR, quadruple-tank) to quantify the stability/latency trade-off
Extend the polynomial approximation to nonlinear or robust/tube MPC formulations, where explicit solutions are typically intractable, using piecewise polynomial fits per critical region
Integrate the encrypted controller with a secure sensor-attestation layer (e.g., TEE-signed measurements) to close the end-to-end trust gap between plant, cloud, and actuator

Replicability: No code or dataset link is mentioned in the abstract. Reproduction would require an FHE library (e.g., Microsoft SEAL, OpenFHE with CKKS), an explicit MPC toolbox (MPT3), a polynomial-fitting routine, and the described laboratory rig. Compute is modest on the client side but FHE evaluation typically demands a multi-core CPU; no GPU required.

Research Gaps:

Abstract omits quantitative results: approximation error bounds, encrypted evaluation latency, and sampling-rate feasibility
Scalability to high-dimensional state spaces where explicit MPC region counts explode is not addressed

💻 COMPUTE

1. Examining QRMI as a Unified Interface for Quantum-HPC Integration

Authors: Thomas Badts, Tim Boyle, Claudio Carvalho... Published: 2026-07-21 | Citations: 0 arXiv | PDF

Research Question: How can quantum computing resources be integrated as first-class schedulable resources across the diverse landscape of HPC workload managers (beyond just Slurm) using a single, vendor-agnostic interface?

Summary: The paper extends QRMI — a vendor-agnostic middleware exposing quantum devices as first-class schedulable resources — from Slurm-only to five additional workload managers (PBS, LSF, Grid Engine, Kubernetes, Flux). It shows the abstraction survives across batch, cloud-native, and graph-based scheduling paradigms with minimal per-scheduler modification, positioning QRMI as the plausible standard interface for quantum-HPC integration.

Key Results: The paper demonstrates that QRMI — a thin middleware layer previously validated only with Slurm — successfully integrates with five additional workload managers spanning three architectural categories: traditional batch schedulers (PBS, LSF, Grid Engine), cloud-native orchestration (Kubernetes), and graph-based scheduling (Flux Framework). No quantitative benchmarks (latency, throughput, overhead percentages) are cited in the abstract; the 'proof' is qualitative demonstration of portability with minimal scheduler-specific modifications across on-premises and cloud environments.

Key Findings:

A single thin middleware layer can bridge quantum backends to structurally distinct workload managers without per-scheduler forks
Integration patterns cluster differently by scheduler category (batch vs. container-orchestration vs. DAG), but QRMI's API surface accommodates all three
Both on-premises HPC and cloud-native environments can consume quantum resources through the same programmatic contract

Technical Novelty: Prior QRMI work only proved integration with Slurm. The novelty here is showing the same abstraction generalizes across categorically different scheduler paradigms — batch (PBS/LSF/SGE), container-orchestration (Kubernetes), and DAG-based (Flux) — validating QRMI as a genuine abstraction rather than a Slurm-specific adapter.

What's New: First demonstration that a quantum-HPC resource-management abstraction generalizes across the full taxonomy of contemporary schedulers, moving QRMI from 'a Slurm plugin' toward 'a candidate standard.'

Extension Opportunities:

Build a QRMI plugin for an emerging scheduler not covered (e.g., HTCondor, Nomad, or Ray) to further stress-test the abstraction's portability claim
Add quantitative benchmarking harness measuring scheduling overhead, queue wait time, and job dispatch latency across the five integrations to identify which scheduler pattern is most efficient for quantum workloads
Extend QRMI with cost-aware and error-budget-aware scheduling policies that route circuits to the cheapest or highest-fidelity backend, exposing quantum-specific metrics (T1/T2, gate fidelity, queue depth) as scheduler-visible resources

Replicability: The abstract does not mention a code release, dataset, or reproducibility artifact. QRMI itself is IBM-affiliated open middleware (likely on GitHub), but reproducing the five integrations would require access to installations of PBS, LSF, Grid Engine, a Kubernetes cluster, and Flux — plus quantum backend credentials (IBM Quantum or simulator). Compute is modest for the classical side; quantum runtime depends on backend access tier.

Research Gaps:

No quantitative performance evaluation (scheduling latency, throughput, overhead) is presented in the abstract — the claim of portability is architectural, not empirical
Quantum-specific scheduling concerns (calibration windows, error rates, hybrid classical-quantum feedback loops) are not addressed as first-class scheduling inputs

2. SynPre-FL: Synthetic data-driven pretraining integrated Federated Learning training framework

Authors: Akarsh K Nair, Muhammad Arifur Rahman, Nicholas Shopland... Published: 2026-07-21 | Citations: 0 arXiv | PDF

Research Question: How can synthetic EHR data generation be systematically integrated with federated learning to overcome data scarcity, non-IID client heterogeneity, class imbalance, and lack of realistic tabular EHR benchmarks in privacy-preserving clinical risk prediction?

Summary: SynPre-FL is a unified framework that uses a latent autoencoder-diffusion model to generate privacy-preserving synthetic EHR cohorts, which warm-start federated training before heterogeneity-aware optimization (class-balanced losses, proximal regularization, adaptive aggregation) is applied. It couples post-hoc calibration and federated-safe SHAP for interpretable clinical risk prediction, showing consistent gains over baselines under non-IID conditions.

Key Results: SynPre-FL demonstrated that a latent autoencoder-diffusion generator preserves univariate, bivariate, and multivariate statistical structure while resisting membership-inference and reconstruction attacks. Downstream utility was validated via TSTR (Train Synthetic Test Real), TRTS (Train Real Test Synthetic), and model-based evaluations. Across federated settings with 5, 10, and 15 heterogeneous clients, it outperformed baselines on robustness and scalability, especially under severe non-IID fragmentation. Calibration improved probability reliability and SHAP produced stable feature attributions across federation sizes.

Key Findings:

Synthetic warm-starting materially improves FL robustness under severe non-IID fragmentation across 5/10/15-client setups
The latent AE-diffusion generator preserves univariate/bivariate/multivariate structure and resists membership-inference and reconstruction attacks
Post-hoc calibration plus federated-safe SHAP yields stable, clinically coherent feature attributions across federation sizes

Technical Novelty: The specific combination of (1) a latent autoencoder-diffusion model for tabular EHR synthesis used as an FL warm-start, (2) heterogeneity-aware optimization stacking class-balanced local objectives + proximal regularization + adaptive server aggregation, and (3) federated-safe SHAP explainability with post-hoc calibration — packaged as a single pipeline. Prior work treated synthetic generation and FL optimization largely in isolation.

What's New: First systematic integration of high-fidelity latent-diffusion tabular EHR synthesis with heterogeneity-aware federated optimization as a single reproducible pipeline, extended with federated-safe explainability — rather than treating synthesis and FL as separate concerns.

Extension Opportunities:

Extend the latent autoencoder-diffusion generator to multi-modal EHR (combining tabular with clinical notes or imaging) and evaluate whether cross-modal synthetic pretraining further stabilizes non-IID FL
Replace the static warm-start with a continual synthetic-data refresh loop where the server periodically regenerates cohorts conditioned on drift signals from client updates
Benchmark SynPre-FL against differentially-private FL baselines with formal (ε, δ) guarantees, quantifying the privacy/utility tradeoff vs. the empirical MIA/reconstruction resistance reported here

Replicability: Abstract does not mention public code or dataset release. Reproduction would require tabular EHR data (likely MIMIC-derived), moderate GPU compute for training the latent diffusion generator, and a federated simulation harness (Flower/FedML) to run 5/10/15-client experiments — feasible on a single multi-GPU workstation.

Research Gaps:

No formal differential-privacy guarantees are claimed; privacy is evaluated only empirically via MIA/reconstruction resistance
Evaluation appears simulation-based on tabular EHR benchmarks; real multi-institution deployment and generalization to multi-modal EHR remain untested

3. Odin: Primitive-Level Synchronization for Distributed Point-Based Neural Rendering

Authors: Zhenxiang Ma, Zeyu He, Yuanzhen Zhou... Published: 2026-07-22 | Citations: 0 arXiv | PDF

Research Question: How can distributed point-based neural rendering (PBNR) training scale efficiently when global task-/iteration-level barriers push synchronization — not rendering compute — onto the critical path?

Summary: Odin is a distributed training system for point-based neural rendering that replaces coarse global barriers with fine-grained primitive-level synchronization. By combining an ahead-of-time scheduler that exploits stable primitive locality with a runtime that validates publication of mutable state, it hides most synchronization latency while preserving reconstruction quality — yielding up to 1.89x throughput over Grendel on 64 GPUs.

Key Results: Across 4 existing PBNR pipelines and 13 non-city scenes on 8 GPUs, Odin delivers 1.22x average throughput improvement and hides 82% of critical-path wait time while preserving reconstruction quality. In a MatrixCity mixed-parallel case study scaling to 64 GPUs, Odin achieves up to 1.89x throughput over Grendel without altering renderer kernels, optimizers, training budgets, or model capacity.

Key Findings:

Synchronization, not rendering compute, is the dominant critical-path bottleneck in modern distributed PBNR as renderers get faster
Primitive-indexed dependencies exhibit stable locality and phase order that a static scheduler can exploit to identify safe overlap windows
A hybrid quality/throughput policy — synchronizing structural changes but permitting bounded stale reads for low-gradient primitives — preserves quality while unlocking 1.22x-1.89x speedups

Technical Novelty: Replaces global barriers with primitive-level synchronization driven by an ahead-of-time scheduler that exploits stable spatial locality and phase order to find low-conflict overlap windows, plus a runtime that validates primitive publication before dependent reads. Introduces a dual-path design: a quality-first path preserving synchronized-training semantics and a throughput-first path that uses gradient evidence to admit only small, low-impact stale reads while keeping structural updates synchronous.

What's New: Prior distributed PBNR systems (e.g., Grendel) rely on iteration-level global barriers borrowed from layered DNN training. Odin is the first to treat PBNR's sparse, view-dependent primitive dependencies as first-class scheduling entities, enabling barrier-free training without kernel or optimizer changes.

Extension Opportunities:

Extend primitive-level synchronization to dynamic/4D Gaussian Splatting where scene primitives evolve over time, requiring temporal conflict detection in the scheduler
Apply the ahead-of-time locality-based scheduler to other sparse-update workloads like NeRF hash grids, neural SDFs, or MoE expert routing where dependencies are also view/input-dependent
Build an adaptive policy that dynamically switches between quality-first and throughput-first paths per-primitive based on live gradient magnitude, rather than a global mode selection

Replicability: No code/data availability mentioned in the abstract. Reproduction requires 8-64 GPUs, MatrixCity dataset plus 13 non-city PBNR scenes, and the ability to integrate with 4 existing PBNR pipelines (likely including 3DGS variants and Grendel as baseline).

Research Gaps:

No treatment of dynamic scenes or streaming primitive insertion/deletion beyond structural-change synchronization
The gradient-evidence threshold for admitting stale reads appears to be a heuristic — a principled convergence guarantee is not claimed in the abstract

⚡ ENERGY

1. Geometric Superconducting Diode Effect in an NbN Nanoring

Authors: Tianyu Li, Peiyuan Huang, Jiong Li... Published: 2026-07-22 | Citations: 0 arXiv | PDF

Research Question: Can a superconducting diode effect (nonreciprocal critical current) be realized in a structurally minimal, single-material device without relying on Josephson junctions, heterostructures, ferromagnets, or gating — using geometry alone to break inversion symmetry?

Summary: The authors demonstrate a superconducting diode effect in a single-material NbN nanoring whose asymmetric geometry alone breaks inversion symmetry, avoiding the Josephson junctions, heterostructures, or ferromagnets used in prior SDE platforms. Magnetic-field- and temperature-dependent transport reveals polarity-switchable nonreciprocity, with diode efficiency following the energy gap and nonreciprocity magnitude following the superfluid density — establishing asymmetric nanorings as a minimal geometric platform for nonreciprocal superconducting transport.

Key Results: Demonstrated a polarity-switchable superconducting diode effect in an asymmetric NbN nanoring where geometry is the sole source of inversion-symmetry breaking. Low magnetic fields asymmetrically redistribute critical current between opposite bias directions without significantly reducing total current-carrying capacity. Temperature-dependent measurements show maximal diode efficiency tracks the superconducting energy gap Δ(T), while maximal nonreciprocity tracks the superfluid density ns(T) — two distinct scaling laws revealing different microscopic origins.

Key Findings:

Geometric asymmetry alone in a single-material NbN nanoring produces a pronounced, polarity-switchable superconducting diode effect
Low magnetic fields redistribute critical current asymmetrically between bias directions without materially degrading the overall superconducting current capacity
Maximal diode efficiency scales with the superconducting energy gap Δ(T), while maximal nonreciprocity scales with the superfluid density ns(T), indicating distinct microscopic drivers

Technical Novelty: Prior superconducting diodes required Josephson junctions, multilayer/heterostructure stacks, ferromagnetic proximity, or gate-defined asymmetries. This work removes all of those: a single-material NbN nanoring with purely geometric asymmetry produces the SDE, and the paper cleanly separates the temperature dependence of diode efficiency (gap-like) from nonreciprocity magnitude (superfluid-density-like), which prior platforms had not disentangled.

What's New: First demonstration (as claimed) of a purely geometry-induced SDE in a structurally minimal single-material device, plus the disentanglement of two different temperature scalings (gap vs. superfluid density) for efficiency and nonreciprocity magnitude.

Extension Opportunities:

Systematically vary ring geometry (asymmetry ratio, arm width, ring diameter) to derive a design-rule map linking geometric parameters to diode efficiency and identify the optimal asymmetry
Fabricate arrays of such nanorings in series/parallel to build rectifier, logic, or memory elements for practical cryogenic/SFQ circuits and benchmark against Josephson-junction diodes
Extend the single-material geometric approach to other superconductors (Nb, MoGe, Al, high-Tc films, topological superconductors) to test universality and probe how pairing symmetry modifies the geometric SDE

Replicability: No code or data release is mentioned in the abstract. Reproduction requires a thin-film NbN deposition capability (sputtering), e-beam lithography for the asymmetric nanoring, and a dilution/He-4 cryostat with low-noise transport electronics and a small magnetic field source — standard for a mesoscopic superconductivity lab but not trivially replicable without cleanroom access.

Research Gaps:

No microscopic theory presented that quantitatively predicts the observed diode efficiency from ring geometry, current distribution, and vortex/Meissner screening
Scaling behavior, reproducibility across many devices, and the geometric parameter space (asymmetry ratio, size, film thickness) are not systematically mapped in the abstract

2. Dirac-vortex modes beyond the continuum limit

Authors: Jiayu Fan, Jiusi Yu, Aoning Luo... Published: 2026-07-22 | Citations: 0 arXiv | PDF

Research Question: Does the continuum Jackiw-Rossi description of Dirac-vortex modes (DVMs) in Kekule-modulated lattices remain valid when the complex mass texture violates the slowly varying envelope approximation, and can the initial phase—treated as a gauge redundancy in the continuum—become a physically observable, tunable parameter in discrete lattices?

Summary: The paper shows that the continuum Jackiw-Rossi description of Dirac-vortex modes fails on discrete lattices when the Kekule mass texture varies rapidly, promoting the initial phase from a gauge redundancy to a physically observable parameter that shifts the mode center. Adding a sublattice-antisymmetric perturbation converts this displacement into a sinusoidal frequency shift spanning nearly the full topological bandgap, opening a route to phase-engineered reconfigurable photonic devices.

Key Results: The authors demonstrate both in simulation and experiment that lattice discreteness promotes the initial phase from a gauge degree of freedom to a physical parameter that shifts the DVM center. By adding a sublattice-antisymmetric perturbation, they convert this center displacement into a continuous spectral shift that tunes the DVM frequency across nearly the entire topological bandgap, with a sinusoidal-like dependence on the initial phase that matches a revised continuum model incorporating mode-center motion.

Key Findings:

Lattice discreteness breaks the gauge freedom of the DVM initial phase, making it an observable that displaces the mode center
A sublattice-antisymmetric perturbation couples this displacement to frequency, enabling tuning across nearly the entire topological bandgap
A revised continuum model with a mode-center-motion term predicts a sinusoidal-like frequency-vs-phase relation that matches simulation and experiment

Technical Novelty: Prior work treated the DVM initial phase as an unobservable gauge degree of freedom via the Jackiw-Rossi continuum model. This paper identifies that discreteness breaks that gauge redundancy, derives a revised perturbative continuum model with a mode-center displacement term, and pairs it with a sublattice-antisymmetric perturbation to turn spatial displacement into a spectral tuning mechanism—an effect invisible in the ideal continuum theory.

What's New: First identification and experimental demonstration that the initial phase of a Kekule mass vortex—universally treated as unphysical in the Jackiw-Rossi continuum limit—becomes a physical, tunable parameter beyond the slowly varying envelope regime, with an accompanying revised continuum theory.

Extension Opportunities:

Build a reconfigurable photonic filter/laser cavity where the initial phase of the Kekule mass texture is electro-optically or thermally tuned to sweep the DVM frequency across the bandgap in real time
Extend the phase-sensitivity analysis to other Dirac-mass topological defects (e.g., 3D Weyl-vortex or higher-order topological bound states) to see whether analogous discrete-lattice gauge-breaking mechanisms yield new tuning knobs
Apply the framework to acoustic, mechanical, or cold-atom Kekule lattices to test universality and to explore many-body/nonlinear regimes where phase-sensitive DVMs could enable programmable interactions

Replicability: The abstract does not mention public code or data. Reproduction would require standard photonic tight-binding/FDTD simulations (feasible on a workstation) plus fabrication of a Kekule-modulated photonic lattice (likely microwave or photonic-crystal platform) with the ability to vary Kekule mass phase and add a sublattice-antisymmetric perturbation—modest experimental infrastructure for a topological-photonics lab.

Research Gaps:

Quantitative limits of the perturbative revised model (how strongly it holds for very fast-varying or large-amplitude mass textures) are not fully characterized
Extensions to nonlinear, gain/loss (non-Hermitian), or many-body settings, as well as to non-photonic platforms, remain open

3. Uni-XAS: Alignment-Driven Bidirectional Multimodal Learning for X-ray Absorption Spectroscopy

Authors: Suyang Zhong, Yuhao Zhao, Boying Huang... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can bidirectional X-ray absorption spectroscopy (XAS) modeling — forward spectrum prediction and inverse 3D structure inference — be unified into a shared representation framework that resolves permutation ambiguity among identical atoms and enables explicit 3D structure generation rather than coarse descriptors?

Summary: Uni-XAS unifies forward and inverse X-ray absorption spectroscopy modeling into a single cross-modal alignment and conditional generation framework, replacing decoupled regression pipelines. It introduces XASLip for spectrum-structure alignment, retrieval-augmented decoding for absolute spectrum prediction, and Permutation-Rectified Flow Matching to enable explicit 3D structure generation despite ligand permutation ambiguity.

Key Results: Uni-XAS was evaluated on a standardized benchmark of 328,839 structure-spectrum pairs, demonstrating strong performance across three tasks: cross-modal retrieval, absolute-spectrum prediction (avoiding physical scale collapse and energy drift), and composition-conditional 3D structure generation. The framework combines XASLip alignment with Permutation-Rectified Flow Matching without requiring heavy high-order equivariant architectures.

Key Findings:

Reframing bidirectional XAS as cross-modal alignment plus conditional generation enables shared representation learning that decoupled regression pipelines cannot achieve
Type-wise optimal transport in flow matching resolves permutation ambiguity for identical ligands without needing heavy high-order equivariant architectures
Retrieval-augmented decoding for absolute-spectrum generation avoids the physical scale collapse and energy drift that plague direct regression approaches

Technical Novelty: Three novel components: (1) XASLip — a CLIP-style alignment between spectra and structures using a physics-aware spectral encoder with absorber-aware manifold optimization; (2) anchored absolute-spectrum generation via retrieval-augmented decoding that prevents scale collapse and energy drift (versus prior regression approaches); (3) Permutation-Rectified Flow Matching integrating type-wise optimal transport into continuous generative flow to handle ligand permutation without heavy equivariant networks.

What's New: Prior work decouples forward and inverse XAS tasks into separate regressors and limits inverse modeling to coarse structural descriptors due to permutation ambiguity. Uni-XAS is the first framework to unify both directions via a shared aligned latent space and to generate explicit 3D structures using a permutation-aware flow matching approach — sidestepping the computational cost of high-order equivariant networks.

Extension Opportunities:

Extend Permutation-Rectified Flow Matching to other permutation-ambiguous inverse problems in scientific ML (e.g., NMR-to-structure, XRD-to-crystal, or molecular property inversion)
Apply the anchored retrieval-augmented decoding pattern to other spectroscopy modalities like IR, Raman, or UV-Vis where absolute scale prediction is challenging
Build a foundation model by pretraining XASLip on the 328k pairs and fine-tuning for downstream tasks like catalyst screening, battery cathode design, or in-situ operando spectroscopy interpretation

Replicability: The paper positions itself as a 'scalable, reproducible, and protocol-consistent foundation' with a standardized benchmark, suggesting code and data release is intended, though not explicitly confirmed in the abstract. Compute needs are moderate — flow matching on 328k pairs with lightweight (non-high-order-equivariant) architectures should be tractable on a small multi-GPU setup.

Research Gaps:

No standardized large-scale benchmark for bidirectional XAS modeling existed prior — this establishes 328,839 pairs as a reference
Explicit 3D structure generation from spectra remained unsolved due to permutation ambiguity, with prior methods restricted to descriptor prediction

🔬 MATERIALS

1. Supercurrent effect in a charge density wave intertwined superconductor

Authors: Zhen Zhu, Wei Cheng, Dang Liu... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How does a supercurrent (Meissner current) modify Bogoliubov quasiparticle excitations in a superconductor that is intertwined with a charge density wave (CDW) order, and can it be used to control the CDW symmetry?

Summary: The paper shows that a Meissner screening current induced by an in-plane magnetic field can Doppler-shift Bogoliubov quasiparticles selectively at precursor CDW wavevectors in NbSe2, breaking the CDW's C3v symmetry down to Cs. By rotating the field, the anisotropy direction can be tuned on demand, providing a momentum-space engineering route to control intertwined charge-superconducting orders.

Key Results: Using scanning tunneling spectroscopic (STS) imaging on superconducting NbSe2, the authors demonstrate a field-driven C3v-to-Cs symmetry breaking of the CDW modulations at the precursor CDW vectors, induced by an in-plane magnetic field. Rotating the field direction produces on-demand tuning of the anisotropic CDW modulations, and model calculations attribute the anisotropy to a selective Doppler-shift-induced reconstruction of the E-k dispersion. No specific quantitative benchmarks (temperatures, field magnitudes, gap values) are cited in the abstract.

Key Findings:

In-plane magnetic fields drive a C3v-to-Cs symmetry breaking of the CDW modulation pattern in superconducting NbSe2, visualized directly with STS.
Model calculations show the anisotropy arises from a selective Doppler shift of the Bogoliubov quasiparticle E-k dispersion at specific CDW-relevant Fermi-surface segments.
Rotating the field direction rotates the CDW anisotropy, enabling on-demand tuning and momentum-space mapping of the CDW modulations.

Technical Novelty: Prior supercurrent-Doppler-shift studies focused on pure superconducting quasiparticle spectra; this work is the first to show that the Doppler shift can act momentum-selectively on the CDW-relevant portions of the Fermi surface, breaking the CDW point-group symmetry (C3v to Cs) and providing a knob for momentum-space engineering of an intertwined order.

What's New: First demonstration that supercurrent-induced Doppler shifts do not just modify the superconducting quasiparticle spectrum but can reshape an intertwined charge-density-wave order through momentum-selective E-k reconstruction, revealing a new control knob for intertwined phases.

Extension Opportunities:

Apply the same Meissner-current/STS approach to other intertwined-order superconductors (cuprates, kagome CsV3Sb5, transition-metal dichalcogenides beyond NbSe2) to test the generality of Doppler-driven CDW symmetry breaking.
Build a device that uses a transport supercurrent (rather than a Meissner screening current) to dynamically switch CDW anisotropy, enabling a supercurrent-controlled electronic nematic switch.
Extend the model calculations into a predictive framework (e.g., coupling Ginzburg-Landau CDW order parameters with Doppler-shifted BdG spectra) to forecast which intertwined phases can be tuned momentum-selectively.

Replicability: The abstract does not mention released code or data. Reproduction would require a millikelvin, vector-magnetic-field STM capable of in-plane fields plus high-quality single-crystal 2H-NbSe2 — specialized experimental hardware rather than compute-intensive resources. The accompanying BdG/CDW model calculations are lightweight and reproducible on a workstation.

Research Gaps:

Quantitative mapping between supercurrent magnitude/direction and CDW order-parameter response is not fully established.
It remains unclear whether the mechanism generalizes to other intertwined-order superconductors (cuprates, kagome metals, twisted moirés) or is specific to NbSe2's precursor CDW.

2. Lifetime effects and satellites in the photoelectron spectrum of platinum metal

Authors: Prajna Bhatt, José Julio Gutiérrez Moreno, Laura E. Ratcliff... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How can the photoelectron spectrum of metallic platinum be comprehensively characterized to disentangle intrinsic line-shape asymmetries, satellite features, and many-body loss processes, providing a unified reference for interpreting Pt-containing materials?

Summary: The paper delivers a unified spectroscopic reference for metallic Pt by combining RHEELS, SXPS, and HAXPES with DFT and G0W0 calculations including spin-orbit coupling. It systematically assigns satellite structures and loss features across the full core-level and valence spectrum, resolving features that were previously ambiguous and demonstrating the necessity of relativistic effects for accurate Pt electronic-structure modeling.

Key Results: The authors combined RHEELS, SXPS, and HAXPES measurements across shallow and deep core levels of Pt with ab initio DFT and G0W0 calculations (with and without spin-orbit coupling). They correlated photoelectron satellites with RHEELS loss features to assign interband transitions, surface/bulk plasmons, plasmonic overtones, and semi-core ionisation losses; identified several previously unresolved satellite features and spin-orbit splittings; and showed that relativistic (spin-orbit) effects are essential for reproducing the Pt valence band structure when compared to orbital-projected DOS.

Key Findings:

RHEELS loss features map directly onto photoelectron satellites, enabling assignment to interband transitions, surface/bulk plasmons, plasmonic overtones, and semi-core losses
Several previously unresolved satellite features and spin-orbit splittings in Pt core-level spectra are identified
Spin-orbit coupling within G0W0 is critical to reproduce the experimental Pt valence band; DFT alone or non-relativistic treatments are insufficient

Technical Novelty: Direct cross-correlation of RHEELS loss features with photoelectron satellites across the full core-level manifold of Pt, combined with G0W0+SOC valence band calculations, yielding assignments of previously unresolved satellites and spin-orbit splittings that prior single-technique studies could not resolve.

What's New: First internally consistent, multi-technique benchmark of Pt photoemission spanning shallow-to-deep core levels with matched EELS-based satellite assignments and relativistic many-body (G0W0+SOC) valence-band comparison.

Extension Opportunities:

Apply the same multi-technique + G0W0-SOC workflow to other 5d transition metals (Ir, Os, Au) to build a broader reference library for heavy-element photoemission analysis
Extend the framework to Pt-based catalysts (Pt/C, Pt alloys, single-atom catalysts) to deconvolve substrate/adsorbate contributions from intrinsic Pt satellites during operando XPS
Develop an automated satellite-fitting tool that uses RHEELS loss functions as priors to decompose HAXPES spectra of Pt-containing electronic materials and quantify chemical states

Replicability: The abstract does not mention public code or data release. Reproduction requires access to synchrotron HAXPES and SXPS beamlines, a RHEELS setup, and clean Pt samples. DFT/G0W0 with SOC on Pt is tractable on a modest HPC cluster (hundreds of core-hours) using codes like BigDFT, VASP, or Quantum ESPRESSO + BerkeleyGW.

Research Gaps:

No operando or environmental-XPS extension to real Pt catalyst systems where satellite structure may shift or overlap with adsorbate features
Absence of comparable unified references for other 5d metals and Pt compounds/alloys where hybridization further complicates satellite interpretation

3. Magneto-Caloric effect and Multiple magnetic phases in Al doped Ni2MnSn0.75Al0.25 Heusler Alloys

Authors: Satya Vijay Kumar, Simran, Madhusmita Jena... Published: 2026-07-23 | Citations: 0 arXiv | PDF

Research Question: How does Al substitution at the Sn site in Ni2MnSn Heusler alloy modify its magnetic ground state, exchange interactions, and magneto-structural coupling?

Summary: The paper investigates Al-substituted Ni2MnSn0.75Al0.25 Heusler alloy and reveals a rich magnetic phase landscape featuring a high Curie temperature (734K), a martensitic transition (263K), and an interacting reentrant cluster glass state with exchange bias. Al substitution modifies Mn-Mn exchange interactions and introduces atomic disorder, producing coexisting ferromagnetic and antiferromagnetic correlations.

Key Results: Ni2MnSn0.75Al0.25 shows a second-order PM-FM transition at TC = 734K, a first-order martensitic transformation near 263K, positive Weiss temperature θCW = 746.4K, and effective magnetic moment of 6.82 μB. ZFC/FCW bifurcation and non-saturating M-H loops confirm coexisting FM/AFM interactions, forming a reentrant cluster glass state with exchange-bias effect.

Key Findings:

High Curie temperature of 734K with positive Weiss temperature (746.4K) confirms dominant ferromagnetic exchange
First-order martensitic transformation at 263K coupled to the magnetic structure indicates strong magnetostructural coupling
Reentrant cluster glass state with exchange-bias effect emerges from competing FM/AFM interactions induced by Al-driven atomic disorder

Technical Novelty: First demonstration that 25% Al substitution at the Sn site simultaneously induces magnetostructural coupling, reentrant cluster glass behavior, and exchange-bias in the Ni2MnSn system — combining three distinct magnetic phenomena in a single Heusler composition.

What's New: Unlike prior work on stoichiometric Ni2MnSn or Ni-Mn-Ga/In/Sb systems, this study shows that partial Sn→Al substitution uniquely stabilizes a cluster glass + exchange-bias regime alongside conventional martensitic and ferromagnetic transitions in the Sn-based family.

Extension Opportunities:

Systematically vary Al concentration (x = 0.1, 0.5, 0.75) to map the phase diagram and locate the critical composition where martensitic transformation and cluster glass behavior optimize
Perform neutron diffraction and DFT calculations to directly resolve the Mn-Mn exchange pathways and atomic disorder driving the reentrant cluster glass state
Measure the magneto-caloric entropy change ΔS(T,H) across TC and the martensitic transition to evaluate this alloy for room-temperature magnetic refrigeration applications

Replicability: No code/data availability mentioned. Reproduction requires arc-melting/annealing facilities for alloy synthesis, XRD, and SQUID/VSM magnetometry down to ~2K with fields up to several Tesla. Compute needs are minimal; primarily an experimental study.

Research Gaps:

No quantitative magneto-caloric ΔS values reported despite the title emphasizing magneto-caloric effect
Microscopic origin of cluster glass state (cluster size, freezing dynamics, atomic-scale disorder mapping) not directly probed

🔥 GitHub Trending

1. fxd20060117/aidanmu

⭐ 2 stars | Python

AI 桌面直播弹幕姬 — 让 AI 一边看你的屏幕一边像直播观众一样发弹幕吐槽

ai danmaku desktop-widget livestream llm pyside6

2. viking-77/GeoPollen-YOLO

⭐ 1 stars | Python

Auditable P3-only YOLOv13 framework with fertility-aware distillation for rice pollen phenotyping

computer-vision deep-learning phenotyping pollen rice yolov13

3. larp-coder/claude-api-key-generator-checker

⭐ 1 stars | Unknown

claude claude-ai claude-api claude-key llm

4. lawrence3699/claude-fable5-7b-distill

⭐ 1 stars | Python

Claude Fable5 distillation data and a reproducible 7B SFT workflow.

claude dataset distillation fable5 llm sft

5. ShreyasThakur11/cipla-cardiac-opportunity-agent

⭐ 1 stars | Python

Ranks the opportunity spaces in the India Cardiac market where Cipla has a sustainable right to win. A deterministic engine computes every figure; the model only narrates what the tools return.

ai-agent amey amey-thakur ameyarc ameythakur cardiac

6. kingtmn/ai-application-design-review

⭐ 1 stars | JSON

Audit-first AI design review skill. Find implementation blockers, define falsifiable experiments, and keep residual risks visible.

agent agent-skills ai ai-agents ai-evaluation ai-governance

7. a-novel/service-genai

⭐ 1 stars | Go

Generative AI proxy, with cost ledger and crash safety baked in.

agora-storyverse cost-tracking generative-ai go golang grpc

8. Lephim/football-intelligence-copilot

⭐ 0 stars | Jupyter Notebook

AI-assisted football tactical analysis platform — xG, xT, and pass network models built on StatsBomb open data

fastapi football-analytics machine-learning sports-analytics

9. CrawlFeeds/movies-datasets

⭐ 0 stars | Unknown

Free movies dataset — 2,647 records from TMDB, Letterboxd & Rotten Tomatoes. CSV format. Multi-language. For ML, recommendation engines & NLP research.

film-data film-database free-datasets machine-learning movies-dataset nlp

10. rcw3712/VsPredict-Geoenergy

⭐ 0 stars | MATLAB

Leakage-free, reproducible cross-well Vs prediction from conventional well logs under severe covariate shift — companion code for our Geoenergy Science and Engineering submission.

cross-validation geomechanics geophysics machine-learning matlab petrophysics

11. Zenith1803/Customer_Churn_Prediction.

⭐ 0 stars | Jupyter Notebook

End-to-end Machine Learning pipeline to predict telecom customer churn using Logistic Regression and Decision Trees, combined with K-Means clustering for customer segmentation.

churn-prediction machine-learning pandas python scikit-learn

12. brian-mwirigi/steam

⭐ 0 stars | TypeScript

Stem Forge - audio → multitrack DAW projects. Stems, drums, chords, arrangement, Ableton/Reaper/FL export. Next.js + Modal Demucs.

audio daw demucs machine-learning modal music

13. navels/neal

⭐ 0 stars | TypeScript

A source-first multi-agent coding loop: independent planner, coder, and reviewer roles — each on the provider you choose — drive scoped, reviewed, crash-safe changes to your repo.

ai-agents claude cli code-review codex developer-tools

14. siliroid/relay

⭐ 0 stars | JavaScript

A file-driven Discord bridge for one agent identity. Two files in, two files out.

agent-tools ai-agents bridge claude discord discord-bot

15. dimondevceo/ChatSpec

⭐ 0 stars | Jinja

Chat templates are untested code. ChatSpec tests them. 🏎️ Fast CI gate for Hugging Face Jinja chat templates — crashes, tool-calling, marker balance, prefix stability. Finds silent regressions with no

ai ai-safety ai-security ai-tools ci-cd cicd

Generated by Research Pulse on 2026-07-26 20:05