2026-06-01·Daily digest

🔬 Research Pulse

Daily Digest

June 01, 2026

🤖 AI

🧠 LLMs

1. Lumos-Nexus: Efficient Frequency Bridging with Homogeneous Latent Space for Video Unified Models

Authors: Jiazheng Xing, Hangjie Yuan, Lingling Cai... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can connector-based unified video models achieve high visual fidelity without the prohibitive cost of training a large high-fidelity generator end-to-end in the unified loop?

Summary: Lumos-Nexus is a training-efficient unified video generation framework that aligns only a lightweight generator with the understanding block during training, then progressively hands off generation to a high-capacity pretrained generator at inference via Unified Progressive Frequency Bridging (UPFB) in a shared latent space. This decoupling produces high-fidelity, temporally coherent videos with strong reasoning-driven control without the prohibitive cost of unified training on a large generator, and the authors introduce VR-Bench to evaluate reasoning-driven video synthesis.

Key Results: Lumos-Nexus demonstrates substantial gains in visual realism and temporal coherence on VBench and strong reasoning-based generative performance on the newly-introduced VR-Bench benchmark. Specific numerical results are not stated in the abstract, but improvements are claimed on both visual fidelity (VBench) and reasoning-driven generation (VR-Bench).

Key Findings:

Decoupling unified training (lightweight generator) from inference (high-capacity pretrained generator) preserves reasoning quality while improving fidelity.
Progressive frequency-domain handoff in a homogeneous latent space enables coarse-to-fine refinement across generators.
Existing video benchmarks underserve reasoning-driven generation, motivating the new VR-Bench evaluation.

Technical Novelty: The key novelty is Unified Progressive Frequency Bridging (UPFB), which decouples training from inference: only a lightweight generator is aligned with the understanding block during training, while inference progressively hands off to a high-capacity pretrained generator in a shared/homogeneous latent space for coarse-to-fine refinement. This avoids training the large generator inside the unified loop, unlike prior connector-based unified models.

What's New: Unlike prior connector-based unified models that either train a large generator end-to-end (expensive) or sacrifice fidelity, Lumos-Nexus introduces a two-stage design where alignment and high-fidelity synthesis happen on different generators sharing a latent space, bridged by progressive frequency handoff. The accompanying VR-Bench is also new for reasoning-driven video generation evaluation.

Extension Opportunities:

Apply the Unified Progressive Frequency Bridging (UPFB) handoff mechanism to other modalities (audio, 3D scenes) where lightweight aligned generators could pass to high-capacity pretrained models in a shared latent space.
Extend VR-Bench with multi-turn or interactive reasoning prompts to evaluate longer-horizon planning in video generation, beyond single-shot intent translation.
Investigate replacing the two-stage handoff with a learned dynamic schedule (e.g., RL-based) that decides per-frequency-band when to switch from the lightweight to the high-capacity generator.

Replicability: Code and models are stated to be available at https://jiazheng-xing.github.io/nexus-lumos-home/. Reproduction would require substantial GPU compute typical of video diffusion training (likely 8+ high-end GPUs for the lightweight alignment stage), plus access to a pretrained high-fidelity video generator at inference — exact compute is not disclosed in the abstract.

Research Gaps:

Lack of benchmarks targeting reasoning-driven (intent-to-video) generation, which VR-Bench partially addresses.
The frequency-bridging handoff schedule appears heuristic; principled or learned policies for when/how to bridge are unexplored.

2. Stateful Online Monitoring Catches Distributed Agent Attacks

Authors: Davis Brown, Samarth Bhargav, Arav Santhanam... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can safety monitors detect cyberattacks where adversaries deliberately split harmful tasks across many user accounts so that each individual agent transcript appears benign in isolation?

Summary: The paper identifies a structural blind spot in current agent safety monitors that only see one transcript at a time, demonstrates the threat by constructing the first distributed multi-agent cyberattack that splits harmful tasks across accounts to evade standard monitors, and proposes an online stateful monitor that clusters weak signals across user accounts and escalates rarely to an LM, catching distributed attacks 30% earlier with negligible added latency for ~99% of traffic.

Key Results: The authors built the first distributed agent attack, a multi-agent scaffold that completes hard cybersecurity tasks while evading a standard monitor that catches it only ~1/5 as often as prior (single-account) agent attacks. Their proposed stateful online monitor, using real-time clustering and selective LM escalation, Pareto dominates standard monitors on large-scale simulated datacenter traffic: catches distributed attacks 30% earlier, flags misuse before the most harmful stages, and adds negligible latency for ~99% of traffic. The detection advantage persists but narrows as benign background traffic grows large. Bonus finding: the defense also catches standard jailbreaks because adaptive attackers reuse variants across accounts.

Key Findings:

A distributed multi-agent attack scaffold reduces detection by a standard monitor to ~20% of the rate at which it catches prior single-account agent attacks
A stateful, cross-account clustering monitor with selective LM escalation Pareto dominates standard monitors and detects distributed attacks 30% earlier, before the most harmful task stages, at negligible latency for ~99% of traffic
The cross-account detection advantage shrinks but persists as benign background traffic scales, and the same monitor incidentally catches standard jailbreaks because adaptive attackers reuse variants across accounts

Technical Novelty: Two contributions: (1) the first concrete distributed agent attack scaffold that decomposes a harmful objective across context-limited subagents on separate accounts, and (2) an online stateful monitor that clusters weak per-transcript suspiciousness signals across accounts in real-time and only escalates rare cluster-level cases to an LM judge — shifting the monitoring unit from a single transcript to a group of users.

What's New: Prior agent safety monitors operate per-transcript and are therefore structurally blind to misuse that is only visible in aggregate. This work reframes monitoring as a population-level inference problem, introduces both the offensive distributed-attack capability and a matching stateful, clustering-based online defense, and validates it at simulated datacenter scale.

Extension Opportunities:

Apply cross-account stateful monitoring to non-cyber domains (bio, CBRN, fraud) where distributed task-splitting could similarly hide intent
Study scaling laws: characterize how detection advantage degrades as benign background traffic grows, and design adaptive clustering thresholds that compensate
Build adversarial co-evolution benchmarks where distributed-attack scaffolds and stateful monitors are trained against each other to map robustness frontiers

Replicability: The abstract does not mention released code, datasets, or model checkpoints. Reproduction would require: a multi-agent scaffold harness, an LM-based judge/monitor (likely a frontier model), simulated datacenter-scale traffic for benign background, and a cybersecurity task suite — compute is moderate-to-high (many concurrent agent rollouts plus selective LM judging), but the selective escalation design keeps the bulk of monitoring cheap.

Research Gaps:

Detection performance still degrades as benign traffic grows very large — operating-point and scaling behavior at true production scale remains open
Adaptive attackers may diversify across accounts to avoid clustering signatures; the long-term arms race against an attacker explicitly optimizing for low cross-account similarity is not yet characterized

3. AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle

Authors: Weitong Qian, Beicheng Xu, Zhongao Xie... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can we build a unified automated scientific research system that supports the full research lifecycle (literature → ideation → experiments → manuscripts → rebuttal), maintains structured persistent memory across projects, and self-improves its research procedures over time — addressing the gap where existing LLM-based scientific agents only partially cover these requirements?

Summary: AutoSci is a memory-centric agentic system that automates the full scientific research lifecycle — from literature understanding through rebuttal — via four coupled modules: a schema-governed memory (SciMem), a five-stage execution harness (SciFlow), DAG-shaped multi-agent operators (SciDAG), and a feedback-driven self-improvement loop (SciEvolve). Its core contribution is treating persistent, structured, evolvable memory as the organizing principle for cross-project scientific agency.

Key Results: The abstract presents AutoSci as a system architecture with four named modules (SciMem, SciFlow, SciDAG, SciEvolve) covering a five-stage lifecycle, but provides no quantitative benchmarks, datasets, or comparative metrics. Proof is structural/qualitative: a code repository at github.com/skyllwt/AutoSci and a claimed end-to-end coverage of memory, execution, augmentation, and evolution that prior systems lack.

Key Findings:

A schema-governed split between reusable Long-Term Knowledge Memory and project-scoped Active Research Memory enables coherent state across long research cycles
A five-stage SciFlow harness controlling state, context, verification, feedback, and orchestration is sufficient to cover lit-review → rebuttal end-to-end
Feedback from users, experiments, reviews, and environments can be converted into versioned updates to memory organization, skills, and DAG templates, enabling system-level evolution

Technical Novelty: The novelty is the memory-centric architectural decomposition: explicit separation of Long-Term Knowledge Memory from Active Research Memory under a schema-governed store (SciMem), paired with a closed evolution loop (SciEvolve) that versions updates back into memory organization, skill harnesses (SciFlow), and DAG operator templates (SciDAG). Prior scientific-agent systems tend to be pipeline-shaped without persistent, schema-governed, evolvable memory spanning projects.

What's New: Unlike prior scientific agents that focus on one or two stages (ideation, experimentation, or writing) with ephemeral context, AutoSci unifies the entire lifecycle around persistent schema-governed memory and introduces an explicit evolution module that versions improvements back into the agent's own skills and templates.

Extension Opportunities:

Benchmark AutoSci against single-stage agents (e.g., AI Scientist, ResearchAgent) on a held-out set of reproducible research tasks to quantify the value of persistent cross-project memory
Extend SciDAG with domain-specific DAG templates (e.g., wet-lab biology, theorem proving, systems-ML empirical studies) and measure transfer of skills via SciEvolve across domains
Add a human-in-the-loop checkpoint layer that records reviewer/PI corrections as first-class signals into SciEvolve, then study how the system's skill templates drift over many projects

Replicability: Code is available at github.com/skyllwt/AutoSci. Compute is not specified in the abstract; based on the description (multi-agent LLM orchestration across long project cycles with DAG-shaped operators), reproduction likely requires sustained access to a strong frontier LLM API and external tool execution environments, plus storage for the structured memory — moderate-to-high cost rather than large training compute.

Research Gaps:

No reported quantitative evaluation: benchmarks, baselines, success rates, or human-acceptance metrics are absent from the abstract
Unclear how SciEvolve avoids skill degradation or memory contamination as feedback accumulates across heterogeneous projects

🦾 ROBOTICS

1. AR Forcing: Towards Long-Horizon Robot Navigation World Model

Authors: Yifei Yang, Zehua Fan, Huan Li... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can diffusion-based robot navigation world models be trained to remain stable over long-horizon autoregressive rollouts, given the train/inference distribution shift caused by parallel (teacher-forced) training versus autoregressive inference?

Summary: AR Forcing is an autoregressive training strategy for diffusion-based robot navigation world models that feeds the model its own rollouts during training while keeping the standard single-step diffusion loss. By aligning training with inference-time distribution without extra discriminators or losses, it produces more consistent long-horizon image predictions and more accurate trajectories across four navigation datasets.

Key Results: On four multi-domain navigation datasets (RECON, SCAND, HuRoN, TartanDrive), AR Forcing improves long-horizon image generation consistency and predicted trajectory accuracy versus strong baselines, demonstrating better robustness in both known and unknown environments. The abstract reports qualitative gains but does not include specific numerical benchmarks.

Key Findings:

Parallel/teacher-forced diffusion training causes a train-inference distribution shift that degrades long-horizon autoregressive rollouts
Folding autoregressive self-rollout into the diffusion training loop measurably improves image consistency and trajectory accuracy over long horizons
The approach generalizes across heterogeneous navigation domains (indoor social, off-road, structured) without architecture or sampler changes

Technical Novelty: Rather than using discriminators (as in GAN-style distribution matching) or auxiliary distribution-matching losses, AR Forcing simply rolls out the model's own predictions during training and applies the standard single-step diffusion noise-prediction loss on this self-generated context — closing the exposure-bias gap while preserving the vanilla diffusion framework and sampler.

What's New: Prior diffusion world models address exposure bias via auxiliary objectives (distribution matching, adversarial losses, consistency regularizers); AR Forcing is a minimal, drop-in training-loop modification that reuses the original diffusion loss and sampler, making it cheap to adopt.

Extension Opportunities:

Combine AR Forcing with action-conditioned policy learning so the world model and planner are co-trained under the same autoregressive distribution
Apply AR Forcing to other autoregressive diffusion domains (video generation, driving world models like GAIA/UniSim) to test generality beyond robot navigation
Investigate scheduled mixing between teacher forcing and AR Forcing (analogous to scheduled sampling) to balance training stability and exposure bias correction

Replicability: Authors state code will be released ('soon'); datasets used (RECON, SCAND, HuRoN, TartanDrive) are public navigation benchmarks. Reproduction likely requires multi-GPU training comparable to a diffusion video world model (tens of GPU-days), but exact compute is not stated in the abstract.

Research Gaps:

Abstract lacks concrete quantitative metrics (FVD, trajectory ADE/FDE, success rate), making the magnitude of improvement unclear
No discussion of computational overhead from autoregressive rollouts during training, which is typically the main cost of such schemes

2. DeMaVLA: A Vision-Language-Action Foundation Model for Generalizable Deformable Manipulation

Authors: Taiyi Su, Jian Zhu, Tianjian Wang... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can a single VLA foundation model generalize across diverse deformable-object folding tasks (varying clothing categories, geometries, materials, scenes) without suffering the task interference that plagues naive multi-task training of category-specific policies?

Summary: DeMaVLA is a Vision-Language-Action foundation model for generalizable cloth folding that combines a VLM backbone with a layer-pruned, flow-matching action expert. It is pre-trained on ~5,000 hours of dual-arm demonstrations and post-trained on a DAgger-aggregated mix of demos and real-failure corrections, replacing category-specific folding policies with a single generalist that performs competitively on RoboTwin and a real household folding benchmark.

Key Results: DeMaVLA, pre-trained on ~5,000 hours of real-world dual-arm demonstrations and post-trained via a human-in-the-loop DAgger pipeline on mixed folding data, achieves competitive performance on the RoboTwin benchmark and strong results on the authors' real-world household folding benchmark. Specific quantitative deltas vs. baselines are not disclosed in the abstract.

Key Findings:

A single VLA can match or exceed category-specific folding policies when post-trained on DAgger-aggregated corrective data, mitigating the task interference of naive multi-task mixing
Pruning every other transformer layer of the action expert while preserving layer-wise alignment with the VLM backbone is sufficient to retain performance while reducing training/inference cost
Scaling real-world dual-arm pre-training to ~5,000 hours yields general manipulation priors that transfer to deformable folding across categories, geometries, materials, and scenes

Technical Novelty: Two coupled ideas: (1) constructing the action expert by pruning every other transformer layer of the VLM backbone while preserving layer-wise alignment, cutting compute without losing VLM grounding; (2) a DAgger-style post-training pipeline that aggregates self-collected demos with corrective trajectories from real-robot failures, applied to a flow-matching action head for continuous control in deformable manipulation.

What's New: Prior VLAs typically train per-category folding policies or fall apart under naive multi-task mixing. DeMaVLA is, per the authors, the first VLA foundation model targeting generalizable deformable manipulation, combining layer-pruned action experts, flow-matching continuous control, large-scale real dual-arm pre-training, and human-in-the-loop DAgger corrective post-training in one stack.

Extension Opportunities:

Apply the layer-pruned action expert + flow matching recipe to other deformable domains (food prep, cable/rope manipulation, surgical suturing) to test whether the priors transfer beyond cloth
Replace the human-in-the-loop DAgger corrections with autonomous failure detection (e.g., VLM-based success classifiers) to scale corrective data collection without operator cost
Investigate alternative pruning schedules (every 3rd/4th layer, learned pruning) and quantify the cost-quality Pareto frontier vs. the every-other-layer choice

Replicability: Abstract does not mention code, model weights, or dataset release. Reproducing the 5,000-hour dual-arm pre-training plus DAgger corrective collection would require a substantial real-robot fleet, teleoperation infrastructure, and likely tens to hundreds of GPUs for VLM+flow-matching training — out of reach for most academic labs without the authors' data.

Research Gaps:

No reported quantitative comparison in the abstract against leading VLAs (pi-0, RDT, OpenVLA) on identical folding splits, leaving the magnitude of the improvement unclear
Generalization is evaluated on folding only; transfer to non-cloth deformables (food, ropes, soft tissue) and to truly novel scenes/embodiments outside the authors' data distribution is untested

3. Adaptive Artificial Time-Delay Control with Barrier Lyapunov Constraints for Euler-Lagrange Robots

Authors: Saksham Gupta, Rishabh Dev Yadav, Sarthak Mishra... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can Euler-Lagrange robotic systems simultaneously compensate for state-dependent uncertainties (without prior model knowledge) while enforcing time-varying constraints on both position and velocity — a combination underserved by existing control designs?

Summary: The paper proposes an adaptive controller for Euler-Lagrange robots that fuses time-delay estimation (TDE) with a barrier Lyapunov function (BLF). It analytically bounds the TDE approximation error as state-dependent, adapts the bound online without model knowledge, and uses the BLF to enforce time-varying position and velocity constraints, with Lyapunov-proven stability and 5-DOF hardware validation.

Key Results: The authors derive an analytical state-dependent upper bound on the time-delay estimation (TDE) approximation error, construct an online adaptive law for its parameters, and prove closed-loop stability via Lyapunov analysis. Validation is experimental on a 5-DOF robotic manipulator, where the framework maintains strict adherence to safety-critical position/velocity constraints under dynamic uncertainties, compared against state-of-the-art baselines (specific quantitative margins not given in abstract).

Key Findings:

The TDE approximation error admits an analytical state-dependent upper bound whose parameters can be estimated online, removing reliance on prior dynamic model knowledge.
Coupling adaptive TDE with a BLF yields provably stable closed-loop behavior while enforcing time-varying position and velocity constraints.
Hardware experiments on a 5-DOF manipulator confirm strict constraint adherence under dynamic uncertainties, outperforming state-of-the-art baselines.

Technical Novelty: Prior TDE controllers assume a constant or loosely-bounded approximation error and do not enforce hard state constraints; prior BLF controllers typically require known model structure. This paper's novelty is the analytically derived state-dependent bound on TDE error combined with an online adaptive law, fused with a BLF that enforces time-varying bounds on both position and velocity — yielding model-free uncertainty compensation under hard safety constraints.

What's New: First framework, to the authors' knowledge, that unifies adaptive TDE with a BLF for simultaneous model-free uncertainty compensation and time-varying state (position + velocity) constraint enforcement in Euler-Lagrange systems.

Extension Opportunities:

Extend the barrier Lyapunov + adaptive TDE framework to multi-robot or cooperative manipulation with coupled constraints (e.g., shared workspace collision bounds).
Integrate the controller with a learning-based residual model (e.g., Gaussian Process or neural network) to handle uncertainties that exceed the analytical state-dependent bound, while preserving the BLF safety guarantees.
Apply the framework to soft or compliant robots and to contact-rich tasks (assembly, polishing) where time-varying force constraints — not just position/velocity — must be enforced.

Replicability: No code or dataset is mentioned in the abstract. Reproduction would require a 5-DOF manipulator (or comparable Euler-Lagrange testbed) with real-time torque-level control (~1 kHz loop); compute requirements are minimal (a standard real-time controller / embedded PC), but hardware access is the binding constraint.

Research Gaps:

Existing TDE controllers do not natively enforce hard safety constraints, while BLF controllers typically require model structure — the joint design has been missing.
Most prior constraint-aware robotic controllers handle only position constraints or constant bounds; time-varying joint velocity constraints with state-dependent uncertainty are rarely addressed.

💻 COMPUTE

1. RAFI -- A Ray/Work Forwarding Infrastructure for Data Parallel Multi-Node/Multi-GPU Computing

Authors: Ingo Wald, Serkan Demirci, Alper Sahistan... Published: 2026-05-28 | Citations: 0 arXiv | PDF

Research Question: How can developers simplify the engineering complexity of building data-parallel multi-GPU/multi-node CUDA applications where work items (e.g., rays in ray tracing) must dynamically migrate between GPUs, without manually handling the low-level CUDA streams, buffers, and MPI communication?

Summary: RaFI is a CUDA+MPI software framework that exposes a simple API for CUDA kernels to forward rays or similar work items to other GPUs across nodes, hiding the underlying stream management, buffering, and MPI plumbing. It targets data-parallel rendering and compute workloads where work migrates across distributed GPU domains, demonstrated through several example applications.

Key Results: The paper presents RaFI as a CUDA+MPI framework abstracting work-item forwarding between GPUs. The abstract does not cite specific benchmark numbers, datasets, or quantitative speedups; it claims demonstrated 'potential' through several example applications (likely ray tracing/visualization given cs.GR categorization), but concrete throughput, scaling efficiency, or latency figures are not provided in the abstract.

Key Findings:

Multi-GPU/multi-node work-item migration can be abstracted into a uniform kernel-side forwarding API rather than re-implemented per application
A combined CUDA + MPI runtime can manage cross-GPU transfers transparently while preserving kernel-level data-parallel programming model
The abstraction is general enough to support multiple example applications (likely distributed ray tracing variants in cs.GR)

Technical Novelty: A kernel-callable forwarding primitive that lets CUDA device code enqueue work items destined for other GPUs (local or remote), with the framework transparently coalescing, buffering, and dispatching them over CUDA peer copies and MPI. Prior multi-GPU ray tracing systems typically hand-rolled this per-application; RaFI factors it into a reusable abstraction.

What's New: Unlike prior distributed ray tracers (e.g., OSPRay, Galaxy, custom MPI codes) where forwarding logic is application-specific and intertwined with rendering code, RaFI isolates the forwarding mechanism into a reusable infrastructure layer callable directly from CUDA device code.

Extension Opportunities:

Extend RaFI beyond rays to general particle/agent-based simulations (SPH, n-body, Monte Carlo transport) where work items migrate between spatial domain decompositions on different GPUs
Integrate RaFI with NVSHMEM or NCCL for finer-grained intra-node GPU peer transfers, bypassing MPI overhead and benchmarking against the current MPI-based forwarding path
Add dynamic load balancing on top of RaFI's forwarding primitive — currently it forwards where the kernel says, but a scheduler layer could redirect work items based on per-GPU queue depth

Replicability: The abstract does not mention an open-source release or specific dataset availability. Reproduction would require a multi-node GPU cluster with CUDA-aware MPI (e.g., OpenMPI + UCX), multiple NVIDIA GPUs per node, and the example applications referenced in the paper. Code availability would need to be checked in the full paper.

Research Gaps:

Abstract provides no quantitative performance comparison against hand-tuned multi-GPU implementations or alternative communication libraries (NCCL, NVSHMEM, GASNet)
Scalability limits, latency overheads, and behavior under load imbalance are not characterized in the abstract

2. HE^2: A Communication-Light Heterogeneous Architecture for Efficient Fully Homomorphic Encryption

Authors: Shangyi Shi, Husheng Han, Zhaoxuan Kan... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can CKKS fully homomorphic encryption (FHE) be accelerated efficiently when it contains both compute-intensive (ComOps) and memory-intensive (MemOps) operators, given that pure ASIC or near-memory-processing (NMP) designs have high hardware overhead, and naive heterogeneous xPU-xMU designs suffer from frequent long-latency communication dominated by the keyswitch operator?

Summary: HE^2 is a heterogeneous xPU-ASIC + xMU-NMP accelerator for CKKS FHE that tackles the communication bottleneck of the keyswitch operator. Through a dataflow-graph optimization that fuses parallel keyswitch blocks via hoisting, plus a group-level pipelined execution that overlaps communication with compute, it delivers 1.66x speedup and 9.23x lower EDAP over the prior SOTA while keeping communication stalls under 7%.

Key Results: HE^2 achieves 1.66x end-to-end speedup and 9.23x lower Energy-Delay-Area Product (EDAP) compared to the state-of-the-art accelerator. Communication stalls are reduced to only 6.67% of total latency, demonstrating effective hiding of cross-component communication through group-level pipelining and DFG fusion of ModUp/ModDown operations.

Key Findings:

Most cross-component communication in heterogeneous FHE accelerators originates at the ModUp/ModDown ↔ neighboring MemOps interface, making it the right optimization target
Parallel keyswitch blocks can be identified and fused via the hoisting algorithm at the DFG level, substantially reducing ModUp/ModDown invocation count
Group-level pipelining across RNS-decomposed groups hides communication latency, reducing communication stalls to 6.67% of total runtime

Technical Novelty: Two co-designed contributions: (1) a DFG-level optimization that identifies parallel keyswitch blocks and fuses them via the hoisting algorithm to reduce ModUp/ModDown communication frequency — going beyond prior single-operator hoisting; (2) a group-level pipelined heterogeneous execution model that hides xPU-xMU communication latency using inherent parallelism across decomposed RNS groups, rather than treating communication as a fixed cost.

What's New: Prior FHE accelerators are either pure ASIC (compute-strong, memory-weak) or pure NMP (memory-strong, compute-weak). HE^2 is among the first to commit to a heterogeneous xPU-xMU split and then directly attack the resulting communication tax via joint DFG-level hoisting fusion and architecture-level group pipelining, rather than treating them as separate concerns.

Extension Opportunities:

Apply the DFG-level hoisting/fusion framework to other FHE schemes beyond CKKS (e.g., BFV, BGV, TFHE) where keyswitch is also a bottleneck
Extend the xPU-xMU heterogeneous co-design to GPU+HBM-PIM commodity hardware to make the approach deployable without custom ASIC fabrication
Integrate with bootstrapping-heavy workloads (deep neural network inference on encrypted data) and study whether group-level pipelining holds under deeper circuits with more frequent bootstraps

Replicability: The abstract does not mention code/dataset release. Reproduction would require either ASIC simulation infrastructure (cycle-accurate simulators for xPU compute units and NMP/PIM memory) plus a CKKS workload suite (e.g., HEAAN, OpenFHE, SEAL benchmarks), or RTL synthesis with a commercial PDK to validate area/energy claims — substantial hardware research compute.

Research Gaps:

Evaluation appears confined to CKKS; generalization to other FHE schemes and to bootstrapping-dominated workloads is not established in the abstract
No discussion of programmability or compiler integration — DFG fusion is presented as an offline optimization rather than part of a general FHE compiler stack

3. HetCCL: Enabling Collective Communication For Mixed-Vendor Heterogeneous Clusters

Authors: Yuejie Wang, Tao Chang, Yuanyuan Zhao... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can collective communication (AllReduce, ReduceScatter, etc.) be performed efficiently across mixed-vendor heterogeneous GPU clusters when existing libraries like NCCL/RCCL assume homogeneous hardware and heterogeneous-capable libraries like Gloo/OpenMPI suffer heavy data-path overhead?

Summary: HetCCL is a collective communication framework for mixed-vendor heterogeneous GPU clusters that combines zero-copy heterogeneous P2P transport, a border-communicator mechanism leveraging each vendor's native reduction, and a hierarchical topology abstraction. It delivers 17-19x bandwidth over Gloo and up to 16.9% end-to-end LLM training speedup, closing the gap between homogeneous-only NCCL-class libraries and slow heterogeneous-capable libraries.

Key Results: HetCCL achieves 17-19x higher bandwidth than Gloo in heterogeneous communications and speeds up end-to-end LLM training by up to 16.9% in per-step time. Implemented and evaluated with 4 vendor backends across 4 heterogeneous cluster settings using both microbenchmarks and end-to-end LLM training workloads.

Key Findings:

Host-device memory copies are the dominant overhead in existing heterogeneous CCLs and can be eliminated via direct P2P with CPU-side control offload
Vendor independence for combining collectives is achievable without reimplementing reductions by exploiting intrinsic reductions already present in vendor CCLs (border-communicator)
A hierarchical decomposition into cluster-level primitives can simultaneously guarantee optimal cross-cluster data transfer volume and optimal bandwidth utilization in heterogeneous topologies

Technical Novelty: Three coupled innovations: (1) heterogeneous P2P transport using direct device-to-device transfers that bypass host memory copies while offloading control to CPUs; (2) a border-communicator mechanism that reuses each vendor's native intrinsic reduction inside vendor CCLs to achieve vendor-independent combining collectives; (3) a hierarchical topology abstraction that decomposes collectives into cluster-level primitives proven optimal for cross-cluster data volume and bandwidth utilization.

What's New: Unlike NCCL/RCCL (homogeneous-only) or Gloo/OpenMPI (heterogeneous but high-overhead), HetCCL is the first framework to combine efficient heterogeneous P2P transport with a portable reduction mechanism that piggybacks on vendor-native primitives, plus a topology abstraction tailored to mixed-vendor clusters.

Extension Opportunities:

Extend the border-communicator mechanism to support emerging accelerators (TPUs, Trainium, Gaudi) beyond the 4 evaluated vendors
Integrate HetCCL with PyTorch FSDP/DeepSpeed ZeRO and develop a heterogeneity-aware parallelism planner that co-optimizes sharding strategy with the hierarchical topology abstraction
Add fault tolerance and elastic scaling so mid-training vendor pool changes (preemptions, spot instances) can be absorbed without restarting the collective topology

Replicability: Abstract does not mention an open-source release. Reproduction requires access to a genuinely mixed-vendor GPU cluster (4 vendors), each vendor's native CCL (NCCL, RCCL, etc.), and an LLM training stack — non-trivial hardware diversity is the main barrier rather than compute scale.

Research Gaps:

No discussion of scaling behavior beyond the 4 evaluated heterogeneous settings or behavior under network contention and failures
Unclear how the approach interacts with advanced parallelism strategies (pipeline, expert, sequence parallelism) and with collectives beyond AllReduce/ReduceScatter such as AllToAll for MoE

⚡ ENERGY

1. Sharp periodic Ge concentration modulations beyond the conduction band valley wavevector $k_0$ in nuclear spin-free Si quantum wells

Authors: Ivo Rahlff, Carsten Richter, Martin Schmidbauer... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: Can periodic Ge concentration modulations in Si quantum wells be grown with sufficient sharpness and periodicity (at or near the Si valley wavevector k₀ = 9.7 nm⁻¹, period 0.64 nm) to deterministically enhance conduction-band valley splitting for scalable spin-qubit quantum computing?

Summary: The paper demonstrates MBE growth of Si quantum wells with periodic Ge concentration modulations matching the Si conduction-band valley wavevector k₀ (period 0.64 nm) using isotopically purified ²⁸Si and ⁷²Ge to create a nuclear-spin-free host. Synchrotron XRD and STEM confirm sharp, laterally homogeneous modulations, and k·p simulations show that 2k₀/3 trapezoidal profiles yield stronger deterministic valley splitting than k₀ profiles — a key step toward scalable Si spin qubits.

Key Results: Using nuclear-spin-free MBE with ²⁸Si and ⁷²Ge, the authors fabricated Ge-modulated Si quantum wells with periods spanning 2.00 nm down to 0.49 nm, including modulations precisely at k₀ and 2k₀/3. Synchrotron XRD and STEM confirmed laterally homogeneous Ge modulations over micrometer scales, with amplitudes up to 10 at-% Ge and concentration gradients reaching 20 at-%/nm. Two-band k·p simulations indicate that steep trapezoidal 2k₀/3 heterostructures yield deterministic valley splitting enhancement, while k₀-period wells produce a much weaker effect.

Key Findings:

Ge-modulated Si quantum wells with periods from 2.00 nm down to 0.49 nm were successfully grown, including the critical k₀ (0.64 nm) and 2k₀/3 conditions
Modulations are laterally homogeneous over micrometer scales with up to 10 at-% Ge amplitude and gradients up to 20 at-%/nm
k·p simulations reveal that steep trapezoidal 2k₀/3 heterostructures, not k₀ wells, give the strongest deterministic valley splitting enhancement — overturning the intuitive k₀-matching assumption

Technical Novelty: First demonstration of sub-nanometer (0.49–2.00 nm) periodic Ge concentration modulations grown via nuclear-spin-free MBE using isotopically purified ²⁸Si and ⁷²Ge sources, achieving the long-targeted k₀ wavevector matching condition along with steep ~20 at-%/nm composition gradients verified by synchrotron X-ray and STEM. Prior work could not simultaneously achieve nuclear-spin-free hosts AND deterministic k₀-scale Ge modulations.

What's New: Combines isotopically purified nuclear-spin-free hosts (²⁸Si/⁷²Ge) with sub-nanometer-period Ge modulations at the valley wavevector — a previously unachieved combination — and identifies 2k₀/3 trapezoidal profiles as more effective than the naively expected k₀ matching condition.

Extension Opportunities:

Fabricate gated single-electron quantum dot devices on these 2k₀/3 trapezoidal heterostructures and directly measure valley splitting via magnetospectroscopy to validate the k·p predictions experimentally
Extend the isotopically purified MBE recipe to grow even sharper trapezoidal profiles (>20 at-%/nm gradients) or higher Ge amplitudes to push valley splitting beyond current thermal-noise thresholds
Integrate these modulated wells with industrial CMOS-compatible processes (e.g., transfer to 300mm wafers, gate stack deposition) to benchmark valley splitting yield and uniformity across a qubit array

Replicability: No mention of public code/data in the abstract. Reproduction requires specialized MBE with isotopically purified ²⁸Si and ⁷²Ge sources (rare and expensive), synchrotron X-ray beamline access, and aberration-corrected STEM — making full replication accessible only to a handful of well-equipped semiconductor/quantum-materials labs. k·p simulations are reproducible with standard tools.

Research Gaps:

No direct electrical measurement of valley splitting in gated qubit devices — predictions remain theoretical via k·p simulations
Scalability to wafer-scale and integration with CMOS-compatible gate stacks not yet addressed

2. Crystallisation kinetics of supercooled liquid palladium

Authors: Zuzanna Kostera, Przemyslaw Dziegielewski, Konstantinos Georgarakis... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How fast does supercooled liquid palladium crystallise, and what cooling rate is required to vitrify it? Specifically, the paper addresses whether homogeneous nucleation alone can explain the supercooling observed in rapidly quenched Pd thin film experiments, bridging atomistic MD simulations with time-resolved X-ray diffraction measurements.

Summary: Classical MD simulations of supercooled liquid Pd quantify nucleation rates, growth velocities, and self-diffusion across 0.38–0.65 T_m, yielding a TTT diagram with a nose at 0.5 T_m / 100 ps and a critical cooling rate of ~10^13 K/s. The simulation results quantitatively reproduce crystallisation onset measured by time-resolved XRD on rapidly quenched Pd thin films, establishing that homogeneous nucleation governs the achievable supercooling.

Key Results: MD simulations across 700–1150 K (0.38–0.65 T_m) yielded: (1) Arrhenius self-diffusion with activation energy 467(6) meV/atom; (2) peak homogeneous nucleation rate ~4×10^35 m^-3 s^-1 near 0.5 T_m; (3) crystal growth velocities of m/s following diffusion-limited Wilson-Frenkel (not collision-limited) kinetics; (4) TTT nose at 0.5 T_m and 100 ps, implying critical cooling rate ~10^13 K/s for vitrification; (5) simulated crystallisation onset matches X-ray diffraction on Pd films quenched at 5×10^11 K/s.

Key Findings:

Self-diffusion in supercooled liquid Pd is Arrhenius with E_a = 467(6) meV/atom
Homogeneous nucleation rate peaks at ~4×10^35 m^-3 s^-1 near 0.5 T_m; TTT nose at ~100 ps implies ~10^13 K/s critical cooling rate for vitrification
Crystal growth follows diffusion-limited Wilson-Frenkel kinetics — not the collision-limited regime often assumed for pure metals
MD predictions match time-resolved XRD on Pd films quenched at 5×10^11 K/s, implying homogeneous (not heterogeneous) nucleation governs experimental supercooling

Technical Novelty: Direct quantitative bridging of atomistic MD-derived TTT diagrams with time-resolved synchrotron XRD on optically molten thin films, with the surprising finding that growth follows diffusion-limited Wilson-Frenkel rather than collision-limited kinetics in a pure metal — and the demonstration that experimentally observed supercooling is governed by homogeneous (not heterogeneous) nucleation.

What's New: Most prior MD work on metallic crystallisation stops at nucleation rates or growth velocities in isolation; this work assembles a full TTT diagram with statistical rigor and directly validates against time-resolved synchrotron experiments on the same material. The finding that pure Pd growth is diffusion-limited rather than collision-limited challenges a long-standing assumption for simple metals.

Extension Opportunities:

Apply the same MD + time-resolved XRD methodology to Pd-based binary/ternary alloys (e.g., Pd-Si, Pd-Cu-Ni-P) to predict glass-forming ability from first-principles TTT diagrams
Couple the simulation framework with machine-learning interatomic potentials (e.g., MACE, NequIP) to extend accuracy to multi-component metallic glass systems while retaining ab initio fidelity
Build a public TTT-diagram database for pure metals using this protocol to benchmark nucleation theories (CNT vs. non-classical pathways) across the periodic table

Replicability: No explicit code/data link mentioned in abstract. Reproduction requires a classical MD package (LAMMPS) with an EAM-type Pd potential, ~100s of independent trajectories of ~10^4–10^5 atoms across 10+ temperatures — feasible on a modest HPC cluster (~10^4–10^5 CPU-hours). XRD validation requires a synchrotron beamline with fast-quench thin-film capability.

Research Gaps:

Pure Pd is not a practical glass-former at experimentally accessible quench rates — extension to alloys is needed to connect to bulk metallic glass design
Classical EAM potentials may underestimate electronic-structure effects near melting; ML-potential or AIMD cross-validation would strengthen the kinetic predictions

3. Direct Observation of Chemical Short-Range Order in CoCrNi Alloy Using Neutron Diffraction

Authors: Vinícius P. Bacurau, Camilo Salvador, Guilherme C. Stumpfa... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: Does chemical short-range order (CSRO) genuinely exist in the equiatomic CoCrNi medium-entropy alloy, and can it be directly observed and quantified experimentally rather than only inferred indirectly or predicted computationally?

Summary: The paper provides direct experimental neutron-diffraction evidence of chemical short-range order in the CoCrNi medium-entropy alloy, identifying a diffuse peak at Q = 1.85 Å⁻¹ whose intensity grows with aging. Combined SANS and atomistic simulations reveal nanoscale Ni-rich disk-shaped domains and D022/Pt2Mo/D1a-like motifs, demonstrating CSRO is an intrinsic, energetically favorable feature of CoCrNi.

Key Results: Neutron diffraction revealed a distinct diffuse CSRO peak at Q = 1.85 Å⁻¹ in CoCrNi. CSRO intensity scaled with aging: prolonged aging (100 h, 240 h) at 748 K and shorter aging (24 h) at 798 K increased ordering. Gas-atomized (least-ordered) samples retained ~70% of the CSRO level of the 240 h/748 K (most-ordered) sample, indicating CSRO is intrinsic. SANS identified Ni-rich, disk-shaped domains of radius ~11 Å and thickness ~1 Å. FT-2D analysis tied (1 ½ 0) reflections in the <001> zone to D022, Pt2Mo, and D1a-like motifs. Atomistic simulations reproduced both the position and presence of the diffuse peak.

Key Findings:

CSRO in CoCrNi manifests as a diffuse neutron-diffraction peak at Q = 1.85 Å⁻¹ whose intensity increases with low-temperature aging (748–798 K).
CSRO is intrinsic — even rapidly solidified gas-atomized samples retain ~70% of the maximum CSRO level achieved by 240 h aging at 748 K.
SANS reveals Ni-rich, disk-shaped CSRO domains (~11 Å radius, ~1 Å thickness); FT-2D analysis attributes reflections to D022, Pt2Mo, and D1a-like local motifs, confirmed by atomistic simulation.

Technical Novelty: Direct experimental detection of CSRO via a diffuse neutron-diffraction peak at Q = 1.85 Å⁻¹, combined with FT-2D motif identification (D022/Pt2Mo/D1a) and SANS domain morphology — prior work relied largely on electron diffraction, APT, or simulation alone, none of which provided bulk, quantitative, isotope-sensitive CSRO measurement.

What's New: First direct, bulk, quantitative neutron-diffraction observation of CSRO in CoCrNi, paired with motif-level structural identification via FT-2D and morphological characterization via SANS — establishing a reusable experimental + computational framework for probing local ordering in MPEAs.

Extension Opportunities:

Apply the same neutron-diffraction + FT-2D + SANS protocol to other multi-principal element alloys (e.g., CrCoNiFe, CrMnFeCoNi) to systematically map CSRO across the MPEA family.
Correlate the quantified CSRO degree with measured mechanical properties (yield strength, stacking-fault energy, fracture toughness) to build a CSRO–property predictive model.
Build a machine-learning interatomic potential trained on the simulation/diffraction data to predict CSRO evolution under arbitrary thermal histories, enabling process-design optimization.

Replicability: No mention of public code or data release in the abstract. Reproduction requires access to a neutron diffractometer and SANS beamline (e.g., national lab facility), gas-atomized CoCrNi powder, controlled aging furnaces (748–798 K, up to 240 h), and standard atomistic simulation tools (e.g., LAMMPS with appropriate MEAM/EAM potentials) — compute for simulations is modest (workstation-to-small-cluster scale); the experimental beam time is the principal barrier.

Research Gaps:

No direct mapping yet between the quantified CSRO degree and the alloy's macroscopic mechanical properties.
Generalization of the diffuse-peak fingerprint to other compositions and processing routes (e.g., additive manufacturing, severe plastic deformation) remains unexplored.

🏥 HEALTHCARE

1. mRNAutilus: Multi-Objective-Guided Discrete Generation of mRNA with Optimized Therapeutic Properties

Authors: Sawan Patel, Sophia Tang, Yesol Kim... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: How can therapeutic mRNA be designed holistically — jointly optimizing coding sequence (codon usage) and UTRs as a coupled system — rather than designing them separately and assembling post hoc, while simultaneously satisfying multiple competing functional objectives (stability, translation efficiency, protein expression)?

Summary: mRNAutilus is a masked discrete diffusion framework that generates complete therapeutic mRNAs (CDS + UTRs together) and uses Monte Carlo Tree Guidance with lightweight property regressors to steer generation toward Pareto-optimal trade-offs across stability, translation efficiency, and protein expression. Zero-shot generated sequences outperform commercial, clinical, and prior ML baselines across luciferase, SARS-CoV-2 Spike, prime editing, and E3 ligase applications.

Key Results: Zero-shot mRNAutilus-generated P. pyralis luciferase mRNAs achieved >400-fold higher expression than wild-type, outperforming commercial and ML-designed baselines including zero-shot generative methods. Zero-shot SARS-CoV-2 Spike mRNAs exceeded clinically used and commercial constructs and matched/surpassed lab-optimized designs with improved durability. Generality was demonstrated on prime editing (PEMax) and uAb E3 ligase constructs that enhanced beta-catenin degradation. The masked discrete diffusion backbone was trained on millions of full-length mRNAs.

Key Findings:

Joint CDS+UTR generation in a single diffusion process outperforms decoupled design-then-assemble pipelines
Monte Carlo Tree Guidance over discrete diffusion trajectories enables tractable multi-objective Pareto optimization for sequence design
Zero-shot generated luciferase mRNAs achieve >400× wild-type expression; Spike constructs match or beat lab-optimized clinical designs with better durability
The framework generalizes to advanced therapeutic modalities including prime editors and peptide-guided E3 ligases

Technical Novelty: The combination of (1) a masked discrete diffusion model over full-length transcripts (CDS + UTRs jointly) with (2) Monte Carlo Tree Guidance using lightweight regressors trained on model embeddings to steer generation toward Pareto-optimal multi-objective solutions — replacing the prior paradigm of separate CDS/UTR design with post hoc assembly and screening.

What's New: Prior methods (e.g., LinearDesign, CodonBERT, UTR-LM) optimize CDS or UTR in isolation and stitch them together. mRNAutilus is the first to (a) generate the full transcript jointly via discrete diffusion and (b) apply Monte Carlo Tree Guidance for multi-objective Pareto-efficient steering, capturing CDS–UTR coupling that single-region methods miss.

Extension Opportunities:

Extend Monte Carlo Tree Guidance with additional objective regressors (e.g., immunogenicity, secondary-structure folding energy, manufacturability/IVT yield) to broaden the Pareto frontier for clinical-grade design
Adapt the framework to tissue- or cell-type-specific expression by fine-tuning the embedding regressors on context-specific ribosome profiling/MPRA datasets
Couple mRNAutilus with a closed-loop wet-lab DBTL pipeline (e.g., MPRA or cell-free expression screening) so that experimental measurements continuously update the guidance regressors

Replicability: The abstract does not mention code/data release. Reproduction would require: a large mRNA corpus (millions of full-length transcripts, likely from RefSeq/GENCODE/Ensembl), substantial GPU compute for training the diffusion backbone (likely multi-GPU, days-to-weeks), labeled datasets for half-life/TE/protein abundance regressors, and wet-lab capacity for validation. Code availability would need to be checked in the full paper.

Research Gaps:

No reported integration of biophysical constraints like RNA secondary structure folding energy or immunogenic motif avoidance into the guidance objectives
Generalization to non-human/non-mammalian hosts and to chemically modified nucleotides (pseudouridine, etc.) is not addressed

🔬 MATERIALS

1. In-situ operation of amorphous circuits under heavy-ion irradiation

Authors: Xuanzhe Sha, Shun Liao, Xiaoxi Li... Published: 2026-05-29 | Citations: 0 arXiv | PDF

Research Question: Can complex digital circuits built on non-silicon (amorphous thin-film) semiconductors operate reliably in situ under heavy-ion irradiation, addressing the gap where radiation tolerance of such platforms has only been characterized at the device level rather than at the operating system level?

Summary: The authors fabricate a ~100-transistor timing circuit on amorphous thin-film semiconductors and demonstrate functional 'Hello World' ASCII output. They then operate the circuit live under tantalum heavy-ion bombardment up to 1 x 10^6 ions/cm^2 fluence, showing that amorphous semiconductor digital logic can sustain prolonged powered operation in extreme radiation environments.

Key Results: Demonstrated a ~100-transistor timing circuit on amorphous thin-film semiconductors producing a functional 'Hello World' ASCII output sequence. Under powered tantalum heavy-ion irradiation at a flux of 2.5 x 10^3 ions/cm^2/s, the circuit maintained stable operation up to a total fluence of 1 x 10^6 ions/cm^2.

Key Findings:

A 100-transistor-scale digital timing circuit can be realized on amorphous thin-film semiconductors with full functional output.
The circuit operates stably under continuous powered heavy-ion (Ta) irradiation at 2.5 x 10^3 ions/cm^2/s.
Total accumulated fluence of 1 x 10^6 ions/cm^2 is tolerated without functional failure, establishing a new milestone for in-situ digital operation.

Technical Novelty: Prior radiation-hardened beyond-silicon work focused on single-device characterization or unpowered post-irradiation testing; this paper presents the first powered, in-situ heavy-ion irradiation of a 100-transistor-scale digital circuit on amorphous thin-film semiconductors, shifting evaluation from device-level to operationally relevant system-level.

What's New: First demonstration of a powered, multi-transistor digital circuit on amorphous (non-crystalline, beyond-silicon) semiconductors operating in situ under heavy-ion irradiation, moving radiation-tolerance assessment from individual devices to integrated circuits.

Extension Opportunities:

Scale the architecture from ~100 to thousands of transistors to build full microcontroller or memory blocks while preserving in-situ radiation tolerance.
Stress-test under different ion species (protons, neutrons, gamma) and broader energy/LET ranges to map the full radiation-tolerance envelope and qualify for space or reactor deployment.
Integrate amorphous-oxide logic monolithically with sensors (e.g., radiation detectors, thermal sensors) for radiation-hardened edge processing on flexible substrates.

Replicability: No mention of open code or data release in the abstract. Reproduction would require thin-film fab capability (amorphous oxide semiconductor deposition, ~100-transistor lithography) plus access to a heavy-ion accelerator beamline (e.g., tantalum ions at ~10^3 ions/cm^2/s flux), making replication accessible only to well-equipped materials/radiation-effects labs.

Research Gaps:

Scaling to larger circuits (thousands+ transistors), more complex logic, and memory remains untested.
Long-term reliability, threshold-voltage drift, and recovery dynamics under broader radiation conditions (different ions, LETs, dose rates) are not yet characterized.

🔥 GitHub Trending

1. PanisHandsome/ai-rules-sync

⭐ 27 stars | JavaScript

Keep one source of truth for your AI coding-agent rules. Convert and sync between AGENTS.md, CLAUDE.md, .cursorrules, Copilot, Windsurf, Cline, Aider & Gemini — or scaffold a fresh AGENTS.md. Zero dep

agents-md ai claude cli codex coding-agent

2. wellkilo/conda-helper

⭐ 2 stars | Python

An enhanced, cross-platform CLI on top of conda. One-liner backup, restore, clone, offline pack, batch cleanup, and an interactive panel — plus friendly error hints that translate raw conda stack trac

backup cli click command-line-tool conda conda-environment

3. ApostolisFalaras/Thesis-Evaluation-of-Modern-RAG-Systems

⭐ 2 stars | Jupyter Notebook

Diploma thesis evaluating Retrieval-Augmented Generation (RAG) systems on MS MARCO and LegalBench-RAG using RAGChecker and RAGAs.

artificial-intelligence chromadb generative-ai information-retrieval jupyter-notebook langchain

4. AhmetErenLapanta/nanoagent

⭐ 2 stars | Python

A minimal AI coding agent in a few hundred lines — no framework, no magic.

agent cli coding-agent llm tool-use

5. peterfei/awesome-ai-agents-cn

⭐ 2 stars | Unknown

中文 AI Agent 落地案例精选 —— 真实场景、真实代码、真实效果。

agent ai-agents awesome-list chinese claude-code llm

6. r-design-j/tri-party-framework

⭐ 2 stars | Shell

Verifiable Codex + Claude + Gemini collaboration framework with source checks, mutual cross-audit, merge gates, HTTP and MCP adapters.

agent-framework ai-agents claude claude-code codex developer-tools

7. Mr-Koala/deepseek-price-calculator

⭐ 2 stars | HTML

deepseek-v4-flash/pro价格计算器，帮助你快速评估自己的预算或者花费情况

api-pricing budget-planner calculator chinese cost-estimator deepseek

8. gmzhang/skillmng

⭐ 2 stars | Python

Agent Skills management platform for creation, validation, versioning, LLM-assisted editing, and Git-backed publishing.

agent-skills codex developer-tools fastapi llm openai

9. ApolloEagle/smooth-brain

⭐ 2 stars | JavaScript

stop AI from being so wordy

ai claude-code llm prompt-engineering

10. Boaxes/Reptile-Central-Database2

⭐ 2 stars | Python

Full-stack CRUD web app for a fictional reptile supply business - Streamlit, MySQL, Google Cloud Run, with an AI chatbot (RAG + natural language SQL) powered by GPT-4o mini and LlamaIndex.

artificial-intelligence ci-cd database google-cloud-run llm mysql

11. myklovenyzforever/chem-pdf-extractor

⭐ 2 stars | Python

LLM-powered scientific PDF data extraction tool for chemistry and chemical engineering literature.

catalysis chemical-engineering data-extraction excel-export literature-review llm

12. jayemscript/llm-systems-from-scratch

⭐ 1 stars | CMake

A hands-on learning project for building the core systems behind Large Language Models using C++, Rust, and optional Python/JavaScript bindings. Includes tensor operations, autograd, neural networks,

ai-systems autograd c-language cpp cuda educational-project

13. rohanmistry231/ML-OS

⭐ 1 stars | Unknown

A curated reference for machine learning concepts, MLOps practices, evaluation metrics, and core algorithms, designed for quick review and practical deployment guidance.

algorithms data-science machine-learning math metrics ml

14. yukthapriya/ARIA-Research-Agent

⭐ 1 stars | JavaScript

Multi-agent AI system for autonomous academic research. Implements agentic pipelines, LLM orchestration, tool-use, context management, and real-time streaming — powered by Claude Sonnet 4.

agentic-ai ai-agent ai-research anthropic autonomous-agents autonomous-agents-

15. kaushikchaturvedula/AI-Voice-Receptionist

⭐ 1 stars | Python

A complete, local, real-time voice receptionist built from scratch: it listens, transcribes, reasons, books appointments, and talks back — and you can interrupt it mid-sentence.

ai-agents machine-learning voice-agents voice-assistant

Generated by Research Pulse on 2026-06-01 06:09