2026-05-19·Daily digest

🔬 Research Pulse

Daily Digest

May 19, 2026

🤖 AI

🧠 LLMs

1. Code as Agent Harness

Authors: Xuying Ning, Katherine Tieu, Dongqi Fu... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How should we conceptualize and design 'agent harnesses' — the infrastructure connecting LLM agents to reasoning, action, and environment — when code itself becomes the operational substrate rather than just an output artifact?

Summary: A survey reframing agentic AI systems by positioning code as the central operational substrate (the 'harness') for agent reasoning, action, environment modeling, and verification. It organizes the field across three layers — harness interface, harness mechanisms (planning/memory/tools/feedback), and multi-agent scaling — and maps these onto application domains from coding assistants to embodied agents and enterprise workflows.

Key Results: This is a survey paper, not an empirical study — no benchmarks or numerical results are reported. Its contribution is a taxonomy organizing the field across three layers (harness interface, harness mechanisms, scaling to multi-agent) and a synthesis of representative methods across application domains: coding assistants, GUI/OS automation, embodied agents, scientific discovery, personalization/recommendation, DevOps, and enterprise workflows.

Key Findings:

Code is shifting from agent output to agent infrastructure, providing executable, verifiable, and stateful scaffolding for long-horizon agent behavior
Harness mechanisms (planning, memory, tool use, feedback control) can be uniformly expressed and optimized through code, enabling adaptive and reliable execution
Multi-agent coordination benefits from shared code artifacts that serve as common ground for review, verification, and stateful collaboration

Technical Novelty: Not a new technique — the novelty is the unifying framing of 'code as agent harness,' treating code as the connective tissue across reasoning, acting, memory, tool use, and multi-agent coordination, rather than as the agent's output. The three-layer organization (interface / mechanisms / scaling) is the contribution.

What's New: Prior surveys treat code generation and agent systems as separate threads; this work unifies them by arguing code is the harness itself, not just what agents produce — turning execution and verification into first-class design dimensions.

Extension Opportunities:

Build an evaluation framework that measures harness quality beyond final task success — e.g., intermediate state correctness, code-artifact reusability, and verification coverage under incomplete feedback
Implement a regression-free harness self-improvement loop where the agent rewrites its own harness code with automated guards against capability regressions on a held-out task suite
Design a shared-state protocol for multi-agent code harnesses (e.g., CRDT-backed code artifacts) so multiple agents can concurrently edit, review, and verify a shared execution substrate

Replicability: N/A — survey paper with no experiments. No code or datasets released. Reproducing the synthesis requires only literature access.

Research Gaps:

Evaluation methodology beyond final task success, including verification under incomplete or absent ground-truth feedback
Regression-free self-improvement of harnesses and consistent shared state across multi-agent systems, plus human oversight for safety-critical actions and multimodal extensions

2. What Does the AI Doctor Value? Auditing Pluralism in the Clinical Ethics of Language Models

Authors: Payal Chandak, Victoria Alkin, David Wu... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: Do large language models used for medical advice exhibit the value pluralism that characterizes good clinical practice, or do they impose systematic ethical preferences that could homogenize patient care at scale?

Summary: The paper introduces a benchmark and attribution framework to audit the ethical value priorities of LLMs giving medical advice, benchmarked against a panel of physicians. It finds that while frontier models collectively span physician value heterogeneity and discuss competing principles in their reasoning, individual models make near-deterministic decisions with systematic value biases — some significantly underweighting patient autonomy — raising the risk that single-model deployment replaces clinical pluralism with a monoculture.

Key Results: Built a clinician-verified benchmark of ethical dilemmas plus an attribution method that recovers value priorities from model decisions. Demonstrated: (1) frontier model ecosystem spans physician-level value heterogeneity; (2) models exhibit Overton pluralism (discuss competing values in reasoning) but commit near-deterministically across repeated sampling and semantic variations; (3) decisions reflect systematic, committed value preferences across cases; (4) most models fall within inter-physician variation, but some significantly underweight patient autonomy relative to physicians.

Key Findings:

Frontier LLMs exhibit Overton pluralism in reasoning (acknowledge competing values) but distributional monism in decisions (near-deterministic across resampling and paraphrases)
Across the model ecosystem, value priorities span the range of inter-physician variation, but specific models systematically underweight patient autonomy
Model decisions on the benchmark reveal stable, committed value preferences rather than case-by-case ethical deliberation — meaning at-scale deployment would propagate one ethical stance to every patient

Technical Novelty: Novel attribution method that recovers latent value priorities (autonomy, beneficence, nonmaleficence, justice) directly from a model's discrete decisions on dilemmas, combined with a distinction between Overton pluralism (in reasoning text) and distributional pluralism (across samples) — prior LLM ethics work largely evaluated surface reasoning rather than decision-level value vectors against a calibrated physician panel.

What's New: First systematic, decision-level audit of pluralism in clinical-ethics reasoning for LLMs, calibrated against a physician panel, and explicitly separating reasoning-surface pluralism from outcome distribution pluralism.

Extension Opportunities:

Build a value-aware ensemble router that selects/blends across multiple LLMs to match a target patient's stated value priorities, restoring distributional pluralism
Develop fine-tuning or RLHF objectives that explicitly induce stochastic pluralism on contested dilemmas while preserving determinism on clear cases
Extend the audit framework to non-Western medical ethics traditions and longitudinal clinical scenarios (e.g., palliative care, pediatric consent) beyond the four-principles canon

Replicability: Abstract does not explicitly mention code/data release. Reproduction would require the clinician-verified dilemma benchmark and physician panel responses (the bottleneck — expert annotation is costly), plus modest API inference compute across frontier models; no training compute needed for the audit itself.

Research Gaps:

No mechanism proposed to actively elicit patient-value-conditioned pluralistic responses from a single deployed model
Scope limited to the Western four-principles framework; cross-cultural and non-principlist medical ethics traditions are unaddressed

3. SkillGenBench: Benchmarking Skill Generation Pipelines for LLM Agents

Authors: Yifan Zhou, Zhentao Zhang, Ziming Cheng... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: Existing benchmarks evaluate whether agents can use provided skills or solve downstream tasks from raw context, but do not isolate skill generation itself as the object of study. How can we systematically benchmark the pipelines that produce reusable, executable skills from repositories and documents?

Summary: SkillGenBench is the first benchmark that treats skill generation — not skill use — as the unit of evaluation for LLM agents. It standardizes inputs, artifacts, and deterministic execution-based checks across a 2x2 matrix of generation regimes (task-conditioned vs task-agnostic) and procedural sources (code repositories vs long-form documents), exposing distinct failure modes in each cell.

Key Results: Introduces SkillGenBench with two generation regimes (task-conditioned and task-agnostic) across two procedural sources (repository-grounded and document-grounded). Experiments across multiple skill-generation methods and backbones show substantial performance variation and reveal distinct failure modes between repo-derived and document-derived skill generation. Specific numbers, dataset sizes, and model scores are not disclosed in the abstract.

Key Findings:

Skill-generation performance varies substantially across methods and backbones, indicating the problem is far from solved.
Task-agnostic reusable skill distillation is markedly harder than task-conditioned generation.
Repository-grounded and document-grounded skill generation exhibit distinct failure modes, suggesting they need different generator designs.

Technical Novelty: Prior work benchmarks skill use or end-to-end task solving from raw context. SkillGenBench is the first to isolate skill generation as an independent unit of evaluation, with standardized artifact formats, pinned environments, fixed execution harnesses, and a 2x2 design separating task-conditioned vs task-agnostic generation and repo-grounded vs document-grounded sources.

What's New: Reframes skill generation as a standalone research problem with its own controlled protocol, rather than entangling it with downstream task-solving evaluation. The 2x2 axis design (regime x source) is a new structuring of the evaluation space.

Extension Opportunities:

Add a multi-modal procedural source (e.g., video tutorials or screencasts) to test whether skill generators can distill procedures from non-textual evidence.
Build a self-improving skill-library agent that uses SkillGenBench's deterministic execution checks as a reward signal to iteratively refine task-agnostic skill distillation.
Extend the harness to measure skill composition — whether generated atomic skills can be chained reliably on multi-step tasks not seen during generation.

Replicability: The abstract describes pinned environments and standardized protocols intended for reproducibility, but does not explicitly state code/data release or compute requirements. Reproducing would require running multiple LLM backbones as generators plus deterministic execution sandboxes — likely moderate API/compute budget but no specialized hardware.

Research Gaps:

No prior benchmark cleanly separates the quality of a generated skill from the agent's ability to use it.
Reusable skill libraries distilled ahead of task knowledge remain an under-evaluated and unsolved capability.

🦾 ROBOTICS

1. Robo-Cortex: A Self-Evolving Embodied Agent via Dual-Grain Cognitive Memory and Autonomous Knowledge Induction

Authors: Nga Teng Chan, Yi Zhang, Yechi Liu... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How can embodied navigation agents overcome 'experiential amnesia' and synthesize generalizable strategies from past interactions in unseen environments, rather than relying on trajectory-driven or reactive policies that fail to transfer?

Summary: Robo-Cortex is a self-evolving embodied navigation framework that combats 'experiential amnesia' by autonomously inducing natural-language navigation heuristics from past trajectories. It pairs a dual-grain memory system (short-term reflective + long-term principle memory) with an Imagine-then-Verify loop using a world model and VLM evaluator, achieving SPL gains up to +15.30% on unseen environments.

Key Results: Robo-Cortex outperforms strong baselines on IGNav, AR, and AEQA benchmarks, with gains of up to +4.16% SPL over the strongest prior method and up to +15.30% SPL under heuristic transfer to unseen environments. Preliminary real-world robotic experiments confirm physical-setting effectiveness.

Key Findings:

Abstracting trajectories into natural-language heuristics enables transfer to unseen environments, yielding up to +15.30% SPL improvement under heuristic transfer
Dual-grain memory (short-term reflective + long-term principle) outperforms purely trajectory-based or reactive policies on IGNav, AR, and AEQA
An Imagine-then-Verify loop combining a world model with a VLM evaluator yields robust action plans and transfers to preliminary real-world deployment

Technical Novelty: The combination of (1) Autonomous Knowledge Induction distilling multimodal trajectories into natural-language heuristics, (2) a Dual-Grain Cognitive Memory split between Short-term Reflective and Long-term Principle Memory, and (3) a multimodal Imagine-then-Verify loop where a world model simulates outcomes and a VLM validates plans — together enabling continuous self-evolution rather than static policy learning.

What's New: Unlike prior trajectory-driven or reactive navigation policies, Robo-Cortex frames experience as inducible, reusable natural-language principles — shifting from passive imitation to active strategy evolution via a continuous reflection-adaptation loop and a structured heuristic library.

Extension Opportunities:

Apply the Autonomous Knowledge Induction (AKI) mechanism to non-navigation embodied tasks like manipulation or multi-agent coordination, building heuristic libraries for object assembly or collaborative tasks
Replace the VLM evaluator in the Imagine-then-Verify loop with a learned reward model fine-tuned on navigation outcomes to reduce inference cost and improve verification reliability
Investigate cross-embodiment heuristic transfer — test whether principles abstracted on one robot platform (e.g., wheeled) generalize to others (e.g., legged or aerial)

Replicability: No code/data availability mentioned in the abstract. Reproduction likely requires significant compute: VLM inference for evaluation, a world model for simulation rollouts, and access to IGNav/AR/AEQA benchmark environments. Real-world experiments would additionally require a physical robot platform.

Research Gaps:

Scalability and quality control of the autonomously induced heuristic library over long horizons (risk of accumulating contradictory or stale heuristics) is not addressed
Limited real-world validation — only 'preliminary' physical experiments, leaving open questions on sim-to-real heuristic robustness and latency of the Imagine-then-Verify loop on embedded hardware

2. Dexora: Open-source VLA for High-DoF Bimanual Dexterity

Authors: Zongzheng Zhang, Jingrui Pang, Zhuo Yang... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How can a single open-source VLA model natively handle dual-arm, dual-hand high-DoF manipulation, given that prior VLAs are limited to dual-gripper control or single-arm dexterous hands?

Summary: Dexora is the first open-source Vision-Language-Action model targeting dual-arm, dual-hand high-DoF manipulation. It combines a hybrid teleoperation pipeline (exoskeleton backpack for arms + Apple Vision Pro for fingers) with a large synthetic + real corpus and a discriminator-weighted diffusion-transformer training recipe to outperform prior VLAs on dexterous benchmarks.

Key Results: Dexora outperforms competitive VLA baselines on both basic and dexterous benchmarks: average dexterous success 66.7% vs. 51.7%, and 90% success on basic tasks. Built on a corpus of 100K simulated trajectories (6.5M frames) plus 10K real teleoperated episodes (2.92M frames). Ablations confirm contributions of real data and the discriminator-based weighting; shows OOD and cross-embodiment generalization.

Key Findings:

Native dual-arm dual-hand VLA achieves 66.7% average dexterous success vs 51.7% for baselines and 90% on basic tasks
Decoupling gross arm kinematics from fine finger motion in teleop yields a usable corpus of 10K real episodes / 2.92M frames alongside 100K sim trajectories
Offline discriminator that down-weights low-quality demo clips is essential — ablations confirm both real data and the discriminator are load-bearing for dexterous performance

Technical Novelty: Three new pieces: (1) first open-source VLA built specifically for dual-arm + dual-hand high-DoF control (not gripper or single-hand), (2) hybrid teleop that decouples gross arm kinematics (exoskeleton backpack) from fine finger motion (markerless Vision Pro tracking) and drives both real platform and a MuJoCo digital twin, and (3) data-quality-aware training where an offline discriminator assigns clip-level weights to down-weight noisy demos during diffusion-transformer policy training.

What's New: Prior VLAs handle either two grippers or a single dexterous hand; Dexora is the first open VLA system to natively support the full dual-arm, dual-hand high-DoF action space, and is paired with a novel decoupled teleop rig and a data-quality-weighted training recipe.

Extension Opportunities:

Replace the offline discriminator with an online RL fine-tuning loop that uses the same quality signal as a dense reward, potentially closing the sim-to-real gap further
Port the hybrid teleop pipeline (exoskeleton backpack + Apple Vision Pro hand tracking) to other dexterous platforms (e.g., Shadow Hand, Allegro) to test embodiment-agnostic scaling
Add tactile/force sensing as an additional input modality and retrain the diffusion-transformer policy to handle contact-rich tasks like deformable object manipulation

Replicability: Explicitly billed as open-source, implying code/data release; reproduction requires a dual-arm dual-hand platform plus exoskeleton + Apple Vision Pro for collection, MuJoCo for the digital twin, and substantial GPU compute to train a diffusion-transformer over ~9.4M frames — likely multi-GPU days-to-weeks scale.

Research Gaps:

No discussion of tactile/force feedback — dexterous tasks involving fine contact dynamics may remain bottlenecked without it
Quality of teleoperation demos is treated as a noise problem to filter rather than improved at the source; better demonstration interfaces could lift the ceiling beyond discriminator weighting

3. REBAR: Reference Ethical Benchmark for Autonomy Readiness

Authors: Jonathan Diller, David Barnes, Rebekah Bogdanoff... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How can we move beyond qualitative ethical AI frameworks and binary safety guardrails to produce objective, computable, repeatable metrics that quantify the ethical and legal readiness of autonomous systems for specific tasks?

Summary: REBAR proposes a quantitative test-and-evaluation framework that converts operational metrics of autonomous systems into a computable Autonomy Readiness Level score. It combines neuro-symbolic LLM reasoning about ethical difficulty, LLM-generated test scenarios, and photorealistic simulation to deliver objective, repeatable ethics benchmarks for white-box autonomy.

Key Results: The abstract does not report concrete numerical results, benchmark scores, or dataset sizes. It introduces the REBAR framework and the Autonomy Readiness Level (ARL) rubric as the contribution, demonstrated qualitatively via a neuro-symbolic LLM pipeline, LLM-driven test generation at scale, and a photorealistic simulation environment for white-box autonomy evaluation.

Key Findings:

Existing ethical embodied-AI evaluation is dominated by qualitative design-time methods (guardrails, red teaming) and lacks computable, comparable metrics.
An Autonomy Readiness Level rubric can map measurable operating metrics to an interpretable ethics score that informs end-user task suitability.
Neuro-symbolic LLMs can both score and explain the ethical difficulty of generated scenarios, enabling accountability and override-friendly transparency.

Technical Novelty: The combination of (1) a neuro-symbolic LLM that both computes and explains scenario ethical difficulty, (2) LLM-driven at-scale generation of ethically-loaded test scenarios, and (3) mapping operational metrics onto a structured ARL rubric — together producing a repeatable quantitative ethics score rather than guardrail pass/fail.

What's New: Unlike prior work that bakes ethics into guardrails or one-off red-team studies, REBAR treats ethical compliance as a measurable, rubric-based readiness level — analogous to TRLs — with LLM-generated test suites and explainable difficulty scoring.

Extension Opportunities:

Apply REBAR to specific domains (autonomous driving stacks like CARLA-integrated planners, surgical robots, or drone swarms) and publish per-domain ARL leaderboards.
Replace the neuro-symbolic ethical-difficulty scorer with a learned reward model trained on human ethicist annotations, then compare calibration against the symbolic baseline.
Extend REBAR to black-box systems by inferring policy behavior via probing-based ARL estimation, removing the white-box requirement.

Replicability: No code, data, or simulator release is mentioned in the abstract. Reproduction would require a photorealistic simulator (Unreal/Unity/CARLA-class), a capable LLM for scenario generation and neuro-symbolic reasoning (likely GPT-4/Claude-tier API access), and GPU compute for white-box autonomy rollouts — moderate-to-high cost.

Research Gaps:

No reported empirical validation, baseline comparisons, or quantitative results in the abstract — unclear how ARL scores correlate with real-world ethical outcomes.
Restriction to white-box autonomy solutions limits applicability to commercial closed systems where ethical accountability is arguably most needed.

💻 COMPUTE

1. LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation

Authors: Yukang Chen, Luozhou Wang, Wei Huang... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How can long video generation be made efficient enough for real-time training and inference, overcoming the speed and memory bottlenecks of current autoregressive diffusion pipelines that rely on multi-stage ODE initialization and distribution matching distillation?

Summary: LongLive-2.0 is an end-to-end NVFP4 parallel infrastructure for long video generation that co-designs sequence-parallel autoregressive training (Balanced SP) with W4A4 quantized inference and asynchronous streaming VAE decoding. It replaces the multi-stage ODE+DMD pipeline used by prior Self-Forcing methods with direct diffusion-to-AR tuning, achieving 2.15x training and 1.84x inference speedups and 45.7 FPS real-time generation at the 5B scale.

Key Results: Demonstrates up to 2.15x training speedup and 1.84x inference speedup using NVFP4 precision. LongLive-2.0-5B achieves 45.7 FPS inference on Blackwell GPUs while maintaining strong benchmark performance. Real-time generation is achieved by reducing denoising steps from 4 to 2 via standalone LoRA weights, with a single-stage tuning pipeline replacing the prior multi-stage ODE+DMD approach.

Key Findings:

Balanced SP enables natural teacher-forcing masks by pairing clean-history with noisy-target chunks per rank, with GEMM acceleration share growing as video length increases
A clean single-stage tuning pipeline can match or exceed the quality of Self-Forcing's ODE+DMD multi-stage approach when paired with high-quality data and infrastructure
NVFP4 KV cache not only reduces memory but also lowers inter-GPU SP communication cost on non-Blackwell architectures

Technical Novelty: First system to integrate NVFP4 W4A4 precision across both training and inference for video generation. Introduces Balanced SP — a co-design pairing clean-history and noisy-target temporal chunks per rank with SP-aware chunked VAE encoding — and eliminates the standard ODE-init + DMD distillation pipeline by directly tuning a diffusion model into a multi-shot interactive AR diffusion model.

What's New: First NVFP4 training+inference system for long video generation; eliminates DMD distillation in favor of direct AR tuning; introduces SP-aware chunked VAE encoding aligned with teacher-forcing layout.

Extension Opportunities:

Apply the Balanced SP teacher-forcing layout to other autoregressive modalities (audio, 3D scene generation) where clean-history/noisy-target chunking could similarly exploit SP
Extend NVFP4 KV-cache quantization techniques to long-context LLM inference, where the same memory and inter-GPU communication bottlenecks apply
Investigate whether the single-stage diffusion-to-AR tuning recipe (skipping ODE init + DMD) generalizes to smaller-scale or non-Blackwell hardware, democratizing long-video training

Replicability: Abstract does not mention code release. Reproduction requires Blackwell-class GPUs (e.g., B200) for native NVFP4 support; non-Blackwell deployment via SP inference is described but still demands significant multi-GPU compute given a 5B-parameter video model.

Research Gaps:

No reported evaluation on perceptual quality trade-offs of NVFP4 quantization versus FP8/BF16 baselines beyond aggregate benchmarks
Reliance on Blackwell-class hardware for full speedups limits accessibility; the non-Blackwell SP fallback path is not quantified in detail

2. EPIC: Abstraction and Polymorphism of In-Network Collectives on Ethernet

Authors: Yitao Yuan, Jianglong Nie, Tianyu Bai... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How can In-Network Collective (INC) acceleration for AI workloads be standardized on open Ethernet despite its cross-layer nature, which has historically blocked vendor investment and ecosystem adoption?

Summary: EPIC proposes a unified Ethernet-compatible protocol specification for In-Network Collective acceleration with polymorphic implementations matched to different hardware capabilities, enabling vendors to adopt INC incrementally. The work combines modular protocol design, formal verification of all polymorphic modes, and a unified resource management model, validated across model checking, simulation, Tofino, and FPGA testbeds.

Key Results: The authors validated EPIC across five complementary methods: formal model checking (proving correctness of all polymorphic modes), packet/flow-level simulations, VM emulation, a Tofino programmable-switch testbed, and FPGA/RTL verification. The paper claims confirmation of correctness, performance gains, and hardware feasibility, though specific speedup percentages, bandwidth numbers, or comparison baselines are not quoted in the abstract.

Key Findings:

A single abstraction can support multiple hardware realizations of INC while preserving correctness, removing a key barrier to open-Ethernet INC adoption
Formal verification (model checking) is tractable for proving correctness across all proposed polymorphic INC modes
The design is feasible in real silicon, demonstrated on both Tofino programmable switches and FPGA/RTL

Technical Novelty: The 'Unified Abstraction, Polymorphic Realization' principle — a single protocol spec with multiple hardware-tier-specific realizations — is the core novelty, combined with aligning functional boundaries to participant roles (host/NIC/switch) rather than fixed layers. Prior INC work (SHARP, ATP, SwitchML) targeted single hardware profiles; EPIC formalizes a modular evolutionary path verified end-to-end.

What's New: Unlike prior INC systems tied to specific proprietary fabrics (e.g., NVIDIA SHARP on InfiniBand) or single switch architectures, EPIC defines a vendor-neutral Ethernet protocol with role-aligned functional boundaries and an evolutionary hardware path, accompanied by formal correctness proofs — a combination not seen in prior INC literature.

Extension Opportunities:

Implement an open-source EPIC-compliant NIC or switch dataplane (P4/FPGA) and benchmark it against NVIDIA SHARP on real distributed training workloads (e.g., Llama-scale all-reduce)
Extend the polymorphic abstraction to cover MoE-style all-to-all collectives and disaggregated inference traffic patterns, which differ from the all-reduce-dominated training case
Build a compiler/runtime layer (e.g., NCCL/Gloo plugin) that auto-selects the appropriate EPIC polymorphic mode based on detected switch/NIC capabilities in heterogeneous clusters

Replicability: No code or artifact link is mentioned in the abstract. Full reproduction would require access to Tofino programmable switches (~$10K+), FPGA boards for RTL verification, and a multi-node GPU cluster — making this a hardware-heavy reproduction. Model-checking and simulation portions are likely tractable on commodity compute.

Research Gaps:

Abstract omits concrete performance numbers (throughput, latency, training-time speedup) versus baselines like SHARP or ring all-reduce, making quantitative comparison impossible
Coverage of non-all-reduce collectives (all-to-all, all-gather for MoE/inference) and behavior under failure/congestion at scale is not addressed

3. iHAC: A Hybrid Cluster Architecture for Enhanced Performance and Resilience

Authors: Siddique Abubakr Muntaka, Edward Danso Ansong, Benjamin Yankson... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How can high-availability clusters overcome single points of failure and inefficient resource allocation inherent to traditional active-active and active-passive configurations?

Summary: The paper proposes iHAC, a hybrid high-availability cluster that integrates active-active and active-passive configurations to address single points of failure and poor resource utilization. OPNET simulations show iHAC reduces HTTP response time by >40% (5s → <3s) versus legacy clusters, alongside lower latency and higher throughput.

Key Results: Simulated in Riverbed Modeler (OPNET), iHAC cut average HTTP page response time by over 40% — from ~5 seconds (traditional active-active) to under 3 seconds — while reducing network latency and increasing overall throughput vs. legacy clusters.

Key Findings:

iHAC reduced average HTTP page response time from ~5s to under 3s — a >40% improvement over a conventional active-active cluster
The hybrid design lowered network latency and raised overall throughput in the same simulation
Combining active-active and active-passive modes mitigates the single-point-of-failure and resource-allocation weaknesses of either approach alone

Technical Novelty: A hybrid topology that fuses active-active load distribution with active-passive standby failover within a single cluster, rather than choosing one paradigm — yielding both better steady-state throughput and resilient failover.

What's New: Prior HA work typically treats active-active and active-passive as alternatives; iHAC unifies them into one integrated architecture, claiming the throughput benefits of A/A with the failover guarantees of A/P.

Extension Opportunities:

Validate iHAC on a real bare-metal/Kubernetes testbed (e.g., Pacemaker/Corosync or kube-vip) rather than OPNET simulation, measuring failover MTTR under chaos-engineered fault injection
Add an ML-driven workload scheduler that dynamically shifts traffic between the active-active and active-passive tiers based on predicted load and node-health telemetry
Extend iHAC to geo-distributed multi-region deployments with quorum-aware consensus (Raft/Paxos) to evaluate split-brain handling and WAN-induced latency tradeoffs

Replicability: No code or dataset is referenced in the abstract; reproduction requires a Riverbed Modeler (OPNET) license and the authors' simulation topology/parameters. Compute needs are modest (single workstation for discrete-event simulation).

Research Gaps:

Evaluation is simulation-only (OPNET) with no real-world deployment, production traffic, or fault-injection validation
No analysis of cost/complexity overhead, consensus behavior, or geo-distributed/split-brain scenarios for the hybrid topology

⚡ ENERGY

1. Defect Control via Cu Enrichment Enhances Multifunctional Properties in the Polar Semiconductor Cu1+xMn1-ySiTe3

Authors: Subrata Ghosh, Yu Liu, Saugata Sarker... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How can the high density of stacking faults that suppress macroscopic ferroelectric polarization in the multiferroic polar semiconductor Cu1-xMn1+ySiTe3 be eliminated to unlock its multifunctional (ferroelectric, magnetic, electronic) properties?

Summary: The paper shows that switching from Cu-deficient to Cu-enriched stoichiometry in the polar chalcogenide semiconductor Cu1+xMn1-ySiTe3 eliminates stacking faults via an emergent interstitial Cu site, while preserving the noncentrosymmetric Pm structure. This defect control unlocks an enhanced SHG response, a new spin-flop transition along the polar axis, and a transition from insulating to doped semiconducting behavior, while retaining antiferromagnetic order at TN ~ 33 K.

Key Results: By inverting the stoichiometry to Cu-enriched Cu1+xMn1-ySiTe3, the authors produced a nearly stacking-fault-free noncentrosymmetric monoclinic (space group Pm) crystal with: (1) pronounced enhancement of second-harmonic generation (SHG) response vs Cu-deficient samples, (2) retained long-range antiferromagnetic order with Néel temperature TN ~ 33 K (without glassy state), (3) a new spin-flop transition along the polar b-axis absent in Cu-deficient compositions, and (4) an electronic ground state shifting from insulating to doped semiconducting behavior.

Key Findings:

Cu enrichment yields nearly stacking-fault-free crystals via an emergent interstitial Cu site while preserving the noncentrosymmetric monoclinic Pm space group
SHG response is markedly enhanced in Cu-enriched vs Cu-deficient samples, indicating restored macroscopic polar order
Long-range AFM order persists at TN ~ 33 K with no glassy state, and a new spin-flop transition emerges along the polar b-axis
Electronic ground state evolves from insulating (Cu-deficient) to doped semiconducting (Cu-enriched)

Technical Novelty: The key insight is that crystal defects (stacking faults) in this multiferroic polar semiconductor originate from Cu non-stoichiometry, and that Cu enrichment (rather than the conventionally studied Cu-deficient regime) stabilizes an interstitial Cu site that suppresses stacking faults while preserving the polar Pm symmetry — flipping the doping direction is itself the novel defect-engineering strategy.

What's New: Prior work on this system focused on the Cu-deficient regime where stacking faults suppress ferroelectric response; this work demonstrates that inverting the non-stoichiometry to Cu-rich compositions is a defect-engineering route that simultaneously enhances polar, magnetic (spin-flop), and electronic functionalities in a single multiferroic platform.

Extension Opportunities:

Systematically map the (x, y) phase diagram with finer Cu/Mn stoichiometry steps to identify a composition that maximizes coupled SHG, ferroelectric polarization, and magnetic ordering simultaneously
Apply pressure, strain, or thin-film epitaxy to engineer the interstitial Cu site and test whether the stacking-fault-free phase persists, enabling device integration
Probe the magnetoelectric coupling directly via polarization-vs-field measurements under applied magnetic field across the spin-flop transition along the polar b-axis to quantify ME coefficients

Replicability: No mention of code or data repositories in the abstract. Reproduction requires crystal growth capability (likely Bridgman or chemical vapor transport for chalcogenides), single-crystal XRD/TEM for stacking-fault characterization, SHG optical setup, SQUID/VSM magnetometry down to ~2 K, and transport measurements — substantial wet-lab materials infrastructure, no significant compute requirement.

Research Gaps:

Direct quantification of ferroelectric hysteresis and magnetoelectric coupling coefficients across the spin-flop transition is not reported in the abstract
Microscopic mechanism by which the interstitial Cu site suppresses stacking-fault nucleation and its energetics is not established

2. Probing Dielectric Screening in van der Waals Heterostructures via Pressure-Tuned Exciton Rydberg Series

Authors: Shalini Badola, Adlen Smiri, Thomas Pelini... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: Can the dielectric properties of hBN encapsulating layers in van der Waals heterostructures be directly measured by exploiting pressure-tuned exciton Rydberg states in monolayer WSe2, addressing the lack of in-situ probes for environmental dielectric screening in 2D semiconductors?

Summary: The paper demonstrates that pressure-tuned exciton Rydberg states in hBN-encapsulated monolayer WSe2 can serve as a direct probe of the dielectric properties of the surrounding hBN. By modeling pressure-induced changes in interlayer spacing together with bulk hBN dielectric behavior, the authors extract the dielectric constant of pressurized hBN, establishing a new in-situ dielectric sensing methodology for van der Waals heterostructures.

Key Results: The authors demonstrate that the exciton Rydberg series in hBN-encapsulated monolayer WSe2 shifts systematically under applied pressure, and by modeling pressure-induced changes in interlayer distances combined with bulk hBN dielectric properties, they extract a direct measurement of the dielectric constant of pressurized hBN. Specific numerical values for pressure ranges and dielectric constants are not provided in the abstract.

Key Findings:

Exciton Rydberg states in monolayer WSe2 are highly sensitive to pressure-induced modifications of the surrounding hBN dielectric environment
A model coupling pressure-dependent interlayer distance with bulk hBN dielectric properties quantitatively reproduces the observed Rydberg-series evolution
The approach yields a direct measurement of the dielectric constant of hBN under pressure, previously inaccessible through standard techniques

Technical Novelty: Prior work used exciton Rydberg states to characterize static dielectric environments, but this paper introduces pressure as a tuning knob and couples it with an explicit model of pressure-induced interlayer distance evolution, allowing extraction of the pressurized hBN dielectric constant — a quantity not directly accessible by conventional means.

What's New: Combines hydrostatic pressure tuning with exciton Rydberg spectroscopy as a quantitative dielectric metrology tool for vdW heterostructures, rather than treating the dielectric environment as a fixed parameter.

Extension Opportunities:

Apply the same Rydberg-series pressure-sensing methodology to other TMD monolayers (MoS2, MoSe2, WS2) encapsulated in alternative dielectrics to build a library of pressure-dependent dielectric constants
Develop a calibrated optical pressure/dielectric sensor device using a packaged hBN/WSe2/hBN stack where exciton peak shifts serve as a readout for local environmental changes
Extend the model to twisted or moiré heterostructures where interlayer distance modulation and dielectric screening interplay non-trivially, enabling probes of moiré-induced dielectric inhomogeneity

Replicability: The abstract does not mention code or data availability. Reproduction requires a diamond anvil cell or pressure setup, hBN-encapsulated WSe2 monolayer samples (fabricated via exfoliation/dry transfer), and low-temperature magneto-optical/reflectance spectroscopy capable of resolving 2s/3s/4s exciton Rydberg states — substantial experimental infrastructure but no heavy compute.

Research Gaps:

No direct method existed for measuring the dielectric constant of hBN under hydrostatic pressure in encapsulated 2D heterostructures
Pressure-induced evolution of interlayer distances and its coupling to nonlocal dielectric screening in TMD/hBN stacks was not quantitatively linked to exciton observables

3. Epitaxial growth and magnetic phase transitions in non-centrosymmetric EuPdSi$_3$ thin films

Authors: Sebastian Kölsch, Alfons G. Schuck, Olena Fedchenko... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: Can EuPdSi$_3$, a non-centrosymmetric magnetic material with potential for hosting chiral spin textures (skyrmions, cycloidal states), be successfully grown as epitaxial thin films, and how do strain/interface effects in this geometry modify its magnetic phase behavior compared to bulk?

Summary: This paper reports the first epitaxial thin films of the non-centrosymmetric magnet EuPdSi$_3$ grown by MBE on MgO(001), confirming tetragonal BaNiSn$_3$ structure and divalent Eu with ~7 $μ_B$ moments. Transport measurements identify two magnetic transitions at 19 K and 15 K and reveal strong field-orientation dependence, establishing the system as a platform for studying chiral spin textures in thin film geometry.

Key Results: First successful MBE growth of epitaxial EuPdSi$_3$ thin films on MgO(001) with confirmed BaNiSn$_3$ tetragonal structure (c-axis out-of-plane, parallel in-plane a-axes via XRD). HAXPES confirms stable Eu$^{2+}$ valence (2.0) yielding ~7 $μ_B$/Eu (Hund's rule). Transport measurements reveal two zero-field magnetic phase transitions at 19 K and 15 K, with distinct field-orientation-dependent phases under H‖[100] and H‖[001].

Key Findings:

Successful epitaxial stabilization of tetragonal EuPdSi$_3$ on MgO(001) with c-axis out-of-plane and parallel in-plane a-axes
Eu valence pinned at 2.0 yielding large ~7 $μ_B$/Eu moment consistent with Hund's rule (S=7/2, L=0)
Two distinct magnetic phase transitions at 19 K and 15 K in zero field, with field-orientation-dependent phase diagrams along [100] vs [001]

Technical Novelty: First demonstration of MBE epitaxial growth of EuPdSi$_3$ in thin film form (prior work limited to bulk crystals), establishing a new thin-film platform where strain and interface effects can be leveraged for chiral magnetism in a divalent Eu (S=7/2, large moment) system.

What's New: Extends EuPdSi$_3$ from bulk single-crystal studies into the epitaxial thin film regime, opening access to strain/interface tuning of a non-centrosymmetric Eu-based magnet — a rare combination of large localized moment, divalent Eu, and broken inversion symmetry suitable for Néel skyrmions.

Extension Opportunities:

Direct imaging of chiral spin textures (skyrmions/cycloids) via Lorentz TEM, MFM, or resonant x-ray scattering to confirm the symmetry-allowed non-collinear states
Systematic strain engineering using different substrates (SrTiO$_3$, LaAlO$_3$, varied buffer layers) to tune the magnetic phase diagram and stabilize specific spin textures
Heterostructure/interface engineering with heavy metals or topological materials to enhance DMI and probe topological Hall effect signatures of skyrmion phases

Replicability: No code/data availability mentioned. Reproduction requires MBE chamber with Eu, Pd, Si sources, MgO(001) substrates, XRD, HAXPES (synchrotron access), and low-temperature (down to <15 K) magnetotransport setup — substantial specialized experimental infrastructure.

Research Gaps:

Direct microscopic confirmation of skyrmion or cycloidal spin textures is absent — only symmetry arguments and bulk transport signatures are presented
No systematic study of film thickness, strain state, or substrate dependence on the magnetic phase diagram

🏥 HEALTHCARE

1. Learning Normal Representations for Blood Biomarkers

Authors: Aashna P. Shah, Michelle M. Li, Yash Lal... Published: 2026-05-18 | Citations: 0 arXiv | PDF

Research Question: How can blood biomarker reference intervals be personalized without overfitting to sparse individual histories, which inflates false positives and may include subclinical disease?

Summary: NORMA is a conditional transformer that generates personalized blood-biomarker reference intervals by anchoring an individual's history to population-level priors on normal variation. Trained on ~2B measurements from 1.6M patients, it outperforms both fixed population intervals and pure personalization in predicting mortality, AKI, and chronic disease, while avoiding the 68% false-abnormal rate of naive personalization.

Key Results: Using ~2 billion longitudinal lab measurements from 1.6M+ individuals across North America, Middle East, and East Asia, the authors show purely personalized intervals classify up to 68% of measurements as abnormal without corresponding adverse outcome associations. Their NORMA conditional transformer achieves higher precision for predicting mortality, acute kidney injury, and chronic disease versus both population-based and pure-personalization baselines.

Key Findings:

Purely personalized reference intervals overfit, flagging up to 68% of measurements abnormal with no link to adverse outcomes
Lab values are highly individual, but anchoring trajectories to population priors beats either personalization or population intervals alone
NORMA intervals yield higher precision for predicting mortality, acute kidney injury, and chronic disease

Technical Novelty: A conditional transformer that jointly conditions on a patient's individual history AND population-level priors over 'normal' variation to generate reference intervals — a hybrid that explicitly regularizes personalization against population structure, rather than choosing one or the other.

What's New: Unlike prior personalization work that uses individual history alone, NORMA explicitly conditions on population-level normal variation as a prior, and operates at unprecedented scale across three continents — providing the first large-scale empirical caution against pure over-personalization in lab medicine.

Extension Opportunities:

Integrate NORMA-derived intervals into EHR clinical decision support and benchmark downstream effects on referral rates and time-to-diagnosis
Extend the conditional transformer to multi-modal inputs (vitals, medications, genomics) to refine 'normal' priors for under-represented subpopulations
Apply the population-anchored personalization paradigm to other longitudinal signals such as wearable physiology, imaging biomarkers, or vital signs

Replicability: Authors publicly release model, code, and an interactive UI. Reproducing training would require access to large multi-region longitudinal lab datasets (proprietary EHR cohorts ~2B measurements) and substantial GPU compute for transformer training; inference and evaluation should be feasible on modest hardware.

Research Gaps:

Limited evaluation across rare conditions and demographic subgroups under-represented in the training cohorts
Unclear how the model handles regimen changes, acute illness, or non-stationary baselines that violate the 'stable individual normal' assumption

🔥 GitHub Trending

1. wanshuiyin/ARIS-in-AI-Offer

⭐ 13 stars | Python

希望大家秋招的时候轻松一点 · Chinese ML/LLM/multimodal/generative-model interview cheat sheets · HTML 排版手机/iPad/电脑随处可读 · auto-generated by ARIS /render-html workflow

ai-interview aris autumn-recruiting cheatsheet chinese claude-code

2. Thomaszhou22/self-refine-skill

⭐ 2 stars | Unknown

🪞 Systematic self-reflection for AI agents. GENERATE → CRITIQUE → REFINE → CHECK. Based on peer-reviewed research. Zero API cost. Works with Claude Code, Cursor, Copilot, Codex CLI, and 10+ more platf

agent-skill ai-agent claude-code copilot cursor llm

3. Yogesh-max2123/NexusOps

⭐ 1 stars | Python

A highly scalable, fault-tolerant distributed Machine Learning pipeline using Celery, Redis, and FastAPI for automated hyperparameter tuning.

celery deep-learning distributed-systems fastapi machine-learning mlops

4. ganesh229999/Life-Expectancy-XGBoost-AI

⭐ 1 stars | Python

Life Expectancy Intelligence | End-to-End ML Pipeline with XGBoost & Tkinter GUI

cross-validation data-science desktop-app feature-engineering healthcare-ai healthcare-analytics

5. marcinwitnik/anonymized-data-Titanic

⭐ 1 stars | Jupyter Notebook

Analiza wpływu anonimizacji i zaszumiania danych Titanic na skuteczność modeli uczenia maszynowego.

classification data data-analysis data-preprocessing data-privacy data-processing

6. ather-techie/rag-interview-questions

⭐ 1 stars | Unknown

A comprehensive interview preparation guide covering all major RAG (Retrieval-Augmented Generation) architectures > 50 questions across 10 types, from Naive RAG to Agentic, Graph, Self-RAG, and beyond

agentic-rag ai deep-learning generative-ai graph-rag interview-preparation

7. mukuldatta/credit-card-fraud-detection

⭐ 1 stars | Jupyter Notebook

End-to-end fraud detection pipeline comparing Isolation Forest vs Random Forest + SMOTE on 284k transactions

anomaly-detection fraud-detection machine-learning python random-forest sckit-learn

8. kishanlad-9/cibil-credit-risk

⭐ 1 stars | HTML

🏦 ML loan default predictor | XGBoost + FastAPI + React | 0.80 AUC

cibil credit-risk fastapi india machine-learning python

9. nameisashish/multimodal-disease-prediction

⭐ 1 stars | Jupyter Notebook

A Multimodal Deep Learning Framework for Symptom-Based Disease Prediction and Clinical Decision Support | Neural Computing and Applications (Springer)

attention-mechanism clinical-decision-support deep-learning disease-prediction keras medical-chatbot

10. Aldenysq/agents-connector

⭐ 1 stars | Rust

Let Claude Code, Codex & Gemini CLI agents talk to each other — a local MCP message bus for multi-agent collaboration, cross-model code review, and fewer hallucinations.

a2a agent-communication agent-orchestration ai ai-agents ai-coding-agents

11. mukuldatta/AutoResearch-Agent

⭐ 1 stars | Python

Multi-agent research pipeline using CrewAI + Groq LLaMA 3.3 70B — autonomously breaks down complex topics and generates structured, cited research reports

crewai docker fastapi groq llama3 llm

12. neylinsomne/Call_Center_AI

⭐ 1 stars | Python

End-to-end self-hosted AI call center: real-time voice STT→LLM→TTS pipeline, multi-tenant agents, voice cloning, RAG, multi-channel (calls/WhatsApp/IG).

ai-call-center asterisk conversational-ai knowledge-destilation kokoro-tts llm

13. zhao1bo-source/harness-engineering

⭐ 1 stars | Unknown

Context Engineering, Agent Architecture & Human-AI Collaboration Systems — notes from real experiments.

agents ai ai-native-systems context-engineering harness llm

14. fushenguang/maestro

⭐ 1 stars | Unknown

AI-powered issue-to-PR automation. Maestro orchestrates Claude agents to autonomously resolve Linear issues — built with TypeScript, Tauri & Supabase.

ai-agent automation claude code-automation developer-tools linear

15. Adonyth/omytea-personal-console

⭐ 1 stars | Python

Probability-calibrated decision support for personal decisions — Streamlit app on the Omytea quantum substrate

apache-2-0 calibration decision-support lindblad llm llm-app

Generated by Research Pulse on 2026-05-19 06:09