2026-07-19·Daily digest

🔬 Research Pulse

Daily Digest

July 19, 2026

🤖 AI

🧠 LLMs

1. RoboTTT: Context Scaling for Robot Policies

Authors: Yunfan Jiang, Yevgen Chebotar, Ruijie Zheng... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How can robot foundation models scale visuomotor context length far beyond single-step or short-history windows without incurring prohibitive inference latency, and does longer context unlock qualitatively new capabilities?

Summary: RoboTTT extends robot foundation models by integrating Test-Time Training as a recurrent fast-weight state, scaling visuomotor context to 8K timesteps without added inference latency. This unlocks one-shot in-context imitation from human videos, on-the-fly improvement, and completion of a 5-minute 10-stage assembly task, while showing pretraining context length is a new scaling axis (+62% at 8K vs 1K).

Key Results: RoboTTT scales visuomotor context to 8K timesteps (3 orders of magnitude beyond SOTA) while keeping inference latency flat. On real-robot manipulation, it improves overall performance by 87% over a single-step context baseline and fully completes a 5-minute, 10-stage assembly task that no baseline completes. Pretraining with 8K-timestep context outperforms the same model pretrained at 1K timesteps by 62%, evidencing context length as a scaling axis. It also demonstrates one-shot in-context imitation from human video demos, on-the-fly policy improvement, and perturbation robustness.

Key Findings:

Fast-weight TTT recurrence enables 8K-timestep context with constant inference latency — 1000x longer than SOTA visuomotor policies.
Longer pretraining context produces steady closed-loop gains (+62% from 1K→8K), establishing context length as a scaling axis for robot policies.
Long context yields emergent behaviors: one-shot imitation from human video, on-the-fly policy improvement, perturbation robustness, and full completion of a 10-stage long-horizon assembly task.

Technical Novelty: Integrating Test-Time Training (fast weights updated by gradient descent during both training and inference) as the recurrent state of a VLA policy, so history is compressed into weight space rather than a growing KV cache. The training recipe pairs sequence action forcing with truncated backpropagation through time to make 8K-timestep contexts tractable.

What's New: First application of Test-Time Training as the recurrent mechanism inside a VLA policy, combined with sequence action forcing + truncated BPTT to train at 8K-timestep contexts — three orders of magnitude beyond prior visuomotor context windows.

Extension Opportunities:

Apply the TTT fast-weights recurrence to open-source VLA policies (e.g., OpenVLA, Pi-0) and measure whether the 62% context-scaling gain transfers to mobile manipulation or dexterous hands.
Combine RoboTTT's fast-weight state with retrieval over an external memory of prior episodes to enable multi-episode lifelong adaptation rather than single-episode context.
Investigate sequence action forcing + truncated BPTT as a general recipe for other embodied long-horizon domains (autonomous driving, surgical robotics, AR/VR agents) where 8K-step context matters.

Replicability: Project page at research.nvidia.com/labs/gear/robottt hosts videos; no code release is confirmed in the abstract. Reproduction would require a real robot platform for 10-stage assembly, VLA-scale pretraining compute (likely multi-node GPU clusters given 8K-timestep BPTT), and large teleoperation/human-video datasets — NVIDIA-lab scale.

Research Gaps:

No public code or benchmark release confirmed, limiting external validation on non-NVIDIA hardware.
Unclear how the fast-weight state degrades or forgets across task boundaries, or across multi-episode deployment beyond a single 8K window.

2. SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

Authors: Yuyao Zhang, Junjie Gao, Zhengxian Wu... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How can multi-agent web-search systems avoid getting trapped in repetitive loops and losing track of task progress as interaction histories grow, so they can complete open-domain information-seeking tasks robustly within a search budget?

Summary: SearchOS is a multi-agent framework that converts fragile implicit search state into explicit, shared artifacts (evidence graph, coverage map, failure memory) via Search-Oriented Context Management, and uses a middleware harness plus pipeline-parallel scheduling to keep agents productive within a search budget. It reports state-of-the-art results across all metrics on WideSearch and GISA versus single- and multi-agent baselines.

Key Results: SearchOS leads all metrics against evaluated single- and multi-agent baselines on two open-domain information-seeking benchmarks: WideSearch and GISA. The abstract does not disclose specific numeric deltas, ablation numbers, or per-metric scores.

Key Findings:

Framing open-domain information seeking as relational schema completion with grounded citations gives agents a concrete, measurable target that exposes coverage gaps.
Externalizing progress into an Evidence Graph, Coverage Map, and Failure Memory measurably reduces repetitive loops and search-budget waste compared to implicit-state baselines.
Pipeline-parallel scheduling that continuously refills freed slots with unresolved-gap tasks improves throughput while a middleware harness enforces reusable strategy/access skills across runs.

Technical Novelty: Rather than relying on implicit chain-of-thought state, SearchOS externalizes search progress into four explicit shared artifacts (Frontier Task, Evidence Graph, Coverage Map, Failure Memory), combined with pipeline-parallel sub-agent scheduling that refills freed slots with coverage-gap tasks, and a Middleware Harness that intercepts model/tool calls to enforce a hierarchical skill library preventing repeated failure patterns.

What's New: Prior tool-integrated agents keep search progress in transient chat context; SearchOS is a system-level design that treats progress tracking, failure memory, and skill reuse as first-class persistent state managed outside the LLM, coupled with an OS-style scheduler for parallel sub-agents.

Extension Opportunities:

Extend the Evidence Graph + Coverage Map into a persistent cross-session knowledge store so repeated queries reuse grounded evidence and Failure Memory across users/runs.
Adapt the Search Tool Middleware Harness as a drop-in wrapper for existing agent frameworks (LangGraph, CrewAI, OpenAI Agents SDK) to add stall detection and skill-based fallback to arbitrary tool-using agents.
Replace the relational-schema-completion formulation with a domain-specific schema (e.g., clinical trials, SEC filings, competitive intel) and evaluate whether SOCM's coverage-gap scheduling generalizes to structured enterprise search.

Replicability: The abstract does not mention a code release, dataset artifacts, or model/compute specs. Reproduction would require a capable tool-using LLM backbone plus paid web-search API budgets across WideSearch and GISA — feasible for a small research team but not turnkey without an open release.

Research Gaps:

No reported numbers, ablations, or cost/latency comparisons in the abstract, making the magnitude of gains and the contribution of each SOCM component unclear.
Unclear how the approach generalizes beyond benchmark schemas to genuinely open-ended research tasks where the target schema itself must be discovered incrementally.

3. In-Place Tokenizer Expansion for Pre-trained LLMs

Authors: Jimmy T. H. Smith, Tarek Dakhran, Alberto Cabrera... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How can a pre-trained LLM's tokenizer be expanded in-place to reduce token fragmentation for newly-prioritized languages, without retraining from scratch or paying the decode-bandwidth penalty of a broad vocabulary on compact on-device models?

Summary: The paper introduces a recipe for expanding a pre-trained LLM's tokenizer in-place by continuing existing BPE merges on multilingual data, initializing new embeddings as means of their source sub-token embeddings, and applying two-stage adaptation. Applied to LFM2-8B-A1B, it produces LFM2.5-8B-A1B with a 128K tokenizer that cuts Hindi/Vietnamese/Thai token counts by 2.4-4.0x and delivers 2.2-3.7x per-character decode speedups on-device.

Key Results: Applied to LFM2-8B-A1B (8B MoE) to produce LFM2.5-8B-A1B with a 128K tokenizer. Expanded tokenizer encodes Hindi ~2.4x, Vietnamese ~2.6x, and Thai up to 4.0x fewer tokens than source. Estimated per-character decode speedups of 2.2-3.7x across reference devices after accounting for larger-vocabulary per-token cost. Two-stage adaptation (embedding-only then full-model continued pre-training) recovers source-checkpoint quality.

Key Findings:

Continuing BPE merges (rather than starting fresh) lets most source tokens carry over unchanged, preserving learned representations
Mean-of-source-sub-token initialization plus embedding-only training before full-model continued pre-training recovers source-checkpoint quality
Fragmentation reduction (up to 4.0x on Thai) translates to real 2.2-3.7x per-character decode speedups on compact on-device models after accounting for larger embedding/LM-head bandwidth cost

Technical Novelty: In-place tokenizer expansion that continues (rather than replaces) existing BPE merges on a multilingual corpus, guaranteeing every new token has an exact decomposition into source tokens. Combined with mean-of-source-sub-token initialization and a two-stage embedding-only-then-full-model adaptation schedule. Prior work typically either retrains tokenizers from scratch (losing checkpoint) or grafts unrelated vocabularies (requiring extensive realignment).

What's New: Frames tokenizer expansion as an in-place upgrade path for compact on-device models where embedding/LM-head bandwidth is a material decode cost — a regime where cloud-scale broad-vocabulary solutions don't transfer. The BPE-continuation strategy guaranteeing exact decomposition into source tokens is the key mechanism enabling clean embedding initialization and stable adaptation.

Extension Opportunities:

Apply the recipe to other open compact models (Llama 3.2 1B/3B, Phi, Gemma Nano) targeting different underrepresented language sets and benchmark on-device decode latency
Explore alternative new-embedding initializations beyond mean-of-sub-tokens (e.g., learned attention over source sub-tokens, frequency-weighted averages) and compare convergence speed in stage-1 embedding-only training
Extend the in-place expansion approach to domain-specific vocabularies (code, math symbols, biomedical terms) rather than just languages, and quantify fragmentation reduction vs. decode overhead trade-offs

Replicability: Model weights and expanded tokenizer are released. Reproduction requires access to an 8B MoE base checkpoint plus multilingual continued pre-training compute — likely tens to hundreds of GPU-days for the two-stage adaptation. The recipe itself is straightforward to port to smaller models with more modest compute.

Research Gaps:

Quality/perplexity impact on original languages after expansion is not quantified in the abstract — the trade-off between new-language gains and potential regression on source languages remains open
The recipe assumes the model producer controls tokenizer design; adapting closed-weight or third-party checkpoints is not addressed

🦾 ROBOTICS

1. Scaling Behavior Foundation Model for Humanoid Robots

Authors: Weishuai Zeng, Kangning Yin, Xiaojie Niu... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How should learning paradigm, behavioral data, and model architecture be coordinated to effectively scale Behavior Foundation Models (BFMs) for humanoid whole-body control?

Summary: The paper presents a scaling recipe for Behavior Foundation Models in humanoid control, coordinating three components: global-frame motion tracking as a unified learning paradigm, balanced scaling of on-policy rollouts with reference motion diversity, and a Humanoid Transformer architecture. The approach achieves >10% and >82% MPKPE reductions in local and global modes respectively, demonstrating that principled scaling yields substantial gains for generalist humanoid control.

Key Results: Demonstrated a coordinated scaling recipe achieving >10% MPKPE reduction in local mode and >82% MPKPE reduction in global mode on test sets vs existing humanoid controllers, validated in both simulation and real-world deployment.

Key Findings:

Motion tracking in the global frame is an effective unifying paradigm that reformulates diverse humanoid control tasks as whole-body behavior reproduction
There is a strategic synergy between on-policy rollout quantity and reference motion diversity — both must scale together for effective BFM training
The Humanoid Transformer architecture enables the natural emergence of structured behavioral representations and scales with data/compute
Global-mode MPKPE improvements (82%) substantially exceed local-mode gains (10%), showing the global-frame formulation especially benefits absolute pose fidelity

Technical Novelty: Reformulating diverse humanoid control problems as global-frame whole-body motion tracking, combined with an explicit synergy between on-policy rollout quantity and reference motion diversity, plus a scalable Humanoid Transformer architecture that yields emergent structured behavioral representations.

What's New: Unlike prior BFMs that mix disparate control paradigms or scale a single axis (data or model), this work identifies the coordination of learning paradigm + data diversity + architecture as the true scaling lever, and introduces a transformer tailored for humanoid morphology with global-frame tracking objectives.

Extension Opportunities:

Extend the Humanoid Transformer to multi-agent humanoid coordination or human-robot interaction scenarios using the global-frame motion tracking paradigm
Investigate data-efficient fine-tuning recipes (LoRA-style adapters) for deploying the BFM to novel humanoid embodiments with different DoF configurations
Combine the BFM with high-level language/vision planners to create a hierarchical generalist embodied agent for long-horizon manipulation and locomotion tasks

Replicability: Abstract does not mention code/data release. Reproducing would likely require large-scale RL infrastructure (GPU cluster for on-policy rollouts across diverse reference motions), a humanoid simulator (e.g., IsaacGym/MuJoCo), a motion capture dataset, and a physical humanoid platform for real-world validation.

Research Gaps:

Prior BFM work scaled individual components without a coordinated recipe, leaving unclear which factors drive generalization
Existing humanoid controllers underperform in global-frame pose fidelity, a gap this paper closes but doesn't fully address for extreme out-of-distribution motions or contact-rich manipulation

2. CosFly-VLA: A Spatially Aware Vision-Language-Action Model for UAV Tracking

Authors: Ruilong Ren, Songsheng Cheng, Yunpeng Zhou... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How can VLA policies for UAV target tracking maintain robust performance during sustained occlusions in urban environments, where existing models lose target state and cascade errors until re-acquisition fails?

Summary: CosFly-VLA is a spatially aware Vision-Language-Action model for UAV target tracking that jointly grounds targets, estimates visibility, and generates flight actions to survive sustained occlusions. It combines spatially grounded continued pretraining, curriculum SFT, chain-of-thought recovery reasoning, and closed-loop RL to substantially outperform OpenVLA on displacement error and success rate.

Key Results: CosFly-VLA-0.8B reduces open-loop Average Displacement Error (ADE) vs OpenVLA by 34.1% on seen-test and 35.3% on unseen-test. Closed-loop RL optimization improves Success Rate (SR) by 29.8% (seen) and 2.5% (unseen). Trained via Spatially Grounded CPT on a 500k mixed pool, three-stage curriculum SFT, CoT training, and closed-loop RL with a multi-component reward.

Key Findings:

Structured joint prediction of grounding + visibility + action reduces cascading errors during occlusion vs. action-only VLA baselines
Injecting UAV-view depth and 3D spatial reasoning via 500k-sample continued pretraining materially improves downstream tracking accuracy
Closed-loop RL yields large SR gains on seen distributions (+29.8%) but marginal gains on unseen (+2.5%), suggesting generalization remains bottlenecked by SFT coverage

Technical Novelty: The joint structured prediction interface that couples target grounding, visibility estimation, and continuous action generation — combined with a four-stage training recipe (spatially grounded CPT → curriculum SFT → CoT recovery reasoning → closed-loop RL with stand-off/grounding/collision/success rewards) specifically engineered for occlusion recovery.

What's New: First VLA tracking policy to explicitly model target visibility as a first-class prediction alongside grounding and action, paired with an occlusion-focused curriculum and CoT recovery traces — a departure from prior visible-frame imitation approaches like OpenVLA.

Extension Opportunities:

Extend the visibility estimation head to handle multi-target tracking and re-identification after longer occlusions using memory-augmented state representations
Integrate real-world sim-to-real transfer with domain randomization on lighting, weather, and sensor noise to move beyond simulated urban benchmarks
Add multi-modal sensor fusion (LiDAR, thermal, event cameras) to the spatial grounding pretraining pool for occlusion-robust tracking in low-visibility conditions

Replicability: Abstract does not mention code/data release. Reproduction would require large-scale UAV-view video with depth/distance annotations (500k pretraining pool), a UAV simulator supporting closed-loop RL, and multi-GPU compute for training a 0.8B VLA model plus RL fine-tuning — likely tens of thousands of GPU-hours.

Research Gaps:

Small closed-loop SR improvement on unseen environments (2.5%) indicates weak generalization of RL-learned recovery behaviors
Evaluation uses a shared oracle state history rather than fully autonomous perception, leaving real-world deployment robustness unproven

💻 COMPUTE

1. Don't Predict, Prioritize: Rethinking GPU Reliability Assessment

Authors: Difeng Ma, Changhua Pei, Yuanwei Lu... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: Can GPU failure timing be accurately predicted in production clusters, and if not, what is a more robust alternative for proactive failure management in large-scale AI training infrastructure?

Summary: The paper argues that predicting exact GPU failure timing is fundamentally hard due to stochastic, low-SNR telemetry, and proposes HeaRank, a Learning-to-Rank framework that instead ranks nodes by relative failure risk. Deployed on a production cluster, it captures 64% of failures in the top 5% of ranked nodes versus 21% for the incumbent system.

Key Results: Analysis of production telemetry shows Double Bit Errors (DBEs) and GPU Lost events exhibit strong stochasticity and low signal-to-noise ratios, making time-based prediction ineffective. HeaRank achieves AUC of 0.83 on a production cluster with thousands of GPUs, outperforming heuristic baselines and SOTA ranking algorithms. In online deployment, it captures 64% of future failures within the top 5% of ranked nodes, versus 21% for the incumbent production system — a ~3x improvement.

Key Findings:

Major GPU failures (DBEs, GPU Lost) show strong stochasticity and low signal-to-noise ratios in telemetry, defeating conventional time-series prediction
Historical failure patterns yield stable relative risk rankings even when absolute timing is unpredictable
HeaRank achieves 0.83 AUC and 3x improvement in top-5% failure capture over the production incumbent

Technical Novelty: Reframes GPU reliability from an absolute time-series prediction problem to a Learning-to-Rank (LTR) problem, exploiting the observation that relative failure risk across nodes is more stable and learnable than absolute failure timing. Prior work applies predictive maintenance techniques that assume learnable temporal signals, which the authors empirically show is invalid for major GPU failure modes.

What's New: Shifts the framing of GPU reliability from predictive maintenance to risk-based prioritization using LTR — a paradigm change validated with production deployment metrics rather than benchmark-only results.

Extension Opportunities:

Integrate HeaRank rankings directly into cluster schedulers (e.g., Kubernetes, Slurm, Ray) to automatically avoid placing synchronous training workloads on high-risk nodes
Extend the LTR framework to other accelerator classes (TPUs, AMD MI300, custom ASICs) and heterogeneous clusters, testing whether stable failure patterns generalize across hardware vendors
Combine relative ranking with checkpointing policies — dynamically adjust checkpoint frequency per node based on its risk rank to minimize wasted compute during synchronous training

Replicability: No explicit mention of open-source code or public dataset release in the abstract. Reproduction would require access to a production-scale GPU cluster (thousands of GPUs) with historical telemetry and failure labels — likely infeasible outside hyperscalers. LTR training itself is lightweight (CPU-scale).

Research Gaps:

No integration with scheduler-level actions (workload placement, preemption, checkpoint scheduling) is evaluated
Generalization across hardware vendors, cluster scales, and workload types (inference vs training) is untested

2. Dynamic Entanglement Distribution for Multi-User and Multi-Protocol Quantum Networking

Authors: Rui Wang, Marcus J. Clark, Obada Alia... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How can entanglement-based quantum networks be made dynamically reconfigurable to support multiple users, protocols, and service requirements over deployed metropolitan fibre infrastructure, rather than being locked into static point-to-point or fixed-topology configurations?

Summary: The paper demonstrates a metropolitan-scale, six-user entanglement-based quantum network built around a quantum reconfigurable optical add-drop multiplexer (q-ROADM) that dynamically distributes polarisation-entangled photon pairs and supports programmable full-mesh, partial-mesh, and sliced topologies. Beyond stable 150+ hour operation, the same infrastructure is shown to host multiple quantum protocols including a Secure Inaugural Authentication-Transfer scheme combined with network flooding, establishing a route toward service-oriented, reconfigurable quantum networking.

Key Results: Demonstrated a metropolitan-scale quantum network serving 6 users via a quantum reconfigurable optical add-drop multiplexer (q-ROADM) distributing polarisation-entangled photon pairs from a broadband source over deployed campus/metro fibre. Achieved stable 6-user full-mesh operation for >150 hours, compared full-mesh vs. time-shared partial-mesh strategies under varying source/detector conditions, realised quantum network slicing with optional interconnection links, and showcased protocol flexibility via Secure Inaugural Authentication-Transfer (SIAT) combined with network flooding over multiple paths for secure user onboarding.

Key Findings:

A q-ROADM can dynamically route polarisation-entangled photon pairs from one broadband source to 6 users over deployed campus and metro fibre while preserving entanglement quality
Stable full-mesh six-user operation was sustained for over 150 hours, and time-shared partial-mesh strategies offer tradeoffs versus full-mesh depending on source and detector conditions
The same physical infrastructure supports quantum network slicing and heterogeneous protocols, including SIAT combined with multi-path network flooding for secure onboarding of new users

Technical Novelty: The q-ROADM itself — bringing classical ROADM reconfigurability into the quantum domain to support programmable full-mesh, partial-mesh, and sliced sub-network topologies from a single broadband entangled source. Prior entanglement-distribution networks were largely static wavelength-assignment schemes; this work adds runtime reconfigurability and demonstrates protocol multiplexing (QKD + SIAT + flooding) on the same physical infrastructure.

What's New: Moves entanglement distribution from static wavelength-multiplexed topologies to a runtime-reconfigurable, service-oriented architecture, and shows heterogeneous quantum protocols (entanglement-based QKD, SIAT, flooding) coexisting on the same deployed fibre plant via a single programmable optical element.

Extension Opportunities:

Scale the q-ROADM architecture beyond 6 users by cascading multiplexers or increasing wavelength channel count, and characterise how key rates degrade with user count
Integrate quantum memory nodes to enable entanglement swapping and extend range beyond metropolitan scale toward true quantum repeater networks
Implement an SDN-style orchestration layer on top of the q-ROADM to enable automated slice provisioning, QoS-aware routing, and dynamic protocol switching (QKD/SIAT/blind quantum computing) based on real-time link telemetry

Replicability: No mention of open code or data release in the abstract. Reproduction would require substantial specialised hardware: a broadband polarisation-entangled photon-pair source, a wavelength-selective switch or programmable filter functioning as the q-ROADM, single-photon detectors (likely SNSPDs), polarisation-analysis modules per user, timing/coincidence electronics, and access to deployed metropolitan fibre — realistically only reproducible by well-funded quantum networking labs.

Research Gaps:

Scalability beyond 6 users and beyond metropolitan distances (no quantum-repeater integration) is not addressed
Automated, policy-driven orchestration and QoS guarantees for slice/protocol allocation remain manual or preconfigured rather than dynamic in response to real-time demand

3. Coulomb blockade in microscopic material defects as a source of decoherence and noise in solid-state quantum circuits

Authors: R. Banerjee, L. P. Lindoy, M. Hegedus... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How can we identify and attribute the microscopic material origins of decoherence in superconducting quantum circuits, particularly mechanisms beyond the widely-assumed two-level system (TLS) defects?

Summary: The authors use scanning gate microscopy on operating superconducting circuits to reveal a new decoherence mechanism: microwave-driven charge tunnelling through Coulomb-blockaded metallic grains introduced during standard thin-film fabrication. These grain defects are shown to be as prevalent and damaging as TLS defects, challenging the dominant TLS-limited coherence paradigm and pointing to grain elimination as a practical fabrication-level fix.

Key Results: Using scanning gate microscopy on live superconducting circuits, the authors identified and characterized a previously unrecognized decoherence mechanism: Coulomb blockade and microwave-driven charge tunnelling in metallic grains. They demonstrated across multiple defects and devices that these grain-based defects are as common and as debilitating to device performance as TLS defects, but arise from a fundamentally different physical mechanism. They also showed that conventional characterisation techniques would misattribute this loss to microwave power-independent processes.

Key Findings:

Metallic grains ubiquitous in lithographically fabricated thin-film devices act as Coulomb-blockade sites that cause microwave-driven charge tunnelling and significant decoherence
Grain-induced defects are comparable in frequency and severity to TLS defects across multiple devices, yet originate from a fundamentally different physical mechanism
Standard characterisation techniques misattribute this loss to microwave-power-independent processes, meaning much of the community's TLS-attributed loss budget may actually be grain-induced

Technical Novelty: The use of scanning gate microscopy on live (operating) superconducting circuits to directly localize and characterize individual defects, combined with the identification of Coulomb-blockade-mediated charge tunnelling in metallic grains as a distinct decoherence channel — previously conflated with TLS or other loss mechanisms.

What's New: First direct identification (via SGM on live circuits) of Coulomb blockade in metallic grains as a widespread, distinct decoherence source in superconducting qubits, breaking from the prevailing TLS-centric explanation of coherence limits.

Extension Opportunities:

Develop fabrication protocols (e.g., alternative lithography, etch chemistries, or annealing steps) that eliminate stray metallic grains and benchmark coherence gains against baseline devices
Build a scanning-gate-microscopy-based screening pipeline for wafer-level qubit QA that flags grain-induced Coulomb-blockade defects before packaging
Extend the theoretical model of microwave-driven charge tunnelling to predict frequency-, temperature-, and power-dependent loss signatures, enabling non-invasive identification of grain defects vs. TLS defects from standard qubit spectroscopy

Replicability: The abstract does not mention public code or datasets. Reproduction requires a cryogenic scanning gate microscopy setup integrated with live superconducting qubit measurement (dilution refrigerator, microwave electronics, cryogenic SGM tip control) — a specialized experimental infrastructure available to only a handful of labs worldwide.

Research Gaps:

No quantitative fabrication recipe is provided for eliminating metallic grains without compromising other device parameters
The relative contribution of grain-induced vs. TLS-induced loss across different superconducting qubit architectures (transmon, fluxonium, cat qubits, etc.) and materials (Al, Nb, Ta) remains uncharacterized

⚡ ENERGY

1. High-Q superconducting microwave resonators using MBE titanium nitride

Authors: Anand Ithepalli, Haoran Lu, Eegene Clara Chung... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How can we grow ultra-crystalline superconducting TiN thin films with minimal defects and dielectric losses to enable high-coherence superconducting qubit platforms and epitaxial Josephson junctions?

Summary: The authors demonstrate MBE-grown epitaxial TiN on c-plane sapphire with record-low structural disorder (18 arcsec XRC FWHM) and use it to fabricate CPW resonators with internal quality factors exceeding 10^6 in the single-photon limit and 2×10^7 at high power. They also identify sub-surface sapphire defects as a previously unrecognized source of structural imperfection in epitaxial superconducting films.

Key Results: MBE-grown (111) TiN on c-plane sapphire achieved the lowest reported X-ray rocking curve FWHM of 18 arcsec. Quarter-wavelength CPW hanger resonators (3/6/3 μm gap/strip/gap) demonstrated internal quality factor Qi > 10^6 at single-photon regime (⟨n⟩ ~ 1) at 5.8 GHz and 10 mK, rising to > 2×10^7 at high power (⟨n⟩ ~ 10^6). The work also identified sub-surface sapphire defects that nucleate structural defects in the epitaxial TiN.

Key Findings:

MBE TiN/sapphire exhibits an 18 arcsec X-ray rocking curve — the lowest reported for this system
Single-photon Qi > 10^6 at 5.8 GHz, 10 mK, scaling to > 2×10^7 at ⟨n⟩ ~ 10^6, competitive with best Ta and Nb resonators
Sub-surface defects in sapphire substrates nucleate structural defects in the epitaxial TiN, revealing a substrate-quality bottleneck

Technical Novelty: First MBE (rather than reactive sputtering or ALD) growth of TiN on sapphire achieving record 18 arcsec rocking curve width, combined with first-time identification of sub-surface sapphire defects as the dominant nucleation source for epitaxial TiN structural defects — linking substrate quality directly to superconducting film performance.

What's New: Prior high-Q TiN resonators used reactive sputtering or ALD; this is the first MBE demonstration reaching competitive Qi values while providing atomically abrupt crystalline interfaces suitable for future epitaxial Josephson junctions — a route to overcoming the amorphous AlOx barrier loss ceiling in transmon qubits.

Extension Opportunities:

Develop epitaxial Josephson junctions with fully crystalline tunnel barriers (e.g., MgO or AlN) grown in-situ on this MBE TiN base, targeting T1 > 1 ms transmon qubits
Investigate sapphire substrate pre-treatment or alternative substrates (e.g., MgO, Si with buffer layers) to eliminate the sub-surface defects that nucleate TiN dislocations
Fabricate 3D integrated qubit arrays and TWPAs using this low-loss MBE TiN and benchmark against sputtered TiN and Ta baselines at scale

Replicability: No code/data availability is mentioned in the abstract. Reproduction requires an MBE system with Ti and active nitrogen sources, c-plane sapphire substrates, a cleanroom for CPW patterning, and a dilution refrigerator (~~10 mK) with low-noise microwave characterization — significant capital investment (~~$1M+ facilities).

Research Gaps:

Sub-surface sapphire defects remain uncharacterized in origin and unmitigated — no substrate preparation protocol yet exists to eliminate them
Epitaxial Josephson junctions with crystalline tunnel barriers on this platform are proposed but not yet demonstrated

2. Growth-controlled suppression of electrically active defects in CrSBr

Authors: Sara R. Tulchinsky, Sergii Grytsiuk, Shen van Hassel... Published: 2026-07-16 | Citations: 0 arXiv | PDF

Research Question: How can defect populations in CrSBr crystals be systematically controlled during chemical vapor transport (CVT) growth, and what is the atomic origin of the dominant electrically active defect (D*)?

Summary: The paper establishes practical CVT growth recipes for reducing electrically active defects in the 2D magnetic semiconductor CrSBr, achieving up to an order-of-magnitude suppression of a dominant defect (D*) by tuning precursor stoichiometry and growth temperature. DFT and thermodynamic modeling identify D* as a sulfur-related vacancy complex, giving both an atomic-scale mechanism and actionable growth levers for producing higher-quality crystals.

Key Results: Using conductive atomic force microscopy (CAFM), the authors measured the density of a characteristic defect (D*) and demonstrated up to a 10x reduction in its concentration via optimized CVT growth. Two tunable levers were identified: (1) sulfur- and bromine-rich precursor stoichiometry, and (2) lower absolute growth temperatures at fixed temperature gradient. Combined thermodynamic modeling + DFT calculations attribute D* to a sulfur-related vacancy complex rather than an isolated point defect.

Key Findings:

Sulfur- and bromine-rich precursor conditions suppress D* defect formation during CVT growth.
Lowering absolute growth temperature (while preserving the temperature gradient) further reduces D* density, achieving up to 10x suppression overall.
D* is most consistent with a sulfur-related vacancy complex, not an isolated point defect, per combined DFT + thermodynamic modeling.

Technical Novelty: First systematic correlation between CVT growth parameters (precursor stoichiometry, absolute temperature at fixed gradient) and a specific electronically active defect signature in CrSBr, combined with a mechanistic identification of D* as a sulfur-vacancy complex via joint thermodynamic + DFT modeling — moving beyond prior work that only characterized as-grown defects without control strategies.

What's New: Prior CrSBr work characterized native defects but did not connect specific CVT growth parameters to defect concentrations or resolve atomic origin of the electrically active species. This paper closes both gaps simultaneously, providing a growth-condition → defect-density → atomic-configuration chain.

Extension Opportunities:

Apply the same CAFM + DFT + thermodynamic modeling workflow to other CVT-grown 2D magnets (CrPS3, NiPS3, FePS3) to map their defect landscapes and growth-condition sensitivities.
Correlate the reduced D* densities with magneto-transport and magnon-photon coupling measurements to quantify how defect suppression improves spintronic and optoelectronic device metrics.
Develop an in-situ growth-monitoring or closed-loop CVT reactor that uses real-time stoichiometry feedback to autonomously target sub-D* defect densities in wafer-scale CrSBr.

Replicability: The abstract does not mention publicly released code or datasets. Reproducing the experimental work requires a CVT growth furnace with controlled precursor ratios, a CAFM setup, and DFT/thermodynamic modeling capability (moderate HPC — typical VASP/QE workloads on ~100–1000 CPU cores for the defect complex calculations). Growth cycles typically take days to weeks per condition.

Research Gaps:

The impact of D* suppression on device-level magneto-electrical and magneto-optical performance is not directly quantified in the abstract.
Other electrically active defects beyond D*, and defect interactions or clustering effects at low temperature, remain unexplored.

🔥 GitHub Trending

1. smixs/mentor

⭐ 28 stars | Python

mentor — a session-insights skill for AI coding agents. This skill reads your local Claude Code and OpenAI Codex history and writes an /insights-style HTML report on how you work: what you build, wher

agent-skills ai-agents claude claude-code claude-skill codex

2. MisterBooo/llm-interview-questions

⭐ 7 stars | Unknown

100 道大模型面试题与 100 张 SVG 图解，覆盖 LLM、RAG、Agent、微调、推理优化和评估。

ai-agents artificial-intelligence chinese deep-learning interview-questions large-language-models

3. hojjatjh/openrouter-php

⭐ 5 stars | PHP

A modern, fully-typed PHP client for the OpenRouter API — chat, streaming, tools, structured outputs, memory & more.

ai api-client llm openai openrouter php

4. Prism-Shadow/penguin-harness

⭐ 4 stars | TypeScript

Efficient Self-Improving Harness for Everyone

agent agentic-ai ai claude-code llm skills

5. Rahulchaube1/Rahul-Chaube-Skills

⭐ 3 stars | Unknown

The most comprehensive open-source AI skills library for LLMs, AI agents, prompt engineering, deep research, and production AI systems.

agents ai ai-engineering artificial-intelligence automation coding

6. miguelestradam3/real_estate_price_prediction

⭐ 3 stars | Jupyter Notebook

This project demonstrates how to build an Artificial Neural Network (ANN) using TensorFlow and Keras to predict real estate prices based on property characteristics.

artificial-neural-networks coursera deep-learning sklearn tensorflow

7. ianustec/neura-office

⭐ 3 stars | Unknown

NEURA Office: native Word, PowerPoint and Excel for Open WebUI. LibreOffice-compatible OOXML. MS365 assistant planned.

docx libreoffice llm microsoft-365 mit office

8. levi-qiao/graphkit

⭐ 3 stars | Shell

A graph of agent nodes, not a drifting loop — a Claude Code skill: an executor node + a clean-context supervisor that watches from outside and corrects drift. 把长跑编码任务跑成一张智能体节点图。MIT, EN/中文.

agent agentic ai-agents anthropic automation claude

9. Ffarha6/Ffarha6

⭐ 2 stars | Unknown

Artificial Intelligence Student | Django Backend Developer

artificial-intelligence backend computer-vision django machine-learning python

10. siddemmohankrishna/snapclass-ai-intelligent-face-voice-attendance-system

⭐ 2 stars | Python

An AI-powered classroom attendance system built with Streamlit, Supabase, Face Recognition, and Voice Recognition for automated student attendance management.

artificial-intelligence attendance-system computer-vision dlib education face-recognition

11. ImMaddd/ImMaddd

⭐ 2 stars | Unknown

My personal GitHub profile README. I'm an IT Educator and Software Developer passionate about building scalable web apps and teaching technology.

django machine-learning nextjs python web-development

12. taotaoboom/pdf2md-agent-skill

⭐ 2 stars | Python

让 LLM agent 高效读懂任何文档：PDF/Word/PPT/Excel 转 Markdown，PDF 智能分流（文本页 PyMuPDF + 公式页/扫描页视觉 OCR）

agent-skill claude-code claude-code-skill document-conversion latex llm

13. AliAkrami1375/Li-Translate

⭐ 2 stars | Vue

Open-source AI-powered subtitle generation and natural language translation platform for video and audio.

agent agentic-ai agentic-workflow ai ai-video-subtitle llm

14. zahidhameed322578-maker/Credit-Scoring-Model

⭐ 1 stars | Python

Machine Learning Credit Scoring Prediction System using Random Forest, Logistic Regression, Decision Tree, and Streamlit.

artificial-intelligence classification credit-scoring data data-science machine-learning

15. Shiv071205/AI-Shop-Assistant

⭐ 1 stars | Python

AI-powered shopping assistant built with FastAPI, Streamlit, Google Gemini, MySQL, and vector search for intelligent product recommendations.

artificial-intelligence chatbot fastapi gemini-api generative-ai google-gemini

Generated by Research Pulse on 2026-07-19 06:05