2026-07-17·Daily digest

🔬 Research Pulse

Daily Digest

July 17, 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 effectively leverage long visuomotor context (thousands of timesteps) without incurring prohibitive inference latency, given that current policies operate on single-step or short-history windows?

Summary: RoboTTT integrates Test-Time Training into Vision-Language-Action robot policies, enabling visuomotor context scaling to 8K timesteps — three orders of magnitude beyond prior work — by treating fast weights as the recurrent state, updated via gradient descent at both training and inference time. This unlocks one-shot imitation from human video, on-the-fly policy improvement, and completion of a five-minute ten-stage assembly task, while establishing context length as a new scaling axis for robot foundation models.

Key Results: RoboTTT scales visuomotor context to 8K timesteps (three orders of magnitude beyond SOTA) without growing inference latency. On real-robot manipulation tasks, it improves overall performance by 87% over single-step context baselines and fully completes a five-minute, ten-stage assembly task that no baseline completes. Pretraining with 8K-timestep context outperforms 1K-timestep pretraining by 62%, demonstrating context length as a new scaling axis.

Key Findings:

8K-timestep context yields 87% overall performance gain over single-step baselines and uniquely completes a five-minute, ten-stage assembly task
Closed-loop performance scales steadily with pretraining context length (8K beats 1K by 62%), a first observation for robot policies
Long context enables new capabilities: one-shot in-context imitation from human video, on-the-fly improvement, and perturbation robustness — with no inference latency penalty

Technical Novelty: Integration of Test-Time Training (TTT) into VLA robot foundation models, where the recurrent state consists of fast weights updated via gradient descent during both training AND inference — compressing long histories into weight space. Combined with sequence action forcing and truncated BPTT to enable practical training at 8K context.

What's New: First application of Test-Time Training to robot foundation models, replacing attention-based or short-history context with fast-weight recurrent state that compresses history into parameters. First demonstration that pretraining context length is a meaningful scaling axis for closed-loop robot performance, analogous to context scaling in LLMs.

Extension Opportunities:

Apply the TTT fast-weights recurrent state to multi-robot or multi-embodiment settings, testing whether cross-embodiment histories compress into shared weight space
Extend one-shot in-context imitation from human video to more diverse demonstration modalities (VR teleoperation, sketch, natural language trajectories) within the 8K context window
Investigate the pretraining scaling law by pushing context beyond 8K timesteps and characterizing diminishing returns or new emergent capabilities

Replicability: Project page with videos exists (research.nvidia.com/labs/gear/robottt/); code/weights availability not stated in abstract. Reproduction likely requires substantial GPU compute (NVIDIA-scale) plus a real robot manipulation setup for the ten-stage assembly benchmark, making full replication difficult outside well-resourced labs.

Research Gaps:

No evaluation on cross-embodiment or sim-to-real transfer with the long-context regime
Unclear how the fast-weight state degrades or catastrophically forgets across task boundaries in continuous multi-day deployment

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 failure loops and losing track of task progress as interaction histories grow during open-domain information-seeking tasks?

Summary: SearchOS reframes multi-agent web search as an operating-systems problem, externalizing fragile implicit search state into persistent structured artifacts (Evidence Graph, Coverage Map, Failure Memory) and using pipeline-parallel scheduling plus a tool middleware harness with reusable skills. It reports leading all metrics on WideSearch and GISA benchmarks against single- and multi-agent baselines.

Key Results: SearchOS leads all metrics among evaluated single- and multi-agent baselines on the WideSearch and GISA benchmarks (specific numbers not provided in abstract). Demonstrates that externalizing implicit search state into structured artifacts (Evidence Graph, Coverage Map, Failure Memory) with pipeline-parallel scheduling outperforms conventional agent architectures.

Key Findings:

Explicit externalized state (schema completion + evidence graph) outperforms implicit context-window state tracking for long-horizon search
Pipeline-parallel scheduling that refills freed slots with coverage-gap tasks improves throughput vs sequential or naive parallel agent execution
A middleware harness intercepting model/tool interactions can detect stalls and enforce anti-repetition via a reusable hierarchical skill library

Technical Novelty: The combination of (1) formulating info-seeking as relational schema completion with grounded citations, (2) SOCM's four externalized state artifacts (Frontier Task, Evidence Graph, Coverage Map, Failure Memory) replacing implicit context, (3) pipeline-parallel scheduling that refills freed agent slots based on coverage gaps, and (4) a middleware harness with reusable hierarchical strategy/access skills — treating agent coordination as a system-level OS problem rather than a prompting problem.

What's New: Prior multi-agent search systems relied on implicit context and prompt engineering for coordination. SearchOS is novel in treating search coordination as a system-level scheduling problem with a formal relational-schema task formulation, persistent shared state artifacts, and an OS-inspired middleware layer with reusable skills across runs.

Extension Opportunities:

Apply the SOCM state-externalization paradigm (Evidence Graph + Failure Memory) to domain-specific vertical search agents (legal, medical, scientific literature) where citation grounding is critical
Extend the Search Tool Middleware Harness into a general-purpose 'agent OS' layer that could sit under other tool-using LLM frameworks like LangGraph or CrewAI to add stall detection and skill reuse
Build a persistent cross-run Failure Memory that learns transferable failure patterns across users/tasks, turning the hierarchical skill library into a shared community knowledge base

Replicability: Abstract does not mention code/data release. Reproduction would likely require significant compute for parallel sub-agent execution plus API access to a tool-integrated LLM and search tools; specific model sizes and search budgets are not specified.

Research Gaps:

Abstract does not quantify improvements or characterize failure modes that remain after externalization
Generalization beyond the two evaluated benchmarks (WideSearch, GISA) to non-tabular or ambiguous-schema information needs is untested

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 producer of a pre-trained LLM upgrade its tokenizer in-place to reduce token fragmentation for languages underrepresented in the original pre-training corpus, without paying the cost of a full re-pretrain — particularly for compact on-device models where embedding/LM-head matrices are a material share of decode bandwidth?

Summary: The paper introduces an in-place tokenizer expansion recipe that upgrades a pre-trained LLM's vocabulary by continuing its BPE merges on a multilingual corpus, preserving source tokens and initializing new embeddings as the mean of their sub-token decompositions. A two-stage adaptation (embedding-only, then full-model continued pre-training) recovers original quality. Applied to LFM2-8B-A1B, it produces LFM2.5-8B-A1B with 2.2-3.7x per-character decode speedups on Hindi, Vietnamese, and Thai.

Key Results: Applied to LFM2-8B-A1B (an 8B MoE checkpoint) to produce LFM2.5-8B-A1B with a 128K tokenizer. The expanded tokenizer encodes Hindi ~2.4x, Vietnamese ~2.6x, and Thai up to 4.0x fewer tokens than the source. Combined with measured per-token cost of the larger vocabulary, this yields an estimated 2.2-3.7x per-character decode speedup on reference devices. A two-stage adaptation (embedding-only training, then full-model continued pre-training) recovers source-checkpoint quality.

Key Findings:

Continuing (rather than replacing) BPE merges preserves source-token identity and gives every new token an exact source-sub-token decomposition, enabling clean embedding initialization via sub-token mean.
A two-stage adaptation — embedding-only training followed by full-model continued pre-training — is sufficient to recover source-checkpoint quality after tokenizer expansion.
For compact/on-device models where embedding and LM-head matrices dominate decode bandwidth, the net effect after accounting for larger-vocabulary overhead is still a 2.2-3.7x per-character decode speedup for target languages.

Technical Novelty: Rather than training a fresh tokenizer or grafting on an unrelated one, the recipe continues the existing BPE merge sequence on a multilingual corpus. This guarantees (a) source tokens carry over unchanged as single tokens, (b) every new token has an exact decomposition into source tokens, enabling principled mean-of-sub-token embedding initialization. Combined with a two-stage embedding-only-then-full-model adaptation, this preserves source-checkpoint quality while dramatically reducing fragmentation. The paper also reports negative findings that shaped the recipe.

What's New: Prior tokenizer-swap or vocabulary-extension approaches either train a fresh tokenizer (losing embedding continuity) or extend the vocabulary without guaranteeing token-level backward compatibility. This work uniquely continues the exact BPE merge sequence, which mathematically guarantees source-token invariance and exact sub-token decomposition of every new token — making mean-of-sub-token initialization principled rather than heuristic. The paper also explicitly documents negative findings, which is rare and valuable.

Extension Opportunities:

Apply the recipe to other compact open-weight models (e.g., Llama 3.2 1B/3B, Gemma 3 small, Qwen small variants) to expand coverage for additional low-resource languages and benchmark against retraining from scratch.
Extend beyond natural language to domain vocabularies — code tokens, math notation, DNA/protein sequences, or chemistry SMILES — using the same BPE-merge-continuation + mean-of-sub-tokens initialization.
Explore adaptive/dynamic tokenizer expansion at deployment time (per-user or per-region vocabularies) with lightweight LoRA-style embedding adapters instead of full re-training.

Replicability: The authors release model weights (LFM2.5-8B-A1B) and the expanded tokenizer. Reproducing the recipe itself requires access to a multilingual corpus for BPE merge continuation plus enough compute for embedding-only training followed by full-model continued pre-training of an 8B MoE — likely on the order of tens to low hundreds of GPU-days, though far cheaper than pre-training from scratch. Recipe-only replication (not the specific checkpoint) is feasible for well-resourced academic labs.

Research Gaps:

The paper focuses on a specific 8B MoE model and three target languages; generalization across model families, sizes, and dozens of low-resource languages simultaneously is untested.
Quality is characterized against the source checkpoint but broader multilingual reasoning/generation benchmarks (e.g., MGSM, XCOPA, Flores translation) versus a from-scratch multilingual baseline are not clearly quantified in the abstract.

🦾 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 enable effective scaling of Behavior Foundation Models (BFMs) for humanoid whole-body control?

Summary: The paper presents a scaling recipe for Behavior Foundation Models for humanoid robots, showing that coordinating a global-frame motion tracking paradigm, scaled on-policy rollouts with diverse reference motions, and a Humanoid Transformer architecture yields large gains in whole-body control fidelity. It reduces keypoint tracking error by 10%+ locally and 82% globally versus prior controllers, and transfers to real hardware.

Key Results: Demonstrated that coordinating three components (motion tracking paradigm, on-policy rollout + motion diversity synergy, Humanoid Transformer architecture) reduces Mean Per-Keypoint Position Error (MPKPE) on the test set by over 10% in local mode and 82% in global mode versus existing humanoid controllers, validated in both simulation and real-world deployment.

Key Findings:

Global-frame motion tracking is a unifying reformulation that subsumes diverse humanoid control tasks
Scaling requires joint scaling of on-policy rollout quantity AND reference motion diversity — neither alone suffices
The Humanoid Transformer architecture enables structured behavioral representations to emerge naturally with scale
Approach transfers from simulation to real humanoid deployment with strong generalization

Technical Novelty: Reformulates diverse humanoid control problems as global-frame integrated whole-body motion tracking (rather than task-specific local objectives), pairs it with a novel Humanoid Transformer that allows structured behavioral representations to emerge, and identifies the specific synergy between on-policy rollout quantity and reference motion diversity as the scaling axis.

What's New: Prior BFM work scaled data or models in isolation without a coordinated recipe; this paper isolates the three coupled ingredients that make scaling actually pay off, and introduces the global-frame tracking reformulation plus Humanoid Transformer as the enabling architecture.

Extension Opportunities:

Extend the Humanoid Transformer to multimodal inputs (vision/language) for instruction-conditioned whole-body control rather than motion tracking alone
Apply the on-policy rollout + reference motion diversity scaling recipe to non-humanoid morphologies (quadrupeds, dexterous hands) to test generality of the recipe
Distill the BFM into a smaller edge-deployable policy while preserving global-frame tracking fidelity for real robot compute budgets

Replicability: Abstract does not mention open-sourced code or data. Reproduction would require large-scale humanoid motion capture datasets, a physics simulator (e.g., Isaac Gym/MuJoCo), substantial GPU compute for on-policy RL rollouts at scale, and access to a physical humanoid platform for sim-to-real validation.

Research Gaps:

Unclear how the recipe interacts with language/vision conditioning for goal-directed tasks beyond motion imitation
Real-world robustness bounds (perturbations, terrain, contact-rich manipulation) not fully characterized in the abstract

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 Vision-Language-Action (VLA) policies for UAV target tracking maintain robust performance during sustained occlusions in complex urban environments, where line-of-sight is repeatedly broken by buildings, vegetation, or roadside objects, causing error accumulation and eventual target loss?

Summary: CosFly-VLA is a compact 0.8B spatially-aware Vision-Language-Action model for UAV target tracking under occlusion, trained via a four-stage recipe (spatial continued pretraining, curriculum SFT, chain-of-thought recovery reasoning, and closed-loop RL). It jointly predicts target grounding, visibility, and continuous flight actions, achieving 34–35% ADE reduction and up to 29.8% success-rate gains over OpenVLA on urban tracking benchmarks.

Key Results: CosFly-VLA-0.8B outperforms OpenVLA baseline with 34.1% reduction in open-loop Average Displacement Error (ADE) on seen-test and 35.3% on unseen-test. Closed-loop RL optimization improved Success Rate (SR) by 29.8% (seen) and 2.5% (unseen). Training used a 500k mixed-source pool for Spatially Grounded Continued Pretraining, evaluated under a shared oracle state history.

Key Findings:

Joint prediction of grounding + visibility + actions via a structured interface outperforms visible-frame imitation baselines
Spatially grounded continued pretraining on UAV-view depth/distance data measurably improves 3D reasoning transfer to tracking
Curriculum SFT (natural → hard/long-occlusion) plus CoT recovery reasoning meaningfully improves occlusion robustness
Closed-loop RL gains generalize weakly to unseen scenes (2.5% SR gain) versus seen scenes (29.8%), exposing a distribution-shift gap

Technical Novelty: The combination of (1) a structured prediction interface that jointly outputs grounding, visibility estimation, and continuous flight actions, (2) Spatially Grounded CPT that injects UAV-view depth/distance/3D reasoning before task specialization, (3) three-stage curriculum SFT progressing from natural to hard/long-occlusion data, and (4) Chain-of-Thought recovery reasoning traces trained explicitly before structured answers — this pipeline is tailored for occlusion-robust aerial tracking rather than generic VLA behavior.

What's New: Prior VLA policies treat tracking as visible-frame imitation and degrade catastrophically under occlusion. CosFly-VLA is the first to combine spatial 3D pretraining, occlusion-aware curriculum learning, recovery-oriented CoT, and closed-loop RL into a single small-model recipe explicitly designed for sustained-occlusion UAV tracking.

Extension Opportunities:

Extend the multi-component RL reward to include energy/battery consumption and wind disturbance robustness for real-world flight deployment
Adapt the three-stage curriculum SFT approach to multi-agent UAV swarm tracking, where partial observations from teammates augment occlusion recovery
Replace the oracle state history assumption with learned state estimation (e.g., Kalman filter or transformer memory) to test real-world viability without ground-truth priors

Replicability: The abstract does not mention code or dataset release. Reproducing would require: a 0.8B-parameter VLA backbone, ~500k mixed spatial pretraining samples, curated UAV tracking data with occlusion annotations, a closed-loop UAV simulator (likely AirSim/Isaac Sim class), and multi-stage training (CPT + SFT + CoT + RL) — estimated multiple A100/H100-days for full pipeline.

Research Gaps:

Large gap between seen (29.8%) and unseen (2.5%) closed-loop SR improvements suggests limited generalization from RL fine-tuning
Evaluation relies on a shared oracle state history, leaving open the question of performance under fully learned/noisy state estimation in real deployments

💻 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 failures in large-scale AI clusters be reliably predicted in absolute time, and if not, what alternative framing better supports proactive reliability management?

Summary: The paper argues that predicting the exact time of GPU failures is fundamentally unreliable due to stochastic, low-SNR telemetry, and instead proposes HeaRank, a Learning-to-Rank framework that produces a relative risk ordering of nodes. Deployed on a production cluster with thousands of GPUs, HeaRank hits AUC 0.83 and captures 64% of future failures in its top-5% risk bucket versus 21% for the incumbent system.

Key Results: Telemetry analysis of a production GPU cluster showed major failures (DBEs, GPU Lost) are highly stochastic with low signal-to-noise ratios, defeating time-based prediction. Their Learning-to-Rank framework HeaRank achieved AUC 0.83 on a cluster with thousands of GPUs, outperforming heuristic and SOTA ranking baselines. In online deployment, HeaRank captured 64% of future failures within the top 5% of ranked nodes vs. only 21% for the incumbent production system.

Key Findings:

DBEs and GPU Lost events show strong stochasticity and low SNR, making time-based prediction ineffective
Relative risk ranking is a more robust and actionable target than absolute failure timing
HeaRank's top-5% risk bucket captures 3x more future failures than the deployed production baseline

Technical Novelty: Reframes GPU reliability from time-series failure prediction to a global Learning-to-Rank problem over nodes, exploiting stable historical failure patterns rather than noisy real-time signals — a paradigm shift for hardware predictive maintenance.

What's New: Prior predictive-maintenance work targets exact failure timing via time-series models; this work is the first to explicitly frame GPU reliability assessment as an LTR problem and validate it at production scale.

Extension Opportunities:

Integrate HeaRank scores into a scheduler that biases synchronous training placement away from high-risk nodes and quantify MTBF/job-restart reduction
Extend the LTR framework to multi-vendor accelerators (AMD MI300, TPU, Trainium) and heterogeneous fleets, testing whether ranking transfers across hardware generations
Combine ranking with active checkpointing policies — dynamically adjust checkpoint frequency per node based on its risk rank to minimize wasted compute

Replicability: Abstract does not mention released code, dataset, or model artifacts. Reproduction would require access to production-scale GPU telemetry (thousands of nodes, months of logs), which is typically proprietary; training LTR models themselves is modest (single-GPU or CPU feasible).

Research Gaps:

No integration story yet with cluster schedulers or checkpointing systems to translate ranking into training-job resilience gains
Limited generality evidence — evaluation is on a single cluster and GPU generation, without cross-vendor or cross-workload validation

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 topologies over deployed fibre infrastructure, rather than being locked into static point-to-point or fixed-topology configurations?

Summary: The paper presents a metropolitan-scale, entanglement-based quantum network built around a quantum reconfigurable optical add-drop multiplexer (q-ROADM) that dynamically routes polarisation-entangled photons from a broadband source to six users. It demonstrates programmable topologies (full-mesh, partial-mesh, sliced), 150+ hours of stable operation, and multi-protocol support including a secure onboarding protocol (SIAT) combined with network flooding.

Key Results: Demonstrated a metropolitan-scale quantum network using a quantum reconfigurable optical add-drop multiplexer (q-ROADM) distributing polarisation-entangled photon pairs from a broadband source to 6 users over deployed campus/metro fibre. Achieved stable six-user full-mesh operation for >150 hours, supported programmable full-mesh, partial-mesh, and sliced sub-network configurations, and executed Secure Inaugural Authentication-Transfer (SIAT) combined with network flooding over multiple paths for secure user onboarding.

Key Findings:

Six-user full-mesh entanglement distribution ran stably for more than 150 hours over deployed campus and metropolitan fibre
The q-ROADM enables on-demand switching between full-mesh, partial-mesh (time-shared), and sliced sub-network topologies, supporting service-oriented entanglement resource allocation
The same physical infrastructure supports heterogeneous quantum protocols, demonstrated via SIAT + multi-path network flooding for secure new-user onboarding

Technical Novelty: The q-ROADM primitive itself — bringing telecom-style reconfigurable add-drop multiplexing to entanglement distribution — enabling programmable network topologies (full-mesh, partial-mesh, sliced) from a single broadband entangled source, plus the demonstration of network slicing and mixed-protocol operation (SIAT + flooding) on the same physical infrastructure. Prior deployed entanglement networks were largely static in their wavelength-to-user allocation.

What's New: Prior entanglement-distribution networks used fixed wavelength-multiplexing plans and static user assignments. This work introduces telecom-inspired reconfigurability (ROADM) into the quantum domain, allowing the same source and fibre plant to serve different topologies, user sets, and protocols on demand.

Extension Opportunities:

Integrate quantum memories at q-ROADM nodes to buffer entanglement and enable multi-hop entanglement swapping beyond single metropolitan hop
Develop an SDN-style control plane / orchestrator on top of the q-ROADM to dynamically allocate entanglement resources based on real-time QKD service demand and link quality metrics
Extend the six-user testbed with heterogeneous protocols (QKD + blind quantum computing + clock sync) simultaneously time-sharing the same q-ROADM to benchmark cross-protocol interference and scheduling policies

Replicability: No code/data availability is mentioned in the abstract. Reproduction would require substantial hardware: a broadband polarisation-entangled photon-pair source, a q-ROADM (wavelength-selective switch adapted for quantum signals), deployed metropolitan-scale fibre, six user nodes with polarisation-analysis and single-photon detection (likely SNSPDs), and time-tagging/coincidence electronics — i.e., a full experimental quantum-networking testbed, not a compute-bound reproduction.

Research Gaps:

No quantum memory or entanglement swapping — the architecture is limited to single-hop metropolitan distances and cannot yet extend to long-haul or continental scales
Scheduling, control-plane, and QoS policies for dynamically arbitrating between users/protocols on the q-ROADM are not deeply explored

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: What microscopic material origins drive decoherence in superconducting quantum circuits beyond the well-studied two-level system (TLS) defects, and how can they be identified in live devices?

Summary: The paper uses scanning gate microscopy on operating superconducting circuits to identify a previously unrecognized decoherence source: Coulomb blockade and microwave-driven charge tunneling in metallic grains left behind by standard lithography. These grain defects are shown to be as prevalent and damaging as TLS defects but arise from a distinct physical mechanism that conventional characterization misattributes.

Key Results: Using scanning gate microscopy on live superconducting circuits, the authors identified and characterized Coulomb-blockade-driven charge tunneling in metallic grains as a distinct decoherence mechanism. They demonstrated across multiple defects and devices that these grain defects are 'as common and as debilitating' as TLS defects, and showed the loss is microwave-power-dependent — a signature conventional characterization would misattribute to power-independent processes.

Key Findings:

Metallic grains in thin-film devices host Coulomb-blockade charge tunneling that dissipates microwave energy and drives decoherence
These grain defects are as common and as harmful to coherence as TLS defects across multiple devices
Conventional characterization misattributes this microwave-power-dependent loss to power-independent processes, obscuring its true origin
Eliminating metallic grains during fabrication offers a direct, practical route to improve coherence

Technical Novelty: Applying scanning gate microscopy to live (operating) superconducting circuits to spatially resolve individual defects, and identifying Coulomb blockade in metallic grains — rather than TLS — as a mechanism for microwave loss.

What's New: Challenges the prevailing TLS-centric paradigm of coherence limits in superconducting qubits by introducing a fundamentally different, materials-origin mechanism (Coulomb blockade in grains) and localizing it directly on live circuits via scanning gate microscopy.

Extension Opportunities:

Develop fabrication protocols (e.g., alternative deposition, etching, or annealing) that eliminate metallic grains and benchmark coherence gains (T1/T2) against grain-containing controls
Build automated scanning-gate-microscopy pipelines that map and classify Coulomb-blockade vs TLS defects across wafer-scale devices to guide yield optimization
Model microwave-driven charge tunneling in metallic grains to derive power-dependent loss signatures, enabling standard qubit spectroscopy to distinguish grain defects from TLS without SGM

Replicability: No code/data availability mentioned in the abstract. Reproduction requires a cryogenic scanning gate microscopy setup coupled to live superconducting qubit/resonator devices — substantial specialized experimental infrastructure, not commodity compute.

Research Gaps:

Quantitative fabrication-process mapping to predict grain density and resulting coherence impact
Scalable, non-microscopy diagnostic that distinguishes grain-Coulomb-blockade loss from TLS loss in standard qubit measurements

⚡ 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 epitaxial superconducting thin films be grown with sufficiently low structural disorder and interfacial defects to yield ultra-high-Q microwave resonators, and what limits the crystalline quality of TiN on sapphire?

Summary: The paper reports MBE-grown epitaxial (111) TiN films on c-plane sapphire with record crystalline quality (18 arcsec rocking curve) and demonstrates CPW resonators with internal Q > 10^6 at single-photon powers and > 2×10^7 at high powers. It also identifies sub-surface sapphire defects as a previously unrecognized source of structural imperfections in epitaxial superconducting films.

Key Results: MBE-grown (111) TiN on c-plane sapphire achieved an X-ray rocking curve FWHM of 18 arcsec (lowest reported). Quarter-wave CPW hanger resonators (3/6/3 μm gap/strip/gap) at 5.8 GHz, 10 mK showed internal Q > 1×10^6 in the single-photon limit (⟨n⟩~1), rising to > 2×10^7 at ⟨n⟩~10^6. Sub-surface sapphire defects were identified as nucleation sites for structural defects in the TiN film.

Key Findings:

MBE TiN/sapphire rocking curve FWHM of 18 arcsec — the lowest reported for TiN thin films
Single-photon internal Q > 1×10^6 at 5.8 GHz and 10 mK, rising above 2×10^7 at high photon numbers
Sub-surface defects in the sapphire substrate nucleate structural defects in the epitaxial TiN, revealing a substrate-quality bottleneck

Technical Novelty: First MBE growth of TiN on sapphire achieving record rocking-curve width (18 arcsec) combined with identification of sub-surface substrate defects as the dominant nucleation source of film disorder — prior high-Q TiN work relied on reactive sputtering or ALD with polycrystalline or lower crystalline quality films.

What's New: Combines record epitaxial crystalline quality of TiN via MBE (vs. the more common sputtered/ALD TiN) with ultra-high-Q resonator performance, and provides the first identification of sapphire sub-surface defects as a limiting factor for epitaxial superconducting films.

Extension Opportunities:

Integrate MBE TiN films as base electrodes for all-epitaxial Josephson junctions with crystalline tunnel barriers (e.g., MgO or AlN) to test qubit T1/T2 improvements
Systematically pre-treat or anneal sapphire substrates to eliminate sub-surface defects, then correlate defect density with TLS loss at single-photon powers
Extend the MBE recipe to other nitride superconductors (NbN, NbTiN) on sapphire and compare Q vs. gap material and kinetic inductance for KIDs and traveling-wave parametric amplifiers

Replicability: No code/data release is mentioned. Reproduction requires an MBE system capable of nitride growth, c-plane sapphire wafers, cleanroom lithography for CPW patterning, and a dilution refrigerator (~10 mK) with low-noise microwave readout for single-photon Q measurement — substantial specialized hardware, not compute-bound.

Research Gaps:

No demonstration yet of epitaxial Josephson junctions or qubits built on this MBE TiN platform
The microscopic origin and mitigation strategy for the observed sub-surface sapphire defects remain uncharacterized

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 be systematically controlled during chemical vapor transport (CVT) growth to enable tailoring of its magneto-electrical properties?

Summary: The paper establishes practical CVT growth strategies to systematically suppress a specific electrically active defect (D*) in the 2D magnetic semiconductor CrSBr by up to 10×, using conductive AFM to quantify defects and DFT plus thermodynamic modeling to identify D* as a sulfur-related vacancy complex. Sulfur/bromine-rich conditions and lower absolute growth temperatures at fixed gradient are the key knobs.

Key Results: Using conductive atomic force microscopy (CAFM), the authors identified a characteristic defect (D*) with a strong electronic fingerprint and reduced its concentration by up to an order of magnitude through optimized CVT growth conditions. They demonstrated that sulfur- and bromine-rich precursor stoichiometry, along with lower absolute growth temperatures (while maintaining the same temperature gradient), suppress defect formation. Thermodynamic modeling and DFT calculations indicate D* is consistent with a sulfur-related vacancy complex rather than an isolated point defect.

Key Findings:

D* defect concentration in CrSBr can be reduced by up to an order of magnitude via optimized CVT growth conditions
Sulfur- and bromine-rich precursor stoichiometry suppresses D* defect formation
Lowering absolute growth temperature while maintaining the same temperature gradient further suppresses defects
D* is consistent with a sulfur-related vacancy complex rather than an isolated point defect, per DFT and thermodynamic modeling

Technical Novelty: First systematic correlation of CVT growth parameters (precursor stoichiometry and absolute temperature at fixed gradient) with atomically-resolved defect densities in CrSBr, combined with identification of D* as a sulfur-related vacancy complex via CAFM electronic fingerprinting plus DFT/thermodynamic modeling — prior work had not established growth-based defect control in this material.

What's New: Prior CrSBr research focused on intrinsic magneto-electrical properties but lacked systematic defect control during growth. This work is the first to link specific CVT parameters to quantitative defect densities in CrSBr and to structurally identify the dominant electrically active defect as a vacancy complex.

Extension Opportunities:

Correlate defect suppression with direct measurements of magneto-transport, magnon lifetimes, or exciton linewidths to quantify how D* reduction improves functional device performance in CrSBr-based spintronic/magnonic devices
Extend the CAFM + DFT + thermodynamic modeling workflow to other 2D magnetic van der Waals materials (e.g., CrPS3, NiPS3, FePS3) to build a general framework for growth-controlled defect engineering
Develop in-situ or post-growth annealing protocols in controlled S/Br atmospheres to further passivate residual D* defects and push toward ultra-clean CrSBr crystals for quantum applications

Replicability: The abstract does not mention public code or data. Reproduction requires a CVT growth setup with controlled precursor stoichiometry and temperature profiles, a conductive AFM system, and DFT capability (likely VASP/QE with moderate HPC resources for defect supercell calculations). Crystal growth cycles typically take days to weeks per condition.

Research Gaps:

The functional impact of D* reduction on magneto-transport, optical, and magnonic device performance is not quantified in the abstract
Whether other defect species (non-D*) become dominant under S/Br-rich or low-temperature growth is not addressed

🔥 GitHub Trending

1. KakoAI/Kako-Code

⭐ 3 stars | TypeScript

Kako Code — an Agent Harness personal assistant. Persistent context, tool execution, sub-agents, memory, permissions, and observability.

agent ai cli harness llm mcp

2. haichen1985/opti-moa

⭐ 3 stars | TypeScript

Optimal LLM proxy: smart routing + conditional MOA + context compression + experience learning

ai-agent llm llm-proxy mixture-of-agents moa multi-model

3. twz007/java-obs-agent

⭐ 2 stars | Java

Framework-agnostic Java observability sidecar agent: chat with a live JVM via MCP/web UI (metrics, slow calls, errors, logs, thread dumps, method traces). Self-hosted, attachable (Arthas-style). Apach

ai-agent apm arthas bytebuddy java javaagent

4. rahulcvwebsitehosting/OpenCluely

⭐ 2 stars | JavaScript

Free, open-source AI meeting/coding copilot - bring your own API key. Electron overlay that sees your screen and hears your calls.

ai-copilot anthropic electron gemini llm macos

5. boddetijayanth22/Enterprise-Knowledge-Assistant

⭐ 2 stars | Python

Production-inspired Retrieval-Augmented Generation (RAG) application for intelligent document search using FastAPI, Streamlit, Qdrant, Sentence Transformers, and Google Gemini.

aiengineering fastapi genai llm python qdrant

6. mohamadali-bl/tehran-house-prices-analysis

⭐ 1 stars | Python

Data preprocessing, feature engineering and visualization for Tehran House Prices.

data-analysis eda feature-engineering house-price-prediction housing machine-learning

7. rishika-glitch06/F1-Podium-Predictor-AI

⭐ 1 stars | Jupyter Notebook

A production-style Explainable AI (XAI) dashboard for Formula 1 podium prediction using XGBoost and SHAP, featuring interactive visualizations, model interpretability, prediction analytics, and automa

classification data-science explainable-ai formula1 machine-learning plotly

8. Johaan-Mannanal/motorsport-telemetry-analytics

⭐ 1 stars | Python

Interactive F1 telemetry dashboard: compare two drivers' laps (speed, delta, sectors, tyres) with a transparent tyre-degradation model. Python, FastF1, Streamlit.

dashboard data-science data-visualization f1 fastf1 formula1

9. LikeACloud7/awesome-personalized-ai-benchmarks

⭐ 1 stars | Python

A living, evidence-based catalog of benchmarks for personalized LLMs and AI agents—covering preference alignment, long-term memory, tool use, safety, privacy, and multimodal adaptation.

agent-evaluation ai-agents ai-safety awesome awesome-list benchmarks

10. AnasHasnain333/Decodelabs_Data-Classification-AI

⭐ 1 stars | Jupyter Notebook

A beginner friendly ML project that trains a K-Nearest Neighbors classifier on the Iris dataset covers data loading, train/test split, training and evaluation.

beginner-project data-science jupyter-notebook machine-learning python scikit-learn

11. tayabhayat/House-Price-Prediction

⭐ 1 stars | Python

House Price Prediction using Linear Regression with Python, Pandas, and Scikit-Learn.

ai data-science linear-regression machine-learning pandas python

12. LOGO127/cs336-2026-assignment1-llm-foundations

⭐ 1 stars | Python

From-scratch BPE, Transformer, AdamW, training, and generation for CS336 2026 Assignment 1.

bpe deep-learning from-scratch language-model llm nlp

13. Ashfaq-Riyaldeen/free-llm-apis

⭐ 1 stars | Unknown

Decision-focused list of free LLM APIs in 2026: which to pick, real limits, and how to stack free tiers so you never pay for inference.

ai free free-api gemini groq llm

14. foreverse-app/character-card-skills

⭐ 1 stars | Python

Agent skills, 15 original character cards and a calibrated AI-flavor detector for roleplay card authors. Skills run in Claude Code / Cursor / Codex / Gemini CLI; cards import into SillyTavern and Fore

agent-skills ai-characters ai-roleplay character-ai character-card character-cards

15. YunyueLi/modelese

⭐ 1 stars | Python

modelese（模型腔）— lexicon, linter & rewrite-gate for the private dialect AI develops on long runs. 新时代文言文检测器：给 AI 黑话打分、打回、翻成人话。CC0 词典。

ai ai-agents chinese claude claude-code jargon

Generated by Research Pulse on 2026-07-17 06:19