2026-07-18·Daily digest

🔬 Research Pulse

Daily Digest

July 18, 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 (VLAs) scale visuomotor context beyond single-step or short histories to unlock long-horizon reasoning, in-context imitation, and on-the-fly adaptation—without blowing up inference latency?

Summary: RoboTTT introduces Test-Time Training into Vision-Language-Action robot policies, using fast-weight recurrent state to scale visuomotor context to 8K timesteps without inference-latency cost. This unlocks one-shot imitation from human video, on-the-fly improvement, and completion of a 5-minute 10-stage assembly task no baseline can finish, while demonstrating context length as a new scaling axis (62% gain from 1K→8K pretraining context).

Key Results: RoboTTT scales context to 8K timesteps (3 orders of magnitude beyond SOTA) with no added inference latency. On real-robot manipulation, it improves overall performance by 87% over the single-step baseline and is the only method to fully complete a 5-minute, 10-stage assembly task. Pretraining at 8K context outperforms 1K-context pretraining by 62%, evidencing context length as a new scaling axis.

Key Findings:

Fast-weight TTT state enables 8K-timestep context with flat inference latency, 3 orders of magnitude beyond current VLAs.
Long context yields emergent capabilities: one-shot in-context imitation from human video, on-the-fly policy improvement, and perturbation robustness.
Pretraining context length is itself a scaling axis for robot foundation models—first demonstration of steady closed-loop gains as context grows.

Technical Novelty: Integrating Test-Time Training into VLAs so the recurrent state is a set of fast weights updated by gradient descent at both train and inference time, compressing history into weight space rather than KV cache. Training recipe pairs sequence action forcing with truncated BPTT to make 8K-step context tractable.

What's New: Prior VLAs (OpenVLA, RT-2, Pi0) rely on single-step or short-horizon transformer context. RoboTTT is the first to bring TTT-style fast-weight recurrence to robot policies and to show context-length scaling laws for closed-loop control, sidestepping the quadratic attention cost that blocks long-context VLAs.

Extension Opportunities:

Apply the TTT fast-weights recurrent-state mechanism to bimanual or mobile manipulation where multi-minute horizons dominate, and measure whether the 87% gain compounds.
Use one-shot in-context imitation from human video as a data-efficient teleoperation replacement—build a pipeline where a human demonstrates once on webcam and the policy executes, benchmarking against SOTA imitation baselines.
Combine sequence action forcing + truncated BPTT recipe with open-source VLAs (OpenVLA, Pi0) to test whether context scaling laws hold outside the NVIDIA stack.

Replicability: Videos posted at research.nvidia.com/labs/gear/robottt/; no code/weights/dataset explicitly mentioned in abstract. Reproduction would need a real robot platform (likely Franka/bimanual), multi-GPU training for 8K-context BPTT (probably 8–64 H100s given VLA scale), and long-horizon manipulation task suites.

Research Gaps:

No evidence yet that the recipe transfers to unseen embodiments or fully out-of-distribution tasks beyond the evaluated assembly suite.
Fast-weight updates at inference raise open questions about safety, drift, and reversibility—no discussion of guarding against catastrophic online updates in the abstract.

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 information-seeking systems avoid getting trapped in repetitive loops and losing track of task progress as interaction histories grow during open-domain web search?

Summary: SearchOS is a multi-agent framework that turns implicit search progress into explicit shared state via Search-Oriented Context Management (Frontier Task, Evidence Graph, Coverage Map, Failure Memory) and a middleware harness that controls tool calls and reuses hierarchical skills. It uses pipeline-parallel scheduling to keep sub-agents busy on unresolved coverage gaps, and reports state-of-the-art results on WideSearch and GISA.

Key Results: SearchOS leads all metrics among evaluated single- and multi-agent baselines on the WideSearch and GISA benchmarks. The abstract does not report specific numerical deltas, only that it dominates across all measured metrics on both benchmarks.

Key Findings:

Framing info-seeking as relational schema completion with grounded citations gives a concrete progress signal that prevents redundant search loops
Externalizing evolving state into four persistent structures (frontier, evidence graph, coverage map, failure memory) outperforms implicit context-based tracking
A tool middleware layer combined with pipeline-parallel scheduling improves search-budget utilization and end-to-end quality on WideSearch and GISA

Technical Novelty: Reformulates open-domain info-seeking as relational schema completion with grounded citations, and externalizes agent state into four explicit shared structures (Frontier Task, Evidence Graph, Coverage Map, Failure Memory) rather than leaving progress implicit in context. Adds a middleware harness that intercepts model-tool calls to enforce anti-repetition and pipeline-parallel scheduling that refills freed agent slots with coverage-gap tasks.

What's New: Prior tool-integrated search agents keep progress implicit in the conversation history; SearchOS elevates it to a system-level shared state with dedicated data structures, a middleware interception layer, and a cross-run skill library that explicitly encodes what has failed before.

Extension Opportunities:

Extend the Evidence Graph and Coverage Map primitives to domain-specific verticals (e.g., legal, biomedical, financial due diligence) where relational schema completion with citation grounding is high-value
Replace or augment the hierarchical skill library with a learned skill-discovery module that mines successful/failed trajectories from Failure Memory to auto-generate new strategy skills
Adapt the Search Tool Middleware Harness as a general orchestration layer for non-search tools (code execution, database queries, API calls) to bring the same stall-detection and budget-control guarantees to broader agentic workflows

Replicability: Abstract does not mention code, data, or model release. Reproducing would require a tool-integrated LLM backbone, web search API access, and orchestration for parallel sub-agents plus the WideSearch and GISA evaluation harnesses — likely moderate cost (API-dominated) rather than heavy training compute.

Research Gaps:

No reported ablations in the abstract quantifying the individual contribution of each SOCM component (evidence graph vs coverage map vs failure memory)
Unclear how the approach generalizes beyond structured, schema-completable queries to open-ended synthesis or reasoning-heavy research tasks

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 languages underrepresented in the original pre-training corpus, without retraining from scratch or sacrificing source-checkpoint quality — 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 continues an existing BPE tokenizer's merges on a multilingual corpus, preserving all source tokens unchanged and initializing new-token embeddings as the mean of their source sub-token embeddings. A two-stage adaptation (embedding-only, then full continued pre-training) recovers source quality. Applied to an 8B MoE checkpoint, it yields 2.4-4.0× fewer tokens for Hindi/Vietnamese/Thai and 2.2-3.7× per-character decode speedup on-device.

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

Key Findings:

Continuing BPE merges (rather than retraining the tokenizer) preserves source tokens as single tokens and gives every new token an exact source-token decomposition, enabling clean embedding initialization
Mean-of-source-subtoken initialization combined with two-stage training (embedding-only then full-model CPT) recovers source-checkpoint quality
Tokenization compression gains (2.4-4.0×) translate to real 2.2-3.7× per-character decode speedups on-device once amortized against the larger vocab's per-token cost

Technical Novelty: Continuing existing BPE merges on multilingual corpus so every new token has an exact decomposition into source tokens (preserving carried-over embeddings unchanged), combined with mean-of-source-subtoken initialization and a two-stage adaptation protocol. Prior tokenizer-swap methods typically retrain the tokenizer from scratch or require heavier realignment; this in-place approach preserves the source vocabulary intact.

What's New: Unlike prior tokenizer-replacement work that swaps or retrains the tokenizer wholesale, this method is strictly additive — it continues the original BPE merge process, guaranteeing source token preservation and structured decomposition of new tokens. The authors also emphasize the on-device economics angle (embedding/LM-head bandwidth) and openly report negative findings that shaped the final recipe.

Extension Opportunities:

Apply the recipe to other on-device models (e.g., Gemma, Phi, Qwen small variants) targeting different low-resource language sets and measure real-world latency/energy on mobile SoCs
Explore alternative initialization strategies beyond mean-of-source-subtoken embeddings (e.g., learned weighted combinations, hypernetwork-generated embeddings) and compare recovery speed
Extend to domain-specific vocabulary expansion (code, math, chemistry SMILES) rather than only multilingual expansion, quantifying compression and downstream task gains

Replicability: Model weights and expanded tokenizer released. Reproducing the full recipe requires MoE-scale continued pre-training compute (8B active-param MoE, multilingual corpus for BPE extension and two-stage adaptation) — likely tens to hundreds of GPU-days on H100-class hardware. The recipe itself is straightforward to port to smaller models at lower cost.

Research Gaps:

No head-to-head comparison reported against alternative vocabulary-expansion or tokenizer-replacement methods on the same base checkpoint
Evaluation focuses on tokenization ratios and decode speedup; downstream task quality across languages (accuracy on multilingual benchmarks post-expansion) is less emphasized 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 the 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 in humanoid control, coordinating three components: global-frame motion tracking as a unified learning paradigm, synergy between on-policy rollouts and motion diversity, and a scalable Humanoid Transformer architecture. The resulting model achieves >10% (local) and >82% (global) MPKPE reduction over prior controllers with successful real-world deployment.

Key Results: The proposed BFM reduces Mean Per-Keypoint Position Error (MPKPE) on the test set by over 10% in local mode and over 82% in global mode compared with existing humanoid controllers. Validated in both simulation and real-world humanoid deployment, showing improved control fidelity and task generalization.

Key Findings:

Reformulating humanoid control as global-frame whole-body motion tracking unifies diverse tasks under one learning paradigm and dramatically improves global tracking (>82% MPKPE reduction).
Scaling on-policy rollout quantity and reference motion diversity together — not independently — is the key driver of BFM performance gains.
The Humanoid Transformer architecture is expressive enough that structured behavioral representations emerge naturally, supporting sim-to-real transfer.

Technical Novelty: Reformulates diverse humanoid control problems as whole-body motion tracking in the global frame (rather than local/root-relative), pairs on-policy rollout quantity with reference motion diversity as a coupled scaling axis, and introduces the Humanoid Transformer architecture that enables structured behavioral representations to emerge naturally.

What's New: Prior BFMs left unclear how paradigm, data, and architecture should co-scale. This work identifies the coupling: a global-frame tracking objective + coordinated rollout/diversity scaling + a transformer built specifically for humanoid embodiment, rather than scaling any one axis in isolation.

Extension Opportunities:

Extend the Humanoid Transformer with multimodal conditioning (vision/language) to enable instruction-following whole-body control from natural language commands.
Investigate scaling laws quantitatively by sweeping model size, on-policy rollout quantity, and reference motion diversity to derive Chinchilla-style compute-optimal recipes for BFMs.
Adapt the global-frame motion tracking paradigm to manipulation-heavy tasks (loco-manipulation) by integrating object-centric reference trajectories alongside whole-body keypoints.

Replicability: The abstract does not mention code or dataset release. Reproduction would require a high-fidelity humanoid simulator (e.g., Isaac Gym/MuJoCo), a large motion capture corpus (likely AMASS-scale), substantial GPU compute for on-policy RL rollouts at scale (likely hundreds to thousands of GPU-hours), and access to a physical humanoid platform for sim-to-real validation.

Research Gaps:

No explicit scaling law formulation (e.g., loss vs. compute/data curves) is derived — the recipe is empirical.
Limited discussion of how the approach handles contact-rich manipulation or dexterous hand control beyond whole-body locomotion tracking.

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 UAV tracking policies maintain target state and recover during sustained visual occlusion in complex urban environments, where existing VLA policies degrade catastrophically when line-of-sight is blocked by buildings, vegetation, or roadside objects?

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, specifically designed to handle sustained occlusion in urban environments. It uses a four-stage training recipe (spatial pretraining, curriculum SFT, chain-of-thought, closed-loop RL) and outperforms OpenVLA by ~34-35% ADE reduction and up to 29.8% success rate improvement.

Key Results: CosFly-VLA-0.8B reduces open-loop Average Displacement Error (ADE) by 34.1% on seen-test and 35.3% on unseen-test relative to OpenVLA. 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 dataset, three-stage curriculum SFT (natural + hard/long-occlusion), CoT reasoning, and closed-loop RL with multi-component reward (stand-off, grounding, collision avoidance, task success).

Key Findings:

Structured joint prediction of grounding + visibility + actions substantially outperforms action-only VLA baselines during occlusion
Occlusion-focused curriculum learning with CoT recovery traces is critical before RL fine-tuning
Closed-loop RL yields large gains on seen distributions (29.8% SR) but marginal generalization to unseen environments (2.5%), exposing a distribution-shift limitation

Technical Novelty: The combination of (1) structured prediction interface jointly outputting grounding + visibility + actions, (2) UAV-view depth/distance-aware spatial pretraining at 500k scale, (3) explicit occlusion-focused curriculum with CoT recovery reasoning traces before RL, is new. Prior VLA work (OpenVLA) treats action generation without explicit visibility estimation or occlusion-recovery reasoning.

What's New: First VLA policy that treats visibility estimation as a first-class output and trains explicit occlusion-recovery reasoning, combined with UAV-specific 3D spatial pretraining — going beyond visible-frame imitation to spatially grounded closed-loop control.

Extension Opportunities:

Extend to multi-target tracking or coordinated swarm scenarios where multiple UAVs share occluded state estimates via inter-agent communication
Integrate active perception — let the policy generate viewpoint-change actions specifically to disambiguate occluded targets rather than passively predicting through occlusion
Replace the oracle state history assumption with learned memory / world-model-based state estimation for fully autonomous long-horizon deployment

Replicability: Abstract mentions no code release. Reproduction would need: OpenVLA-scale base (0.8B params), 500k spatial pretraining pool, UAV simulator for closed-loop RL, and multi-stage pipeline (CPT + 3-stage SFT + CoT + RL). Estimated compute: multi-GPU (8x A100) for weeks; UAV sim likely required (AirSim/Isaac).

Research Gaps:

Weak unseen-environment generalization after RL (only 2.5% SR gain) suggests closed-loop optimization overfits to training distributions
Evaluation relies on a shared oracle state history rather than fully self-consistent perception, leaving real-world deployability under sensor noise 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 failures in large-scale AI clusters be reliably predicted in time, or should reliability assessment be reframed as a ranking problem given the stochastic nature of failures like Double Bit Errors (DBEs) and GPU Lost events?

Summary: The paper argues that predicting exact GPU failure timing is fundamentally hard due to stochastic, low-SNR telemetry, and instead proposes HeaRank, a Learning-to-Rank framework that ranks nodes by relative failure risk. Deployed on a production cluster with thousands of GPUs, HeaRank captures 64% of failures in the top 5% of ranked nodes versus 21% for the incumbent system, with an AUC of 0.83.

Key Results: On a production cluster with thousands of GPUs, HeaRank achieves AUC 0.83, outperforming heuristic baselines and state-of-the-art ranking algorithms. In online deployment, it captures 64% of future failures within the top 5% of ranked nodes, versus only 21% for the incumbent production system — a ~3x improvement in top-k recall. The authors also empirically demonstrate strong stochasticity and low signal-to-noise ratios in GPU telemetry that undermine time-based prediction.

Key Findings:

Major GPU failures (DBEs, GPU Lost) exhibit strong stochasticity and low signal-to-noise in time-series telemetry, making conventional predictive maintenance ineffective
Reframing reliability as a ranking rather than prediction problem yields a ~3x improvement in top-5% failure capture (64% vs 21%) over the production baseline
Stable historical failure patterns are more informative than short-window telemetry signals for identifying risky nodes

Technical Novelty: Paradigm shift from absolute failure-time prediction to relative risk ranking via Learning-to-Rank on stable historical failure patterns, rather than noisy time-series telemetry. Prior predictive-maintenance work in HDD/CPU domains assumes learnable temporal precursors; HeaRank explicitly rejects that assumption for GPUs and reframes the task.

What's New: Most GPU/hardware reliability literature focuses on time-to-failure prediction using time-series models. This work is the first (in its domain) to explicitly demonstrate the limits of that framing for GPUs and to apply Learning-to-Rank at cluster scale for proactive maintenance and scheduling.

Extension Opportunities:

Integrate HeaRank scores directly into distributed training schedulers (e.g., Kubernetes/Slurm/Ray) as a risk-aware placement constraint that avoids top-percentile risky nodes for long synchronous jobs
Extend the LTR framework with checkpointing policies — dynamically adjust checkpoint frequency per node based on relative risk rank to minimize wasted compute from failures
Generalize the ranking approach to heterogeneous accelerators (TPUs, AMD MI300, custom ASICs) and multi-vendor clusters, testing whether stable historical failure patterns transfer across hardware generations

Replicability: The abstract does not mention released code or datasets — telemetry is from a production cluster, which is typically proprietary. Reproducing the exact results would require access to a comparable multi-thousand-GPU cluster with DBE/GPU-Lost telemetry history; the LTR method itself is standard and reimplementable on public rankings benchmarks with modest compute (single GPU/CPU for training).

Research Gaps:

No evaluation of how ranking-based risk scores integrate with actual scheduling/checkpointing decisions and their end-to-end impact on training job completion time or cost
Limited to a single production cluster/vendor — generalization across GPU generations, workloads, and datacenter environments remains 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 dynamically and flexibly distribute entanglement resources to multiple users while supporting diverse protocols over deployed fibre infrastructure, rather than relying on static point-to-point links?

Summary: The paper demonstrates a metropolitan-scale entanglement-based quantum network where a quantum reconfigurable optical add-drop multiplexer (q-ROADM) dynamically distributes polarisation-entangled photon pairs to six users over deployed fibre, supporting programmable full-mesh, partial-mesh, and sliced topologies. Beyond 150+ hours of stable operation, it showcases multi-protocol capability including a Secure Inaugural Authentication-Transfer scheme combined with network flooding for secure user onboarding.

Key Results: Demonstrated a metropolitan-scale entanglement 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 and metropolitan fibre. Achieved stable six-user full-mesh operation for over 150 hours, showed programmable full-mesh, partial-mesh, and sliced sub-network configurations, and demonstrated Secure Inaugural Authentication-Transfer (SIAT) combined with network flooding across multiple paths for secure user onboarding.

Key Findings:

A single q-ROADM plus one broadband entangled source can serve six users in a fully reconfigurable full-mesh, partial-mesh, or sliced topology over deployed fibre
The network sustained stable six-user full-mesh operation for more than 150 hours, showing practical viability for real-world deployment
The same physical infrastructure supports heterogeneous quantum protocols (e.g., SIAT + network flooding for secure onboarding), proving service-oriented quantum networking is feasible

Technical Novelty: The q-ROADM as a programmable wavelength-routing element for entanglement distribution — unlike prior static WDM entanglement networks, this enables on-demand reallocation of entangled photon pairs across full-mesh, partial-mesh, and sliced topologies from a single broadband source, plus multi-protocol support (QKD + SIAT authentication) on shared infrastructure.

What's New: Moves entanglement distribution from static, point-to-point demonstrations toward SDN-like dynamic reconfigurability using a q-ROADM, enabling network slicing and multi-protocol coexistence on shared quantum infrastructure — a paradigm borrowed from classical optical networking but adapted for polarisation-entangled photons.

Extension Opportunities:

Scale the q-ROADM architecture beyond 6 users by cascading multiplexers or integrating on-chip photonic switches to test 10+ user meshes
Integrate quantum memories at nodes to enable entanglement swapping and extend beyond metropolitan (~50km) distances toward intercity links
Build an SDN-style orchestration layer that dynamically reconfigures q-ROADM topology based on real-time traffic demands and link fidelity telemetry

Replicability: No code/data availability mentioned in the abstract. Reproduction requires substantial physical infrastructure: a broadband polarisation-entangled photon source, a reconfigurable ROADM adapted for quantum signals, deployed metropolitan fibre (multi-km), and six-user detector stations with polarisation analysis — likely a multi-institution testbed effort.

Research Gaps:

Scalability beyond 6 users and beyond metropolitan distances is not addressed; longer-range operation likely requires quantum repeaters/memories not integrated here
No standardized control-plane or orchestration protocol for dynamically negotiating quantum network slices across administrative domains is proposed

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 link microscopic material origins of decoherence in superconducting quantum circuits, beyond the dominant two-level system (TLS) paradigm, to enable targeted mitigation?

Summary: The paper uses scanning gate microscopy on live superconducting circuits to identify Coulomb blockade and microwave-driven charge tunneling in metallic grains as a widespread, previously unrecognized source of decoherence. These grain-based defects are shown to be as common and as damaging as TLS defects but arise from a distinct mechanism, challenging the dominant TLS-limited-coherence paradigm and suggesting grain elimination during fabrication as a mitigation path.

Key Results: Using scanning gate microscopy on live superconducting circuits, the authors identified a previously unrecognized decoherence mechanism from Coulomb blockade and microwave-driven charge tunneling in metallic grains. They characterized multiple defects across different devices, demonstrating these defects are as common and as debilitating as TLS defects, while arising from a fundamentally different physical mechanism. Conventional characterization techniques would misattribute this loss to microwave-power-independent processes.

Key Findings:

Metallic grains in thin-film superconducting devices host Coulomb blockade defects that cause microwave-driven charge tunneling and decoherence
These grain defects are comparable in prevalence and severity to TLS defects but originate from a fundamentally different physical mechanism
Standard characterization techniques misattribute this loss to microwave-power-independent processes, obscuring the true origin

Technical Novelty: First application of scanning gate microscopy on live superconducting circuits to spatially localize decoherence sources, revealing Coulomb blockade in metallic grains as a distinct microwave-driven loss mechanism separate from the TLS paradigm.

What's New: Directly localizes and identifies a non-TLS microscopic decoherence source using scanning gate microscopy on operational superconducting circuits, challenging the prevailing assumption that TLS defects dominate coherence limits.

Extension Opportunities:

Develop fabrication process modifications (e.g., improved lithography, annealing, or alternative deposition methods) that systematically eliminate metallic grains and quantify the resulting coherence improvements
Extend scanning gate microscopy protocols to map metallic-grain defect densities across wafers as a QA tool for qubit fabrication lines
Build power-dependent loss models that jointly account for TLS and Coulomb-blockade grain tunneling to enable correct attribution during standard qubit characterization

Replicability: No code/data availability is mentioned in the abstract. Reproduction requires a specialized cryogenic scanning gate microscopy setup coupled to operational superconducting qubit devices — significant experimental infrastructure (dilution refrigerator, cryogenic SGM, qubit fabrication), not compute-bound.

Research Gaps:

No quantitative fabrication protocol yet demonstrated for reliably eliminating metallic grains at scale
Lack of standard characterization methodology that can distinguish grain-Coulomb-blockade loss from TLS loss in routine qubit benchmarking

⚡ 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: Can molecular beam epitaxy (MBE) produce titanium nitride (TiN) films with sufficient crystalline quality to yield ultra-high-Q superconducting microwave resonators, enabling a path toward epitaxial Josephson junctions for high-coherence qubits?

Summary: The authors use molecular beam epitaxy to grow (111)-oriented TiN films on c-plane sapphire with record structural quality (18 arcsec XRD rocking curve FWHM) and demonstrate CPW resonators with internal Q > 10^6 at single-photon powers and > 2×10^7 at high powers. They also identify sub-surface sapphire defects as the limiting source of epitaxial TiN defects, opening a path toward fully crystalline Josephson junctions for high-coherence superconducting qubits.

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 (3μm/6μm/3μm gap/strip/gap) demonstrated internal quality factor Qi > 10^6 in the single-photon limit (⟨n⟩ ~ 1) at 5.8 GHz and 10 mK, rising to Qi > 20×10^6 at ⟨n⟩ ~ 10^6. Sub-surface defects in sapphire substrate were identified as nucleation sites for structural defects in the epitaxial TiN.

Key Findings:

MBE TiN on c-plane sapphire achieves 18 arcsec X-ray rocking curve FWHM, the lowest reported for TiN
Single-photon internal quality factor Qi > 10^6 at 5.8 GHz and 10 mK, rising above 2×10^7 at high photon numbers
Sub-surface defects within the sapphire substrate — not the film growth process — nucleate structural defects in the epitaxial TiN layer

Technical Novelty: First demonstration of MBE-grown TiN on sapphire with record-low 18 arcsec rocking curve, combined with the first observation that sub-surface sapphire defects (not the TiN growth itself) are the dominant nucleation source of structural defects in epitaxial TiN. Prior TiN resonator work relied on reactive sputtering or ALD with lower crystallinity.

What's New: Combines the highest-crystallinity TiN films yet reported with ultra-high-Q resonator performance, and provides the first identification of sub-surface substrate defects as the structural bottleneck, shifting the materials-improvement focus from film deposition to substrate preparation.

Extension Opportunities:

Develop epitaxial Josephson junctions using this MBE TiN as the base electrode with a crystalline barrier (e.g., epitaxial AlN or MgO) to build high-coherence transmon qubits with reduced TLS loss
Investigate alternative substrate preparation techniques (annealing, chemical-mechanical polishing, or buffer layers) to eliminate the sub-surface sapphire defects that nucleate TiN structural defects
Benchmark these MBE TiN resonators against sputtered/ALD TiN and Ta/Nb films in full qubit devices to quantify T1/T2 improvements from the crystalline interface

Replicability: No code/data availability mentioned in the abstract. Reproduction requires an MBE system with nitrogen plasma source, c-plane sapphire substrates, cleanroom lithography for CPW fabrication, and a dilution refrigerator (10 mK) with low-noise microwave measurement setup for single-photon resonator characterization — significant capital equipment (~$1M+) but standard for superconducting qubit labs.

Research Gaps:

Substrate preparation strategies to eliminate sub-surface sapphire defects are not addressed
The link between crystalline TiN quality and full qubit coherence (T1, T2) remains to be experimentally established

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 nature of the dominant electrically active defect (D*)?

Summary: The paper links CVT growth conditions to defect populations in CrSBr, identifying a dominant electrically active defect D* via conductive AFM and suppressing its density by up to 10x through sulfur/bromine-rich precursors and lower absolute growth temperatures. DFT and thermodynamic modeling attribute D* to a sulfur-related vacancy complex, providing an actionable recipe for growing higher-quality CrSBr crystals.

Key Results: Using conductive atomic force microscopy (CAFM), the authors identified a characteristic defect D* and reduced its concentration by up to an order of magnitude via optimized CVT growth. Sulfur- and bromine-rich precursor stoichiometry plus lower absolute growth temperatures (with fixed temperature gradient) suppressed defect formation. Thermodynamic modeling and DFT calculations indicate D* is most consistent with a sulfur-related vacancy complex rather than an isolated point defect.

Key Findings:

A characteristic electrically active defect D* in CrSBr can be resolved and quantified with conductive AFM.
Sulfur- and bromine-rich precursor stoichiometry combined with lower absolute growth temperatures (at fixed gradient) reduces D* density by up to an order of magnitude.
DFT plus thermodynamic modeling indicates D* is a sulfur-related vacancy complex, not an isolated point defect.

Technical Novelty: First systematic mapping of CVT growth parameters (precursor stoichiometry, absolute temperature at fixed gradient) to defect densities in CrSBr, combined with a CAFM-resolved electronic fingerprint (D*) whose atomic identity is pinned down by DFT as a sulfur-vacancy complex rather than a simple point defect.

What's New: Prior CVT-grown CrSBr studies lacked systematic defect control; this work is the first to correlate growth parameters with a specific, atomically-assigned defect species and demonstrate order-of-magnitude tunability.

Extension Opportunities:

Correlate the reduced D* density with magneto-transport and magneto-optical measurements to quantify how defect suppression improves magnon coherence, exciton linewidths, or spin-lattice coupling.
Apply the same CAFM + DFT + thermodynamic-modeling workflow to other emerging 2D magnetic van der Waals crystals (e.g., CrPS3, NiPS3, CrI3) to establish a general defect-engineering pipeline.
Fabricate devices (tunnel junctions, spin-FETs) from the low-D* crystals and benchmark performance/yield vs standard CrSBr to demonstrate a direct payoff of growth-controlled defect suppression.

Replicability: The abstract does not mention code or data release. Reproduction requires a CVT growth setup with controlled precursor stoichiometry, a CAFM instrument, and moderate compute for DFT (defect-formation energies in a magnetic 2D system — typically hundreds to low-thousands of CPU-hours on a small HPC cluster). Wet-lab access is the main gating factor, not compute.

Research Gaps:

No direct measurement in the abstract of how D* suppression translates to improved magnetic, excitonic, or transport device metrics.
The exact atomic structure of the sulfur-vacancy complex (composition, charge state, spin) remains inferred from DFT rather than directly imaged.

🔥 GitHub Trending

1. Emmimal/memory-decay-engine

⭐ 2 stars | Python

A pure-Python, zero-dependency memory decay engine for AI agent context — implements Ebbinghaus forgetting-curve retention with usage-based reinforcement, benchmarked against naive recency-window prun

ai-agents ai-memory artificial-intelligence context-window ebbinghaus forgetting-curve

2. ozgursntrk/ddpm-ddrm-fmri-superres

⭐ 2 stars | Jupyter Notebook

Diffusion-based super-resolution for 7T fMRI-EPI, with an MR-realistic k-space degradation operator for DDRM

ddpm ddrm diffusion-models fmri mri pytorch

3. Anjali-creater/House-Price-Prediction

⭐ 1 stars | Python

Machine Learning web application that predicts house prices using Python, Scikit-learn, and Streamlit.

data-science house-price-prediction linear-regression machine-learning ml numpy

4. zahidhameed322578-maker/HandwrittenDigitRecognizer

⭐ 1 stars | Python

Handwritten Digit Recognition using CNN (TensorFlow & Python)

artificial-intelligence cnn computer-vision deep-learning keras machine-learning

5. prateek2k5/Digit_Prediction_App2.0

⭐ 1 stars | Jupyter Notebook

A CNN-based handwritten digit recognition web application built with TensorFlow, OpenCV, and Streamlit. Supports image upload, drawing canvas, and camera input with advanced image preprocessing.

ai cnn computer-vision deep-learning handwritten-digit-recognition image-processing

6. Ciprian-LocalPulse/vitagraph

⭐ 1 stars | Python

Decentralized Bio-Signal Knowledge Graph for Longevity & Precision Medicine. An open-source research initiative integrating multi-omic data, real-time biometrics, and AI to map human health trajectori

artificial-intelligence bioinformatics biometrics computational-biology data-science digital-twin

7. shaheen-shaik25/churn_classification_ANN

⭐ 1 stars | Jupyter Notebook

Customer Churn Prediction using Artificial Neural Networks (ANN) to identify customers likely to leave a business using deep learning techniques.

artificial-neural-networks classification customer-churn-prediction data-science deep-learning keras

8. devtantai-coder/nexus

⭐ 1 stars | Python

🧠 Nexus AI - Open Source Vietnamese LLM. Multi-Head Latent Attention (MLA), Mixture-of-Experts 64 experts, FP8 Quantization bằng Triton, YaRN RoPE 128K+ context. Mô hình ngôn ngữ Transformer mã nguồn

artificial-intelligence attention-mechanism cuda deep-learning fp8 large-language-model

9. snehkris/IEEE-ISI3P-AI-Based-Pranayama-Assessment

⭐ 1 stars | Jupyter Notebook

A research project for real-time Pranayama assessment using breathing audio analysis, pose estimation, and machine learning.

ai audio-processing breathing-pattern breathinganalysis computer-vision deep-learning

10. Omer-Hussien/orion-ai

⭐ 1 stars | Unknown

AI-powered executive assistant for intelligent productivity and task automation.

ai assistant automation calender fastapi flutter

11. SergiiRudniev/cerberus-colony-lab

⭐ 1 stars | Python

Neuroevolutionary multi-agent colonies in Minecraft.

artificial-life emergent-communication machine-learning minecraft mineflayer multi-agent-systems

12. StanleyChanH/PixelAlgo

⭐ 1 stars | JavaScript

极致直观的机器学习/深度学习算法可视化与实战平台。硬核拆解 20+ 核心算法（含 Transformer, Diffusion, GAN, GCN），提供“手写底层推导”与“主流框架调用”双版本代码，配合沉浸式 Web 交互动画，让算法不仅能跑，更能被“看见”。

algorithm-visualization computer-vision data-science deep-learning echarts from-scratch

13. Sathyaaa27/Brain-Tumor-Detection

⭐ 1 stars | Python

A Deep Learning Project for Detecting Barin Tumor from MRI images using Convolutional Neural Networks (CNN).

braintumorclassification braintumordetection classification cnn computer-vision deep-learning

14. manishklach/attnres-kernel-lab

⭐ 1 stars | Python

Hydra-2P, KDA, and CADR: fused PyTorch/Triton kernels for routing across model depth and sequence time.

attention-residuals autograd cross-axis-routing cuda deep-learning delta-routing

15. like041123/arch-vision-skill

⭐ 1 stars | Unknown

Vibe coding 架构图生成与代码库理解

ai-agent architecture llm mermaid prompt-engineering vibe-coding

Generated by Research Pulse on 2026-07-18 06:05