2026-08-07·Daily digest

🔬 Research Pulse

Daily Digest

August 07, 2026

🤖 AI

🧠 LLMs

1. Tracing the Heart: An Evidence-Linked Pipeline for Heart-Failure Feature Engineering

Authors: Soorya Ram Shimgekar, Michelle Hu, Dorisa Shehi... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can we automate heart-failure EHR feature engineering in a way that is auditable, evidence-linked, and grounded in clinical guidelines, given that feature engineering consumes 39-45% of data scientists' workload and existing rule-based/LLM approaches lack maintainability and provenance?

Summary: The paper introduces nMAS, a multi-agent LLM pipeline for automated, evidence-linked feature engineering on heart-failure EHR data, producing 132 structured and 70 aggregated rubric-scored features with provenance and audit. Adding the aggregated features materially boosts phenotyping AUROC for both HFrEF and HFpEF, and an independent rubric assessment scores the features at 81.5% of max.

Key Results: nMAS was evaluated on 500 dummy patient records from 9 EHR source tables, generating 132 structured and 70 rubric-scored aggregated features. Adding aggregated features improved held-out AUROC from 0.895→0.963 for HFrEF and 0.870→0.910 for HFpEF phenotyping. An independent LLM-based rubric scored features at 81.5% of maximum points for evidence support and methodological soundness.

Key Findings:

Aggregated rubric-grounded features raised HFrEF phenotyping AUROC from 0.895 to 0.963 and HFpEF from 0.870 to 0.910
The pipeline generated 132 structured plus 70 aggregated features from 9 EHR source tables while maintaining provenance and structural integrity
Independent LLM-based rubric evaluation scored the generated features at 81.5% of maximum for evidence support and methodological soundness

Technical Novelty: A multi-agent LLM system that couples feature engineering with explicit evidence provenance and rubric-grounded scoring, plus a restricted-LLM auditor for structural integrity and compliance checks — going beyond prior rule-based or single-LLM approaches that lacked traceability.

What's New: Unlike prior rule-based systems or monolithic LLM prompts, nMAS combines multi-agent orchestration with explicit clinical rubrics, evidence linking, and a restricted-LLM auditor — making the feature engineering process both automated and traceable back to guideline evidence.

Extension Opportunities:

Adapt the nMAS pipeline to other chronic diseases with guideline-based reasoning (e.g., diabetes, CKD, COPD) by swapping the clinical rubric layer
Perform external multi-institutional validation on real (non-dummy) EHR cohorts like MIMIC-IV or eICU to test generalization
Add a human-in-the-loop clinician review interface where cardiologists can accept/reject/edit rubric-scored features, feeding corrections back into the agent's prompting

Replicability: Abstract does not mention public code or dataset release; evaluation used 500 dummy patient records so raw data may not be shareable. Compute requirements are modest — multi-agent LLM inference over hundreds of records is feasible on standard API access without specialized hardware.

Research Gaps:

Evaluation was single-institution and used dummy patient records rather than real multi-site EHR data, leaving external validity unproven
The rubric grounding is heart-failure specific; generalization of the agent framework to other diseases and to less-structured guideline domains is untested

2. TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories

Authors: Yunjia Qi, Zehua Yin, Xintong Shi... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can we accurately identify the earliest critical error step in long-horizon LLM agent trajectories, given that (1) evidence for judging steps is scattered across distant context, and (2) failed trajectories contain multiple local errors with varying downstream impact — only some of which actually cause final failure?

Summary: TrajDebug is an error-lifecycle tracing framework that finds the critical (earliest causally-responsible) error in failed long-horizon LLM agent trajectories using multi-granularity history compression and evidence-based identification, distinguishing recovered local errors from truly failure-causing ones. The authors release TrajErrBench, a 486-trajectory benchmark from Tau2Bench and SWE-Bench Pro, and show the method beats baselines while its diagnoses actionably improve downstream agent success.

Key Results: TrajDebug achieves best overall performance vs existing baselines on critical error detection across diverse agent benchmarks. Authors constructed TrajErrBench: 486 manually annotated failed trajectories drawn from Tau2Bench (tool-use) and SWE-Bench Pro (coding). Application studies show diagnoses provide actionable feedback that improves downstream agent success rates (specific deltas not stated in abstract).

Key Findings:

Distinguishing 'local errors that got recovered' from 'errors that terminally caused failure' is a distinct problem from step-level error classification, and requires tracing resolution status across the trajectory
Multi-granularity history compression is necessary because relevant evidence for a step's correctness is scattered across distant instructions, observations, and prior context
Critical-error diagnoses translate to measurable downstream agent improvement, suggesting error attribution is not just a diagnostic tool but a training/feedback signal

Technical Novelty: Two coupled innovations: (1) multi-granularity history compression tailored to error-discovery (versus generic summarization), and (2) explicit error-lifecycle tracing — tracking each candidate error's resolution status and terminal impact — which lets the system distinguish locally-wrong-but-recovered steps from truly critical failures. Prior baselines treat error localization as single-step classification without lifecycle context.

What's New: First framework to formalize error 'lifecycle' (introduction → resolution status → terminal impact) rather than treating agent errors as isolated per-step judgments. Also fills a benchmark gap: TrajErrBench appears to be the first sizable manually-annotated dataset for critical error attribution in realistic tool-use + coding agent trajectories.

Extension Opportunities:

Integrate TrajDebug diagnoses into an online self-correction loop where the agent replans from the identified critical step rather than restarting, reducing token cost on retries
Extend the error-lifecycle tracing framework beyond Tau2Bench/SWE-Bench Pro to multi-agent collaborative trajectories where error attribution must span across agents
Use the 486-trajectory benchmark to train a specialized small-model critic (distilled from the multi-granularity compression + evidence-based identification pipeline) for cheap runtime monitoring

Replicability: Authors state code and data will be released. TrajErrBench (486 annotated trajectories) is the reproducibility asset. Compute requirements not specified but inference-time framework over existing LLMs suggests modest cost — likely a few dollars of API calls per trajectory analyzed, no training required for the core method.

Research Gaps:

No mention of real-time / online critical error detection during trajectory execution — the framework appears post-hoc on completed failed trajectories
Benchmark is bounded to Tau2Bench and SWE-Bench Pro; generalization to open-ended web agents, multi-agent systems, or long-running background agents is untested

3. RRC: Unlocking Generative Reward Models in LLM Reinforcement Learning via Ranking-Based Reward Construction

Authors: Chenglong Wang, Ziming Zhu, Yifu Huo... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: Why do generative reward models (GRMs), despite excelling at response ranking, underperform when used as reward signals in RL fine-tuning of LLMs, and how can their comparative strength be converted into usable RL signals?

Summary: RRC addresses the underuse of generative reward models in RL by constructing scalar rewards from rankings rather than raw scores, via self-competitive ranking among sampled responses and anchor-guided ranking against a small reference set. This preserves the comparative signal GRMs are trained on and consistently improves RL training on chat and reasoning benchmarks.

Key Results: The authors identify a mismatch: GRMs are comparative/pairwise while RL algorithms (e.g., PPO/GRPO) expect scalar rewards. They propose RRC with two strategies (self-competitive ranking over sampled responses, and anchor-guided ranking against a small reference set) and show on open-ended chat and reasoning benchmarks that RRC yields consistent gains over existing reward construction methods when using generative reward models in RL. Specific numeric deltas are not quoted in the abstract.

Key Findings:

The scalar-vs-comparative mismatch — not judge quality — is the primary reason GRMs underperform in RL
Self-competitive ranking within a sampled group turns GRM comparisons directly into GRPO-style relative advantages
Anchor-guided ranking with a small reference pool scales GRM-based RL without quadratic judge calls, and generalizes across chat and reasoning tasks

Technical Novelty: Prior work either uses GRMs only for offline ranking/DPO-style preference learning, or converts a single GRM verdict into a scalar for RL. RRC is novel in explicitly deriving RL scalar rewards from rankings over a set — either self-competitive (rankings among the policy's own samples in a group) or anchor-guided (rankings against a small fixed reference pool) — preserving the comparative semantics GRMs are trained on.

What's New: First systematic framing of the GRM-to-RL gap as a comparative/scalar mismatch, plus a ranking-based reward construction that is drop-in for standard RL algorithms and works with either intra-batch or anchor-set comparisons.

Extension Opportunities:

Combine RRC with process reward models (PRMs) so step-level ranking signals guide reasoning-trace RL rather than only final-response ranking
Investigate anchor-set curation strategies (adaptive/curriculum-selected anchors, difficulty-matched anchors) to make anchor-guided ranking robust across domains like code and math
Apply RRC to multi-modal or tool-use RL settings where scalar reward design is especially brittle, testing whether pairwise judgments from a VLM judge scale similarly

Replicability: Code is released at https://github.com/wangclnlp/RRC. Compute is not specified in the abstract but is implicitly RL fine-tuning of an LLM policy with a GRM as judge — typically multi-GPU (8x A100/H100-class) for 7B-scale models over chat + reasoning benchmarks. Anchor-guided variant is explicitly designed to be scalable with a small reference set, reducing judge-call cost vs full pairwise self-competition.

Research Gaps:

No abstract-level evidence on how anchor set size, composition, or drift affects long-horizon RL stability
Unclear whether RRC transfers to agentic/tool-use or multi-turn settings where per-response ranking is harder to define

🦾 ROBOTICS

1. $ω$-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation

Authors: Zhe Li, Zhenzhe Zhang, Yangyang Wei... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can humanoid robots perform concurrent loco-manipulation (moving, balancing, and manipulating simultaneously) as a single coordinated behavior, rather than decomposing locomotion and manipulation into separate policies as existing humanoid systems and arm-centric/video-centered world-action models do?

Summary: ω-0 is a latent predictive world-action model that unifies whole-body locomotion and manipulation for humanoid robots by predicting compact future visual embeddings alongside diffusion-based action latents, rather than reconstructing future video or treating arms/legs separately. Trained on the new 40+ hour ω-HOME dataset, a single model outperforms IL, VLA, humanoid, and WAM baselines on 11 real-world household tasks.

Key Results: Demonstrated on 11 real-world household tasks that a single ω-0 model produces smooth manipulate-while-moving behaviors and consistently outperforms representative imitation learning, VLA, humanoid, and WAM baselines. Introduced ω-HOME dataset: 40+ hours of real-world household humanoid data with synchronized multi-view observations, whole-body SMPL motions, robot states, and action latents.

Key Findings:

Latent future-embedding prediction is a viable lightweight substitute for pixel-level video prediction in world-action models, enabling real-robot execution
A single unified whole-body policy can handle diverse concurrent loco-manipulation tasks without decomposing into separate locomotion/manipulation controllers
Controller-based simulation replay effectively grounds human/public visual-motion priors into robot-executable action latents

Technical Novelty: Replaces pixel-level future video reconstruction (used in typical world models) with compact future observation embeddings as a lightweight predictive objective, coupled with diffusion-based whole-body action generation. Unifies whole-body loco-manipulation (rather than arm-centric) with multi-view egocentric+exocentric RGB+depth inputs, and uses controller-based simulation replay to bridge human motion priors to robot-executable action latents.

What's New: First whole-body (not arm-centric) latent predictive world-action model for real humanoid loco-manipulation; substitutes future video reconstruction with lightweight latent foresight; combines egocentric+exocentric RGB+depth in one model; introduces a large synchronized multimodal humanoid household dataset.

Extension Opportunities:

Extend the latent predictive objective to multi-robot coordination scenarios where multiple humanoids must share visual foresight for collaborative loco-manipulation tasks
Adapt the controller-based simulation replay pipeline to ground large-scale internet video (e.g., cooking/cleaning videos) into robot-executable latents for zero-shot skill acquisition
Add tactile/force modalities alongside RGB and depth inputs to improve contact-rich manipulation while moving (e.g., carrying heavy or fragile objects)

Replicability: Abstract mentions the ω-HOME dataset (40+ hours) but does not confirm public release of code or data. Reproduction would likely require a real humanoid platform, multi-view RGB-D capture setup, SMPL motion capture infrastructure, and substantial GPU compute for diffusion policy training — likely out of reach for individual researchers without institutional support.

Research Gaps:

Existing humanoid policies decompose locomotion and manipulation, preventing truly concurrent whole-body behaviors
Prior world-action models are either arm-centric or video-centered, unsuitable for real-robot whole-body humanoid execution

2. DyPES-VLA: Learning Shared Dynamics Priors and Embodiment-Specific Control for Cross-Embodiment Manipulation

Authors: Junfeng Li, Junjie He, Zhide Zhong... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can a single generalist VLA policy be trained across heterogeneous robot embodiments without extensive manual action-format alignment, while still exploiting shared visual/interaction dynamics priors for cross-embodiment transfer?

Summary: DyPES-VLA is a cross-embodiment vision-language-action model that learns shared dynamics priors via a future-prediction objective on the VLM and decodes actions through an embodiment-specific Mixture-of-Experts head operating in each robot's native action space. This removes the need to manually unify heterogeneous action formats while achieving state-of-the-art generalist performance on LIBERO, RoboCasa-GR1, and RoboTwin 2.0.

Key Results: DyPES-VLA achieves SOTA as a generalist policy: 98.0% success on LIBERO, 59.25% on RoboCasa-GR1, and 89.02% on RoboTwin 2.0, spanning both simulation and real-world evaluations across distinct embodiments.

Key Findings:

A future-prediction pretraining objective on the VLM yields shared query representations that capture object motion, contact, and interaction dynamics, boosting cross-embodiment transfer.
MoE action heads with shared attention but embodiment-specific FFN experts can decode actions directly in native action spaces, bypassing manual action alignment.
The approach delivers SOTA generalist results (98.0% LIBERO, 59.25% RoboCasa-GR1, 89.02% RoboTwin 2.0) spanning simulation and real-world.

Technical Novelty: Two coupled ideas: (1) using a future-prediction objective on cross-embodiment vision-language data to force VLM query tokens to encode object motion, contact, and interaction-induced scene changes as shared dynamics priors; and (2) an embodiment-specific MoE action head that shares attention layers (common temporal action structure) but uses per-embodiment FFN experts to decode in each robot's native action space — eliminating manual action-format harmonization used by prior cross-embodiment VLAs.

What's New: Prior cross-embodiment VLAs either flatten heterogeneous actions into a common schema (costly preprocessing) or underuse cross-embodiment visual dynamics. DyPES-VLA is the first, per the authors, to combine an explicit dynamics-prior pretraining signal (future prediction) with a per-embodiment MoE decoder that natively speaks each robot's action language.

Extension Opportunities:

Add new embodiments (e.g., humanoids, soft grippers, mobile manipulators) by training only new MoE experts while keeping shared dynamics-prior VLM frozen, testing few-shot embodiment onboarding.
Replace the future-prediction objective with richer world-model targets (e.g., 3D point-cloud flow, contact/force prediction) to strengthen physical dynamics priors and measure downstream manipulation gains.
Investigate expert routing dynamics in the MoE head — e.g., analyze expert specialization across kinematic classes and enable dynamic expert sharing across similar embodiments to improve data efficiency.

Replicability: Abstract does not mention a code/data release. Reproducing likely requires a pretrained VLM backbone plus multi-embodiment datasets (LIBERO, RoboCasa-GR1, RoboTwin 2.0), on the order of multi-GPU (8×A100-class) training typical for VLA fine-tuning; exact compute not disclosed.

Research Gaps:

Underutilization of dynamics priors shared across diverse visual and interaction data in existing VLAs.
Reliance on manual preprocessing to align heterogeneous embodiment action spaces into a common format.

3. GeniWorld: A Generalizable Interactive World Model for Robotic Manipulation via Visual Actions

Authors: Chenghao Gu, Hanyang Yu, Jingbo Zhang... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can we build an action-conditioned world model for robotic manipulation that generalizes to out-of-distribution scenes and provides precise action controllability, avoiding the scene-overfitting and weak action-following of prior video-based world models?

Summary: GeniWorld is an interactive, action-conditioned video world model for robotic manipulation that renders the robot's URDF at commanded joint states as a visual action signal, decoupling known embodiment kinematics from learned scene dynamics. Trained on limited fixed-scene data atop a pretrained video generator, it generalizes zero-shot to randomized unseen scenes, serves as a robust policy evaluator, and augments sparse real demos with diverse synthetic rollouts that improve downstream policy performance.

Key Results: GeniWorld, built on pretrained video generative models with URDF-based visual action rendering, demonstrates: (1) superior in-domain video prediction quality when trained only on limited fixed-scene data, (2) zero-shot generalization to highly randomized unseen environments, (3) reliable policy evaluation under environmental perturbations, and (4) downstream policy improvement via synthetic trajectory augmentation from limited real demos. Specific numerical benchmarks (FVD/PSNR/success-rate deltas) are not disclosed in the abstract.

Key Findings:

URDF-rendered visual actions give tighter spatial control than numerical action conditioning and reduce scene overfitting
Fixed-scene training data is sufficient for zero-shot generalization to heavily randomized environments when kinematics are decoupled from dynamics
Autoregressive rollouts run fast enough for closed-loop interaction with policies and teleoperators, enabling use as a live simulator
Synthetic trajectories sampled from the world model improve downstream policy robustness under environmental perturbation

Technical Novelty: The core novelty is representing actions visually — rendering the robot's URDF at commanded joint states and conditioning the video generator on this rendered overlay — rather than feeding raw numerical action vectors. This spatial grounding explicitly decouples embodiment kinematics (deterministic, known) from environmental dynamics (learned), which prior action-conditioned world models like UniPi, RoboDreamer, or IRASim conflate. Paired with an autoregressive predictor at high control frequency for closed-loop use.

What's New: Unlike prior action-conditioned world models that inject numerical actions via cross-attention or token concatenation, GeniWorld converts actions into a visually rendered robot pose using the URDF, giving the video model a spatially grounded, embodiment-aware signal. This is the key mechanism behind its OOD generalization from limited scene data.

Extension Opportunities:

Extend the URDF-based visual action representation to bimanual/humanoid and mobile manipulator embodiments, testing whether decoupling kinematics from dynamics scales to higher-DoF systems
Use GeniWorld as a differentiable world model for model-based RL or MPC rather than only for evaluation and data augmentation — closing the loop for policy learning directly against the imagined rollouts
Combine URDF-projected action conditioning with physics-aware losses or diffusion priors on object dynamics to address the abstract's implicit gap: environmental dynamics modeling is still learned purely from video

Replicability: Abstract does not mention code, model weights, or dataset release. Reproduction would likely require: a pretrained video diffusion backbone (e.g., CogVideoX/Wan-scale, tens of GB VRAM), URDF assets for the target robot, a real-robot manipulation demonstration dataset, and multi-GPU training (order 8+ A100/H100s for fine-tuning video generators). Autoregressive inference for closed-loop control at high frequency is non-trivial and may require distillation.

Research Gaps:

Physical accuracy of predicted object dynamics (contact, friction, deformation) is still learned implicitly from video and not evaluated against ground-truth physics
Long-horizon compounding error in autoregressive rollouts and its impact on the reliability of policy evaluation is not quantified in the abstract

💻 COMPUTE

1. Dual-Faraday-laser-pumped cesium beam clock with $7.7\times 10^{-13}/\sqrtτ$ frequency stability

Authors: Xiaomin Qin, Suyang Wei, Haijun Chen... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can the short-term frequency stability of compact cesium beam clocks be improved beyond the SNR limitations imposed by laser-induced frequency-to-amplitude noise conversion in two-laser optical pumping schemes?

Summary: The authors demonstrate a compact cesium beam atomic clock that uses two Faraday-laser-stabilized optical pump beams to overcome the long-standing SNR limitation of two-laser optical pumping. By exploiting an intracavity FADOF for inherent Cs D₂ alignment and modulation transfer spectroscopy for noise suppression, they achieve 7.7×10⁻¹³/√τ short-term stability — a record for compact Cs beam clocks.

Key Results: Demonstrated a compact dual-Faraday-laser-pumped (DFP) Cs beam clock achieving: (1) Lorentzian laser linewidth of 2.12 kHz, (2) clock SNR of 46,365 in 1-Hz bandwidth, (3) fractional Allan deviation of 7.7×10⁻¹³/√τ, and (4) Hadamard deviation of 7.7×10⁻¹⁵ at 10,000 s — pushing compact Cs beam clock stability into the 10⁻¹³/√τ regime for the first time.

Key Findings:

The atom-referenced laser architecture achieves a Lorentzian linewidth of 2.12 kHz with turnkey robustness suitable for deployable systems
Clock SNR of 46,365 in 1-Hz bandwidth translates to 7.7×10⁻¹³/√τ Allan deviation, entering the 10⁻¹³ regime for compact Cs clocks
Hadamard deviation of 7.7×10⁻¹⁵ at 10,000 s demonstrates strong long-term stability suitable for field precision timing

Technical Novelty: Combines an intracavity Faraday anomalous dispersion optical filter (FADOF) — which self-aligns to Cs D₂ resonances — with modulation transfer spectroscopy for frequency stabilization, producing a low-noise atom-referenced dual-laser pump system that overcomes the frequency-to-amplitude noise conversion bottleneck of prior two-laser optical pumping schemes.

What's New: First demonstration of a compact Cs beam clock reaching 10⁻¹³/√τ stability by resolving the laser-induced FM-to-AM noise conversion problem that has historically capped two-laser optical pumping performance, using a Faraday-filter + MTS approach.

Extension Opportunities:

Integrate the DFP architecture into miniaturized/chip-scale packaging for portable GNSS-denied navigation systems
Extend the Faraday-laser + modulation transfer spectroscopy approach to other alkali species (Rb-87, K) for comparative stability studies
Combine with Ramsey interrogation or optically detected microwave transitions to further push long-term stability below 10⁻¹⁵

Replicability: No code or dataset mentioned in the abstract; replication requires specialized atomic physics hardware (Cs beam tube, Faraday filters with Cs vapor cells, ECDLs, microwave cavity, magnetic shielding). Compute requirements are minimal — this is an experimental hardware demonstration, not a computational one.

Research Gaps:

Long-term stability limits, environmental sensitivity (magnetic/thermal), and aging behavior beyond 10,000 s are not established in the abstract
Trade-off between compactness/SWaP and stability versus lab-scale primary standards is not quantified for deployment envelopes

2. PLoRA: An NDP-Enhanced Pooled-Memory System for Cost-Efficient Multi-LoRA Serving

Authors: Zhongkai Yu, Ohm Rishabh Venkatachalam, Zheng Wang... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can multi-LoRA serving (1000+ adapters on one base model) escape the PCIe/DRAM bottleneck by leveraging emerging pooled-memory fabrics (CXL/NVLink) with near-data processing, rather than continuing to stage adapters from CPU DRAM over PCIe?

Summary: PLoRA rearchitects multi-LoRA serving around pooled memory with near-data processing: adapters and KV cache stay in a CXL/NVLink-attached pool, an NDP unit performs reductions locally, and the GPU pulls only results via native load/store — eliminating PCIe kernel-stop overhead. A cost-model-driven scheduler chooses among 4 LoRA and 2 attention strategies per adapter, delivering 6.6x lower decode latency than S-LoRA on 1000 adapters with <3.4% area overhead.

Key Results: On one H100 serving 1000 adapters, PLoRA achieves the lowest decode latency across every model and workload tested, averaging 6.6x below a real-machine S-LoRA baseline while adding under 3.4% device area. Throughput saturates at just 32 GB/s on short contexts (a quarter of CXL 3.1 bandwidth). Per-GPU adapter-traffic demand falls from 7B to a modeled 1.2T deployment when sharded with tensor parallelism, and the design works unchanged from CXL-class to NVLink-class fabrics.

Key Findings:

Serving 1000 adapters on an H100 achieves 6.6x lower average decode latency than S-LoRA with under 3.4% added device area
Bandwidth is not the bottleneck: throughput saturates at 32 GB/s on short contexts — only a quarter of CXL 3.1's capacity — so surplus bandwidth can be spent buying more pooled capacity rather than speed
The design scales from 7B to a modeled 1.2T deployment and runs unchanged across CXL-class and NVLink-class fabrics because adapter traffic shards naturally with tensor parallelism

Technical Novelty: First system to co-design multi-LoRA serving with pooled memory + NDP: adapters and KV cache live in the pool, only reduced results traverse the link, and the GPU drives access via its own load/store instructions (no kernel stops, no host copies). Combined with a link-parameterized cost model that selects among 4 LoRA and 2 attention execution strategies per adapter, plus a criticality-based GPU cache — none of which exist in PCIe-staged systems like S-LoRA or Punica.

What's New: Prior multi-LoRA systems (S-LoRA, Punica, dLoRA) all stage adapters from CPU DRAM over PCIe, incurring kernel stops and host-driven copies with capacity bounded by DIMM slots. PLoRA is the first to treat pooled memory as the primary adapter store and to push reduction compute into the pool itself, converting the multi-LoRA problem from a PCIe-bandwidth problem into a memory-capacity problem.

Extension Opportunities:

Extend the four-LoRA / two-attention strategy selector into an online RL-based scheduler that adapts to shifting adapter popularity distributions in production traffic
Apply the NDP pooled-memory read-compute interface to other parameter-efficient methods (prefix tuning, IA3, DoRA) or to MoE expert offloading, where the same many-small-tensor access pattern dominates
Prototype the NDP compute unit on a real CXL 3.x device (e.g., using Samsung/SK Hynix CMM-H boards) to validate the modeled area/bandwidth numbers against silicon

Replicability: The abstract does not mention public code or datasets. Reproduction would require either an H100 GPU with a CXL 3.x pooled-memory testbed and a custom NDP device, or a cycle-accurate simulator modeling the memory fabric — substantial hardware/simulation infrastructure well beyond commodity setups.

Research Gaps:

No prior work has explored serving LLM adapter workloads on memory-semantic fabrics with NDP — the paper explicitly identifies this as unexplored territory
Existing systems are architecturally capped by motherboard DIMM slots, blocking the scaling path to tens of thousands of adapters that agent and per-user deployments demand

3. MCHA: A Memory-Centric Hierarchical Architecture for Parallel-Sequential Computing

Authors: Daijing Shi, Hongxiao Zhao, Yihan Fu... Published: 2026-08-05 | Citations: 0 arXiv | PDF

Research Question: How can hardware architectures efficiently execute parallel-sequential workloads (MARL, neuromorphic computing, probabilistic graphical models) that suffer from global buffer saturation and memory-bound bottlenecks due to irregular access patterns centralized to main memory?

Summary: MCHA is a reconfigurable memory-centric hierarchical accelerator co-designed with an event-driven programming model to execute parallel-sequential workloads without saturating global memory. It delivers up to 2456x speedup over A100 GPUs on MARL tasks while shrinking main-memory access from 96% to 5.44%, at only 2.92mm^2 and 115mW in 28nm.

Key Results: MCHA achieves 153.06x to 2456.96x speedup over NVIDIA A100 GPUs on MARL workloads, reduces main memory access from 96% to 5.44%, and when synthesized in 28nm occupies 2.92mm^2 with 115.36 mW power consumption at 200 MHz. Validated via an open-source cycle-accurate simulator across MARL, motor variable control, and Markov random field benchmarks.

Key Findings:

Hierarchical inter-core distributed routing largely eliminates the global buffer bottleneck for irregular access patterns
Event-driven conditional triggers effectively hide data transmission latency in the pipeline for sequential dependencies
Massive 153x–2457x speedup over A100 on MARL with modest area (2.92mm^2) and power (115.36mW) at 200MHz in 28nm

Technical Novelty: Combines a hierarchical inter-core routing fabric (avoiding global buffer bottleneck) with an event-driven conditional-trigger programming model that hides data-transmission latency inside the execution pipeline — specifically co-designed for parallel-sequential (not purely dataflow or SIMT) patterns, unlike prior CGRAs, systolic arrays, or GPU SIMT approaches.

What's New: First architecture to explicitly target the parallel-sequential execution class as a first-class citizen, unifying a hierarchical communication fabric with an event-driven trigger-based programming abstraction — as opposed to treating these workloads as either pure dataflow or GPU-style SIMT.

Extension Opportunities:

Port MCHA to advanced process nodes (7nm/5nm) and evaluate scaling on larger MARL agent populations to test bandwidth-hierarchy limits
Extend the event-driven conditional trigger programming model to support graph neural network inference or sparse transformer attention workloads with similar irregular access patterns
Integrate MCHA as a chiplet/accelerator alongside CPU/GPU hosts and build a compiler pass that automatically partitions PyTorch/JAX programs into parallel-sequential kernels

Replicability: Yes — open-sourced at https://github.com/carabdis/MCHA including a cycle-accurate simulator. Reproduction requires the simulator (likely commodity CPU for functional runs) and standard EDA tools for 28nm synthesis if physical validation is desired; GPU baselining requires an NVIDIA A100.

Research Gaps:

No comparison against other domain-specific accelerators (TPU, neuromorphic chips like Loihi, or CGRAs) — only GPU baseline reported
Silicon validation is absent; results rely on cycle-accurate simulation and post-synthesis estimates rather than measured tape-out data

⚡ ENERGY

1. Pulse-Duration Control of Subcycle Multiband Electron Dynamics Extends the High-Harmonic Cutoff in a Light-Driven Insulator

Authors: Hortense Allegre, Simon V. B. Jensen, Joseph J. Broughton... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can pulse duration and intensity be jointly tuned to control high-harmonic generation (HHG) pathways in solids and extend the extreme-ultraviolet (XUV) cutoff energy beyond limits imposed by decoherence and single-band dynamics?

Summary: The paper shows that jointly tuning laser pulse duration and intensity selects between two distinct HHG pathways in a light-driven insulator: multi-cycle cumulative carrier transfer at moderate intensity vs. subcycle multiband dynamics at few-cycle, high intensity. The latter regime outraces decoherence and extends the coherent XUV harmonic cutoff to 25–50 eV, establishing pulse duration as a first-class control knob for solid-state HHG.

Key Results: The authors experimentally demonstrated pathway-selective HHG control by varying pulse duration (5–29 fs) and intensity (0.8–74 TW/cm²). Many-cycle pulses at ~6 TW/cm² promote cumulative multi-cycle carrier transfer to higher conduction bands, while few-cycle pulses at ~22 TW/cm² drive subcycle multiband dynamics that reach photon energies of 25–50 eV before decoherence suppresses coherent emission.

Key Findings:

Many-cycle pulses (~6 TW/cm²) drive cumulative cycle-by-cycle promotion of carriers into progressively higher conduction bands
Few-cycle pulses (~22 TW/cm²) trigger subcycle multiband dynamics that produce coherent harmonics up to 25–50 eV before decoherence sets in
Pulse duration, not just intensity, is a decisive knob enabling band-structure-guided pulse design for higher-energy XUV sources

Technical Novelty: Prior solid-state HHG work treated intensity as the dominant knob and typically operated in either the multi-cycle or few-cycle regime separately. This paper is the first to systematically map the joint (duration, intensity) parameter space and identify two distinct physical regimes — cumulative multi-cycle interband climbing vs. subcycle multiband coherent emission — with the latter beating decoherence to extend the XUV cutoff.

What's New: Reframes solid-state HHG as a two-dimensional (duration × intensity) control problem and identifies subcycle multiband dynamics as a distinct, decoherence-outrunning route to higher cutoffs — moving beyond intensity-only scaling laws that dominated prior work.

Extension Opportunities:

Apply this pulse-duration/intensity control framework to other wide-bandgap insulators (e.g., diamond, MgO, LiF) to map material-specific cutoff-extension regimes
Combine with waveform-shaped or two-color pulses to further steer subcycle interband pathways and push the XUV cutoff into the soft-X-ray water window
Build a TDDFT or semiconductor Bloch equation simulation pipeline that predicts optimal pulse parameters from a material's band structure, enabling inverse-design of pulses for target harmonic orders

Replicability: The abstract does not mention released code or data. Reproducing the experiment requires a tunable few-to-many-cycle IR laser system (5–29 fs, up to ~74 TW/cm²), an XUV spectrometer covering 25–50 eV, high-vacuum sample handling for the insulator crystal, and TDDFT/SBE simulation compute (typically a modest GPU cluster) for theory support.

Research Gaps:

No quantitative decoherence-time measurement or material-specific model that predicts the crossover between the two regimes
Demonstrated on a single insulator; generality across bandgap, crystal symmetry, and orientation remains open

2. Mask-free fast patterning of organic light-emitting diode pixels using laser-assisted close-space sublimation

Authors: Subhamoy Sahoo, Jain Jose, Mani R... Published: 2026-08-05 | Citations: 0 arXiv | PDF

Research Question: How can OLED display pixels be patterned at micrometer-scale precision without shadow masks while achieving high throughput suitable for large-area AMOLED substrates, overcoming the slow processing times of existing patterning methods?

Summary: The paper introduces a mask-free OLED pixel patterning technique that combines laser heating with close-space sublimation, using a donor substrate pre-patterned with absorber and reflector regions to selectively transfer organic material to a receiver. The authors model the optical and thermal dynamics governing selectivity and demonstrate micrometer-precision pixels with OLED performance matching conventional vacuum thermal evaporation.

Key Results: The authors demonstrated a mask-free laser-assisted close-space sublimation (LA-CSS) process using an engineered donor substrate with patterned absorber/reflector layers, achieving selective one-step or two-step organic transfer with micrometer-scale spatial fidelity. Fabricated OLED devices exhibited optoelectronic performance comparable to control devices produced via conventional vacuum thermal evaporation (VTE), validating the method. Specific numerical benchmarks (efficiencies, resolutions, transfer times) are not disclosed in the abstract.

Key Findings:

Spatially patterned absorber/reflector donor substrates enable selective, mask-free organic transfer under laser-assisted close-space sublimation
One-step and two-step transfer modes both achieve micrometer-scale pixel fidelity suitable for AMOLED displays
Optoelectronic performance of LA-CSS-fabricated OLEDs is comparable to devices made via conventional vacuum thermal evaporation

Technical Novelty: Prior laser-induced thermal imaging (LITI) and radiation-induced sublimation transfer (RIST) methods use uniform donor films and rely on laser scanning to define pixels. This work spatially pre-patterns the donor with absorber and reflector layers, so a broad-area (rather than tightly focused scanning) laser can perform selective transfer in one or two steps — shifting pixel definition from the laser optics to the donor architecture, which promises higher throughput on large substrates.

What's New: Encodes pixel geometry into the donor substrate itself via patterned absorber/reflector layers, replacing scanning-laser or shadow-mask approaches and enabling faster, mask-free patterning without sacrificing resolution or device performance.

Extension Opportunities:

Extend the donor substrate design to enable simultaneous multi-color (RGB) sub-pixel patterning in a single laser pass by encoding wavelength-selective absorbers
Couple the process with in-line optical metrology and closed-loop laser power control to compensate for donor degradation across repeated transfers, enabling roll-to-roll manufacturing
Apply the same donor-engineered LA-CSS approach to transfer perovskite emitters or quantum dot layers, which are notoriously difficult to pattern via shadow masks or photolithography

Replicability: No mention of open code, CAD files, or public datasets in the abstract. Reproduction requires a cleanroom with vacuum deposition (for donor fabrication with patterned absorber/reflector stacks), a close-space sublimation chamber, a suitable pulsed/CW laser source, and OLED device characterization equipment — capital-intensive rather than compute-intensive.

Research Gaps:

Lack of quantitative throughput, resolution limit, and lifetime data (efficiency, luminance, T50) to benchmark against LITI, FMM, and inkjet approaches
No demonstration of full RGB sub-pixel co-patterning, donor reusability, or scaling to Gen-8+ substrate sizes

3. CCAT: Characterization of the first science-grade MKID array for the Prime-Cam 850 GHz module

Authors: Anthony I. Huber, Jordan Wheeler, James Burgoyne... Published: 2026-08-05 | Citations: 0 arXiv | PDF

Research Question: Can a science-grade 850 GHz MKID array meet the sensitivity, uniformity, and yield requirements needed to deploy ~38,000 polarization-sensitive detectors in the Prime-Cam 850 GHz module on FYST, using a novel two-octave design multiplexed via an RFSoC readout?

Summary: The paper reports the fabrication and cryogenic characterization of the first science-grade 850 GHz TiN LEKID array for CCAT's Prime-Cam module on FYST, demonstrating 99% fabrication yield and validating a novel two-octave design that maximizes RFSoC multiplexing. Measured resonator, optical, and noise properties support the module's planned 2027 deployment of ~38,000 polarization-sensitive detectors across three arrays.

Key Results: The first full science-grade TiN LEKID array for the 850 GHz module was fabricated and cryogenically characterized, achieving a fabrication yield of 99%. The team measured resonator frequency mapping, quality factors, optical load sweeps, and noise performance, and reported optical efficiency, sensitivity, and uniformity metrics consistent with the module's on-sky requirements. The design leverages a two-octave bandwidth to maximize multiplexing density on the RFSoC readout, targeting deployment across three arrays totaling ~38k detectors.

Key Findings:

99% fabrication yield on the first full science-grade 850 GHz TiN MKID array
Two-octave resonator design successfully increases multiplexing density on the RFSoC readout
Measured optical efficiency, sensitivity, and uniformity are consistent with the on-sky performance budget for the 850 GHz Prime-Cam module

Technical Novelty: A two-octave lumped-element TiN MKID design that broadens usable resonator bandwidth to maximize channels per RFSoC readout line, applied at 850 GHz — the highest frequency and largest submm MKID count in a single Prime-Cam module to date.

What's New: First demonstration of a science-grade, two-octave TiN LEKID array at 850 GHz, and the largest submillimeter MKID count planned in a single Prime-Cam instrument module.

Extension Opportunities:

Scale the characterization pipeline to the two remaining arrays and quantify wafer-to-wafer uniformity before 2027 deployment
Develop end-to-end simulations tying measured NEP and optical efficiency to expected map-depth for CMB B-mode and [CII] intensity-mapping science cases at 850 GHz
Extend the two-octave multiplexing design and RFSoC firmware to even higher-frequency (>1 THz) MKID modules, potentially enabling submillimeter cameras with >100k pixels

Replicability: No code/data release mentioned. Reproduction requires a superconducting-detector fab line (TiN deposition, lithography), a sub-K cryostat (dilution or ADR), an RFSoC-based multi-tone readout, and submillimeter optical test infrastructure — accessible only to specialized instrumentation labs.

Research Gaps:

Long-term stability, cosmic-ray response, and on-telescope systematics of the two-octave design remain untested
Full three-array integration, focal-plane assembly performance, and end-to-end polarimetric fidelity are not yet characterized

🏥 HEALTHCARE

1. IL-10 rs1800896 polymorphism predicts biochemical remission in IBD patients undergoing biologic therapy

Authors: Michela Helga Falzone, Davide Giuseppe Ribaldone, Martina Buglione... Published: 2026-08-05 | Citations: 0 arXiv | PDF

Research Question: Can cytokine gene SNPs (TNF-alpha, TGF-beta, IL-6, IL-10) predict clinical phenotype and biochemical response to biologic/molecular-targeted therapy in IBD patients, enabling personalized treatment selection?

Summary: A prospective study of 197 IBD patients found that the IL-10 rs1800896 -1082 G>A variant allele independently predicts biochemical remission at 12 months on biologic therapy (adjusted OR 4.15). The work supports cytokine-SNP genotyping as a low-cost stratification tool for personalized biologic selection in IBD.

Key Results: In 197 IBD patients (142 CD, 55 UC), the IL-10 rs1800896 variant allele predicted biochemical remission at 12 months (CRP <5.0 mg/L, fecal calprotectin <250 mcg/g, no steroids) with unadjusted OR 2.15 (95% CI 1.03-4.44, p=0.041) and adjusted OR 4.15 (95% CI 1.49-11.56, p=0.007) in multivariable analysis. 41.0% of the 134 per-protocol patients achieved remission. Secondary findings: IL-6 rs1800795 C allele linked to younger diagnosis age (p=0.049); TNF-alpha rs1800629 A allele more common in CD vs UC (p=0.036).

Key Findings:

IL-10 rs1800896 variant allele carriers had ~4x higher odds of biochemical remission at T12 after multivariable adjustment (p=0.007)
Only 41.0% of per-protocol patients (n=134) achieved combined CRP + calprotectin remission at 12 months, quantifying the substantial biologic non-response burden
TNF-alpha rs1800629 A allele distinguishes CD from UC phenotype (p=0.036), and IL-6 rs1800795 C allele associates with earlier disease onset (p=0.049)

Technical Novelty: First study to link IL-10 rs1800896 -1082 G>A variant specifically to biochemical (CRP + calprotectin) remission at 12 months in a mixed-biologic IBD cohort, with a strong adjusted effect size (OR ~4). Prior pharmacogenomic IBD studies focused mostly on TNF SNPs and anti-TNF response alone.

What's New: Extends IBD pharmacogenomics beyond the well-studied TNF axis by demonstrating IL-10 as the dominant predictor of biologic response, using a stringent dual-marker (CRP + calprotectin) steroid-free biochemical remission endpoint rather than symptom-based clinical scores.

Extension Opportunities:

Build a multi-SNP polygenic risk score combining all four cytokine variants plus clinical covariates into a decision-support tool for pre-treatment biologic selection
Validate the IL-10 rs1800896 finding in a larger, drug-stratified cohort (anti-TNF vs anti-integrin vs anti-IL-23) to determine if the effect is class-specific or universal to biologics
Integrate SNP genotypes with baseline transcriptomic/microbiome data via ML models to improve remission prediction beyond the OR 4.15 achieved by IL-10 alone

Replicability: No code or genotype data availability mentioned in the abstract. Reproduction requires: prospective enrollment of ~~200 biologic-treated IBD patients, TaqMan/PCR SNP genotyping for 4 loci (~~$10-20/sample), 12-month CRP + fecal calprotectin follow-up. Minimal compute — standard logistic regression. Cost is clinical, not computational.

Research Gaps:

Single-center cohort with modest UC representation (n=55) and no drug-class stratification — cannot determine whether IL-10 effect generalizes across anti-TNF, anti-integrin, and anti-IL-23 agents
No functional/mechanistic validation of how the -1082 G>A promoter variant modulates IL-10 expression to drive remission, and no external replication cohort

🔬 MATERIALS

1. Correlated topological-polarization surface states in the narrow-gap insulator FeSb2

Authors: Takahiro Iwagaki, Hideki Matsuoka, Ginta Hoshino... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How can strong electron correlations (3d transition metals) be unified with band topology, given that conventional topological phases require spin-orbit coupling from heavy elements incompatible with strong correlations?

Summary: The authors show that epitaxial FeSb2 thin films host metallic polar surface states of topological-polarization origin that are controlled by bulk 3d electron correlations, evidenced by nonreciprocal transport that switches on below a correlation-driven orbital reconstruction. Electrostatic gating pushes this correlated surface into a ferromagnetic (possibly altermagnetic) phase, establishing topological polarization as a route to correlated topological phases without spin-orbit coupling.

Key Results: Demonstrated in epitaxial FeSb2 thin films that metallic polar surface states of topological-polarization origin emerge, with nonreciprocal surface transport appearing only below the onset temperature of a correlation-driven Fe 3d orbital reconstruction — providing direct evidence of bulk-edge correspondence in a correlated topological system. Electrostatic gating drives the surface across a quantum phase transition into a ferromagnetic (possibly altermagnetic) state.

Key Findings:

FeSb2 thin films host metallic polar surface states arising from topological polarization rather than SOC-driven band inversion
Nonreciprocal surface transport emerges only below the bulk correlation-driven Fe 3d orbital reconstruction temperature, evidencing bulk-edge correspondence in a correlated system
Gate-induced quantum phase transition drives the correlated surface into a ferromagnetic or possibly altermagnetic state

Technical Novelty: Uses topological polarization (bonding-charge-derived polar surface states) as the topological mechanism rather than SOC-driven band inversion, enabling topology in a strongly correlated 3d compound; combines epitaxial thin-film growth of FeSb2 with nonreciprocal transport and electrostatic gating to reveal a correlation-locked bulk-edge correspondence.

What's New: Extends band topology to correlated 3d transition-metal compounds via topological polarization — a SOC-free mechanism — and provides the first transport-level demonstration of bulk-edge correspondence tied to a correlation-driven orbital reconstruction, plus gate-tunable access to altermagnetism.

Extension Opportunities:

Apply topological-polarization design principle to other narrow-gap 3d transition-metal compounds (e.g., FeSi, FeGa3) to search for correlated topological surface states without SOC
Perform spin- and angle-resolved ARPES on gated FeSb2 films to directly confirm the predicted altermagnetic surface state and map its spin texture
Engineer heterostructures pairing FeSb2 surfaces with superconductors to probe whether the correlated polar surface states host unconventional (e.g., topological) superconductivity

Replicability: No code/data availability mentioned in abstract. Reproduction requires an MBE/PLD lab capable of epitaxial FeSb2 thin films, low-temperature nonreciprocal transport setup, and ionic-liquid or solid-electrolyte gating; substantial materials-physics infrastructure rather than compute.

Research Gaps:

Direct spectroscopic (ARPES/STM) confirmation of the polar surface band structure and its correlated character is not established
Microscopic nature of the gated ferromagnetic/altermagnetic phase and its topological classification remains ambiguous

2. Strongly Enhanced Charge-Density Waves and Correlated Insulating State in Atomically Thin 1$T$-TaS$_2$

Authors: Gan Liu, Yulu Liu, Qiling Luo... Published: 2026-08-06 | Citations: 0 arXiv | PDF

Research Question: How does dimensional confinement (thickness reduction to the monolayer limit) modify the charge-density-wave (CDW) transitions, correlated insulating state, and electron-electron interactions in 1T-TaS2, and what mechanism drives thickness-dependent enhancement?

Summary: The paper demonstrates that the three characteristic CDW phases of 1T-TaS2 survive down to the monolayer, but with enhanced transition temperatures, orders-of-magnitude larger sheet resistance, and — uniquely — no hysteresis at the CCDW-NCCDW transition in the monolayer. First-principles calculations attribute this enhancement to strengthened Coulomb interactions arising from reduced out-of-plane screening, elevating the role of nonlocal electron correlations in the 2D limit.

Key Results: Temperature-dependent Raman spectroscopy and electrical transport show that IC, NC, and C-CDW phases persist down to the monolayer, with transition temperatures rising as thickness decreases. Sheet resistance increases by orders of magnitude, carrier localization length drops sharply, and the first-order hysteretic CCDW-NCCDW transition vanishes uniquely in the monolayer. DFT/many-body calculations attribute the enhancement to strengthened Coulomb interactions from reduced out-of-plane dielectric screening, especially in the nonlocal component.

Key Findings:

IC, NC, and C-CDW phases persist all the way to the monolayer, with transition temperatures increasing as thickness decreases
Sheet resistance rises by orders of magnitude and carrier localization length shrinks sharply in the thin limit, indicating a strongly enhanced correlated insulating state
The first-order hysteretic CCDW-NCCDW transition disappears in the monolayer, suggesting a qualitative change in the transition mechanism
Reduced out-of-plane dielectric screening — particularly in the nonlocal Coulomb component — is identified as the microscopic origin of the enhancement

Technical Novelty: First systematic Raman + transport study tracking all three CDW phases down to the true monolayer limit of 1T-TaS2, combined with first-principles calculations isolating the nonlocal Coulomb screening as the microscopic driver — prior work largely focused on few-layer regimes and could not disentangle interlayer coupling from screening effects.

What's New: Extends CDW physics of 1T-TaS2 to the strict monolayer limit and provides a screening-based microscopic explanation for the enhancement, rather than invoking interlayer coupling changes alone; the disappearance of hysteresis in the monolayer is a distinctive experimental signature.

Extension Opportunities:

Fabricate monolayer 1T-TaS2 heterostructures with tunable dielectric environments (hBN vs SrTiO3 vs graphene) to directly test the screening-driven enhancement hypothesis
Use gate-tunable dual-gated devices to modulate carrier density and map the phase diagram of the correlated insulating state vs Mott/CDW crossover in the monolayer
Combine with pump-probe or ultrafast optical spectroscopy to investigate whether the absence of hysteresis in the monolayer enables reversible, non-thermal CDW switching for neuromorphic devices

Replicability: No code/data availability mentioned in the abstract. Reproduction requires a glovebox for air-sensitive exfoliation, a low-temperature Raman setup with variable-temperature stage, electrical transport measurement infrastructure, and DFT + constrained RPA computational resources (~modest HPC cluster).

Research Gaps:

Direct experimental probe of the nonlocal Coulomb interaction strength (e.g., via STM/STS or dielectric-environment engineering) is not yet performed
The nature of the monolayer ground state — Mott insulator vs band insulator vs disorder-driven Anderson insulator — is not fully resolved by transport + Raman alone

🔥 GitHub Trending

1. c-zerooo/AegisAgent

⭐ 3 stars | Python

Autonomous Multi-Agent Security Operations Center (SOC) powered by Agentic AI, FastAPI, and Orchestrated AI Agents.

agentic-ai ai-agents cybersecurity fastapi llm multi-agent

2. psyhetao/human-ai-interaction-psychology

⭐ 2 stars | Unknown

人机交互心理学知识库 | HRI × 社会心理：人如何感知、归因、信任、依赖非人智能体（不限载体），以及这件事对人本身意味着什么。

anthropomorphism awesome-list embodied-ai hri human-ai-interaction human-robot-interaction

3. AubakirovArman/WAL-Ternary-8B

⭐ 2 stars | Python

2.898-BPW Qwen3-8B with direct-packed CPU/CUDA inference

cuda llm-inference low-bit quantization qwen3 ternary

4. Kartikpatil726/loan-bias-auditor

⭐ 1 stars | TypeScript

ML-powered loan approval auditor with fairness metrics, SHAP explanations, and DiCE counterfactuals

counterfactual-explanations docker explainable-ai fairness-ai fastapi machine-learning

5. fallenangel1990/quant-trading

⭐ 1 stars | Python

面向 Python 新手的量化交易完整教程：数据获取、技术指标、回测、多因子、机器学习、深度学习、风控与模拟盘，全部代码可运行并附测试。

backtesting deep-learning factor-model finance machine-learning pandas

6. armin5872/DataLith

⭐ 1 stars | TypeScript

The open-source operating system for AI datasets — import, explore, clean, label, version, augment, and export all in one place.

byok-ai data-engineering data-tooling dataset-management developer-tools local-first

7. DevMasteryHub245/Image-Classification-CNN

⭐ 1 stars | Jupyter Notebook

Image Classification using TensorFlow & CNN | Cats vs Dogs Classification | Deep Learning Project

cats-vs-dogs cnn computer-vision deep-learning image-classification jupyter-notebook

8. alenaktzv/event-driven-lob-research-clean

⭐ 1 stars | Python

Event-driven limit-order-book research pipeline with session-safe features, purged walk-forward validation and queue-aware market-making simulation.

algorithmic-trading-quantitative docker duckdb lightgbm limit-order-book machine-learning

9. Lilbarisx/turkey-microclimate-subspace-clustering

⭐ 1 stars | MATLAB

Unsupervised detection of microclimate zones in Turkey using Subspace Clustering (BDR) on high-dimensional meteorological data.

bdr climatology graph-algorithms graph-theory graph-theory-algorithms machine-learning

10. kartikmorya860-eng/Word-prediction-

⭐ 1 stars | Jupyter Notebook

Developed an LSTM-based prediction model using Python, TensorFlow, and Keras. The project involved data preprocessing, feature scaling, sequence creation, model training, and evaluation to learn seque

deep-learning

11. taesaksit/resnet18-cifar10-pytorch

⭐ 1 stars | Python

PyTorch implementation of ResNet18 on CIFAR-10 dataset achieving 96.22%

cifar10 classification cnn deep-learning restnet18

12. thaicn1712/agentflow

⭐ 1 stars | Rust

AgentOps for Rust: orchestrate, guard, validate, evaluate — in one crate

agentic-ai agentops ai-agents llm llmops rust

13. mokrosi/Agent2WhatsApp

⭐ 1 stars | HTML

A lightweight, open-source Node.js microservice that connects any AI agent directly to your personal WhatsApp using a session-based approach.

ai-agents automation llm llm-tools microservice nodejs

14. 2021Null/intent-driven-video-summarization

⭐ 1 stars | Python

Natural-language intent-driven personalized video summarization with Qwen, Chinese-CLIP, Visual-BiLSTM and QFVS evaluation.

chinese-clip llm multimodal pytorch qfvs video-summarization

15. ecoma-io/winnow

⭐ 1 stars | TypeScript

Triage GitHub issues with any OpenAI-compatible model — a cheap screening tier so spam never reaches the expensive one, verdicts checked against your taxonomy in code, and an explicit allowlist for wh

automation github-action github-actions issue-triage issues labels

Generated by Research Pulse on 2026-08-07 06:08