2026-08-03·Daily digest

🔬 Research Pulse

Daily Digest

August 03, 2026

🤖 AI

🧠 LLMs

1. Zero-Mem: Zero-Token Memory Operations for LLM Agents

Authors: Yilin Xiao, Zhehan Zhu, Yujing Zhang... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: Can structured memory access for LLM agents be performed without any intermediate LLM generation calls, eliminating the recurring token and latency costs of memory operations while preserving original interaction traces?

Summary: Zero-Mem proposes a memory architecture for LLM agents that performs zero LLM calls during memory operations, using an entity-context graph and temporal hierarchy over raw interaction traces with deterministic calibration. Only the final QA step invokes an LLM, yielding a 57.6% reduction in memory-operation time versus the fastest baseline while maintaining competitive QA performance.

Key Results: Zero-Mem achieves competitive performance on long-memory and long-context QA benchmarks while eliminating all LLM calls/tokens from memory operations (only the final-QA reader invokes an LLM). With the same reader and context budget, it reduces memory-operation time cost by 57.6% relative to the fastest compared baseline. Ablations confirm contributions of both the entity-context graph view and temporal hierarchy view, plus their query-dependent coordination.

Key Findings:

Structured agent memory does not require generating intermediate representations of past interactions
Dual complementary views (entity-context graph + temporal hierarchy) with query-dependent coordination outperform single-view approaches in ablations
Deterministic conflict-resolution calibration can keep the reader grounded without LLM-based reconciliation
57.6% memory-operation time reduction versus the fastest compared baseline at equivalent reader/context budget

Technical Novelty: Prior memory systems (MemGPT, Mem0, generative summarization approaches) use LLM calls to write/summarize/merge memory records. Zero-Mem eliminates all generative steps outside final QA by (1) preserving raw traces as source-of-record, (2) organizing them via an entity-context graph plus temporal hierarchy, (3) query-dependent weighing of the two views, and (4) deterministic (non-LLM) calibration to discard conflicting evidence.

What's New: First memory system to strictly enforce zero LLM/token consumption for all memory operations (write, organize, retrieve, calibrate) while retaining structured multi-view retrieval — contrasting with generative memory approaches that pay recurring LLM costs for summarization and record mediation.

Extension Opportunities:

Apply the zero-token memory paradigm to multi-agent systems where memory sharing across agents currently multiplies LLM overhead
Extend the deterministic calibration step to handle multimodal traces (images, code, tool outputs) rather than only text interactions
Integrate Zero-Mem's dual-view retrieval as a drop-in memory layer for existing agent frameworks (LangGraph, AutoGen) and benchmark cost/latency reductions in production workloads

Replicability: Code will be released post-peer-review at github.com/TheMoon0815/Zero-mem (not yet available). Compute needs are modest for memory operations since only encoder computation (not LLM generation) is used; the final-QA reader requires standard LLM inference. Full reproduction depends on unreleased implementation details and benchmark specifics not enumerated in the abstract.

Research Gaps:

Abstract does not specify which benchmarks, dataset sizes, or baselines were used, limiting comparability
No discussion of how the entity-context graph scales with very long interaction histories or handles entity disambiguation without an LLM

👁️ Vision

1. RayViT: Ray-Conditioned Visual Representations for Viewpoint-Robust Imitation Learning

Authors: Qian Wang, Longrui Chen, Peiran Sun... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How can visual imitation learning policies be made robust to camera viewpoint perturbations, given that RGB observations lack explicit geometric cues and pretrained ViT backbones have no notion of camera geometry?

Summary: RayViT injects camera geometry into pretrained Vision Transformers by patchifying Plücker ray maps and using gated cross-attention to produce a geometry-aware class token, plus dense ray positional embeddings. Combined with a cosine similarity auxiliary loss, this yields ~13pp robustness gains under camera perturbations on RoboCasa and +1.78 stages on real multi-task manipulation.

Key Results: RayViT improves robustness by ~13 percentage points under camera perturbations on the multi-task RoboCasa benchmark, and achieves +1.78 average completed stages in real-world multi-task success rate versus baselines. Gains are attributed to the ray-conditioned class token plus auxiliary cosine similarity loss.

Key Findings:

Explicit Plücker ray conditioning meaningfully improves imitation policy robustness to camera perturbations without abandoning pretrained ViT features.
Replacing the ViT CLS token with a ray-conditioned class token via gated cross-attention outperforms simple positional-embedding approaches.
An auxiliary cosine similarity loss on geometry-aware tokens consistently boosts both performance and robustness in sim and real.

Technical Novelty: Rather than retraining a ViT with geometry from scratch or concatenating camera params as scalars, RayViT patchifies a Plücker ray map, uses gated cross-attention to fuse ray features into a learnable class token, and injects ray features as dense positional embeddings — replacing the ViT's original CLS token with a geometry-aware summary. The auxiliary cosine similarity loss on ray-conditioned tokens is a new regularizer for geometric consistency.

What's New: Prior viewpoint-robust IL either augments views heavily, uses point clouds/depth, or trains geometry-aware backbones from scratch. RayViT is novel in treating Plücker rays as a first-class token stream fused into a pretrained ViT via gated cross-attention, preserving pretrained visual priors while adding camera-aware geometry.

Extension Opportunities:

Extend Plücker ray conditioning to multi-camera or stereo setups where ray maps from multiple viewpoints are fused into a single geometry-aware token for richer 3D reasoning.
Apply the ray-conditioned encoder to diffusion policy or VLA (vision-language-action) architectures such as OpenVLA/Pi-0 to test whether geometry-aware tokens improve viewpoint generalization at scale.
Investigate self-supervised pretraining of the gated cross-attention module on large-scale unlabeled robot video with known camera intrinsics/extrinsics to reduce reliance on task-specific imitation data.

Replicability: The abstract does not mention a code release. Reproduction would require a pretrained ViT backbone, RoboCasa simulation setup, and a real robot with calibrated cameras. Compute is likely modest (single/multi-GPU fine-tuning) since RayViT is described as lightweight and built atop frozen/pretrained ViTs, though real-robot evaluation is the bottleneck.

Research Gaps:

Evaluation appears limited to a single benchmark (RoboCasa) plus one real-robot suite; generalization to dexterous, mobile, or bimanual settings is untested.
The method assumes accurate camera calibration to build Plücker maps — robustness to calibration noise or unknown intrinsics is not addressed in the abstract.

📊 Other AI

1. Development of FDD-ON: an Ontology for VAV HVAC System Fault Detection and Diagnostics

Authors: Yimin Chen, Brian Fricke, Bo Shen... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How can fragmented FDD knowledge for VAV HVAC systems be formally structured to enable interoperability across heterogeneous data sources, equipment types, and diagnostic outputs — particularly for digital twin and AI-driven maintenance applications?

Summary: The paper introduces FDD-ON, a modular OWL ontology for VAV HVAC fault detection that formalizes components, faults, symptoms, and impacts along with their causal relations. It aims to break down FDD information silos by providing a machine-interpretable vocabulary that can bridge heterogeneous BMS data, diagnostic tools, and downstream AI/digital-twin applications.

Key Results: The authors developed FDD-ON, a modular ontology capturing VAV HVAC components, fault types, symptoms, and impacts with explicit cause-fault-symptom-impact relations. It was evaluated using publicly available VAV HVAC datasets and demonstrated via FDD development applications, though the abstract cites no specific quantitative benchmarks (accuracy, coverage counts, or dataset sizes).

Key Findings:

A dedicated FDD ontology with cause-fault-symptom-impact links enables interoperable diagnostic knowledge representation absent from existing building ontologies
Comprehensive fault, symptom, and impact libraries can capture a broad spectrum of VAV operational abnormalities under one controlled vocabulary
FDD-ON functions as a semantic substrate for digital-twin FDD frameworks and AI-driven maintenance decision systems

Technical Novelty: Prior FDD ontologies (e.g., SAREF4BLDG, Brick, Haystack) focus on equipment/point metadata but lack rich fault semantics. FDD-ON's novelty is the explicit contributing-cause → fault → symptom → impact relational chain plus dedicated fault/symptom/impact libraries with controlled vocabulary — making diagnostic reasoning machine-interpretable rather than requiring hand-coded rules per deployment.

What's New: Unlike Brick or Haystack which model building topology and metadata, FDD-ON specifically encodes diagnostic semantics — fault taxonomies, symptom manifestations, cause chains, and consequential impacts — closing the interpretability gap between raw BMS data and actionable diagnosis.

Extension Opportunities:

Extend FDD-ON beyond VAV to other HVAC subsystems (chillers, AHUs, boilers, heat pumps) and unify with Brick/Haystack schemas for whole-building coverage
Build an LLM-powered diagnostic agent that uses FDD-ON as a retrieval/reasoning backbone, mapping raw BMS telemetry to fault hypotheses via SPARQL queries
Integrate FDD-ON with digital twin platforms (e.g., Ptolemy, Azure Digital Twins) to enable real-time fault impact simulation and maintenance prioritization

Replicability: The abstract does not mention a public code repository or ontology release URL. Reproducing the ontology itself needs no GPU compute — standard OWL/RDF tooling (Protégé, RDFLib, SPARQL endpoints) suffices. Evaluation datasets are described as publicly available VAV HVAC datasets (likely LBNL/NIST fault datasets), but exact provenance isn't given.

Research Gaps:

Absence of quantitative benchmarks (coverage %, query performance, cross-dataset mapping accuracy) makes it hard to compare against alternative schemas
Scope limited to VAV systems; no demonstration of how the ontology composes with adjacent equipment or whole-building energy models

🦾 ROBOTICS

1. HAM-VLN: Harnessing Hierarchical Agentic Memory for Zero-Shot Vision-and-Language Navigation

Authors: An Liu, Bingxi Liu, Hongyu Ding... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How can training-free MLLM-based vision-and-language navigation agents overcome the memory and reasoning bottleneck that grows unboundedly with long-horizon navigation using image streams or dense maps?

Summary: HAM-VLN is a zero-shot vision-and-language navigation method that gives an MLLM agent a persistent, depth-grounded topological world graph whose semantic and reflective annotations are authored by the agent itself during its normal action-selection call. Combined with a bounded recency window and relevance/recency/salience retrieval, it beats prior training-free VLN methods on R2R, RxR, and HM3D ObjectNav while cutting context length by over 65%.

Key Results: HAM-VLN achieves 61.0% Success Rate on VLN-CE R2R, 52.7% SR on VLN-CE RxR, and 79.7% SR on HM3D-v2 ObjectNav — all without any training. It also reduces context length by more than 65% versus prior training-free methods, while adding zero extra LLM calls beyond the per-waypoint decision (memory writes are folded into the same call).

Key Findings:

Folding memory-writes into the action-selection call eliminates the extra LLM overhead that prior agentic-memory VLN methods incur
A depth-grounded world graph with topological one-hop expansion gives more useful history than raw image streams or dense maps
The approach delivers strong zero-shot results across three distinct benchmarks (R2R 61.0%, RxR 52.7%, HM3D ObjectNav 79.7%) while shrinking context >65%

Technical Novelty: The 'decision-coupled, agent-authored memory' — the MLLM writes semantic/reflective notes (room type, objects, progress, failure notes) into a persistent depth-grounded world graph within the same call that selects the next action, avoiding auxiliary LLM invocations. Retrieval combines relevance/recency/salience scoring with one-hop topological expansion, and a bounded verbatim window preserves recent waypoints.

What's New: Prior training-free VLN either stuffs image histories into context (which blows up token budgets) or uses dense maps (which are expensive to reason over), and agentic-memory approaches typically require separate summarizer/reflector LLM calls. HAM-VLN is novel in unifying memory authoring with the decision call and grounding the memory in a depth-derived topological graph rather than raw frames or occupancy grids.

Extension Opportunities:

Replace the retrieval scoring (relevance + recency + salience) with a learned scorer distilled from agent trajectories to improve retrieval precision under long horizons
Extend the depth-grounded world graph to multi-agent settings where several robots share and merge topological memory during collaborative navigation
Port the decision-coupled memory pattern to embodied manipulation or mobile-manipulation tasks (e.g., ALFRED, BEHAVIOR) where semantic + reflective notes on object states matter

Replicability: The abstract does not mention a code release. Reproduction requires a capable multimodal LLM (likely GPT-4V/Gemini-class) plus the VLN-CE R2R, VLN-CE RxR, and HM3D-v2 ObjectNav simulators (Habitat). No training compute is needed since the method is zero-shot; cost is dominated by MLLM API/inference calls per waypoint across evaluation episodes.

Research Gaps:

No mention of robustness to MLLM hallucinated memory writes or how erroneous 'failure notes' propagate across long trajectories
Evaluation is in simulation only (Habitat-style); real-world transfer with noisy depth and imperfect localization is not addressed

2. Diagnosing Compositional Generalization in Sequential Robot Tasks

Authors: Yixiao Wang, Cheng-En Wu, Lingfeng Sun... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How much and what structure of instruction-space coverage is needed for robot policies to compositionally generalize to novel sequential task combinations without exhaustive demonstration collection?

Summary: The paper diagnoses why sequential robot manipulation policies fail on novel instruction combinations by decomposing the generalization gap into three shift types. It empirically shows that ~25% structured coverage of the task space plus minimal per-task finetuning suffices, and that most OOD failures stem from instruction steering rather than absent low-level skills.

Key Results: A structured subset covering only ~1/4 of the full task space recovers strong OOD performance. Sparse training fails primarily due to instruction steering, not missing skills: finetuning with just one demonstration per task lifts OOD success from 0.4% to 54.7%. The generalization gap is decomposed into three measurable components: marginal instruction shift, instruction-compositional shift, and context-action shift.

Key Findings:

Exhaustive tuple enumeration is unnecessary — one-quarter structured coverage recovers strong OOD performance when action-relevant dependencies are covered
OOD failure is dominated by instruction steering: one demo per task boosts success from 0.4% to 54.7%
For semantically dependent tasks, coverage must preserve relational structure between instruction components, not just diversity of individual factors

Technical Novelty: Formal decomposition of the compositional generalization gap into three distinct shift sources (marginal, compositional, context-action), which enables diagnostic attribution of failures rather than treating OOD failure as monolithic. Prior compositional-generalization work in robotics typically measures aggregate success without isolating whether failure stems from skill gaps vs. instruction grounding.

What's New: Reframes robot compositional generalization as an instruction-space coverage problem with a formal three-part gap decomposition, and empirically separates 'missing skill' failures from 'instruction grounding' failures — a distinction largely conflated in prior imitation-learning benchmarks.

Extension Opportunities:

Build an active data-collection algorithm that greedily selects demonstrations to maximize dependency coverage in instruction space rather than uniform tuple sampling
Extend the three-way decomposition to continuous or long-horizon language instructions (e.g., LLM-generated task graphs) beyond discrete tuple compositions
Test whether the 'instruction steering' failure mode holds for VLA models (OpenVLA, RT-2) and design a lightweight steering-only adapter instead of full finetuning

Replicability: Project website exists (yixiaowang7.github.io); code/data availability not stated in abstract. Compute likely moderate — sequential manipulation benchmarks with policy finetuning typically run on 1-8 GPUs. Requires access to their instruction-tuple benchmark suite.

Research Gaps:

No principled algorithm yet for selecting the minimal dependency-covering subset a priori
Findings are on discrete instruction tuples; unclear how the decomposition transfers to free-form natural language or long-horizon LLM-planned tasks

💻 COMPUTE

1. Quantum tomography of inelastic electron scattering \emph{via} orbital angular momentum states

Authors: Amir H. Tavabi, Alessio D'Errico, Paolo Rosi... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How can quantum state tomography be practically implemented in transmission electron microscopy to characterize inelastically scattered electron beams, given that conventional approaches suffer from quadratic scaling of required measurements with phase-space dimensionality?

Summary: The paper introduces OAM-subspace quantum state tomography for transmission electron microscopy, using an OAM sorter to discretize the electron phase space and make density-matrix reconstruction of inelastically scattered electrons experimentally tractable. Demonstrated on volume-plasmon excitation in a carbon film with structured probes, the method reveals symmetry-breaking, coherence properties, and the nature of induced quantum transitions.

Key Results: The authors demonstrated OAM-restricted quantum tomography on inelastic scattering of a structured electron probe exciting volume plasmons in a carbon film. Using an OAM sorter to discretize phase space into a finite measurable basis, they reconstructed density matrices of scattered states and, via diagonalization, resolved induced state transitions and symmetry-breaking effects. Specific quantitative benchmarks (fidelity, OAM basis size, sample thickness) are not stated in the abstract.

Key Findings:

Restricting tomography to the OAM subspace via an electron OAM sorter converts an intractable continuous-phase-space problem into a manageable discrete measurement set
Structured (OAM-carrying) electron probes exciting volume plasmons in carbon exhibit measurable symmetry-breaking signatures in the reconstructed density matrix
Diagonalization of the reconstructed density matrix identifies the dominant eigenstates and reveals the character of inelastic state transitions

Technical Novelty: Prior electron tomography attempts scaled quadratically in transverse phase-space dimension, making them impractical. This work restricts tomography to the OAM subspace using an electron-optical OAM sorter that projects the beam onto a discrete, finite basis, reducing measurement count while still capturing the essential coherence structure of inelastic scattering — an approach borrowed conceptually from quantum optics but not previously realized in TEM for inelastic events.

What's New: First demonstration of quantum state tomography of inelastically scattered electrons in TEM using OAM as the measurement basis, bridging quantum-optics tomography methodology with electron microscopy and overcoming the dimensionality bottleneck that has blocked prior attempts.

Extension Opportunities:

Apply OAM-resolved tomography to other inelastic excitations (phonons, core-loss edges, excitons in 2D materials) to map symmetry-selective coupling channels
Combine the OAM sorter with energy-filtered TEM (EELS) for joint (OAM, energy-loss) density-matrix reconstruction of individual excitation modes
Build a compressed-sensing or machine-learning reconstruction pipeline that reduces the number of OAM projections needed for high-fidelity density matrices of magnetic or chiral samples

Replicability: The abstract does not mention public code or data releases. Reproduction requires a TEM equipped with an aberration-corrected probe former, a programmable phase plate or holographic mask to generate structured (OAM-carrying) probes, and a custom OAM sorter (typically two conformal-mapping optical elements). This is specialized hardware available only in a handful of electron microscopy labs; classical compute demands for the density-matrix reconstruction are modest.

Research Gaps:

The OAM subspace captures only rotational-mode coherence; radial and energy-resolved degrees of freedom of the scattered state are not fully characterized
Scalability to larger OAM bases, mixed-state noise limits, and quantitative fidelity benchmarks versus ground truth remain to be established

2. Rethinking AI Cloud Infrastructure for Agentic Serving Systems with the Aries Experimentation Framework

Authors: Leonid Kondrashov, Hongrui Liu, JooYoung Park... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How should AI cloud infrastructure be redesigned to serve autonomous agents, which couple repeated LLM inference with persistent context and sandboxed tool execution — a workload that breaks assumptions of conventional LLM serving systems?

Summary: The paper introduces Aries, an experimentation framework designed specifically for agentic LLM serving systems, and uses it (alongside production traces) to show that conventional LLM-serving assumptions — token-centric metrics, unbounded context, and snapshot-based sandbox management — are misaligned with real agent workloads. It argues for a new class of agent-native serving systems built around trajectory-level metrics, adaptive context management, elastic sandbox resources, and minimized sandbox attack surfaces.

Key Results: Using the Aries framework on open agent harnesses/benchmarks plus production traces from a commercial platform, the authors demonstrate three concrete findings: (1) token-centric metrics fail to capture non-inference bottlenecks in agent trajectories, (2) retaining additional context yields diminishing accuracy benefits while measurably reducing serving capacity, and (3) tool sandboxes exhibit bursty behavior — long idle periods punctuated by short resource bursts — and current snapshot-based state management makes aggressive suspension costly. A security analysis further quantifies sandbox attack surface concerns. Specific numerical thresholds are not provided in the abstract.

Key Findings:

Token-centric serving metrics (TTFT, TPOT) miss dominant non-inference bottlenecks in agent trajectories
Growing context beyond a certain point produces diminishing accuracy gains while cutting serving throughput
Tool sandboxes are highly bursty (long idle + short spikes), and existing snapshot suspension is too expensive to exploit this, leaving significant resource-efficiency headroom

Technical Novelty: Aries is the first full-stack experimentation framework that (a) separates task semantics from execution configurations, (b) reconstructs cross-component agent trajectories correlated with system-level telemetry, and (c) exposes stateful tool execution through a consistent interface across heterogeneous sandbox substrates — enabling reproducible systems research on agents rather than isolated inference benchmarking.

What's New: Prior LLM serving work (vLLM, Sarathi, DistServe, etc.) optimizes single-request inference; prior agent work optimizes reasoning quality. This paper is one of the first to treat the full agent stack — inference + persistent context + stateful sandboxes — as a joint systems problem, grounded in both open benchmarks and real production traces from a commercial agent platform.

Extension Opportunities:

Build trajectory-level SLO schedulers that optimize end-to-end agent task completion rather than per-token latency, using Aries telemetry as ground truth
Design lightweight sandbox suspension mechanisms (e.g., CRIU-lite, unikernel snapshots, or memory-diff checkpoints) that beat current snapshot-based approaches for bursty tool workloads
Develop adaptive context compression/eviction policies that trade off accuracy vs serving capacity dynamically per agent trajectory

Replicability: The abstract does not explicitly announce a code release, but the framing as an 'experimentation framework' on 'open agent harnesses and benchmarks' strongly implies open-source release. Reproduction would require GPU inference capacity for the LLM backend plus container/VM infrastructure for sandbox substrates; production traces from the commercial partner likely remain proprietary.

Research Gaps:

No standard benchmark or metric suite for evaluating agent-serving systems end-to-end at trajectory level
Snapshot/checkpoint mechanisms for tool sandboxes are not designed for sub-second suspension/resume at agent-workload scale, and sandbox attack surfaces remain large

3. A Taxonomy of Performance Metrics for the Distributed Computing Continuum

Authors: Praveen Kumar Donta, Boris Sedlak, Alfreds Lapkovskis... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can performance evaluation of Distributed Computing Continuum Systems (DCCS) be standardized when existing practices only capture isolated dimensions (compute, network, energy, or app-level) and miss cross-layer behavior across edge-to-cloud environments?

Summary: The paper proposes a structured taxonomy of performance metrics for Distributed Computing Continuum Systems, organizing them into computing-, network-, and application/user-level categories while adding emerging dimensions like sustainability, adaptability, and continuum fragmentation. It also formalizes each metric mathematically and specifies acquisition requirements (scope, phase, method), giving practitioners a consistent vocabulary for evaluating cross-layer DCCS behavior.

Key Results: The paper is a taxonomy/survey rather than an empirical study — no benchmark numbers, datasets, or experimental measurements are reported in the abstract. Its 'proof' is the structured organization itself: metrics grouped into computing-level, network-level, and application/user-level categories, plus emerging dimensions (sustainability, observability, adaptability, data locality, migration awareness, continuum fragmentation), with mathematical formulations and acquisition requirements (scope: single-node/multi-node/full-system; phase: operational monitoring vs experimental evaluation; measurement method).

Key Findings:

Existing DCCS evaluation is fragmented — metrics are typically reported in a single layer, obscuring cross-layer trade-offs
Six emerging dimensions (sustainability, observability, adaptability, data locality, migration awareness, continuum fragmentation) are underrepresented in classical taxonomies but essential for AI/edge workloads
Metric usefulness depends not just on definition but on acquisition scope (node/multi-node/full-system) and phase (operational vs experimental), which the taxonomy makes explicit

Technical Novelty: Prior surveys treat compute, network, energy, and QoS metrics in silos. This work unifies them into a single cross-layer taxonomy specifically scoped to the continuum, and adds a second axis — acquisition requirements (scope, phase, method) — that tells you how a metric can actually be collected in a heterogeneous distributed setting. The emerging dimensions (migration awareness, continuum fragmentation, data locality) are the genuinely new categories vs. classical distributed-systems taxonomies.

What's New: First taxonomy to explicitly unify cross-layer metrics for the continuum and pair each metric with acquisition-feasibility metadata, rather than presenting definitions in isolation.

Extension Opportunities:

Build an open-source DCCS benchmarking harness that instruments the taxonomy's metrics end-to-end across a real edge-cloud testbed (e.g., k3s + cloud k8s) and publishes reference traces
Develop a Prometheus/OpenTelemetry exporter library that maps directly onto the taxonomy's acquisition-scope categories, so operators get cross-layer dashboards out of the box
Extend the taxonomy with AI-workload-specific metrics (LLM inference tail latency, KV-cache locality, token/joule) and validate via a distributed inference deployment split between DGX and edge nodes

Replicability: N/A for empirical reproduction — no code, datasets, or experiments. The taxonomy itself is reusable directly from the paper; adopting it requires only reading effort. Building tooling on top would need a modest edge-cloud testbed.

Research Gaps:

No empirical validation or case study applying the taxonomy to a real deployment
AI-workload-specific metrics (inference tail latency, model-swap cost, KV-cache locality, token/joule) are not deeply covered despite the abstract's motivation from AI/data-driven applications

⚡ ENERGY

1. Magnetic properties of a quasi-two-dimensional spin-1/2 antiferromagnet Y2CuGe4O12

Authors: J. Khatua, Changhyun Koo, Suyoung Kim... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How do competing exchange interactions and frustration shape the ground state of a spatially anisotropic spin-1/2 distorted triangular-lattice antiferromagnet, and can Y2CuGe4O12 realize a frustrated quantum magnet where further-neighbor couplings dominate?

Summary: The authors characterize Y2CuGe4O12, a spin-1/2 distorted triangular-lattice antiferromagnet, using susceptibility, specific heat, ESR, and DFT, finding that further-neighbor AFM exchange (J3≈−3.22 K) dominates over nearest-neighbor FM coupling. No long-range order occurs down to 0.4 K, and a modest 2.6 T field drives the system into a gapped field-polarized state, establishing the compound as a tunable platform for frustration-driven quantum magnetism.

Key Results: Through susceptibility, specific heat, ESR, and DFT on polycrystalline Y2CuGe4O12: identified competing exchanges J1≈0.138 K (FM), J2≈0.01 K (FM), J3≈−3.22 K (AFM), J4≈−1.56 K (interplanar AFM); Curie-Weiss θ_CW=−1.8 K; no long-range order down to 0.4 K despite dominant AFM couplings; broad maxima in χ(T) and C_mag(T) indicate short-range correlations; critical ESR linewidth broadening consistent with low-D frustration; saturation field μ₀H_s=2.6 T drives field-polarized state; exponential C_mag(T) above H_s consistent with gapped magnon excitations.

Key Findings:

Cu²⁺ ions form a distorted triangular lattice with competing FM (J1, J2) and AFM (J3, J4) exchanges, with the third-neighbor J3≈−3.22 K being dominant
No long-range magnetic ordering down to 0.4 K despite dominant AFM interactions; broad χ and C_mag maxima and critical ESR linewidth broadening reveal short-range correlations characteristic of low-D frustrated magnets
A saturation field of only 2.6 T yields a field-polarized state with exponential C_mag(T), signaling gapped magnon excitations

Technical Novelty: Identification of a distorted triangular Cu²⁺ lattice where third-neighbor AFM exchange (J3) — not nearest-neighbor coupling — dominates the magnetism, combined with a small saturation field (2.6 T) making the full H-T phase diagram experimentally accessible with standard lab magnets.

What's New: Unlike typical triangular-lattice antiferromagnets where nearest-neighbor coupling sets the physics, Y2CuGe4O12 is a rare case where further-neighbor exchange dominates, and its unusually low saturation field makes the entire field-tuned phase diagram experimentally accessible.

Extension Opportunities:

Grow single crystals of Y2CuGe4O12 to enable neutron scattering measurements that directly map the spin correlations and confirm the DFT-derived J3-dominated exchange topology
Perform sub-0.4 K μSR or dilution-refrigerator NMR to determine whether a true spin-liquid ground state or an ultra-low-T ordered phase exists below current measurement range
Chemically substitute Y³⁺ or Ge⁴⁺ (e.g., with La, Si, Sn) to tune lattice distortion and the J3/J1 ratio, mapping a phase diagram between triangular and 1D-chain regimes

Replicability: No code/data explicitly mentioned in the abstract. Reproduction requires solid-state synthesis of polycrystalline Y2CuGe4O12, SQUID magnetometry, low-T (≤0.4 K) specific heat (dilution fridge or 3He), ESR spectrometer, and DFT software (e.g., VASP/WIEN2k) for exchange parameter extraction — feasible for a well-equipped condensed-matter lab.

Research Gaps:

Absence of single-crystal samples limits direct probes (neutron scattering, anisotropy) of the magnetic ground state
The nature of the state between 0.4 K and long-range order (spin liquid, spin glass, or ordered) remains undetermined

2. Layer- and Field-Dependent Magnetic Order in 2D CrSBr Revealed by Pulsed Nanocalorimetry

Authors: Hugo Gomez-Torres, Roop K. Mech, Alessandra Canetta... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How can we directly measure the thermodynamic properties (heat capacity, magnetic entropy) of atomically thin van der Waals magnets like CrSBr, given that exfoliated flakes have femtogram-scale masses that defeat conventional calorimetry, and how does magnetic order evolve as thickness approaches the monolayer limit?

Summary: The paper introduces microsecond pulse-heating nanocalorimetry as a thermodynamic probe capable of measuring heat capacity and magnetic entropy of exfoliated CrSBr flakes down to the monolayer. It reveals thickness-dependent suppression of interlayer AFM order, a layer-parity effect from uncompensated layers, and extended magnetic fluctuations above the ordering transition — establishing entropy-based access to 2D magnetism previously inaccessible at femtogram scales.

Key Results: Using microsecond pulse-heating nanocalorimetry, the authors measured heat capacity and magnetic entropy of CrSBr flakes down to the monolayer. They demonstrated: (1) a reduction of the interlayer antiferromagnetic transition temperature with decreasing thickness, (2) an entropy-derived effective magnetic moment approaching the monolayer limit, (3) a crossover from bulk-like interlayer AFM to intralayer FM-dominated correlations, (4) a layer-parity effect where odd-layer samples show an additional high-temperature entropy contribution from uncompensated layers, (5) extraction of a thickness-dependent critical in-plane field along the easy axis showing weakened interlayer exchange, and (6) magnetic fluctuations persisting well above the interlayer ordering transition.

Key Findings:

Interlayer AFM transition temperature decreases as CrSBr thickness is reduced toward the monolayer, with a crossover to intralayer FM-dominated correlations
Odd-layer flakes exhibit a distinctive high-temperature entropy contribution attributable to uncompensated magnetic layers — a clear layer-parity signature
In-plane fields along the easy axis suppress AFM order with a thickness-dependent critical field, quantifying weakened interlayer exchange in thin samples, and entropy analysis shows magnetic fluctuations persisting well above the ordering transition

Technical Novelty: The core novelty is adapting microsecond pulse-heating nanocalorimetry to resolve heat capacity of femtogram-scale exfoliated flakes — a regime where conventional calorimetry fails. This gives direct thermodynamic access (entropy, exchange, effective moments) to 2D magnets that prior work could only probe indirectly via transport, optics, or magnetometry.

What's New: First application of pulsed nanocalorimetry to obtain full thermodynamic (entropy) characterization of a van der Waals magnet down to the monolayer, overcoming the femtogram-mass barrier that has kept 2D magnets outside the reach of conventional calorimetric methods.

Extension Opportunities:

Apply the pulsed nanocalorimetry technique to other 2D van der Waals magnets (CrI3, Fe3GeTe2, NiPS3) to build a comparative thermodynamic database of dimensional crossover behavior
Combine nanocalorimetry with in-situ optical/magneto-optical probes (e.g., MOKE or Raman) on the same flake to correlate entropy signatures with spin texture and magnon dynamics
Extend measurements to twisted or heterostructured CrSBr bilayers to thermodynamically characterize moiré-induced magnetic phases and proximity effects

Replicability: The abstract does not mention public code or data release. Reproduction requires specialized nanocalorimeter chips with integrated micro-heaters/thermometers, cleanroom fabrication for device patterning, exfoliation and deterministic transfer of CrSBr flakes, cryogenic + magnetic field setup, and microsecond-resolved pulse electronics — a substantial experimental infrastructure rather than a compute-bound reproduction.

Research Gaps:

Thermodynamic characterization has been largely absent for 2D vdW magnets — prior work relied on indirect probes (transport, MOKE, Raman) that don't directly yield magnetic entropy or exchange energies
Layer-parity effects and the extent of fluctuation regimes above ordering transitions in 2D magnets were not directly quantified thermodynamically

3. Fast and Accurate Foundation Models for Equivariant Machine-Learned Interatomic Potentials

Authors: Seán R. Kavanagh, Chuin Wei Tan, Menghang Wang... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can equivariant machine-learned interatomic potentials (MLIPs) simultaneously achieve high inference/training speed AND accuracy at foundation-model scale, particularly when trained on extremely large datasets where data efficiency matters less than throughput?

Summary: The paper releases a family of foundation-scale equivariant MLIPs in the NequIP and Allegro architectures that push the speed–accuracy Pareto frontier, enabled by acceleration work in the NequIP infrastructure that makes ultra-large-dataset training tractable. It also argues that further accuracy gains in materials discovery are gated by dataset diversity and better transition-metal energy surface descriptions rather than model architecture.

Key Results: The authors present a family of foundation potentials built on NequIP and Allegro equivariant architectures that achieve leading inference speeds, strong scalability, and competitive accuracy across community benchmarks spanning materials discovery, thermal conductivity prediction, and near-equilibrium mechanical/thermodynamic property prediction. Accelerations implemented in the NequIP infrastructure dramatically reduce training cost for ultra-large datasets. Specific numerical benchmark values are not disclosed in the abstract.

Key Findings:

Equivariant MLIPs can be made fast enough for MD-scale deployment while retaining accuracy, dispelling the speed-vs-symmetry trade-off assumption at foundation scale
The released NequIP/Allegro foundation potentials achieve leading inference speed and strong scalability across materials discovery, thermal conductivity, and mechanical/thermodynamic benchmarks
Remaining accuracy bottleneck for materials discovery is dataset diversity and consistent treatment of transition-metal compound energetics, not model capacity

Technical Novelty: Speed/scalability optimizations inside the NequIP infrastructure that make equivariant architectures viable at foundation-model scale, along with reformulations of NequIP and Allegro trained on very large datasets — challenging the common assumption that equivariant models are inherently too slow for foundation deployment relative to invariant or lower-symmetry alternatives.

What's New: Prior equivariant foundation MLIPs (e.g., MACE-MP, SevenNet, Orb) traded speed for symmetry-preserving accuracy; this work explicitly targets and closes that gap for NequIP/Allegro at foundation scale, and empirically reframes the accuracy bottleneck as a data problem rather than an architecture problem.

Extension Opportunities:

Fine-tune the released foundation potentials on targeted transition-metal chemistries (where the paper identifies persistent accuracy weaknesses) and quantify downstream MD accuracy gains
Integrate the accelerated NequIP/Allegro potentials into large-scale MD workflows (e.g., LAMMPS, ASE) for high-throughput screening of catalysts or battery materials, benchmarking wall-clock cost vs. MACE/SevenNet
Build an active-learning loop that specifically expands dataset diversity for transition metal compounds, addressing the dataset-quality bottleneck the paper flags for materials discovery

Replicability: The paper is framed around NequIP/Allegro, both of which are established open-source frameworks, and the phrase 'accelerations implemented within the NequIP infrastructure' strongly implies code release. Training foundation potentials on ultra-large datasets historically requires multi-GPU clusters (tens to hundreds of GPU-days); the paper's contribution is precisely to reduce this. Exact code/data URLs and compute budgets are not stated in the abstract.

Research Gaps:

Transition-metal compound energy surfaces are inconsistently described across current foundation training datasets, limiting materials-discovery accuracy
Lack of dataset diversity in existing large-scale MLIP training corpora bounds achievable universal-potential accuracy independent of architecture

🏥 HEALTHCARE

1. EasyBCI Agent: Towards Universal Neural Data Preprocessing for Brain-Computer Interfaces

Authors: Yu Zhu, Runkai Zhao, Zhimin Zhou... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How can neural signal preprocessing for brain-computer interfaces be automated in a way that is reproducible, domain-safe, and covers multiple signal modalities—overcoming the manual, expert-dependent bottleneck that current LLM coding agents cannot solve due to lacking modality coverage, raw-data isolation, and domain oversight?

Summary: EasyBCI is a two-phase LLM agent (Plan + Execution) that automates preprocessing of neural signals across six BCI modalities without exposing raw data to the model, using a text-only Data Fingerprint, a self-correcting code loop, an experience-accumulation skill library, and two human decision gates. It preserves more task-relevant EEG separability than manual pipelines and beats general-purpose coding agents on 4 of 5 configurations.

Key Results: On EEG with a fixed linear classifier, all 5 EasyBCI backbones preserved more task-relevant separability than manual pipelines. In same-backbone comparisons, EasyBCI outperformed general-purpose coding agents on both label schemes for 4 of 5 configurations. The system generalized to 5 additional modalities spanning ~3 orders of magnitude in sampling rate, producing complete reproducible pipelines with decision provenance.

Key Findings:

All five LLM backbones inside EasyBCI outperformed manual EEG preprocessing on fixed-classifier separability
Domain-specific orchestration beat general-purpose coding agents in 4 of 5 same-backbone comparisons
The framework generalized across six modalities spanning ~1000× range in sampling rate while producing auditable, reproducible pipelines

Technical Novelty: The Data Fingerprint abstraction—a text-only recording profile that lets the LLM plan pipelines without ever seeing raw neural data—combined with a two-phase Plan/Execute split, a quality-gated experience system that promotes validated pipelines into reusable skills, and two explicit human-in-the-loop gates. Prior coding agents lack all four of these together.

What's New: Unlike prior scientific coding agents, EasyBCI enforces raw-data isolation via a Data Fingerprint, embeds two mandatory expert intervention gates, and retains validated pipelines as reusable skills—design principles tailored for high-stakes scientific preprocessing rather than generic code generation.

Extension Opportunities:

Extend the two-phase Plan/Execution agent pattern to downstream BCI stages (feature extraction, decoder training, online calibration) using the same Data Fingerprint abstraction
Port the quality-gated experience/skill retention system to other scientific domains (genomics, MS spectra, astronomy pipelines) where raw-data isolation matters
Replace the two human decision gates with a learned critic model trained on accumulated expert interventions to move toward full autonomy

Replicability: Abstract does not mention a code or data release. Reproduction would need the six signal-type datasets, five LLM backbones (API access, likely GPT/Claude-class), and the linear-classifier evaluation harness. Compute is modest—LLM API calls plus standard signal-processing runs—so a single workstation would suffice if code and prompts are published.

Research Gaps:

Evaluation quality relies on a single fixed linear classifier, which may not reflect downstream performance with modern nonlinear decoders
Human-in-the-loop gates limit scalability and were not quantified for expert time cost or inter-rater agreement

2. Self-Play Meets Skill Evolution: Self-Evolving Search Agents that Pose, Solve, and Remember

Authors: Zenghuang Fu, Zhaoyang Li, Qiuyuan Ai... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: How can self-play curricula and external skill memories be unified so that procedural memory becomes an evolving state that co-adapts with task generation, rather than a static plug-in learned from a fixed task distribution?

Summary: SESA turns procedural skill memory into an evolving state inside a tool-augmented search self-play loop: a challenger poses problems, a separately parameterized solver retrieves skills, and distilled failures are written back to a bank that in turn reshapes both solver behavior and the challenger's future task distribution. This bidirectional co-evolution beats SSP by 1.2–3.2 points and SkillRL by 0.9 points on seven QA benchmarks, and — crucially — the benefits partially internalize into model weights, enabling memory-free deployment.

Key Results: Across seven open-domain and multi-hop QA benchmarks, SESA improves average accuracy over SSP by 1.2–3.2 points across multiple backbones, and surpasses the skill-augmented SkillRL baseline by 0.9 points under a unified evaluation protocol. On Qwen3 models, SESA-Off (memory-free deployment) retains 1.8–2.2 points over SSP, and adding the final skill bank yields a further 0.5–1.0 points.

Key Findings:

Co-evolving skill memory with a self-play challenger outperforms both static-curriculum skill memories (SkillRL, +0.9) and memory-less self-play (SSP, +1.2–3.2) across seven QA benchmarks
Skill-augmented on-policy training internalizes skills into parameters: SESA-Off (no memory at inference) still retains 1.8–2.2 points over SSP on Qwen3
The external skill bank remains valuable as an optional inference-time plug-in, adding a further 0.5–1.0 points on top of the internalized gains

Technical Novelty: Prior self-play (SSP) leaves failures only in gradients; prior skill-augmented RL (SkillRL) learns skills against fixed task distributions. SESA closes the loop: the challenger's reward depends on the solver's skill-augmented success, so the task distribution evolves against the current skill bank while the bank evolves against current failures — and because retrieved skills shape on-policy trajectories, skills are internalized into weights, not just stored externally.

What's New: The bidirectional loop where task distribution and skill memory mutually shape each other — most prior work fixes one side. Additionally, treating retrieved skills as trajectory-shaping context during RL (rather than only inference-time retrieval) is what enables memory-free deployment, a novel dual on-policy/off-policy value for a skill bank.

Extension Opportunities:

Apply the challenger/solver + skill-memory co-evolution loop to code-generation or tool-use agents (e.g., SWE-bench) where procedural skills are naturally reusable functions or scripts
Study skill-bank compression / forgetting policies — the paper writes distilled failures back, but a builder could add retrieval-frequency-weighted pruning or hierarchical skill abstraction to prevent memory bloat
Extend the bidirectional loop to multi-agent settings where multiple solvers with divergent skill banks compete, letting the challenger target skill gaps across a population

Replicability: Code is available at https://github.com/Zenghuang-Fu/SESA-Self-Evolving-Search-Agents. Reproduction requires RL fine-tuning of Qwen3-scale backbones with two separately parameterized policies (challenger + solver) plus a retrieval-augmented tool loop — realistically multi-GPU (A100/H100-class) for days; the benchmarks (7 open-domain / multi-hop QA sets) are standard and cheap to evaluate.

Research Gaps:

No analysis of skill-bank scaling behavior — how memory quality, redundancy, or interference evolve as training continues, and whether the loop eventually collapses or plateaus
Evaluation is confined to open-domain/multi-hop QA; whether the co-evolution dynamic transfers to domains with richer procedural skills (code, math proofs, multi-tool workflows) is untested

🔬 MATERIALS

1. Beyond Stoner-Wohlfarth: Machine-Learning Models and Symbolic Regression of Hard-Magnet Properties

Authors: Samuel J. R. Holt, Christina Winkler, Timoteo Colnaghi... Published: 2026-07-31 | Citations: 0 arXiv | PDF

Research Question: Can machine learning and symbolic regression predict hard-magnet extrinsic properties (coercive field, remanent magnetisation, maximum energy product) from intrinsic micromagnetic parameters more accurately than analytical models like Stoner-Wohlfarth, while avoiding the compute cost of direct micromagnetic simulation?

Summary: The paper trains ML models and applies symbolic regression to 12,012 micromagnetic simulations to predict permanent-magnet extrinsic properties, outperforming analytical models like Stoner-Wohlfarth. Symbolic regression rediscovers the Kronmüller coercivity law with a material-dependent demagnetising factor and produces new compact expressions for remanence and BHmax, released via the mammos-ai package for rapid screening.

Key Results: Trained ML models on 12,012 micromagnetic simulations of an idealised cubic grain spanning saturation magnetisation, exchange constant, and uniaxial anisotropy constant. On held-out data, ML models predict all three extrinsic properties with substantially lower errors than established analytical baselines. Symbolic regression rediscovers the Kronmüller form for coercive field (with a material-dependent effective demagnetising factor) and derives new closed-form expressions for remanence and maximum energy product, each with at most two fitted constants yet approaching ML accuracy. Inverse recovery works well for saturation magnetisation and anisotropy, but fails for the exchange constant due to weak influence on extrinsic properties.

Key Findings:

ML models substantially outperform analytical baselines on all three extrinsic properties on identical held-out data
Symbolic regression yields interpretable laws (≤2 fitted constants) that nearly match ML accuracy, including a Kronmüller-form coercivity with material-dependent demagnetising factor
The inverse problem recovers Ms and K accurately but not the exchange constant A, because A weakly influences extrinsic hysteresis observables

Technical Novelty: First application of symbolic regression to recover interpretable closed-form laws for all three key extrinsic magnet properties from a large micromagnetic dataset, revealing a material-dependent demagnetising factor inside the Kronmüller form and yielding novel remanence/BHmax equations with ≤2 constants that rival black-box ML.

What's New: Combines large-scale micromagnetic simulation, ML surrogates, and symbolic regression to produce both accurate predictors and human-readable physical laws, going beyond prior analytical/Stoner-Wohlfarth approaches while remaining interpretable — unlike pure black-box ML surrogates.

Extension Opportunities:

Extend the dataset beyond idealised cubic grains to realistic polycrystalline microstructures with grain-boundary phases and defects, retraining the ML/symbolic pipeline
Couple mammos-ai with a Bayesian optimisation loop to inverse-design novel rare-earth-lean permanent magnets by screening millions of candidate compositions
Add temperature-dependent intrinsic parameters and hysteresis simulations to enable prediction of thermal stability curves for motor-grade magnets

Replicability: Trained models released via the mammos-ai Python package. The 12,012-simulation dataset is described but availability isn't confirmed in the abstract. Reproducing training is inexpensive; regenerating the micromagnetic dataset would require substantial GPU/CPU time (hours-to-days per simulation × 12k runs).

Research Gaps:

Idealised cubic grain geometry ignores realistic polycrystalline microstructure, grain boundaries, and defects that dominate real magnets
Exchange constant A is not identifiable from extrinsic properties alone, limiting fully-inverse material design

🔥 GitHub Trending

1. banye-technology/AbaoPal

⭐ 6 stars | Kotlin

开源 Android 手机自动化 Agent：用自然语言驱动多模态大模型理解屏幕、规划任务并跨应用执行操作，支持可复用 Skills。

ai-agent android automation jetpack-compose kotlin llm

2. xorcise-ai/xorcise

⭐ 6 stars | Python

Run cyber-AI agents against real missions in isolated environments. Record everything as OpenTelemetry evidence. Grade the evidence, not the claim.

agent-evaluation ai-agents benchmark ctf cybersecurity evaluation

3. lamhotsiagian/pytorch-ai-engineering

⭐ 4 stars | Python

A production-grade, open-source AI Engineering codebase and educational curriculum built with PyTorch 2.x. Designed for Software Engineers, ML Engineers, Backend Engineers, and AI professionals who bu

pytorch

4. jshn9515/deep-learning-notes

⭐ 3 stars | Python

Personal deep learning study notes and tutorial-style notebooks

deep-learning deep-learning-tutorial jupyter-notebooks pytorch quarto tutorial

5. danishnaseer00/SLM-cot-Eval

⭐ 2 stars | Python

Reproducibility artifact for the TMLR paper on quantization and cue-conditioned reasoning: frozen trials, 15 generation files, 3,078 automated judgments, human annotations, and analysis code

chain-of-thought cot deepseek evaluation faithfulness llm

6. RightNow-AI/autoevolve

⭐ 2 stars | Python

Agent-native evolutionary optimization. Say the goal in english, evolve code toward a measured target with a population of coding agents.

ai-agents claude code-optimization developer-tools evolutionary-algorithms llm

7. BlumFinancialLab/Blum

⭐ 1 stars | Python

Open-source financial decision intelligence for equities and Forex: evidence-bound AI agents, paper trading, backtesting and benchmark validation.

ai ai-agents algorithmic-trading backtesting equities fastapi

8. Mayank-589/Portfolio-Website

⭐ 1 stars | CSS

Personal portfolio website showcasing my data science & ML projects — built with HTML, CSS, and JavaScript.

csss data-science html javascript machine-learning portfolio

9. Anil-matcha/minimax-h3-comfyui

⭐ 1 stars | Python

ComfyUI custom nodes for MiniMax H3 text-to-video, image-to-video, and multimodal reference-to-video generation through Muapi.

ai-video api artificial-intelligence comfyui comfyui-custom-nodes creative-ai

10. twomathematicians-code/archon

⭐ 1 stars | Python

AI/ML Meta-Platform: YAML/DSL-driven compiler for AI intent. Describe data → models → pipelines → policies. 6-layer architecture with code generation, guardrails, GenAI, and causal ML.

ai causal-inference code-generation compiler dsl guardrails

11. AtomicHalifax/ChestXRay-Disease-Classification

⭐ 1 stars | Jupyter Notebook

Multi-label chest X-ray disease classification using DenseNet121 on the Stanford CheXpert dataset with Grad-CAM explainability and comprehensive evaluation.

artificial-intelligence chest-xray chexpert computer-vision deep-learning densenet121

12. ZhouYinLong-lab/Pocket-World-Model

⭐ 1 stars | Python

A tiny, observable 2D world model that learns motion dynamics, imagines future trajectories, and plans through its learned environment.

gymnasium machine-learning model-based-reinforcement-learning pytorch reinforcement-learning world-model

13. AbhishekChandra961/ipl-impact-player-win-probability

⭐ 1 stars | Jupyter Notebook

Data-driven investigation into whether the IPL Impact Player rule improves win probability — statistics + XGBoost + SHAP on ball-by-ball data.

cricket-analytics data-science machine-learning python shap sports-analytics

14. Shriya-Guptaa/Pima-diabetes-prediction-app

⭐ 1 stars | Jupyter Notebook

deployment docker docker-image dockerfile gcp gcp-cloud-build

15. Anil-matcha/flux-3-comfyui

⭐ 1 stars | Python

ComfyUI custom nodes for Muapi's FLUX 3 API: text-to-image, Dev, image editing, text/image-to-video, native audio, media savers, and example workflows.

ai-image-generation ai-video-generation api black-forest-labs comfyui comfyui-custom-nodes

Generated by Research Pulse on 2026-08-03 06:07