2026-05-28·Daily digest

🔬 Research Pulse

Daily Digest

May 28, 2026

🤖 AI

🧠 LLMs

1. OmniVerifier-M1: Multimodal Meta-Verifier with Explicit Structured Recalibration

Authors: Xinchen Zhang, Bowei Liu, Jiale Liu... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can multimodal verifiers be trained to produce reliable, fine-grained verification with interpretable rationales, rather than only binary decisions, in a way that scales without depending on auxiliary judge models?

Summary: OmniVerifier-M1 is a generalist multimodal verifier trained with meta-verification feedback in which bounding-box-style symbolic rationales serve as the explanation signal, allowing rule-based RL rewards without auxiliary judge models. The authors further show that decoupling the RL objectives for binary judgment and meta-verification beats joint optimization, and they leverage the resulting verifier to drive M1-TTS, an agentic generation system that performs dynamic region-level self-correction.

Key Results: The abstract reports two empirical findings rather than headline numbers: (1) symbolic verifier outputs (e.g., bounding boxes) outperform textual explanations as meta-verification rationales and enable rule-based RL rewards without auxiliary judge models; (2) decoupling RL objectives for binary judgment vs. meta-verification substantially outperforms joint reward optimization. The resulting model, OmniVerifier-M1, delivers robust verification plus fine-grained error localization, and powers M1-TTS, a verifier-driven agentic generation system with dynamic region-level self-correction. No specific benchmark scores, dataset sizes, or compute figures are disclosed in the abstract.

Key Findings:

Symbolic outputs (bounding boxes) are stronger meta-verification rationales than textual explanations and enable judge-free, rule-based RL rewards.
Decoupling RL objectives for binary judgment vs. meta-verification substantially outperforms joint reward optimization due to differences in output structure and learning dynamics.
A verifier with fine-grained localization can drive an agentic generation loop (M1-TTS) that performs region-level self-correction at inference time.

Technical Novelty: Two coupled innovations: (1) using symbolic verifier outputs (bounding boxes) as meta-verification rationales so RL rewards become rule-based and judge-model-free, and (2) decoupling the RL objectives for binary judgment and meta-verification rather than jointly optimizing — motivated by their different output structures and learning dynamics. The downstream M1-TTS agentic system that uses the verifier's localized errors for region-level self-correction during generation is also novel.

What's New: Prior multimodal verifiers focus on decision-only signals or rely on model-based rewards from auxiliary LLM judges; this work reframes verification as meta-verification with symbolic rationales, replaces judge-model rewards with rule-based ones, and decouples the two RL objectives — then closes the loop by using the verifier itself to steer generation via region-level correction.

Extension Opportunities:

Apply the symbolic-rationale + decoupled-RL recipe to video verification, where symbolic outputs become spatio-temporal tubes enabling frame-level self-correction in video generation pipelines.
Integrate OmniVerifier-M1 as a reward model in RLHF/RLVR loops for image generation models (e.g., diffusion or autoregressive image models) to drive region-level reward shaping instead of whole-image preference signals.
Extend M1-TTS-style agentic self-correction to 3D or UI/document domains, where bounding-box-like symbolic outputs (e.g., element boxes, layout regions) can supervise iterative regeneration.

Replicability: The abstract does not mention released code, weights, or datasets. Reproduction would likely require a strong multimodal base model (e.g., a Qwen-VL- or InternVL-class backbone), grounded annotation data with bounding-box rationales, and a multi-stage RL pipeline (separate reward heads for binary and symbolic outputs) — plausibly tens to low-hundreds of GPU-days on A100/H100-class hardware, though exact compute is not stated.

Research Gaps:

No quantitative benchmark numbers, ablations, or comparisons to prior verifiers are disclosed in the abstract, making the magnitude of gains hard to assess.
Symbolic rationales are demonstrated via bounding boxes; it is unclear how the approach generalizes to non-spatial errors (semantic, compositional, temporal, or attribute-binding errors not localizable as boxes).

2. Rethinking Memory as Continuously Evolving Connectivity

Authors: Jizhan Fang, Buqiang Xu, Zhixian Wang... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can memory-augmented LLM agents move beyond static repositories with fixed retrieval pipelines to handle dynamic agentic environments where feedback, task variation, and heterogeneous signals continuously reshape what should be remembered and how it connects?

Summary: FluxMem reframes agent memory as a heterogeneous graph whose topology continuously evolves through three stages — initial connection, feedback-driven refinement, and long-term consolidation — rather than a static retrieval store. It repairs links, prunes interference, and distills recurrent successful trajectories into reusable procedural circuits, achieving SOTA on LoCoMo, Mind2Web, and GAIA.

Key Results: FluxMem achieves state-of-the-art performance across three fundamentally distinct benchmarks: LoCoMo (long conversational memory), Mind2Web (web agents), and GAIA (general assistant tasks). The abstract reports 'consistent SOTA' but does not disclose specific delta percentages versus baselines.

Key Findings:

Treating memory connectivity as evolvable (not fixed) yields consistent gains across three structurally different benchmarks spanning conversation, web navigation, and general assistant tasks
Distilling recurrent successful trajectories into procedural circuits enables skill reuse without explicit fine-tuning
A single generalizability/maturity metric can govern heterogeneous graph operations (link repair, pruning, abstraction alignment) coherently

Technical Novelty: Three-stage evolutionary lifecycle for memory graphs (initial connection → feedback-driven refinement → long-term consolidation) combined with procedural circuit distillation from recurrent successful trajectories, governed by a unified metric for generalizability and evolutionary maturity. Prior memory frameworks (MemGPT, Mem0, A-Mem) treat memory as static stores with fixed retrieval; FluxMem makes topology itself the learnable substrate.

What's New: Shifts the memory abstraction from 'what to store and retrieve' to 'how connections should evolve.' Unlike Mem0/MemGPT/A-Mem which optimize storage representation and retrieval policy over a fixed structure, FluxMem makes the graph topology itself the object of continuous refinement, and introduces procedural circuits as a distinct memory tier beyond episodic/semantic.

Extension Opportunities:

Apply the connectivity-evolving graph approach to multi-agent shared memory systems where agents collaboratively prune/repair links across a federated memory graph
Replace the single generalizability/maturity metric with learned reward signals (RL or DPO) to optimize topology refinement end-to-end
Extend procedural circuit distillation into a code-generation pipeline that compiles recurrent successful trajectories into executable tool/skill libraries

Replicability: Code promised at https://github.com/zjunlp/LightMem (ZJU-NLP lab, established repo with prior releases). Reproducing requires running LoCoMo, Mind2Web, and GAIA evaluations — moderate compute (LLM API calls for agent rollouts plus graph storage); no model training implied, so a single workstation with API access should suffice.

Research Gaps:

No reported ablations in the abstract isolating contributions of each of the three stages or the procedural circuit mechanism
Unclear how the framework scales as the graph grows unboundedly over long-horizon deployments, and what the consolidation cost is

⚡ CUDA/Systems

1. CubePart: An Open-Vocabulary Part-Controllable 3D Generator

Authors: Yiheng Zhu, Kangle Deng, Jean-Philippe Fauconnier... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can generative 3D models produce meshes decomposed into application-specific semantic parts (aligned to user-defined schemas) rather than monolithic meshes or arbitrary, non-controllable part decompositions?

Summary: CubePart is a text-to-3D generative framework that produces meshes decomposed according to a user-supplied open-vocabulary parts schema. It combines a scalable part-labeled 3D data pipeline with a two-stage architecture that separates global shape synthesis from part-level decoding, yielding assets directly usable in game engines without manual cleanup.

Key Results: The paper introduces CubePart, demonstrating open-vocabulary part-controllable mesh generation conditioned on a text prompt plus a user-defined parts schema. The abstract reports qualitative demonstration that generated assets integrate directly into game engines and are driven by animation/behavior scripts without manual post-processing. No quantitative benchmarks, dataset sizes, or comparison metrics are stated in the abstract.

Key Findings:

Part decomposition can be exposed as an explicit inference-time control rather than a fixed taxonomy, enabling open-vocabulary user schemas.
Decoupling global shape synthesis from per-part decoding produces parts that still assemble into a coherent whole.
Generated assets are claimed to drop into game engines and respond to animation/behavior scripts without post-processing.

Technical Novelty: Treating part structure as an explicit inference-time control signal via an open-vocabulary schema, combined with a two-stage architecture that decouples global shape synthesis from per-part decoding — and a scalable pipeline to build the open-vocabulary part-labeled 3D dataset that makes this trainable.

What's New: Prior generative 3D systems output monolithic meshes or unconstrained part splits; CubePart conditions generation on a user-defined open-vocabulary parts schema and supplies the data pipeline + architecture needed to make schema-controllable generation feasible.

Extension Opportunities:

Extend the parts schema from open-ended names to richer constraints (kinematic joints, mass, collision primitives) so generated parts come pre-rigged for physics/animation.
Add a feedback loop where a game engine evaluates animation/collision quality and fine-tunes the part decoder via RL or DPO on engine-level rewards.
Generalize the two-stage architecture to scenes (multi-object, part-aware) or to other modalities like image-to-part-controllable-3D for asset retopology of scanned objects.

Replicability: A project page (cubepart.github.io) is linked, but the abstract does not mention released code, weights, or the dataset. Compute is not disclosed; comparable open-vocabulary 3D generators typically require multi-GPU training over days to weeks on large mesh corpora, so independent reproduction would be expensive without released artifacts.

Research Gaps:

No quantitative metrics, baselines, or ablations reported in the abstract — unclear how part fidelity and global coherence compare to existing part-aware 3D generators.
Limits of the schema (e.g., very fine-grained parts, novel categories outside the dataset distribution) and physical/kinematic correctness of parts are not characterized.

🦾 ROBOTICS

1. PrimitiveVLA: Learning Reusable Motion Primitives for Efficient and Generalizable Robotic Manipulation

Authors: Yutai Li, Shaohui Peng, Jiaming Guo... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can VLA models overcome data inefficiency and poor generalization caused by Direct Instruction-to-Control Mapping that forces memorization of monolithic trajectories instead of reusable motion patterns?

Summary: PrimitiveVLA reframes VLA training as a disassemble-and-assemble process: demonstrations are automatically broken into reusable motion primitives sharing a Multimodal Canonical Representation, then reassembled at inference by a VLM planner and an LLM-generated switch module. This shift away from monolithic instruction-to-control mapping yields improved data efficiency and stronger zero-shot generalization on unseen and long-horizon manipulation tasks.

Key Results: The paper demonstrates that PrimitiveVLA improves data efficiency and achieves superior zero-shot generalization across unseen and long-horizon tasks compared to standard VLA approaches. The abstract claims 'extensive experiments' but does not cite specific numerical benchmarks, datasets, or quantitative success rates.

Key Findings:

Direct Instruction-to-Control Mapping is identified as the root cause of VLA data inefficiency and poor generalization
Disassembling demonstrations into reusable primitives lets models learn invariant motion patterns rather than task-specific trajectories
A VLM planner combined with an LLM-generated switch module enables robust closed-loop primitive assembly at inference, improving long-horizon task performance

Technical Novelty: The Primitive-Centric Disassemble & Assemble paradigm itself — specifically the combination of (1) a Multimodal Canonical Representation (MCR) shared across phases, (2) an automated demonstration-disassembly pipeline producing reusable primitives at fine-tuning time, and (3) inference-time assembly via a VLM planner plus LLM-generated switch module for closed-loop execution. This contrasts with prior VLA work that maps instructions directly to control.

What's New: Unlike prior VLA models that learn monolithic instruction-to-trajectory mappings, PrimitiveVLA introduces an explicit primitive abstraction with a shared multimodal canonical representation bridging automated disassembly during training and VLM/LLM-driven assembly during inference, treating primitives as first-class reusable units.

Extension Opportunities:

Extend the automated disassembly pipeline to handle bimanual or multi-robot coordination tasks where primitives must be synchronized across agents
Build a cross-embodiment primitive library that transfers MCR-encoded primitives between different robot morphologies (arms, humanoids, mobile manipulators)
Replace the LLM-generated switch module with a learned reinforcement-learning controller that adapts primitive transitions from real-world failure feedback

Replicability: The abstract does not mention code, model weights, dataset releases, or compute requirements. Reproduction would likely require VLA-scale fine-tuning compute (multi-GPU), access to a VLM and LLM for the planner/switch modules, and robotic demonstration data — making full reproduction non-trivial without released artifacts.

Research Gaps:

Abstract lacks concrete benchmark numbers, baselines compared against, and dataset specifics, making the magnitude of improvement unclear
Unclear how the automated disassembly pipeline handles ambiguous or continuous motions that lack natural primitive boundaries, and how robustly the LLM-generated switch module performs under distribution shift

2. SPRINT: Efficient Spectral Priors for Humanoid Athletic Sprints

Authors: Yantong Wei, Kaihong Huang, Hainan Pan... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can humanoid robots achieve stable, high-speed athletic sprinting when humanoid-viable kinematic reference data is scarce and existing frameworks fail to maintain stability at sprint speeds?

Summary: SPRINT introduces frequency-domain spectral priors derived from just 5 reference motion sequences to train humanoid sprint policies that generalize beyond their training distribution. The approach achieves zero-shot sim-to-real transfer on the Unitree G1, reaching 6 m/s with biomimetic gait transitions, establishing spectral priors as a data-efficient foundation for humanoid athletics.

Key Results: SPRINT achieves zero-shot sim-to-real transfer on the Unitree G1 humanoid platform, reaching a peak sprinting velocity of 6 m/s using only 5 discrete reference motion sequences. The frequency-adaptive spectral priors successfully extrapolate beyond the reference velocity distribution and enable seamless gait transitions while preserving biomimetic naturalness.

Key Findings:

Frequency-domain priors built from only 5 motion sequences enable extrapolation to sprint velocities exceeding the reference distribution
Zero-shot sim-to-real transfer on Unitree G1 reaches 6 m/s peak velocity in field experiments
The policy demonstrates seamless gait transitions (walk/run/sprint) while preserving biomimetic naturalness

Technical Novelty: The core novelty is characterizing human locomotion via frequency-domain spectral priors rather than time-domain motion capture trajectories. By exploiting the fundamental periodicity of gait, the method extracts compact frequency-adaptive priors from just 5 reference sequences that extrapolate to unseen velocities, avoiding the data-hungry imitation learning paradigm.

What's New: Prior humanoid locomotion work relies on large mocap datasets in the time domain and struggles at sprint speeds. SPRINT is the first to use frequency-adaptive spectral characterization as the prior, dramatically reducing data requirements (5 sequences) while enabling velocity extrapolation that time-domain imitation cannot achieve.

Extension Opportunities:

Extend spectral priors to other athletic skills (jumping, cutting, lateral sprints) by building frequency-domain libraries for non-periodic or quasi-periodic motions
Apply the frequency-adaptive prior approach to quadruped or bipedal animals with different morphologies to test generalization of spectral characterization
Combine spectral priors with terrain-aware perception modules to enable sprinting on uneven outdoor surfaces rather than flat fields

Replicability: An anonymous project page (anonymous.4open.science/w/SPRINT-138A/) is referenced suggesting code/materials availability. Reproduction would require a Unitree G1 humanoid (~$16K+), GPU compute for RL policy training (likely Isaac Gym/Sim with 1-4 GPUs), and a safe outdoor sprint testing area.

Research Gaps:

Lack of humanoid-viable kinematic reference data for high-speed athletic motions
Existing RL/imitation frameworks fail to maintain stability and naturalness at sprint velocities

3. Digital-Based Potentiostat and Mesoporous Microelectrode Co-Design for Non-Enzymatic Glucose Detection at 0.3V-VDD and 1.65nW-Power

Authors: Andrea De Gregorio, Mara Serrapede, Danilo Kaddouri... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can a chronoamperometric electrochemical sensor for non-enzymatic glucose detection be designed to operate at ultra-low voltage and ultra-low power while maintaining the dynamic range and linearity required for physiological-level sensing at the point of care?

Summary: The paper introduces a digital-based potentiostat co-designed with mesoporous microelectrodes for non-enzymatic glucose sensing in 130nm CMOS, operating at a record-low 0.3V supply and 1.65nW power. It provides the first frequency-domain analytical model of the DB architecture and demonstrates physiological glucose detection with selectivity against ascorbic acid under aerobic conditions.

Key Results: Demonstrated a 130nm CMOS Digital-Based (DB) reconfigurable potentiostat that detects electrochemical currents from 600pA to 650nA with R²=0.991 linearity, consuming only 1.65nW at Vdd=300mV (53.5nW at Vdd=500mV). Successfully detected glucose at physiological levels using nanostructured mesoporous microelectrodes, even in the presence of ascorbic acid interferent and under aerobic conditions — claimed as the lowest reported voltage and power for this application.

Key Findings:

DB potentiostat spans 600pA–650nA detection range with R²=0.991 linearity at sub-µW power.
Operates at 300mV Vdd consuming 1.65nW — orders of magnitude below conventional analog potentiostats.
Non-enzymatic glucose detection works at physiological levels with selectivity against ascorbic acid in aerobic conditions.

Technical Novelty: First analytical frequency-domain linearized model for signal transfer and noise in a digital-based potentiostat, combined with co-design of the IC and nanostructured non-enzymatic mesoporous microelectrodes — enabling sub-nW operation at 0.3V, well below conventional analog transimpedance/switched-capacitor potentiostats that typically require >0.8V.

What's New: Combines a digital-based (rather than analog op-amp-based) potentiostat with non-enzymatic mesoporous electrodes, and provides the first linearized small-signal frequency-domain model of this DB topology — enabling principled noise/bandwidth design rather than purely empirical tuning.

Extension Opportunities:

Co-integrate with energy harvesting (e.g., body-heat thermoelectric or biofuel cells) to build a fully self-powered implantable/wearable glucose patch, exploiting the 1.65nW budget.
Extend the digital-based potentiostat to a multi-analyte array (lactate, cortisol, urea) by reconfiguring the bias and reusing the same DB architecture for parallel mesoporous electrodes.
Develop a closed-loop insulin delivery prototype by pairing this readout with a low-power MCU and BLE radio, validating drift and calibration stability over days of continuous wear.

Replicability: No mention of open code, RTL, or measurement datasets in the abstract. Reproduction would require a 130nm CMOS tapeout (significant cost/access via MPW shuttles), fabrication of mesoporous microelectrodes, and an electrochemical test bench with glucose/ascorbic acid solutions — moderate-to-high barrier for academic replication.

Research Gaps:

No long-term stability, drift, or in-vivo validation reported; physiological detection is demonstrated in controlled buffer conditions only.
Selectivity is shown against a single interferent (ascorbic acid); real biofluids contain many electroactive species (uric acid, acetaminophen, dopamine) not addressed.

💻 COMPUTE

1. A cryogenic apparatus for coupling two-dimensional materials to a confocal multimode optical cavity

Authors: Han S. Hiller, Pranav Parakh, Samuel H. Aronson... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can one build a cryogenic, vibration-stabilized apparatus that couples 2D van der Waals materials (e.g., TMDs) to a high-finesse confocal multimode optical cavity to enable cavity-enhanced CW Raman driving of phonons/CDWs while preserving sample tunability and exchangeability?

Summary: The paper describes a cryogenic UHV apparatus that couples 2D van der Waals materials to a length-tunable, mm-scale confocal Fabry-Pérot cavity, preserving light-matter coupling enhancement via a localized confocal supermode while permitting sample alignment, exchange, and electrical gating. The platform is designed to enable cavity-enhanced CW Raman driving of phonons or charge density waves in materials such as TMDs.

Key Results: The authors demonstrate an ultrahigh-vacuum apparatus housing a length-tunable Fabry-Pérot cavity with intracavity TMD sample, cryogenically cooled and vibration-stabilized. They show that operating near confocal geometry concentrates the optical field into a localized supermode enhancing light-matter coupling, and that this enhancement is preserved at millimeter-scale cavity lengths. A four-axis nanopositioner provides sample alignment plus electrical leads for carrier density modulation; transmission through the multimode cavity enables in situ imaging for alignment. No quantitative finesse, cooperativity, or temperature numbers are given in the abstract.

Key Findings:

Near-confocal operation concentrates the optical field into a localized supermode that substantially enhances light-matter coupling even at mm-scale cavity lengths
A four-axis nanopositioner enables in situ alignment plus electrical contact for carrier-density tuning and transport measurements on the intracavity sample
Multimode cavity transmission doubles as an imaging channel for sample alignment, integrated with cryogenic UHV and vibration isolation

Technical Novelty: Combines four normally-incompatible requirements in one platform: (1) millimeter-scale length-tunable confocal cavity that still concentrates field into a localized supermode, (2) UHV + cryogenic operation, (3) vibration stabilization sufficient for cavity locking, and (4) four-axis nanopositioning with electrical leads for in situ gating/transport. Prior cavity-2D-material work used either fiber/micro-cavities (no sample exchange, no gating) or free-space cavities lacking cryogenic UHV integration.

What's New: First integrated platform combining a length-tunable confocal multimode cavity with cryogenic UHV, vibration stabilization, and gated 2D-material sample handling — bridging the gap between fiber-microcavity QED setups (no gating, limited tunability) and standard cryostat optics (no cavity enhancement).

Extension Opportunities:

Integrate pulsed/multi-tone drive schemes to study transient phonon condensation dynamics or Floquet-engineered CDW phases beyond steady-state CW operation
Swap the TMD for moiré heterostructures (twisted bilayer graphene, WSe2/MoSe2) and use the gate-tunable carrier density to map cavity-modified correlated phases vs filling
Add cavity-coupled magneto-transport by integrating a superconducting magnet, enabling exploration of cavity-renormalized quantum Hall or excitonic insulator states

Replicability: The paper is an instrumentation/apparatus description, so the 'reproducibility' question is hardware build cost rather than compute. Expected to require a custom cryostat with optical access, UHV chamber, low-vibration mounts, piezo nanopositioners, high-reflectivity mirrors, and length-locking electronics — likely several hundred thousand USD in capital. No code/data availability mentioned in the abstract.

Research Gaps:

Quantitative characterization (finesse, cooperativity, base temperature, vibration spectrum) is not given in the abstract — needed to benchmark against alternative cavity-QED platforms
Demonstration of an actual cavity-driven collective phase (superradiant phonon population, CDW manipulation) remains future work; the paper establishes the tool, not the physics result

2. How Far Can Disaggregation Go? A Design-Space Exploration of Attention-FFN Disaggregation for Efficient MoE LLM Serving

Authors: Hanjiang Wu, Abhimanyu Rajeshkumar Bambhaniya, Sarbartha Banerjee... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: When does each level of LLM inference disaggregation (chunked-prefill, P/D, AFD) actually pay off for MoE serving, and how should attention/FFN be partitioned across GPUs as a function of workload and architecture under strict TTFT/TPOT SLOs?

Summary: The paper systematically explores when operator-level Attention-FFN Disaggregation (AFD) outperforms shallower disaggregation strategies (chunked-prefill, P/D) for serving MoE LLMs. Combining on-device kernel profiling with network simulation, it shows AFD sustains ~4k tokens/s on DeepSeek-V3.2 under strict TTFT/TPOT SLOs across chat, coding, and agentic workloads where non-AFD configurations cannot meet SLOs, and distills partitioning principles for attention vs FFN at rack and cluster scale.

Key Results: Using a framework fusing on-device kernel measurements with high-fidelity network simulation, the authors show AFD sustains ~4k tokens/s system throughput on DeepSeek-V3.2 across chat, coding, and agentic-coding workloads under strict TTFT/TPOT SLOs — regimes where non-AFD deployments are infeasible. The study spans realistic workloads varying input/output sequence lengths, prefix-KV reuse, and per-user latency constraints.

Key Findings:

AFD enables feasible serving under strict TTFT/TPOT SLOs in regimes where chunked-prefill and P/D disaggregation cannot meet targets
DeepSeek-V3.2 achieves ~4k tokens/s system throughput with AFD across chat, coding, and agentic-coding workloads
Optimal attention/FFN partitioning depends jointly on workload (sequence length, prefix-KV reuse, latency SLO) and model architecture, not a single fixed split
MoE inference exposes three distinct resource profiles — memory-bound attention, compute-intensive expert FFN, and dispatch/combine communication — that benefit from independent scaling

Technical Novelty: First systematic design-space exploration that treats operator-level Attention-FFN Disaggregation as a tunable knob alongside chunked-prefill and P/D — characterizing trade-offs jointly across workload, resource allocation, and interconnect topology rather than proposing a single point design. The hybrid methodology (real kernel measurements + high-fidelity network simulation) is also a distinguishing contribution for cluster-scale serving studies.

What's New: Prior work proposed AFD as a point design; this paper is the first to map the full disaggregation design space (chunked-prefill → P/D → AFD) against realistic MoE workloads and SLOs, identifying when each level pays off rather than advocating one. The unified kernel-measurement + network-simulation methodology bridges single-node profiling and cluster-scale studies.

Extension Opportunities:

Extend the kernel+network simulation framework to other MoE models (Mixtral, Qwen-MoE, GPT-OSS) and emerging architectures (hybrid SSM-attention) to test generality of AFD partitioning heuristics
Build an online scheduler that dynamically switches between chunked-prefill, P/D, and AFD modes based on live workload signals (sequence-length distribution, prefix-cache hit rate, SLO headroom)
Explore AFD over alternative interconnect topologies (CXL, optical switches, scale-up Ethernet) and quantify break-even points vs NVLink/InfiniBand assumed in the paper

Replicability: The abstract does not mention released code or simulator artifacts. Reproduction would require access to H100/B200-class GPUs for kernel measurements on DeepSeek-V3.2 (671B params, MoE) plus a high-fidelity network simulator — likely tens of GPUs minimum and significant simulation compute. Realistically reproducible only at a well-resourced lab.

Research Gaps:

No dynamic/adaptive policy for switching disaggregation levels online as workload mix shifts
Limited coverage of interconnect alternatives beyond current NVLink/IB-class fabrics, and of non-DeepSeek MoE architectures with different expert granularity

3. Heterogeneous Parallelism for Multimodal Large Language Model Training

Authors: Yashaswi Karnati, Kamran Jafari, Akash Mehra... Published: 2026-05-26 | Citations: 0 arXiv | PDF

Research Question: How can multimodal LLM training escape the throughput limits imposed by a single LLM-centric TP/CP/PP/DP/EP layout, where encoders are forced to inherit LLM sharding choices that add communication overhead and constrain parallelism — especially at long contexts where the LLM needs context parallelism for fused sequences but encoder inputs remain bounded?

Summary: The paper introduces heterogeneous parallelism for multimodal LLM training, letting encoders and the LLM each use independent TP/CP/PP/DP/EP layouts and GPU placements within one end-to-end graph, connected by boundary communicators that handle forward/backward layout transforms. Implemented as a Megatron-LM extension, it delivers up to 49.3% TFLOPS/GPU gains (colocated) and 13.0% token-throughput gains (non-colocated) while preserving loss convergence.

Key Results: Implemented as an open-source Megatron-LM extension and evaluated across multimodal workloads at varying GPU scales: colocated heterogeneous parallelism improves TFLOPS/GPU by up to 49.3% over optimized homogeneous baselines; non-colocated heterogeneity improves aggregate token throughput by up to 13.0% and TFLOPS/GPU by up to 9.6%. Loss convergence parity vs. homogeneous baselines is validated.

Key Findings:

Forcing encoders to inherit LLM-driven sharding wastes throughput, and the mismatch grows at long contexts where the LLM needs context parallelism but encoder inputs do not
Colocated heterogeneity (same GPUs, different layouts per module) yields the largest per-GPU compute gains (up to +49.3% TFLOPS/GPU)
Non-colocated heterogeneity (disjoint rank sets per module) gives smaller per-GPU gains (+9.6%) but improves aggregate token throughput by up to +13.0%, and convergence matches homogeneous baselines

Technical Novelty: The abstraction of per-module independent parallelism layouts and rank placements within a single end-to-end training graph, plus 'boundary communicators' that implement forward activation materialization and backward gradient routing across layout transforms — supporting both colocated (shared GPUs) and non-colocated (disjoint rank sets) execution modes with corresponding scheduling extensions. Prior work largely assumed a unified layout across the whole model.

What's New: Unlike prior multimodal training stacks that impose one global parallelism layout, this work treats each module as a first-class parallelism citizen with its own layout and placement, and solves the resulting boundary-tensor problem with explicit forward/backward layout transforms plus scheduler support for both shared and disjoint GPU placements.

Extension Opportunities:

Add an auto-tuner / cost model that searches the joint encoder×LLM layout space (TP/CP/PP/DP/EP per module + colocated vs disjoint placement) given a workload and cluster topology, rather than requiring manual configuration
Extend boundary communicators to support dynamic, per-step layout reconfiguration (e.g., curriculum learning where modality mix or context length shifts mid-training) and to MoE routing across heterogeneous expert placements
Generalize beyond vision+text to >2 modalities (audio, video, point clouds) with per-encoder heterogeneous layouts, and evaluate on inference/serving where prefill vs decode also benefit from disjoint placements

Replicability: Released as an open-source Megatron-LM extension, so code is available. Reproducing the headline numbers requires multi-node GPU clusters at the scales they sweep (multimodal LLM pretraining-class workloads with long contexts), likely tens to hundreds of H100/A100-class GPUs — out of reach for individuals but tractable for industry/academic clusters.

Research Gaps:

No automated search/cost model for choosing layouts and placements — users must still pick the heterogeneous configuration manually
Evaluation focuses on training throughput; inference/serving, fault tolerance under disjoint placements, and >2-modality scaling are not characterized

⚡ ENERGY

1. Ultrafast dynamics of excitons in black phosphorus

Authors: Geoffroy Kremer, Juan F. P. Mosquera, Joël Morf... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: What microscopic mechanism causes decoherence of optically generated excitons during early non-equilibrium dynamics in single-valley direct-gap semiconductors, and what limits coherent light-induced band-structure engineering?

Summary: The authors resonantly generate excitons in black phosphorus with mid-infrared pulses and use trARPES to follow their picosecond-scale dynamics with momentum resolution. A quantum-kinetic model identifies phonon-mediated scattering from bright to dark excitons as the dominant decoherence channel, establishing intravalley phonon scattering as a fundamental bottleneck for coherent exciton control in single-valley semiconductors.

Key Results: Using resonant mid-infrared photoexcitation of black phosphorus combined with time- and angle-resolved photoemission spectroscopy (trARPES), the authors directly track exciton dynamics on few-picosecond timescales. They develop a quantum-kinetic framework modeling decoherence via phonon-mediated scattering of bright excitons into dark excitons, and by fitting simulation to experiment quantify key parameters of the intravalley scattering process. Specific numerical values (rates, coupling constants) are not stated in the abstract.

Key Findings:

Excitons in BP decohere on few-picosecond timescales after resonant mid-IR excitation
Bright-to-dark exciton conversion via phonon scattering is the dominant decoherence pathway, captured quantitatively by a quantum-kinetic model
Intravalley phonon scattering — not intervalley — limits coherent exciton phenomena in single-valley direct-gap semiconductors like BP

Technical Novelty: Direct momentum-resolved observation of exciton decoherence dynamics in black phosphorus via resonant mid-IR pumping in trARPES, paired with a quantum-kinetic theory that explicitly models bright-to-dark exciton conversion through phonon scattering — rather than treating decoherence phenomenologically.

What's New: Prior exciton decoherence studies in 2D materials (notably TMDs) emphasized intervalley scattering between K/K' valleys; this work isolates intravalley phonon scattering in a single-valley system and pairs momentum-resolved trARPES with a first-principles-style quantum-kinetic treatment of bright/dark exciton populations.

Extension Opportunities:

Apply the same trARPES + quantum-kinetic framework to other single-valley 2D semiconductors (e.g., InSe, GaSe) to test universality of phonon-mediated intravalley decoherence
Engineer suppression of intravalley phonon scattering via strain, dielectric environment, or heterostructure encapsulation of BP to extend exciton coherence times
Combine with Floquet driving experiments to determine whether decoherence rates extracted here predict the threshold field for observing light-induced band renormalization

Replicability: No code/data availability mentioned in the abstract. Reproduction requires a trARPES endstation with mid-IR OPA pump (~0.3 eV photon energy for BP gap), high-quality cleaved BP single crystals, UHV/cryogenic capability, and expertise implementing semiconductor Bloch / quantum-kinetic equations — capital-intensive, only feasible at specialized ultrafast spectroscopy facilities.

Research Gaps:

Quantitative role of specific phonon modes (acoustic vs optical, in-plane anisotropy of BP) in the scattering channel is not fully resolved
Implications for actually achieving Floquet-engineered band structures — i.e., the field strength and pulse duration regime where coherent control survives decoherence — remain to be mapped

2. Magneto-Optical Detection of Anisotropic Spin Currents in Altermagnetic RuO2

Authors: Joongwon Lee, Jeonglyul Kim, Sreejith Nair... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: Does RuO2 sustain robust altermagnetic order beyond the ultrathin limit, and can its predicted anisotropic spin-splitter effect be optically detected for efficient charge-to-spin conversion?

Summary: The authors use polarization-resolved SHG and polar MOKE on a 12-nm (101)-RuO2/sapphire film to optically detect a current-induced, crystallographically anisotropic spin polarization consistent with the altermagnetic spin-splitter effect. The anisotropy (strong along [010], suppressed along [-101]) cannot be explained by Rashba or non-magnetic mechanisms, and persistent strain seen by STEM stabilizes the altermagnetic order beyond the ultrathin limit.

Key Results: In a 12-nm-thick (101)-oriented RuO2 film on sapphire, polarization-resolved second-harmonic generation (SHG) confirmed surface symmetry and Néel order. Under applied current, both SHG and polar MOKE detected directionally anisotropic current-induced spin polarization: enhanced signals for current along [010] and strongly suppressed signals along [-101], matching the altermagnetic spin-splitter symmetry. STEM revealed substantial residual strain even in thicker films, ruling out Rashba/non-magnetic origins.

Key Findings:

SHG polarization patterns match the surface symmetry and Néel order expected for altermagnetic RuO2
Current-induced spin polarization shows pronounced anisotropy: enhanced along [010], suppressed along [-101], matching spin-splitter symmetry
Substantial residual strain persists in 12-nm films per STEM, plausibly stabilizing altermagnetic order beyond the ultrathin regime

Technical Novelty: First use of combined polarization-resolved SHG and polar MOKE as symmetry-selective optical probes to detect the anisotropic spin-splitter effect in a relatively thick (12 nm) altermagnetic RuO2 film, rather than relying on transport or neutron measurements limited to ultrathin samples.

What's New: Provides optical (rather than transport) symmetry-resolved evidence of the altermagnetic spin-splitter effect in RuO2 at a thickness where bulk magnetic order in RuO2 has been actively contested, and links the observation to strain visualized by STEM.

Extension Opportunities:

Integrate the (101)-RuO2 film into spin-orbit-torque MRAM-style stacks to quantify field-free switching efficiency vs. heavy-metal references like Pt/Ta
Systematically vary film thickness (5-50 nm) and strain (via different substrates) to map the phase boundary where altermagnetic order survives
Combine the SHG/MOKE optical readout protocol with terahertz emission spectroscopy to time-resolve spin-current generation dynamics in altermagnets

Replicability: No code/data availability mentioned in abstract. Reproduction requires PLD/MBE growth of epitaxial (101)-RuO2 on sapphire, ultrafast laser SHG setup, polar MOKE rig with current injection, and aberration-corrected STEM — substantial experimental physics infrastructure, no significant compute needed.

Research Gaps:

Quantitative charge-to-spin conversion efficiency vs. established spin-Hall materials is not established
The role and tunability of epitaxial strain in stabilizing altermagnetism across thicknesses and substrates remains open

3. Orbital Altermagnetic Photonic Crystal

Authors: Sichang Qiu, Huichang Li, Yan Meng... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: Can altermagnetism — a momentum-dependent spin-splitting phenomenon discovered in electronic systems — be realized in photonic systems despite the fundamental fermion-boson divide that makes direct translation impossible?

Summary: The paper reports the first experimental realization of altermagnetism in a photonic system via an orbital mechanism enforced by C_{4z}T antiunitary symmetry, producing momentum-dependent pseudospin splitting with a d_{xy}-wave form factor. The crystal supports pseudospin-selective wave transport including pseudospin splitting and filtering, extending altermagnetism beyond fermionic systems into photonics.

Key Results: Experimentally demonstrated the first orbital altermagnetic photonic crystal using antiunitary C_{4z}T symmetry coupling a local p-orbital σ/π doublet to crystal momentum. Measured band structures and iso-frequency contours confirm momentum-dependent pseudospin splitting with a d_{xy}-wave form factor and alternating pseudospin polarization. Demonstrated pseudospin-selective transport including photonic pseudospin splitting and pseudospin filtering.

Key Findings:

C_{4z}T symmetry enforces a correspondence between local p-orbital σ/π doublets and crystal momentum, enabling altermagnetic-like behavior in photons
Measured band structures and iso-frequency contours exhibit d_{xy}-wave momentum-dependent pseudospin splitting with alternating polarization
The crystal enables pseudospin-selective transport phenomena including photonic pseudospin splitting and pseudospin filtering

Technical Novelty: Prior altermagnetism work was confined to fermionic electronic systems; this paper introduces an orbital-based mechanism leveraging p-orbital σ/π doublets as pseudospin and uses C_{4z}T symmetry to enforce momentum-spin coupling in bosonic photonic systems — bypassing the lack of intrinsic spin in photons.

What's New: First photonic realization of altermagnetism, bridging a fundamental gap between condensed-matter altermagnetism (fermionic) and bosonic systems via an orbital symmetry-protected mechanism rather than attempting to mimic electron spin directly.

Extension Opportunities:

Design 3D orbital altermagnetic photonic crystals with d-wave or g-wave form factors using higher-order rotation symmetries (C_{6z}T, C_{3z}T) for richer pseudospin textures
Integrate the pseudospin-filtering functionality into on-chip silicon photonic circuits to build polarization-routing devices and pseudospin demultiplexers for optical communications
Explore nonlinear or non-Hermitian extensions where gain/loss interacts with the C_{4z}T symmetry to create reconfigurable or topologically robust altermagnetic phases

Replicability: No code/data availability stated in the abstract. Reproduction requires microwave/photonic crystal fabrication (likely dielectric resonator arrays), network analyzer for band structure measurement, and near-field scanning setup for iso-frequency contour mapping — standard photonic metamaterial lab equipment, no significant compute required.

Research Gaps:

Lack of theoretical/experimental frameworks for translating spintronic concepts (altermagnetism, spin Hall effects) into photonic platforms
Absence of devices that leverage pseudospin as a degree of freedom for electromagnetic wave routing and filtering

🔬 MATERIALS

1. A variability-aware simulation and design workflow for wafer-scale, heterogeneously integrated lithium niobate modulators

Authors: Patrick Nenezic, Ewoud Vissers, Arno Moerman... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can wafer-scale heterogeneously integrated lithium niobate (LN) traveling-wave modulators be designed to maintain stable performance despite fabrication variability inherent to micro-transfer printing on silicon photonics platforms?

Summary: The paper introduces a variability-aware simulation and design workflow for heterogeneously integrated lithium niobate traveling-wave modulators that leverages real fabrication-variation data from a dedicated pilot line. By optimizing geometric parameters against measured manufacturing tolerances, the authors theoretically establish that wafer-scale LN-on-silicon-photonics integration via micro-transfer printing can be systematically engineered for stable performance.

Key Results: The paper theoretically demonstrates feasibility of reliable wafer-scale LN modulator integration on silicon photonics via micro-transfer printing by incorporating measured fabrication-variation data from a dedicated pilot line into a simulation framework that optimizes geometric parameters. Specific quantitative benchmarks (e.g., Vπ·L, bandwidth, yield percentages) are not disclosed in the abstract.

Key Findings:

Fabrication-variation data from a real pilot line can be directly embedded into modulator design optimization workflows
Geometric parameter optimization can compensate for variability inherent to micro-transfer-printing-based heterogeneous integration
Wafer-scale reliable LN modulator integration on silicon photonics is theoretically feasible when design is variability-aware

Technical Novelty: The integration of empirically measured fabrication-variation statistics from a dedicated micro-transfer-printing pilot line directly into a co-simulation/optimization workflow for traveling-wave LN modulators — moving beyond nominal-value design toward statistically robust, manufacturing-aware geometric optimization at wafer scale.

What's New: Unlike prior LN modulator design work that uses nominal geometric parameters or generic tolerance assumptions, this work closes the loop between an actual heterogeneous-integration pilot line and the modulator design optimizer, enabling manufacturing-statistics-driven robust design specifically for the micro-transfer-printing process.

Extension Opportunities:

Extend the variability-aware framework to incorporate temperature and aging drift models for long-term reliability prediction of deployed LN modulators
Apply the same methodology to other heterogeneously integrated materials (e.g., InP, BTO, polymer EO) to create a unified variability-aware design platform for multi-material PICs
Build a machine-learning surrogate model trained on the pilot-line variation data to enable real-time design-space exploration and inverse design of robust modulator geometries

Replicability: The abstract does not mention public code, datasets, or pilot-line variation data release. Reproduction would require access to a heterogeneous LN-on-Si micro-transfer-printing pilot line plus EM/RF/optical co-simulation tools (e.g., Lumerical, COMSOL, ADS); compute is modest (workstation-scale Monte Carlo / parametric sweeps) but the empirical variation dataset is the critical bottleneck.

Research Gaps:

Lack of experimental validation — the demonstration is theoretical; measured wafer-scale yield and performance distributions on fabricated devices are not presented
No discussion of how the framework handles correlated variations across the wafer or process drift over time

2. Simultaneous Measurement of Circular Dichroism and Circular Differential Scattering

Authors: Qiang Hao, Pathum Wathudura, Huy Pham... Published: 2026-05-27 | Citations: 0 arXiv | PDF

Research Question: How can circular dichroism (CD) and circular differential scattering (CDS) spectra be acquired simultaneously from the same sample under identical conditions, given that prior chiroptical setups measure them separately and CDS baseline correction is notoriously difficult?

Summary: The paper introduces a dual-channel spectrometer that simultaneously measures circular dichroism and circular differential scattering from the same solution, addressing the long-standing inability to compare these complementary chiroptical signals under identical conditions. Validation on achiral-scatterer/chiral-absorber mixtures and plasmonic gold helicoids demonstrates the instrument can distinguish whether CD and CDS originate from coupled or independent physical mechanisms.

Key Results: The authors built a dual-channel spectrometer that concurrently captures CD and CDS spectra from a single solution, validated on two model systems: (1) ammonium d-10 camphor sulfonate mixed with achiral polystyrene nanoparticles (PSNPs), where CDS showed opposite sign to CD (because PSNPs are achiral scatterers modulated by chiral absorption), and (2) plasmonic gold helicoid nanoparticles, where CD and CDS displayed matched resonance wavelengths and stronger right-handed CPL response, confirming both signals arise from shared plasmonic modes. A scattering spectral matching method is introduced to solve CDS baseline correction.

Key Findings:

Concurrent ensemble-averaged CD and CDS acquisition is feasible with a dual-channel design, eliminating sample/condition variability between measurements
For achiral scatterers mixed with chiral absorbers, CDS exhibits opposite sign to CD, indicating scattering inherits a chirality signature solely through modulation by chiral absorption
For gold helicoid nanoparticles, matched CD/CDS resonance wavelengths and shared handedness preference prove chiral absorption and scattering originate from the same plasmonic resonance modes

Technical Novelty: First reported instrument for concurrent ensemble-averaged CD and CDS acquisition from the same solution, plus a novel scattering spectral matching method for CDS baseline correction — prior work required separate measurements with sample/condition mismatches.

What's New: First experimental demonstration of simultaneous ensemble-averaged CD + CDS acquisition, combined with a new scattering spectral matching baseline-correction method that resolves the persistent CDS baseline problem.

Extension Opportunities:

Extend the dual-channel architecture to add circularly polarized luminescence (CPL) as a third channel, enabling simultaneous CD/CDS/CPL on the same sample for fully correlated chiroptical fingerprints
Apply the technique to biological chiral assemblies (amyloid fibrils, chiral supramolecular peptides, virus capsids) to disentangle absorption vs scattering contributions in turbid biosamples where conventional CD is corrupted by scattering artifacts
Couple the spectrometer with stopped-flow or temperature-jump kinetics to capture time-resolved CD+CDS during chiral self-assembly, distinguishing nucleation (scattering-dominated) from growth (absorption-dominated) phases

Replicability: Abstract does not mention code or data release. Reproduction would require custom optics hardware (polarization modulators, dual detection channels, monochromator) rather than significant compute; primary cost is instrumentation (~lab-scale optical bench) and synthesis of validation samples (camphor sulfonate, PSNPs, gold helicoids).

Research Gaps:

CDS baseline correction was unsolved in prior ensemble measurements, limiting quantitative comparison with CD
Previous CD and CDS data came from separate runs, introducing concentration/aggregation/temperature mismatches that obscured whether differences were physical or experimental

🔥 GitHub Trending

1. ATOM00blue/machine-learning-library

⭐ 35 stars | Unknown

A hand-curated library of the best machine learning education — 590 docs (78 arXiv papers, 474 course lectures from Stanford/MIT/Karpathy/fast.ai, 38 explainer articles), normalized to Markdown with f

arxiv corpus dataset deep-learning education llm

2. marvinified/voctar

⭐ 6 stars | TypeScript

Simple TypeScript library with RAG primitives for embeddings, chunking, storage, and semantic retrieval.

ai chunking llm open-source qdrant-vector-database rag

3. charlotte-12s/paper-craft

⭐ 2 stars | JavaScript

PaperCraft — From idea to accepted paper at top CS conferences. 17 rules + 12 skills that guide CS researchers through the full research pipeline.

academic ai-coding claude-code cs-research cursor llm

4. yenlikgaisina/home-heat-risk-ai

⭐ 1 stars | Jupyter Notebook

UK overheating risk dashboard combining climate, housing, EPC, and carbon-intensity data

climate data-science housing machine-learning portfolio python

5. derrickhabibii/kronos-trading

⭐ 1 stars | Python

Automated algorithmic trading system using Kronos (Tsinghua University financial foundation model) — data pipeline, model inference, MT5 execution bridge, and MQL5 Expert Advisor

algorithmic-trading apple-silicon huggingface machine-learning metatrader5 mql5

6. MithraKL/spam-detector

⭐ 1 stars | Jupyter Notebook

SMS Spam Detection System using ML

deep-learning logistic-regression machine-learning naive-bayes nlp scikit-learn

7. sumith25-dev/Rheumatoid-Arthritis-Clinical-Diagnosis-system

⭐ 1 stars | Jupyter Notebook

Rheumatoid Arthritis (RA) Diagnosis Support Using AI & Machine Learning A dual-modal clinical decision support system combining blood test analysis and hand X-ray interpretation to assist in RA diagn

ann artificial-intelligence deep-learning diagnosis doctor healthcare

8. anvshka/cloudportal-cppe

⭐ 1 stars | HTML

Self-service cloud resource portal with Flask, ML demand forecasting, and AI integration

artificial-intelligence cloud flask machine-learning python tailwindcss

9. Mohamad-101/Vehicle_classifier

⭐ 1 stars | Python

A real-time Streamlit web application that uses pre-trained computer vision (MobileNetV2) to detect and classify vehicles in user-uploaded images.

computer-vision image-classification machine-learning python streamlit tensorflow

10. yiheng870106/breast-cancer

⭐ 1 stars | Jupyter Notebook

Breast cancer classification using SVM, Logistic Regression, and Random Forest with scikit-learn.

breast-cancer classification data-science logistic-regression machine-learning python

11. RaghavapriyanSaravanapriyan/Machine-Learning

⭐ 1 stars | Python

my personal ml space

machine-learning

12. ibrahimSalihreisoglu/CO2-Emission-Prediction-Asia-Graduation-Project

⭐ 1 stars | Python

Predicting per capita CO2 emissions across 16 Asian countries using CatBoost, SHAP, and Google Gemini LLM

catboost co2-emissions environmental-sustainability graduation-thesis machine-learning panel-data

13. simodai/ParkingSystem

⭐ 1 stars | Python

基于 PyTorch + EasyOCR + OpenCV 的深度学习车牌自动识别与停车计费管理系统。支持摄像头实时识别、MySQL/JSON 双模存储、Excel 报表导出。

cnn computer-vision database deep-learning easyocr license-plate-recognition

14. SpeedAssassinNet/VisionAI-Enhancer

⭐ 1 stars | Python

Advanced AI-powered video enhancement platform with realtime upscaling, cinematic rendering and neural frame interpolation.

4k-upscaling ai-upscaler ai-video cinematic-video deep-learning frame-interpolation

15. MartinCastroAlvarez/django-admin-mcp-api

⭐ 1 stars | Python

MCP (Model Context Protocol) adapter for the Django admin. v1.0.0 stable — wire-protocol-only layer over django-admin-rest-api, no new logic, no parallel permissions.

agents ai django django-admin json-rpc llm

Generated by Research Pulse on 2026-05-28 06:09