2026-05-18·Daily digest

🔬 Research Pulse

Daily Digest

May 18, 2026

🤖 AI

🧠 LLMs

1. Argus: Evidence Assembly for Scalable Deep Research Agents

Authors: Zhen Zhang, Liangcai Su, Zhuo Chen... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can deep research agents scale inference-time compute effectively when parallel rollouts duplicate evidence rather than complete it, causing diminishing returns and context overflow during aggregation?

Summary: Argus reframes deep research as jigsaw assembly: a Navigator maintains a shared evidence graph and dispatches ReAct Searchers to fill specific gaps, rather than running redundant parallel rollouts. Only the Navigator is RL-trained (to verify, dispatch, and synthesize), while Searchers remain standard agents, enabling flexible scaling from 1 to 64 workers and achieving 86.2 on BrowseComp with sub-21.5K-token reasoning context.

Key Results: Argus, built on a 35B-A3B MoE backbone, gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers averaged across eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every benchmarked proprietary agent, while the Navigator's reasoning context stays under 21.5K tokens.

Key Findings:

Parallel rollouts in prior systems duplicate evidence rather than complement it, causing diminishing returns and aggregation-context blowup
Decoupled training (RL Navigator + frozen ReAct Searcher) enables scaling from 1 to 64 parallel Searchers with no retraining
Evidence-graph-mediated aggregation keeps Navigator context under 21.5K tokens even at 64-way parallelism, while reaching 86.2 on BrowseComp

Technical Novelty: Decoupling deep research into a Navigator that maintains a shared evidence graph (verifying gaps, dispatching workers, synthesizing) from independent ReAct Searchers — and RL-training only the Navigator on verify/dispatch/synthesize while keeping Searchers as standard agents. This allows the same Navigator to scale from 1 to many Searchers without retraining, treating research as jigsaw assembly rather than parallel brute force.

What's New: Unlike parallel-sampling-and-aggregate systems that scale by brute-forcing more trajectories, Argus introduces explicit evidence-gap reasoning via a shared graph and assigns the orchestration role to a separately RL-trained Navigator. This shifts the scaling axis from trajectory count to evidence completeness, and makes the worker count a runtime knob rather than a training-time choice.

Extension Opportunities:

Apply the Searcher/Navigator evidence-graph decomposition to multi-modal research (images, tables, PDFs) where complementarity is even more structured
Replace the shared evidence graph with a typed knowledge graph backed by entity linking to enable cross-query reuse and persistent research memory
Train a learned scheduler that predicts marginal information gain of dispatching another Searcher, optimizing the cost/accuracy frontier dynamically instead of fixed parallel counts

Replicability: Abstract does not mention code or data release. Reproduction would require a 35B-A3B MoE backbone, RL training infrastructure for the Navigator, and substantial inference compute (64-way parallel Searchers for top BrowseComp results) — likely a multi-GPU cluster.

Research Gaps:

No discussion of how the evidence graph handles conflicting or low-confidence evidence pieces during verification
Unclear how the approach generalizes beyond text-based benchmarks (BrowseComp etc.) to long-horizon tasks with stateful tool use or multi-modal sources

2. VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation

Authors: Yiming Zhao, Yu Zeng, Wenxuan Huang... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can LVLMs perform precise instance-level spatiotemporal video understanding when text prompts alone cannot specify fine-grained spatial/temporal references and reasoning is decoupled from visual perception?

Summary: VideoSeeker introduces a visual-prompt-driven agentic paradigm for instance-level video understanding, internalizing tool-calling and proactive perception into an LVLM through a four-stage automated data pipeline plus cold-start SFT and RL. It outperforms GPT-4o and Gemini-2.5-Pro by +13.7% on instance-level benchmarks while transferring to general video tasks.

Key Results: VideoSeeker achieves an average +13.7% improvement over baselines on instance-level video understanding tasks, surpassing closed-source models including GPT-4o and Gemini-2.5-Pro, and demonstrates transferability to general video understanding benchmarks.

Key Findings:

Visual prompts outperform text prompts for specifying fine-grained spatiotemporal instances
Native tool invocation (model proactively retrieves segments) beats language-centric reasoning over static features
Automated four-stage synthesis can produce instance-level video data at scale sufficient for SFT+RL training
Instance-level training transfers positively to general video understanding benchmarks

Technical Novelty: Native agentic tool invocation internalized into the LVLM via cold-start SFT + RL, where the model proactively calls visual tools to retrieve segments rather than reasoning purely in language — paired with visual prompts (not text) as the interaction modality for instance specification.

What's New: Shifts the interaction paradigm from text-only prompts to visual prompts for instance grounding, and shifts reasoning from language-centric chain-of-thought to agentic proactive perception via internalized tool calls — combined in a single end-to-end trained LVLM.

Extension Opportunities:

Extend the visual-prompt paradigm to 3D/egocentric video where instance tracking is harder and text references fail more severely
Apply the four-stage automated data synthesis pipeline to other domains needing instance grounding (medical imaging, surveillance, robotics demonstrations)
Combine VideoSeeker's tool-invocation RL with streaming/online video for real-time agentic perception in interactive applications

Replicability: Authors state datasets and code will be released publicly. Reproduction likely requires multi-GPU RL training on top of an open LVLM backbone (typical scale: 8+ A100/H100s) plus the synthesized instance-level video corpus.

Research Gaps:

No reported latency/cost analysis of repeated tool invocation during inference
Unclear how the approach scales to long-form (hour-plus) videos or multi-instance scenes with dense interactions

👁️ Vision

1. End-to-end plaque counting and virus titration from laboratory plate images with deep learning

Authors: Eugenia Moris, Alicia Costábile, Sebastián Rey... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can virus plaque assay counting—the gold standard for measuring virus infectivity—be automated end-to-end to eliminate labor-intensive manual counting and inter-operator variability across heterogeneous plate formats and imaging conditions?

Summary: The paper presents an end-to-end deep-learning workflow that automatically segments wells and counts viral plaques in laboratory plate images using two SAM-derived models, then converts counts to PFU/mL through a web platform. Validated on three viruses across two plate formats, it achieves Pearson correlations of 0.88–0.92 against manual counts and matches expert annotators.

Key Results: Two-stage SAM-based pipeline (SAM2 for well segmentation + SAM for plaque segmentation) evaluated on held-out plates: 17 MAYV/CVB3 plates achieved Pearson r=0.92 vs manual annotation; 22 VACV plates achieved r=0.88. Generalized across 6-well and 12-well formats and three viruses (Mayaro, Coxsackievirus B3, Vaccinia). Showed high concordance against four independent expert annotators. Automatically computes PFU/mL.

Key Findings:

SAM2-based well segmentation generalizes across heterogeneous imaging conditions and 6-well/12-well formats without per-plate calibration
Pearson correlations of 0.92 (MAYV/CVB3) and 0.88 (VACV) on held-out plates indicate strong agreement with manual counts
Automated counts show concordance comparable to inter-expert variability across four independent annotators, suggesting the tool can replace manual counting in production

Technical Novelty: First end-to-end SAM/SAM2-derived pipeline tailored to plaque assays that couples well-localization and plaque-instance segmentation, rather than treating plaque counting as a generic blob/CNN detection problem. Combines a foundation-model backbone with PFU/mL computation and an audit-ready web interface.

What's New: Repurposes foundation segmentation models (SAM and SAM2) specifically for virology readouts, rather than training a virus-specific CNN from scratch as in most prior plaque-counting work; delivers a deployable, audit-ready web platform—not just a model.

Extension Opportunities:

Extend to additional virus families (influenza, herpesviruses, SARS-CoV-2) and non-standard plate formats (24/96-well) by fine-tuning the SAM plaque head on small labeled sets
Add plaque morphology classification (size, clarity, syncytial vs lytic) to distinguish virus strains or quantify phenotypic variants relevant to antiviral drug screening
Integrate active learning in the web platform so expert corrections automatically refine the segmentation models, closing the loop between annotation and deployment

Replicability: Authors commit to open-sourcing code and the web platform upon acceptance. Uses public VACVPlaque dataset plus private MAYV/CVB3 images. Compute needs are modest—SAM/SAM2 inference runs on a single consumer GPU; fine-tuning SAM heads typically fits in 16–24 GB VRAM.

Research Gaps:

Evaluation is limited to three viruses and two plate formats; performance on small/diffuse plaques or atypical cytopathic effects is unquantified
No reported throughput, latency, or cost figures for lab-scale deployment, and no comparison against existing tools like ViralPlaque or CellProfiler pipelines

🦾 ROBOTICS

1. DexJoCo: A Benchmark and Toolkit for Task-Oriented Dexterous Manipulation on MuJoCo

Authors: Hanwen Wang, Weizhi Zhao, Xiangyu Wang... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: Existing dexterous manipulation benchmarks lack tasks that exploit the unique capabilities of dexterous hands (vs parallel grippers) and lack comprehensive, standardized evaluation pipelines for tool-use, bimanual coordination, and long-horizon reasoning.

Summary: DexJoCo is a MuJoCo benchmark and toolkit of 11 task-oriented dexterous manipulation tasks targeting capabilities unique to dexterous hands — tool-use, bimanual coordination, long-horizon execution, and reasoning. It ships 1.1K teleoperated trajectories, domain randomization, and a unified evaluation across visual/dynamics randomization, multi-task training, and action-head adaptation, exposing systematic weaknesses of current policies.

Key Results: Introduces DexJoCo: 11 functionally grounded tasks on MuJoCo, 1.1K collected trajectories via a low-cost data collection system, with domain randomization support. Benchmarks modern policies under visual/dynamics randomization, multi-task training, and action-head adaptation, surfacing limitations of current dexterous policies.

Key Findings:

Current policies struggle on dexterous-hand-specific tasks (tool-use, bimanual, long-horizon) even when they handle parallel-gripper benchmarks
Visual and dynamics randomization meaningfully degrade learned policies, indicating poor robustness
Action-head adaptation and multi-task training shift performance enough to matter, suggesting architecture and data-mix choices are first-order

Technical Novelty: First MuJoCo benchmark explicitly designed around dexterous-hand-specific affordances (tool-use, bimanual, long-horizon, reasoning) rather than reusing parallel-gripper tasks, paired with a low-cost data pipeline and a unified evaluation protocol covering visual/dynamics randomization, multi-task, and action-head adaptation.

What's New: Unlike prior dexterous benchmarks (e.g., DexArt, Bi-DexHands) that focus on isolated skills, DexJoCo defines tasks specifically chosen because parallel grippers cannot do them, and bundles the data collection rig plus a comparative evaluation across multiple axes (randomization, multi-task, action-head) into one toolkit.

Extension Opportunities:

Add tactile/force-sensor modalities to the benchmark and evaluate whether contact-rich feedback closes the gap on tool-use and bimanual tasks
Scale trajectory collection beyond 1.1K via teleop crowdsourcing or sim-to-real distillation, then test scaling laws for diffusion/VLA policies on dexterous tasks
Build a sim-to-real bridge from DexJoCo's MuJoCo tasks to a physical dexterous hand (e.g., LEAP, Allegro) to validate that benchmark rankings transfer

Replicability: Project page (dexjoco.github.io) suggests code/data release; MuJoCo-based so compute is modest for simulation, but policy training (diffusion/VLA-style models) likely requires multi-GPU setups. 1.1K trajectories is small enough to redistribute easily.

Research Gaps:

Dataset is still small (1.1K trajectories) relative to what modern VLA/diffusion policies likely need
MuJoCo-only — no sim-to-real validation that benchmark performance predicts real-hand performance

2. WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation

Authors: Baining Zhao, Jiacheng Xu, Weicheng Feng... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can aerial vision-language navigation (VLN) agents in 3D environments anticipate latent world evolution and act on predicted consequences in a closed-loop manner, rather than relying on full-sequence video generation or reactive policies?

Summary: WorldVLN reformulates aerial vision-language navigation as a prediction-driven world-action problem, using an autoregressive video backbone to predict short-horizon latent world transitions and decode them directly into waypoint actions in a closed loop. A two-stage training scheme grounds the video prior in navigation dynamics and then applies Action-aware GRPO to optimize waypoints by downstream rollout outcomes, yielding 12%+ success-rate gains over VLA baselines and zero-shot real-drone transfer.

Key Results: WorldVLN outperforms existing Vision-Language-Action baselines on public outdoor and indoor benchmarks with 12%+ success-rate gains, shows larger advantages on challenging cases, and transfers zero-shot to real drone deployment.

Key Findings:

Short-horizon autoregressive world-state prediction outperforms full-sequence video-generation world models for navigation action selection
Action-aware GRPO enables RL fine-tuning of autoregressive WAMs by optimizing through downstream rollout consequences
The model transfers zero-shot from simulation to real drone deployment, indicating practical viability

Technical Novelty: First autoregressive world-action model for aerial VLN that adapts a latent autoregressive video backbone to predict short-horizon world transitions and directly decode them into waypoint actions, plus Action-aware GRPO — the first RL method tailored for autoregressive WAMs that optimizes waypoints via downstream rollout consequences.

What's New: Unlike full-sequence video-generation world models, WorldVLN predicts only short-horizon latent transitions and decodes them into executable waypoints, enabling tight closed-loop control. It also introduces the first RL method (Action-aware GRPO) tailored to autoregressive world-action models.

Extension Opportunities:

Extend Action-aware GRPO to other autoregressive world-action models in ground robotics or manipulation domains
Investigate longer-horizon world-state rollouts and hierarchical planning on top of the short-horizon predictions
Combine WorldVLN with multi-agent aerial coordination or dynamic obstacle environments to test robustness

Replicability: Code and demos available at https://embodiedcity.github.io/WorldVLN/. Compute requirements not explicitly stated, but training a latent autoregressive video backbone plus GRPO-style RL likely requires multi-GPU setups (estimated 8+ A100/H100 GPUs).

Research Gaps:

Lack of efficient closed-loop world models for aerial 3D navigation that avoid the cost of generating full video sequences
Absence of reinforcement learning methods specifically designed for autoregressive world-action architectures

3. Dynamic Plasma Shape Control with Arbitrary Sensor Subsets

Authors: D. Sorokin, M. Stokolesov, A. Granovskiy... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can a tokamak plasma shape controller simultaneously track dynamically changing shape targets and remain robust to arbitrary magnetic diagnostic failures, without relying on the classical pipeline of equilibrium reconstruction followed by a linear controller with a fixed sensor set?

Summary: The authors train a single RL policy in the NSFsim DIII-D simulator that performs dynamic plasma shape tracking while tolerating random failure of 30% of magnetic sensors per episode, eliminating the need for separate equilibrium reconstruction and mode-switching backup controllers. The policy uses an asymmetric actor-critic with privileged equilibrium information and an auxiliary shape-reconstruction head, achieves 2.01 cm mean shape error on a held-out static configuration, and transfers zero-shot to physical DIII-D shots and the independent GSevolve simulator.

Key Results: An RL agent trained in NSFsim (DIII-D configuration) on 120 experimental plasma shapes with shape targets resampled as random step changes every 0.25s achieves a mean shape error of 2.01 cm on a held-out static configuration. It tolerates 30% random magnetic sensor masking per episode with a single policy (no backup controllers/mode-switching), demonstrates qualitative dynamic trajectory following in simulation and on physical DIII-D shots commanding coil actuators on two dynamic maneuvers, and transfers zero-shot to the independent GSevolve simulator.

Key Findings:

A single RL policy can replace the classical equilibrium-reconstruction + linear-controller stack while simultaneously handling dynamic shape targets and sensor dropout.
Random 30% magnetic-sensor masking during training yields a policy robust to arbitrary sensor subsets without explicit fault-detection or backup logic.
Asymmetric actor-critic with privileged equilibrium information for the critic and an auxiliary shape-reconstruction head on the actor improves learning under partial observability and provides interpretability.
The simulation-trained policy transfers zero-shot both to the independent GSevolve simulator and to real DIII-D shots driving coil actuators on dynamic maneuvers.

Technical Novelty: Unifies dynamic shape tracking and arbitrary-sensor-subset robustness in a single RL policy via (1) random 30% magnetic-sensor masking during training, (2) an asymmetric actor-critic where the critic sees privileged equilibrium state to stabilize value learning under partial observability, and (3) an auxiliary shape-reconstruction head on the actor that doubles as an interpretability probe — eliminating the classical equilibrium-reconstruction + linear-controller decomposition and mode-switching backup logic.

What's New: Prior tokamak shape controllers either assume a fixed, fully-operational sensor set or handle faults via separate backup controllers and mode-switching; recent RL controllers (e.g., DeepMind/TCV) focused on shape control but not on arbitrary sensor-subset robustness. This work is the first to demonstrate a single end-to-end policy that jointly tracks dynamic shape targets and tolerates arbitrary magnetic-sensor subsets on a major device (DIII-D), with cross-simulator and sim-to-real transfer.

Extension Opportunities:

Port the asymmetric actor-critic + sensor-dropout training recipe to other tokamaks (ITER, JT-60SA, SPARC) by retraining on their equilibrium datasets and diagnostic layouts.
Extend the auxiliary shape-reconstruction head into a full real-time equilibrium reconstruction service usable by other plant systems, leveraging its interpretability for operator-facing diagnostics.
Increase the dropout rate or simulate correlated sensor failures (e.g., coil-bank outages, calibration drift) to stress-test and harden the policy for routine experimental campaigns.
Combine the controller with disruption-avoidance or MHD-stability objectives in a multi-task RL setup so a single policy handles shape + stability + actuator-limit constraints.

Replicability: The abstract does not mention released code or data. Reproduction would require access to NSFsim (and ideally GSevolve) configured for DIII-D, the 120-shape experimental dataset, and DIII-D experimental time for hardware validation — substantial barriers. Compute is likely modest by RL standards (single-machine GPU training over days), but the simulator and device access are the real gating factors.

Research Gaps:

No quantitative dynamic-tracking error metrics are reported — dynamic trajectory following is only shown qualitatively in simulation and on hardware.
Robustness is characterized only at 30% random magnetic-sensor masking; behavior under higher dropout rates, correlated sensor failures, or non-magnetic diagnostic loss is unaddressed.
The approach is demonstrated only on DIII-D with NSFsim; generalization to other tokamak geometries, divertor configurations, and burning-plasma regimes (e.g., ITER, SPARC) remains open.

💻 COMPUTE

1. Sub-picosecond inter-core skew characterization in multicore fibers via Hong--Ou--Mandel interference

Authors: L. Lira Tacca, L. Marques Fagundes, M. Morales Lillo... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can inter-core skew (ICS) in multicore fibers be measured with sub-picosecond precision across length scales spanning laboratory to field-deployed fibers, where classical interferometric methods fail due to path fluctuations?

Summary: The paper introduces a Hong-Ou-Mandel two-photon interference technique to measure inter-core skew in multicore fibers with ±0.11 ps precision, a ~180× improvement over the standard C-OTDR method. By measuring all 12 core-pair combinations in a commercial 4-core fiber from 7.7 m to 1300 m, it provides the first direct validation of √L random-walk scaling for ICS across laboratory-to-field length scales.

Key Results: Demonstrated ICS measurement in a commercial 4-core fiber via two-photon HOM interference in a fiber-integrated 4×4 multiport beam splitter, achieving ±0.11 ps precision across all 12 core-pair combinations. Validated stochastic random-walk scaling σ_τ(L) = κ√L + c with κ = 48.7 ± 2.5 ps/√km and c = 9.76 ± 1.2 ps over fiber lengths from 7.7 m to 1300 m. This represents a ~180× improvement over C-OTDR, the prior standard for long-fiber ICS characterization.

Key Findings:

HOM interference enables ±0.11 ps ICS precision, ~180× better than C-OTDR, with the limit set by delay-stage positioning rather than the interference method itself
ICS follows stochastic random-walk scaling σ_τ(L) = κ√L + c with κ = 48.7 ± 2.5 ps/√km across three orders of magnitude in length
Fisher information analysis shows a fundamental Cramér-Rao precision floor in the femtosecond range, leaving substantial room for improvement with better delay control

Technical Novelty: First use of HOM two-photon interference in a fiber-integrated 4×4 multiport beam splitter to simultaneously characterize ICS across all 12 core-pair combinations. The key novelty is exploiting HOM's immunity to first-order path-length fluctuations, enabling precision measurements on long installed fibers that are impractical for classical interferometry.

What's New: Prior ICS characterization on long fibers relied on C-OTDR or classical interferometry, the latter being defeated by path-length fluctuations in installed fibers. This work is the first to leverage HOM's intrinsic insensitivity to first-order phase fluctuations for ICS, and the first direct experimental confirmation of √L stochastic scaling across both lab and field length regimes.

Extension Opportunities:

Push toward the femtosecond Cramér-Rao bound by replacing the mechanical delay stage with piezo-driven or electro-optic delay control
Extend the technique to fibers with 7, 19, or higher core counts and characterize correlations between adjacent vs. distant core pairs
Apply HOM-based skew characterization in situ for active synchronization/calibration of MCF-based quantum key distribution or entanglement distribution networks

Replicability: No code or dataset mentioned in the abstract. Reproduction requires significant specialized hardware: a heralded single-photon or entangled photon-pair source, a fiber-integrated 4×4 multiport beam splitter, sub-ps delay stages, single-photon detectors, coincidence electronics, and access to MCF segments of varied lengths (7.7 m to 1300 m).

Research Gaps:

Demonstration limited to a 4-core fiber; scaling behavior and cross-core correlations in higher-count MCFs (7, 12, 19 cores) remain uncharacterized
Precision is currently limited by delay-stage hardware rather than the quantum measurement itself — closing the gap to the Cramér-Rao femtosecond limit is unexplored

2. Demonstration of a Multiplexing Trapped Ion Quantum Processing Unit

Authors: F. Anmasser, M. Abu Zahra, K. Schüppert... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can trapped-ion quantum processors scale to thousands of qubits without requiring thousands of dedicated control lines, which becomes a major wiring bottleneck for fault-tolerant architectures?

Summary: The paper demonstrates a trapped-ion quantum processing unit that uses a time-multiplexed sample-and-hold scheme to drive many trap electrodes from few input control lines, addressing the wiring scalability bottleneck. Experimental characterization shows the approach preserves low motional heating (<1 phonon/s) and induces gate errors below 10^-4 for sampling intervals under 50 ms.

Key Results: Demonstrated a surface ion trap integrated with a time-multiplexed sample-and-hold control system. Measured motional heating rates below 1 phonon/second in both open and closed switch configurations. Showed that sampling intervals below 50 ms keep expected gate errors from decaying charges during hold phase below 10^-4, indicating compatibility with high-fidelity quantum operations.

Key Findings:

Motional heating rates remain below 1 phonon/second whether the multiplexer switch is open or closed, showing the integration does not degrade trap quality
A sample-and-hold interval below 50 ms keeps charge-decay-induced gate errors under 10^-4, meeting fault-tolerance-relevant error budgets
Time-multiplexed control is experimentally compatible with high-fidelity ion-trap quantum operations, validating a path around the I/O scaling wall

Technical Novelty: Combines a surface ion trap with a time-domain multiplexer using sample-and-hold: electrodes are charged to fixed voltages then disconnected during qubit operations, decoupling control wiring count from electrode count — a hardware-level solution distinct from prior software/DAC-multiplexing approaches.

What's New: Prior trapped-ion scaling proposals discussed multiplexing conceptually; this work delivers a physical demonstration of a sample-and-hold multiplexer integrated with a real surface ion trap and quantifies its impact on heating and gate-error budgets.

Extension Opportunities:

Scale the multiplexing ratio (input signals to electrodes) further and characterize fidelity at higher fan-outs to validate the approach at fault-tolerant qubit counts
Integrate the sample-and-hold multiplexer with two-qubit entangling gate benchmarks to directly measure operational fidelity rather than inferred error bounds
Co-design CMOS-integrated multiplexing circuits with cryogenic ion trap chips to enable on-chip control electronics for modular QCCD architectures

Replicability: No mention of open code/data in the abstract. Reproduction requires a surface ion trap setup, vacuum/cryogenic infrastructure, laser systems for ion cooling and readout, and custom sample-and-hold multiplexing electronics — a major experimental quantum hardware lab investment.

Research Gaps:

No direct measurement of one- or two-qubit gate fidelity under multiplexed operation — errors are inferred from charge decay rather than observed in benchmarking experiments
Scalability to the thousands-of-electrodes regime needed for fault tolerance is argued rather than demonstrated; crosstalk and switch density at scale remain open

3. Quantum Feature Amplification Network (QFAN) as An Autoregressive Quantum Generative Model

Authors: Jamal Slim, Saverio Monaco, Florian Rehm... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can quantum generative models for calorimeter shower simulation scale beyond tiny image sizes when the quantum register size is otherwise tied to the full image dimension, creating a hard bottleneck far below detector-scale geometries used in high-energy physics?

Summary: QFAN reframes quantum generative modeling of calorimeter showers as an autoregressive block-by-block process, where the same small parameterized quantum circuit is reused conditioned on a compressed summary of already-generated pixels. This decouples qubit count from image size, fixing it by block size instead, and is demonstrated with just 3 qubits and 12 shared parameters on both simulator and IBM hardware to reproduce intensity, correlation, and energy statistics of showers.

Key Results: A 3-qubit circuit with only 12 shared variational parameters, combined with closed-form ridge decoders and a post-hoc residual sampler, reproduces per-pixel intensity distributions, inter-pixel correlations, and total energy distributions of calorimeter showers on both classical simulator and IBM quantum hardware. The authors derive a conservative worst-case bound on shot-noise propagation through the autoregressive chain and provide an empirical decoder-capacity heuristic for reachable sequential depth. The hardware-simulator gap at this scale is consistent with optimization-budget limits dominating over device noise.

Key Findings:

Block-wise circuit reuse with conditioning removes the register-size bottleneck while keeping per-step quantum cost independent of image size for Pauli observables
A 3-qubit, 12-parameter circuit with ridge decoders and residual sampling reproduces pixel intensities, inter-pixel correlations, and total energy distributions
On IBM hardware, the gap from simulator is consistent with optimization-budget limits rather than device noise, though the paper does not causally separate these

Technical Novelty: Prior quantum generative models for calorimetry tie qubit count to image dimension, or only partially offload to hybrid latent models. QFAN reuses a single small parameterized quantum circuit per block, conditioned on a compressed summary of previously generated pixels, decoupling qubit requirements from image size entirely. The per-step quantum cost is independent of total image size for the Pauli-observable family used.

What's New: First quantum generative architecture for calorimetry that achieves register-size independence from image size via autoregressive block reuse, rather than offloading dimensions into a hybrid classical latent model. Adds a conservative shot-noise propagation bound and a decoder-capacity heuristic specific to this sequential setting.

Extension Opportunities:

Scale the autoregressive block-generation scheme to full detector-scale calorimeter geometries (e.g., ATLAS/CMS HGCAL) and empirically test whether the shot-noise bound and decoder-capacity heuristic predict the breakdown point
Replace closed-form ridge decoders with learned neural decoders (e.g., small transformers) to push the reachable sequential depth and compare against the derived capacity heuristic
Apply the QFAN block-reuse paradigm to other structured generative tasks beyond calorimetry — e.g., medical imaging patches, weather field tiles, or molecular conformer sequences — where register-size bottlenecks similarly limit quantum models

Replicability: Abstract does not mention a public code or data release. Reproduction requires a 3-qubit simulator (trivial on a laptop) and optionally IBM Quantum hardware access for the hardware validation. Calorimeter shower datasets such as CaloChallenge or Geant4-simulated samples would be needed. Compute is minimal due to the tiny 12-parameter circuit and closed-form ridge decoders.

Research Gaps:

Experiments do not causally separate optimization-budget limits from device noise in explaining the hardware-simulator gap
Results are a proof of principle and do not validate that the approach actually scales to detector-scale geometries — extrapolations are motivated but unproven

🏥 HEALTHCARE

1. Look Before You Leap: Autonomous Exploration for LLM Agents

Authors: Ziang Ye, Wentao Shi, Yuxin Liu... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can LLM agents overcome 'premature exploitation' — acting on prior knowledge before gathering environment-specific information — to perform reliably in unfamiliar environments?

Summary: The paper identifies premature exploitation as a core failure mode of LLM agents and proposes Exploration Checkpoint Coverage as a verifiable metric for autonomous exploration. It introduces an interleaved RL training scheme with separate verifiable rewards for exploration and task-execution rollouts, instantiated as an Explore-then-Act paradigm that decouples information gathering from task solving.

Key Results: The paper introduces Exploration Checkpoint Coverage (ECC) as a verifiable metric for measuring discovery of key states, objects, and affordances. Systematic evaluation shows standard task-oriented RL produces narrow, repetitive behaviors that hurt downstream performance. Their interleaved training (task-execution + exploration rollouts with separate verifiable rewards) under the Explore-then-Act paradigm improves both exploration breadth and task success. Specific numeric benchmarks and datasets are not stated in the abstract.

Key Findings:

Standard task-oriented RL induces narrow, repetitive agent behavior that impedes downstream task performance in unfamiliar environments
Exploration Checkpoint Coverage provides a verifiable, task-independent signal that exposes this narrowness and can be optimized directly
Interleaving exploration and task rollouts under the Explore-then-Act paradigm yields more generalizable agents than task-reward-only training

Technical Novelty: Two coupled novelties: (1) ECC, a verifiable, checkpoint-based exploration reward that decouples 'did the agent learn about the environment' from 'did the agent solve the task', and (2) interleaved RL that optimizes task and exploration rollouts with their own verifiable rewards, operationalized via the Explore-then-Act inference paradigm with an explicit interaction budget.

What's New: Prior agent-RL work treats exploration as an implicit byproduct of task reward shaping or entropy bonuses; this paper makes exploration a first-class objective with its own verifiable reward (ECC) and a dedicated inference-time phase, rather than entangling it with task success.

Extension Opportunities:

Apply Explore-then-Act to web/browser agents (e.g., WebArena, Mind2Web) where premature exploitation on cached UI patterns is a known failure mode
Extend ECC into a curriculum signal: weight exploration rollouts toward checkpoints with lowest coverage, yielding adaptive exploration budgets per environment
Combine the exploration phase with a learned world model so gathered knowledge persists across episodes rather than being re-acquired each task

Replicability: Abstract does not mention code, data, or model releases. Reproduction would require an RL training stack for LLM agents (likely PPO/GRPO-style), a checkpoint-annotated interactive environment (ALFWorld/ScienceWorld-class), and multi-GPU compute typical of agent RL fine-tuning runs (order of 8+ GPUs for days).

Research Gaps:

No clear treatment of how the exploration budget should scale with environment complexity or transfer across environments
Unclear how ECC checkpoints are defined at scale without human annotation, which limits applicability to open-ended or real-world environments

2. Fully Open Meditron: An Auditable Pipeline for Clinical LLMs

Authors: Xavier Theimer-Lienhard, Mushtaha El-Amin, Fay Elhassan... Published: 2026-05-15 | Citations: 0 arXiv | PDF

Research Question: How can clinical LLM pipelines be made fully auditable and reproducible end-to-end (data, curation, training, evaluation), rather than merely open-weight, while still matching or exceeding state-of-the-art domain performance?

Summary: Fully Open Meditron is the first end-to-end auditable pipeline for clinical LLMs, releasing not just weights but the full corpus, curation procedures, training framework, and physician-validated evaluation. Applied to five fully-open base models, it sets a new FO SoTA on medical benchmarks and beats MedGemma head-to-head, showing auditability and SoTA performance are not in tension.

Key Results: Built Fully Open Meditron across 5 FO base models. Apertus-70B-MeditronFO improved +6.6 points over base (47.2% → 53.8%) on aggregate medical benchmarks, a new FO SoTA. Gemma-3-27B-MeditronFO beat MedGemma in 58.6% of LLM-as-a-judge comparisons and on HealthBench (58% vs 55.9%). Corpus unifies 8 public medical QA datasets plus synthetic extensions grounded in 46,469 clinical practice guidelines, validated by a 4-physician panel and calibrated against 204 human raters.

Key Findings:

All five MeditronFO variants are preferred over their base models, demonstrating the recipe generalizes across architectures and scales.
Apertus-70B-MeditronFO sets a new Fully Open SoTA with +6.6 points (47.2% → 53.8%) on aggregate medical benchmarks.
Gemma-3-27B-MeditronFO outperforms the open-weight-only MedGemma on HealthBench (58% vs 55.9%) and is preferred 58.6% of the time by LLM-judge, showing full openness need not cost performance.

Technical Novelty: First end-to-end Fully Open medical LLM stack: clinician-audited corpus construction, system-wide decontamination, gold-label resampling of teacher generations, guideline-grounded synthetic QA from 46k+ CPGs, and a human-calibrated LLM-as-a-judge protocol — prior 'open' medical LLMs (MedGemma, Meditron-v1) released weights but not the full data/curation pipeline.

What's New: Unlike prior medical LLMs that are open-weight only (MedGemma, original Meditron), this releases the complete training stack: clinician-vetted corpus with guideline-grounded synthesis, system-wide decontamination, gold-label teacher resampling, and a human-calibrated evaluation protocol — making the entire CDSS pipeline reproducibly auditable.

Extension Opportunities:

Extend the guideline-grounded QA generator to non-English guidelines (e.g., WHO, regional ministries) to build multilingual clinical models with the same auditable provenance chain.
Add a longitudinal patient-record reasoning track (EHR-style multi-turn vignettes) to the corpus, since current synthetic extensions are exam/vignette-shaped and miss real chart-review workflows.
Replace the LLM-as-a-judge with a calibrated ensemble-of-judges plus per-specialty rubrics, and publish the rater disagreement distribution to make eval itself auditable.

Replicability: Paper positions itself as fully open (corpus, construction framework, training recipe, eval protocol released); code/data availability implied but not confirmed from abstract alone. Reproducing the 70B SFT run requires substantial GPU compute (likely 100s of H100-hours); smaller 8B/9B variants are tractable on a modest multi-GPU node.

Research Gaps:

No fully open medical LLM pipeline existed prior — 'open' had effectively meant open-weights only, blocking reproducible clinical validation.
Evaluation of clinical LLMs has lacked calibration against substantial human rater pools tied to use-aligned vignettes.

🔥 GitHub Trending

1. Sakura66/sagesched

⭐ 7 stars | Python

SageSched: Intelligent LLM Request Scheduler with Workload Prediction — QoS-aware dual-queue scheduling for black-box LLM APIs (OpenAI/Azure/Doubao/Gemini)

api-gateway faiss fastapi gittins-index llm llm-inference

2. MikeMengTR/whats-this-function

⭐ 4 stars | TypeScript

WTF? What's This Function? 这是什码？— Double-tap Ctrl → instantly explain any code. Desktop AI code reader powered by DeepSeek/OpenAI — select code anywhere → Ctrl+C → double-tap Ctrl → Chinese explanati

ai-assistant ai-tools chinese code-explainer code-tutor coding-assistant

3. markadelnawar/json-schema-sketch

⭐ 3 stars | TypeScript

Infer compact schema descriptions from any JSON value. For AI/LLM tool-use contexts.

agents context json llm llm-inference token

4. swaylq/master-skill

⭐ 2 stars | Shell

大师.skill — 输入行业，自动调研 6 轨[行业大佬 / 工具地图 / 工作流 / 知识正典 / 信息源 / 术语标准] → 提炼为可运行的行业 Master OS skill；装到任意 Claude Code / OpenClaw / Codex / Hermes agent 即让其进入「这一行的资深人」模式。MIT，Python + Shell。

agent ai-agent chinese-nlp claude-code claude-skill industry-knowledge

5. AEUS-06/Teselcore

⭐ 1 stars | C

TeselCore es un framework experimental con propuestas y curiosidades matematicas, fisicas o cosas raras que planteo

c deep-learning experiment framework machine-learning mathematics

6. aakarsh-hub/learning-analytics-engine

⭐ 1 stars | Python

Student engagement analytics with ML-based dropout risk prediction (GBM), cohort retention heatmaps, and Plotly interactive dashboards

analytics edtech machine-learning pandas plotly python

7. aakarsh-hub/edtech-ai-tutor

⭐ 1 stars | Python

Adaptive AI tutoring engine — personalized lessons, adaptive quizzes, and learning paths using FastAPI + OpenAI GPT-4o

ai edtech fastapi machine-learning openai python

8. Ashik-AI-DataScientist/Customer_churn_prediction_engine

⭐ 1 stars | Jupyter Notebook

End-to-end churn prediction engine | XGBoost · K-Means · Kaplan-Meier · Power BI

churn-prediction-models data-science eda machine-learning power-bi python

9. kishan-arya/rail-track-defect-detection

⭐ 1 stars | Jupyter Notebook

Rail track defect classification using pretrained FCCN (ICCV 2023) — feature extraction + ML classifiers

computer-vision feature-extraction machine-learning python pytorch scikit-learn

10. pablo-reyes8/dual-aging-diffusion

⭐ 1 stars | Python

Identity-preserving facial aging with dual-scale diffusion: global low-frequency residuals for coarse age structure and local crop-level diffusion for high-frequency aging details.

computer-vision deep-learning diffusion-models face-aging facial-analysis generative-ai

11. liquidos-ai/FastLLM

⭐ 1 stars | Rust

Unified LLM Gateway with Local Models in Rust

agent agentic-ai llm llm-gateway

12. Moshigs/Repo-PR-Agent

⭐ 1 stars | Python

多 Agent 流水线：Git 仓库体检（TODO+Ruff）→ Planner 任务拆解 → Coder 补丁草案 → Reviewer 评审｜适配 R&D 申报材料与 CI 演进

code-review devtools llm multi-agent python repository-health

13. MedHbibHlel/bugscalpel

⭐ 1 stars | Unknown

Surgical Python debugging with zero side-effects. Operates like a 40-year senior engineer—one fix.

ai-agent ai-agents ai-skills claude claude-code llm

14. teguh407/warungbrain

⭐ 1 stars | Python

🏪 Multi-agent AI assistant for Indonesian UMKM. Riset kompetitor → strategi harga → konten → CS auto-reply. Powered by Xiaomi MiMo.

ai-agent content-generation ecommerce fastapi indonesia llm

15. Drnaive/CodePRM-DataKit

⭐ 1 stars | Python

Step-level preference data construction toolkit for code-agent process reward models

code-agent llm process-reward-model reward-modeling rlhf swe-bench

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