2026-05-03·Weekly digest

🔬 Research Pulse

Weekly Digest

May 03, 2026

📈 Emerging Trends

🚀 RL Now Optimizes Reasoning, Not Just Actions

Reinforcement learning is migrating from end-task reward optimization to directly shaping intermediate reasoning processes. Both LaST-R1 (latent physical reasoning for robotics) and PRISM (distribution alignment for multimodal models) apply RL to optimize how models think before they act, yielding 4-6 point gains over SFT-only baselines.

Signals:

LaST-R1 uses LAPO to jointly optimize latent CoT and action generation
PRISM inserts an RL-based distribution alignment stage between SFT and RLVR
Both report substantial gains from optimizing reasoning quality rather than just final outputs

🔨 Builder opportunity: A modular post-training toolkit that lets teams insert RL-optimized reasoning stages into existing pipelines — particularly for multimodal and embodied AI models where SFT alone causes distributional drift.

🚀 Agent Reliability Gap Demands Recovery Infrastructure

The best frontier models achieve only 66.7% on realistic workflow automation tasks, revealing that production-grade agents need fault tolerance, not just better prompting. Crab's checkpoint/restore system for agent sandboxes and Claw-Eval-Live's structured failure analysis both point to reliability engineering as the critical missing layer.

Signals:

Claw-Eval-Live: best model scores 66.7% across 105 real workflow tasks
Crab discovers 75%+ of agent turns need no checkpoint, enabling cheap selective recovery
Structured failure patterns cluster by task family, suggesting systematic rather than random breakdowns

🔨 Builder opportunity: An agent observability and recovery platform that profiles which execution steps are state-critical, auto-checkpoints only those, and provides structured failure dashboards by task type — the 'Datadog for agents' layer.

🌱 Interpretability Abandons Linear Directions for Geometry

The assumption that concepts are independent linear directions in representation space is being formally challenged. New theoretical and empirical work shows SAE features fragment concept manifolds into a suboptimal 'dilution' regime, motivating a shift toward geometric objects — subspaces and tilings — as the true atomic units of meaning in neural networks.

Signals:

Theoretical framework proves SAEs can capture manifolds globally or locally but in practice produce diluted representations
The paper motivates treating geometric objects, not isolated directions, as basic interpretability units
This undermines the foundation of current sparse probing and steering vector approaches

🔨 Builder opportunity: Manifold-aware interpretability tools that cluster SAE features into geometric structures and let researchers interact with concept subspaces rather than individual neurons — enabling more faithful model editing and steering.

🌱 Methodology Knowledge Graphs as Research Infrastructure

Intern-Atlas demonstrates that structuring the causal relationships between research methods — not just citations — into queryable graphs enables automated idea generation and evaluation. With 9.4M semantically typed edges grounded in verbatim evidence, this represents a shift from passive literature review to active research acceleration tooling.

Signals:

Intern-Atlas covers 1M+ papers with causally typed method-evolution edges
Self-guided temporal tree search traces method lineages across time
Enables downstream applications: idea evaluation, novelty scoring, automated hypothesis generation

🔨 Builder opportunity: A developer-facing API or IDE plugin that, given a method description, returns its evolutionary lineage, identifies unexplored combination opportunities, and scores novelty against the full graph — turning literature review into a structured query.

📈 Quantum Advantage Remains Gated by Classical Preprocessing

Across quantum computing applications — from CFD simulation to circuit scheduling — classical preprocessing consistently dominates performance while quantum modules provide only conditional marginal gains. The scalability bottleneck for quantum lattice Boltzmann methods is density readout, not quantum gate fidelity, suggesting the real engineering challenge is hybrid orchestration.

Signals:

Disentanglement scheduling: classical preprocessing dominates; quantum circuit is conditionally useful compact representation
QLBM identifies density readout/reloading as key scalability bottleneck, not qubit count
CV quantum network achieves practical rates but requires 1.25B state exchanges for finite-size security

🔨 Builder opportunity: Hybrid quantum-classical orchestration frameworks that intelligently route subproblems to quantum hardware only when the classical preprocessing indicates genuine advantage — avoiding the current pattern of forcing quantum solutions where classical suffices.

🌱 Physics-Aware Latent Reasoning for Embodied AI

Vision-language-action models are gaining an explicit internal physics simulation step: reasoning over physical dynamics in latent space before committing to actions. LaST-R1's adaptive reasoning depth mechanism shows that variable-length 'thinking' about physical consequences, optimized end-to-end with RL, bridges the sim-to-real gap far more effectively than direct imitation.

Signals:

LaST-R1 achieves near-perfect simulation results and substantial real-world gains over IL baselines
Adaptive reasoning depth adjusts computation to task difficulty
Joint RL optimization of latent reasoning and action generation outperforms separate training

🔨 Builder opportunity: A latent world-model reasoning module that can be plugged into existing VLA architectures, pretrained on physics simulations, and fine-tuned with LAPO-style RL for specific robot embodiments — selling 'physical intuition as a service' for robotics teams.

🤖 AI

🧠 LLMs

1. Intern-Atlas: A Methodological Evolution Graph as Research Infrastructure for AI Scientists

Authors: Yujun Wu, Dongxu Zhang, Xinchen Li... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: Existing research infrastructure is document-centric (citation graphs) and lacks explicit representations of methodological evolution — how methods emerge, adapt, and build on each other. This is especially problematic for AI research agents, which cannot reliably reconstruct method lineage from unstructured text. The paper asks: can we build a structured, queryable causal graph of methodological development across AI research?

Summary: Intern-Atlas introduces a methodological evolution graph — a structured, queryable causal network that captures how AI research methods emerge, adapt, and build upon one another across over 1M papers. Unlike citation graphs, each of its 9.4M edges is semantically typed and grounded in verbatim evidence, enabling applications like idea evaluation and automated idea generation. The paper also proposes a self-guided temporal tree search algorithm for tracing method evolution chains over time.

Key Results: Built a methodological evolution graph from 1,030,314 papers (AI conferences, journals, arXiv preprints) containing 9,410,201 semantically typed edges, each grounded in verbatim source evidence. Evaluated the graph against expert-curated ground-truth evolution chains and observed 'strong alignment' (specific metrics not reported in abstract). Demonstrated downstream utility in idea evaluation and automated idea generation tasks.

Key Findings:

A graph of 9,410,201 semantically typed edges can be automatically constructed from 1,030,314 AI papers, capturing method-level lineage relationships with verbatim source grounding
The self-guided temporal tree search algorithm produces evolution chains that align strongly with expert-curated ground-truth method progressions
The methodological evolution graph enables practical downstream tasks: idea evaluation (assessing novelty/positioning of proposed methods) and automated idea generation

Technical Novelty: Two core contributions: (1) an automated pipeline that identifies method-level entities, infers lineage relationships between methods, and captures the bottlenecks driving transitions between successive innovations — going beyond citation links to semantically typed causal edges; (2) a self-guided temporal tree search algorithm for constructing evolution chains that trace method progression over time, rather than relying on flat citation traversal.

What's New: Shifts the unit of scientific knowledge representation from documents/citations to methods/evolution. Prior work treats papers as nodes and citations as edges; Intern-Atlas treats methods as nodes and typed causal relationships (emergence, adaptation, bottleneck-driven transition) as edges. The temporal tree search is a new retrieval algorithm purpose-built for navigating this structure.

Extension Opportunities:

Extend beyond AI/CS to other scientific domains (biology, chemistry, materials science) where methodological lineage is equally critical but even less structured
Build an AI research agent that consumes Intern-Atlas as a live knowledge layer to generate novel method proposals by identifying unexplored combinations at graph frontiers
Integrate the evolution graph with code repositories (e.g., Papers With Code) to link methodological lineage to runnable implementations, enabling automated reproducibility audits

Replicability: Abstract does not mention code or data release. The pipeline ingests 1M+ papers, so reproduction would require substantial compute for NLP extraction and graph construction at scale, plus access to the same corpus of AI conference/journal/arXiv papers. The evaluation against expert-curated ground truth suggests a human annotation effort that would also need replication.

Research Gaps:

Evaluation is described as showing 'strong alignment' with expert ground truth, but the abstract lacks specific quantitative metrics — the rigor of the graph quality assessment is unclear without precision/recall/F1 numbers
Coverage is limited to AI conferences, journals, and arXiv — cross-disciplinary methodological transfer (e.g., methods borrowed from physics or statistics into ML) is likely underrepresented

2. Claw-Eval-Live: A Live Agent Benchmark for Evolving Real-World Workflows

Authors: Chenxin Li, Zhengyang Tang, Huangxin Lin... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: Existing LLM agent benchmarks freeze tasks at release time and primarily grade final responses, failing to evaluate agents against evolving real-world workflow demands or verify whether tasks were actually executed (not just answered). The paper addresses the gap between static, response-focused evaluation and the dynamic, action-verified evaluation needed for workflow automation agents.

Summary: Claw-Eval-Live introduces a live, refreshable benchmark for LLM workflow agents that decouples evolving task demand from reproducible evaluation snapshots, and grades agents on execution traces and artifacts rather than just final responses. Evaluating 13 frontier models on 105 tasks, it demonstrates that reliable workflow automation is far from solved (best model: 66.7%) with structured failure patterns across task families.

Key Results: Evaluated 13 frontier models on 105 tasks spanning business services and local workspace repair. The leading model achieved only 66.7% pass rate; no model reached 70%. Failures are structured: HR, management, and multi-system business workflows are persistent bottlenecks, while local workspace repair is comparatively easier but still unsaturated. Models with similar pass rates diverge in overall completion behavior, and task-level discrimination concentrates in a middle band of difficulty.

Key Findings:

No frontier model exceeds 70% pass rate on real-world workflow tasks, with the best achieving 66.7% — indicating workflow automation remains substantially unsolved
Failure patterns are structured by task family and execution surface: HR, management, and multi-system business workflows are persistent bottlenecks across all models, while local workspace repair is easier but still not saturated
Leaderboard rank is insufficient for comparing agents — models with similar aggregate pass rates diverge significantly in completion profiles, and discriminative power concentrates in a middle band of task difficulty

Technical Novelty: Two-layer benchmark architecture separating a refreshable signal layer (updated from public workflow-demand signals like ClawHub Top-500 skills) from reproducible time-stamped release snapshots. Grading combines deterministic checks on execution traces, audit logs, service state, and workspace artifacts with structured LLM judging only for semantic dimensions — moving beyond final-response-only evaluation to verifiable action evaluation.

What's New: Unlike static benchmarks (SWE-bench, WebArena, etc.) that freeze tasks at release and grade final outputs, Claw-Eval-Live is the first to combine live demand-driven task refresh with execution-trace-based grading. The two-layer architecture (refreshable signal layer + frozen release snapshot) is a novel design pattern for benchmarks that balances ecological validity with reproducibility.

Extension Opportunities:

Build domain-specific workflow agents targeting the identified bottleneck categories (HR, management, multi-system business workflows) where all 13 frontier models consistently fail
Develop a continuous integration pipeline that auto-refreshes benchmark tasks from live workflow-demand signals (e.g., job postings, support tickets, tool changelogs) to stress-test agents on emerging tool APIs and process changes
Create execution-trace-aware training curricula that use the audit logs and trace data from Claw-Eval-Live to fine-tune agents on verifiable multi-step action sequences rather than just final outputs

Replicability: The benchmark is described as public with a shared public pass rule and time-stamped release snapshots with fixed fixtures, services, workspaces, and graders. The current release contains 105 tasks. Compute requirements include running 13 frontier models end-to-end against controlled services and workspaces, which likely demands significant API costs but the deterministic infrastructure should be reproducible.

Research Gaps:

The current release covers 105 tasks — expanding to broader workflow categories (e.g., finance, legal, DevOps pipelines) would test generalization further and reveal whether identified bottleneck patterns are universal
The paper evaluates existing frontier models but does not explore whether execution-trace feedback or the benchmark's structured failure analysis can be used to improve agent training — closing this loop between evaluation and learning is an open direction

3. PRISM: Pre-alignment via Black-box On-policy Distillation for Multimodal Reinforcement Learning

Authors: Sudong Wang, Weiquan Huang, Xiaomin Yu... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: SFT introduces distributional drift that degrades both the model's original capabilities and its alignment with the supervision distribution. This drift is especially harmful in multimodal reasoning, where perception errors and reasoning failures follow distinct patterns that compound during subsequent reinforcement learning. The paper asks: can an explicit distribution-alignment stage between SFT and RLVR close this gap?

Summary: PRISM introduces a three-stage post-training pipeline (SFT → on-policy distribution alignment → RLVR) for large multimodal models that addresses the distributional drift problem caused by SFT. It uses a black-box adversarial game with a MoE discriminator to provide disentangled corrective signals for perception and reasoning errors, improving downstream RL performance by 4-6 points across multiple algorithms and benchmarks on Qwen3-VL.

Key Results: PRISM improves average accuracy over the SFT-to-RLVR baseline by +4.4 points on Qwen3-VL 4B and +6.0 points on Qwen3-VL 8B across diverse multimodal benchmarks. The gains hold across three RL algorithms (GRPO, DAPO, GSPO), demonstrating the stage is algorithm-agnostic rather than tuned to one optimizer.

Key Findings:

Inserting an on-policy distillation (OPD) alignment stage between SFT and RLVR consistently improves downstream RL performance, regardless of which RL algorithm is used (GRPO, DAPO, GSPO)
A Mixture-of-Experts discriminator with dedicated perception and reasoning experts provides disentangled corrective signals, addressing the distinct drift patterns of visual perception errors vs. reasoning failures
High-fidelity curated data matters: 1.26M public demos suffice for broad SFT, but 113K additional dense-grounding, step-by-step reasoning demonstrations (from Gemini 3 Flash) are needed for effective distribution alignment on the hardest unsolved problems

Technical Novelty: The core novelty is casting distribution alignment as a black-box, response-level adversarial game using a MoE discriminator with separate perception and reasoning expert heads. Unlike standard distillation, this requires no access to teacher logits (black-box), operates on-policy (generated from the current model), and provides disentangled corrective gradients for two fundamentally different failure modes in multimodal reasoning.

What's New: Prior work either skips alignment entirely (SFT → RLVR) or uses standard KL-based distillation requiring white-box teacher access. PRISM is novel in three ways: (1) it introduces an explicit third stage in the post-training pipeline, (2) it uses adversarial discrimination rather than likelihood matching for alignment, and (3) it disentangles perception vs. reasoning drift via MoE expert specialization — all without needing teacher logits.

Extension Opportunities:

Apply PRISM's alignment stage to other modalities (video, audio, embodied agents) where perception-reasoning drift decomposition would similarly apply
Replace the binary perception/reasoning MoE split with a finer-grained expert taxonomy (e.g., spatial reasoning, OCR, mathematical reasoning) for domain-specific multimodal tasks
Investigate whether the OPD alignment stage can be applied iteratively during RL training (online re-alignment) rather than as a one-shot pre-alignment step

Replicability: Code, data, and model checkpoints are publicly available at https://github.com/XIAO4579/PRISM. Reproduction would require compute for training Qwen3-VL 4B and 8B models through three stages (SFT, OPD alignment, RLVR) — likely multiple nodes of A100/H100 GPUs. The 113K curated dataset from Gemini 3 Flash would incur API costs to regenerate but is released directly.

Research Gaps:

The 113K high-fidelity alignment dataset is curated from Gemini 3 Flash — the method's sensitivity to the quality and domain coverage of this curated dataset is not fully characterized; it's unclear how performance degrades with weaker or smaller curation sources
The adversarial MoE discriminator architecture choices (number of experts, expert specialization strategy) appear manually designed; no ablation or principled method for determining the optimal expert decomposition is described in the abstract

4. LaST-R1: Reinforcing Action via Adaptive Physical Latent Reasoning for VLA Models

Authors: Hao Chen, Jiaming Liu, Zhonghao Yan... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: Existing VLA models either use slow/discrete linguistic reasoning or continuous latent reasoning but are limited to static imitation learning. Online RL methods for VLAs only optimize the action space without improving the underlying physical reasoning process. How can we jointly optimize latent reasoning and action generation in VLA models through reinforcement learning?

Summary: LaST-R1 introduces a VLA framework that performs latent Chain-of-Thought reasoning over physical dynamics before generating actions, then uses a novel RL algorithm (LAPO) to jointly optimize both the reasoning and action generation. The adaptive reasoning depth mechanism and RL-based optimization of the latent reasoning space yield near-perfect simulation results and substantial real-world improvements over imitation-learning-only baselines.

Key Results: 99.8% average success rate on the LIBERO benchmark with only one-shot supervised warm-up, significantly improving convergence speed over prior SOTA. In real-world deployments, LAPO post-training yields up to 44% improvement over the initial warm-up policy across four complex tasks (single-arm and dual-arm settings).

Key Findings:

Joint RL optimization of latent reasoning and action generation (LAPO) substantially outperforms methods that only optimize the action space, achieving 99.8% on LIBERO with one-shot supervised warm-up
Adaptive latent CoT that scales reasoning depth to task complexity is more effective than fixed-length reasoning, suggesting that not all manipulation steps require equal deliberation
The approach transfers from simulation to real-world with up to 44% improvement over warm-up policy across both single-arm and dual-arm tasks, demonstrating practical generalization

Technical Novelty: Two core contributions: (1) Latent-to-Action Policy Optimization (LAPO) — an RL algorithm that jointly optimizes the latent CoT reasoning process and action generation, rather than treating reasoning as frozen and only optimizing actions. (2) An adaptive latent CoT mechanism that dynamically adjusts reasoning depth based on environment complexity, avoiding fixed-length reasoning overhead.

What's New: Prior VLA+RL work treats reasoning as fixed (optimizing only actions); prior latent reasoning work uses only imitation learning. LaST-R1 is the first to close this loop — making the latent physical reasoning process itself improvable through online RL, and doing so with variable-length reasoning chains rather than fixed computation budgets.

Extension Opportunities:

Apply LAPO to mobile manipulation or navigation tasks where physical dynamics reasoning over longer horizons is critical (e.g., BEHAVIOR-1K or Habitat benchmarks)
Combine the adaptive latent CoT mechanism with hierarchical RL to enable multi-level reasoning — task-level planning in language space and low-level control in latent space
Extend the framework to multi-agent robotic coordination where each agent maintains its own latent reasoning chain but must align on shared physical dynamics priors

Replicability: No code/data availability mentioned in the abstract. Reproducing would likely require significant compute: VLA model training with RL post-training on LIBERO simulation suite, plus real-robot hardware for the physical experiments (single-arm and dual-arm setups). The one-shot warm-up claim suggests relatively low data requirements for the SFT phase, but the RL phase likely needs substantial simulation interaction.

Research Gaps:

The evaluation is on LIBERO (tabletop manipulation) — generalization to more diverse embodiments, longer-horizon tasks, or unstructured environments remains unvalidated
The adaptive CoT mechanism's behavior is not characterized in the abstract — it's unclear what triggers deeper reasoning or whether the policy can learn when reasoning is unnecessary (e.g., simple grasps), raising questions about interpretability of the latent reasoning

📊 Other AI

1. Do Sparse Autoencoders Capture Concept Manifolds?

Authors: Usha Bhalla, Thomas Fel, Can Rager... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: Do sparse autoencoders (SAEs) faithfully capture concept manifolds in neural network representations, given that many concepts are organized along low-dimensional manifolds rather than independent linear directions? Specifically: what does it mean for an SAE to capture a manifold, when do existing architectures succeed, and how?

Summary: This paper challenges the dominant assumption in mechanistic interpretability that concepts correspond to independent linear directions in SAE feature spaces. It provides a theoretical framework showing SAEs can capture concept manifolds either globally (via subspace spanning) or locally (via geometric tiling), but in practice produce a suboptimal 'dilution' regime that fragments manifold structure across features — motivating a shift toward treating geometric objects, not isolated directions, as the basic units of interpretability.

Key Results: The paper develops a theoretical framework showing SAEs can capture manifolds in two distinct modes — global (compact atom group whose linear span contains the full manifold) and local (tiling features that each cover a restricted geometric region). Empirically, they demonstrate that current SAEs suboptimally recover continuous structures, instead producing a fragmented 'dilution' regime that mixes global and local solutions. No specific benchmark numbers are cited in the abstract, but the theoretical characterization of the three regimes (global, local tiling, dilution) and their empirical verification on trained SAEs constitutes the core proof.

Key Findings:

SAEs can capture manifolds in two fundamentally different ways: globally (allocating a compact group of atoms whose span contains the manifold) or locally (distributing across features that tile restricted regions of the geometry)
In practice, SAEs suboptimally mix both strategies in a 'dilution' regime that fragments manifold structure, explaining why continuous geometric relationships are rarely visible at the level of individual SAE features
Post-hoc unsupervised methods that search for coherent groups of atoms (rather than interpreting isolated directions) can recover manifold structure that individual feature analysis misses

Technical Novelty: A formal theoretical framework distinguishing two fundamentally different ways SAEs can represent manifolds (global subspace capture vs. local geometric tiling), plus the identification and characterization of the 'dilution' failure mode where SAEs mix both strategies suboptimally. Prior SAE interpretability work implicitly assumed concepts = independent linear directions; this paper is the first to rigorously formalize manifold capture.

What's New: First paper to formally define what it means for an SAE to 'capture' a manifold, taxonomize the mechanisms (global vs. local), and identify the dilution failure mode. Shifts the interpretability paradigm from 'one feature = one concept' toward 'coherent atom groups = one geometric concept.'

Extension Opportunities:

Build SAE training objectives that explicitly penalize dilution — e.g., add a manifold-coherence regularizer that encourages atom groups to span contiguous geometric regions rather than fragmenting across features
Develop automated post-hoc discovery pipelines that cluster SAE atoms into coherent manifold-aligned groups, turning the paper's theoretical insight into a practical interpretability tool
Extend the framework to non-linear manifolds and hierarchical/compositional concept structures (e.g., using topological data analysis to detect when SAE features tile a manifold vs. span a subspace)

Replicability: No code/data availability is mentioned in the abstract. Reproducing theoretical results requires expertise in representation geometry and SAE internals. Empirical validation likely requires training SAEs on standard language model internals (e.g., GPT-2/Pythia-scale), which is moderate compute (single GPU). The post-hoc discovery methods would need implementation from the paper's descriptions.

Research Gaps:

No prescriptive SAE architecture or training objective is proposed to fix dilution — the paper diagnoses the problem and motivates post-hoc fixes, but doesn't offer a new SAE variant that natively respects manifold structure
The framework appears focused on linear/low-dimensional manifolds; extension to higher-dimensional, non-linear, or compositionally structured concept manifolds remains open

💻 COMPUTE

1. Observation of Vinen turbulence during far-from-equilibrium Bose-Einstein condensation

Authors: Sebastian J. Morris, Martin Gazo, Simon M. Fischer... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: Can decaying quantum (Vinen) turbulence—theoretically predicted to mediate far-from-equilibrium Bose-Einstein condensation and the emergence of long-range order—be directly observed and characterized in a homogeneous 3D atomic Bose gas?

Summary: This paper provides the first direct observation of Vinen (ultraquantum) turbulence during far-from-equilibrium Bose-Einstein condensation in a homogeneous 3D atomic gas. Using matter-wave magnification and thin-slice imaging, they measure vortex line-length density decay consistent with theoretical predictions, revealing that even weakly interacting compressible quantum gases exhibit universal incompressible-fluid turbulence dynamics identical to superfluid helium.

Key Results: Measured the vortex line-length density L and demonstrated its decay follows the predicted scaling for Vinen 'ultraquantum' turbulence. Critically showed that L decay is independent of interatomic interaction strength, and that weakly interacting, highly compressible atomic gases exhibit large-scale dynamics consistent with incompressible hydrodynamic behavior—matching strongly interacting superfluid helium.

Key Findings:

Vortex line-length density L decays following the scaling law predicted for Vinen ultraquantum turbulence (L ~ 1/t) rather than Kolmogorov-type quasi-classical turbulence
The decay dynamics are independent of interatomic interaction strength, demonstrating universality across weakly and strongly interacting superfluids
Highly compressible dilute Bose gases exhibit large-scale hydrodynamic behavior indistinguishable from incompressible superfluids like helium-4

Technical Novelty: Matter-wave magnification technique combined with thin-slice imaging of the expanded cloud, enabling direct visualization of randomly oriented vortex line imprints in a 3D homogeneous gas—previously inaccessible due to density averaging along the line of sight in standard absorption imaging.

What's New: First experimental observation of Vinen turbulence in an atomic BEC context, bridging the theoretical prediction that turbulent vortex tangles mediate condensation with direct measurement. The universality finding—that interaction strength is irrelevant to the decay—was theoretically expected but never experimentally confirmed across such different superfluid systems.

Extension Opportunities:

Develop real-time feedback control of turbulent tangle density to study critical dynamics at the BEC phase transition boundary
Extend matter-wave magnification imaging to 2D Bose gases to probe the BKT transition and compare vortex-antivortex pair annihilation rates with Vinen decay
Use tunable interactions (Feshbach resonances) to map the crossover between compressible and incompressible turbulence regimes quantitatively

Replicability: Requires ultracold atom apparatus capable of producing homogeneous (box-trapped) 3D Bose gases with tunable interactions and time-of-flight matter-wave magnification optics. No code/data availability mentioned. Reproduction demands a specialized AMO physics lab with optical box traps and high-resolution imaging—accessible to ~20-30 groups worldwide.

Research Gaps:

The microscopic mechanism by which vortex reconnections drive energy dissipation at scales below the intervortex spacing in compressible gases remains uncharacterized experimentally
The early-time buildup and saturation of the vortex tangle (before decay onset) is not addressed—understanding the non-equilibrium steady state preceding decay would complete the condensation picture

2. Quantum Lattice Boltzmann Solutions for Transport under 3D Spatially Varying Advection on Trapped Ion Hardware

Authors: Sayonee Ray, Jezer Jojo, Jason Iaconis... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: How can the Quantum Lattice Boltzmann Method be extended to handle transport under spatially non-uniform (3D varying) advection velocity fields on real quantum hardware, and what are the bottlenecks preventing scalability?

Summary: This paper advances Quantum Lattice Boltzmann Methods toward realistic CFD by demonstrating, for the first time, transport under spatially non-uniform velocity fields on IonQ trapped-ion quantum hardware (up to 64 qubits). The authors identify density readout/reloading as the key scalability bottleneck, propose MPS shadow tomography as a mitigation path, and introduce a novel wall boundary implementation for advection-diffusion problems.

Key Results: Demonstrated QLBM solving advection-diffusion with non-uniform velocity fields on IonQ trapped-ion hardware for the first time. Implemented on Forte generation systems and a 64-qubit Barium development system (precursor to IonQ Tempo line). Identified density readout and reloading as the primary algorithmic bottleneck and validated MPS shadow tomography as a mitigation strategy for scaling to large systems with complex density distributions.

Key Findings:

First demonstration of QLBM with non-uniform (spatially varying) advection fields on quantum hardware, bridging the gap between toy problems and realistic CFD requirements
Density readout and subsequent reloading of fluid density is identified as the primary algorithmic bottleneck; MPS shadow tomography is proposed as a promising scalable solution
A novel wall boundary condition method for advection-diffusion in QLBM is introduced and validated through simulation, enabling more physically realistic problem setups

Technical Novelty: First quantum hardware demonstration of QLBM with spatially varying (non-uniform) velocity fields in 3D, advancing beyond prior work limited to uniform advection. Additionally introduces a novel wall boundary condition method for advection-diffusion in QLBM, and identifies MPS shadow tomography as a scalable readout strategy.

What's New: Prior QLBM work on quantum hardware was limited to uniform velocity fields and simple geometries. This paper is the first to tackle spatially varying 3D advection on real hardware, uses cutting-edge 64-qubit trapped-ion systems, and addresses the practical scalability question head-on by analyzing readout bottlenecks and proposing concrete mitigations.

Extension Opportunities:

Implement full Navier-Stokes via QLBM by coupling momentum and energy transport on trapped-ion hardware, moving beyond the scalar advection-diffusion model problem
Develop hybrid classical-quantum pipelines where MPS shadow tomography handles mid-circuit readout while classical solvers manage boundary conditions at scale
Benchmark QLBM with non-uniform advection against superconducting qubit platforms (IBM/Google) to compare error profiles and identify hardware-optimal regimes for CFD

Replicability: No code/data availability mentioned in abstract. Reproduction requires access to IonQ trapped-ion systems (Forte or 64-qubit Barium-class hardware). The algorithmic framework (QLBM with streaming/collision operators) is well-documented in literature, but hardware access is the primary barrier. Classical simulation of the quantum circuits would be feasible for small qubit counts.

Research Gaps:

The readout bottleneck (density extraction and reloading) remains unsolved at scale — MPS shadow tomography is proposed but not fully validated for large, complex distributions on hardware
Extension to turbulent flows, multi-phase transport, or full Navier-Stokes coupling is not addressed, leaving the gap between advection-diffusion and production CFD still open

3. Deep Strong light-matter Coupling in 3D Kane Fermions

Authors: Dmitriy Yavorskiy, David Hagenmuller, Noureddine Charrouj... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: Whether a superradiant quantum phase transition (spontaneous dipole ordering and photon condensation) is possible in relativistic-like electronic systems hosting Dirac and Kane fermions, where the diamagnetic A² term was believed to be formally absent from the low-energy Hamiltonian.

Summary: This paper achieves record deep-strong light-matter coupling (g/ω > 1.6) in Kane fermions hosted in bulk HgCdTe coupled to a cavity, continuously tunable from weak to deep-strong regimes via thermal carrier density control. Crucially, it resolves a long-standing theoretical controversy by proving that the diamagnetic A² term naturally emerges in these relativistic-like systems, ruling out the superradiant quantum phase transition.

Key Results: Achieved a record normalised light-matter coupling ratio exceeding 1.6 (g/ω > 1.6) above room temperature using Landau polaritons in bulk HgCdTe coupled to a Fabry-Perot resonator. Continuously tuned coupling from weak to deep-strong regime via thermally tunable carrier density. Demonstrated that a diamagnetic A² term naturally emerges despite the nonlinear Landau level structure of relativistic Kane fermions, precluding a superradiant phase transition. Experimental polariton spectra matched gauge-invariant microscopic theory.

Key Findings:

Record normalised coupling ratio g/ω > 1.6 achieved above room temperature in HgCdTe Landau polaritons
A diamagnetic A² term naturally emerges in Kane fermion systems despite its formal absence in the low-energy Hamiltonian, precluding superradiant phase transitions
Continuous tuning from weak to deep-strong coupling regime demonstrated using thermally tunable carrier density in a single device

Technical Novelty: First realisation of deep-strong light-matter coupling with Kane fermions specifically, using HgCdTe's ultralow effective mass to reach g/ω > 1.6. Developed a rigorous gauge-invariant microscopic theory showing the A² term emerges naturally in Kane fermion systems despite its apparent absence in the bare low-energy Hamiltonian. Thermal tuning of carrier density as a knob to continuously sweep coupling regimes.

What's New: Extends deep-strong-coupling cavity QED to a new material platform (Kane fermions in HgCdTe), achieves the highest reported normalised coupling ratio, and definitively resolves the decades-long theoretical debate about superradiant phase transitions in Dirac/Kane systems by showing the A² term is not truly absent but emergent.

Extension Opportunities:

Engineer polaritonic semiconductor devices exploiting deep-strong coupling in HgCdTe for THz photonics or quantum sensing applications
Explore whether other narrow-gap semiconductors or topological semimetals with ultralow effective mass can achieve even higher coupling ratios or novel ground-state properties
Investigate cavity-modified transport or non-equilibrium phenomena in Kane fermion systems at deep-strong coupling for cavity-controlled electronics

Replicability: No code/data availability mentioned. Requires bulk HgCdTe epitaxial layers, Fabry-Perot microresonator fabrication, cryogenic-to-room-temperature magneto-optical spectroscopy with Landau level resolution, and variable-temperature sample environment. Moderate experimental infrastructure (MBE growth, cleanroom, FTIR + magnet system).

Research Gaps:

Ground-state quantum correlations predicted in the deep-strong regime (virtual photon populations, entanglement) have not been directly measured or exploited in this system
Whether cavity-driven phase transitions of a different nature (not superradiant but topology- or correlation-driven) could still occur in these extreme coupling conditions remains unexplored

4. Learning quantum disentanglement scheduling from reduced states via modular hybrid policies

Authors: Y. -X. Xiao, J. -Z. Han, Z. Zheng... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: How to design effective hybrid quantum-classical control policies for multiqubit disentanglement scheduling when only partial state information (two-qubit reduced density matrices) is available, as is typical on near-term quantum devices without full wavefunction access.

Summary: This paper introduces a modular hybrid quantum-classical policy framework for multiqubit disentanglement scheduling under partial state observations (two-qubit reduced density matrices only). Through systematic benchmarking on 4-6 qubit tasks, it reveals that classical preprocessing dominates performance while the quantum circuit module serves as a conditionally useful compact representation, and that circuit width matters more than depth.

Key Results: Benchmarked on 4-, 5-, and 6-qubit disentanglement tasks. Demonstrated that classical preprocessing is the dominant performance factor under reduced-state observations. Showed that increasing quantum circuit width is generally more beneficial than increasing depth. Identified a concrete performance-efficiency trade-off across different policy families. The quantum module provides conditional compact representation whose utility depends on input features and model budget.

Key Findings:

Classical preprocessing is the dominant factor governing policy performance under reduced-state observations, not the quantum circuit component
The parameterized quantum circuit provides a conditional compact nonlinear representation whose utility depends on input feature structure and model budget constraints
Increasing quantum circuit width is generally more beneficial than increasing circuit depth for these hybrid control policies

Technical Novelty: A modular hybrid quantum-classical policy architecture with three distinct stages — classical preprocessing, a parameterized quantum circuit as a compact nonlinear latent block, and classical postprocessing — specifically designed for control under reduced-state (partial) observations. The decomposition enables systematic ablation of which component drives performance, revealing preprocessing as dominant rather than the quantum module.

What's New: Shifts focus from full-state quantum control to the practically relevant partial-observation setting, and provides the first systematic decomposition of hybrid policy architectures into modular components to isolate where quantum advantage (or lack thereof) actually resides in the control pipeline.

Extension Opportunities:

Scale the modular hybrid policy framework to larger qubit counts (10+) and characterize how the preprocessing dominance finding holds or breaks down at scale
Apply the reduced-state observation framework to other quantum control tasks beyond disentanglement, such as state preparation or error correction scheduling
Integrate noise-aware training by incorporating realistic device noise models into the partial-observation pipeline to bridge the sim-to-real gap on actual quantum hardware

Replicability: No code or data availability mentioned in the abstract. Reproducing would require a quantum circuit simulator capable of handling parameterized circuits on 4-6 qubits, a reinforcement learning or policy optimization framework, and infrastructure to sweep over circuit width/depth configurations. Moderate compute — the qubit counts are small enough for classical simulation.

Research Gaps:

Results limited to small qubit counts (4-6); unclear whether the preprocessing dominance and width-over-depth findings generalize to larger, more practically relevant system sizes
No demonstration on actual quantum hardware — the partial-observation setting is motivated by near-term devices but validation appears simulation-only

5. Adaptable Continuous Variable Quantum Network with Finite Size Security

Authors: Runjia Zhang, Akash nag Oruganti, Huy Q Nguyen... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: How can continuous-variable quantum key distribution (CV-QKD) be practically scaled from point-to-point links to multi-user quantum access networks while maintaining finite-size security guarantees and allowing adaptable operation per user?

Summary: This paper experimentally demonstrates a 1:4 active continuous-variable quantum network operating under finite-size security, achieving 0.19 bits/channel use over 11 km links with 1.25 billion exchanged coherent states. It introduces adaptable protocols allowing individual users to trade off between security levels and key generation rates, establishing practical compatibility with existing telecom infrastructure.

Key Results: Experimental demonstration of an active 1:4 multi-user CV quantum network over 11 km quantum channels, exchanging 1.25×10⁹ coherent states per channel, achieving a highest secret key generation rate of 1.9×10⁻¹ bits/channel use in the finite-size security regime.

Key Findings:

An active 1:4 CV quantum network achieves secret key rates of 1.9×10⁻¹ bits/channel use over 11 km channels under finite-size security with 1.25×10⁹ coherent states
Adaptable CV-QN protocols enable per-user configuration of security and key rate requirements within the same network infrastructure
CV-QKD with reverse reconciliation naturally supports scaling from point-to-point to quantum access networks via passive quantum broadcasting channels

Technical Novelty: The combination of an active (switched) 1:4 network topology with finite-size security analysis and adaptable protocols that allow per-user tuning of security level vs. key rate trade-offs. Prior CV-QKD network work largely addressed passive broadcasting or assumed asymptotic security; this work addresses finite-size effects in a multi-user active network with user-configurable operation modes.

What's New: First experimental demonstration combining active multi-user CV-QKD networking with rigorous finite-size security analysis and user-adaptable protocol flexibility — bridging the gap between theoretical CV-QKD network proposals and practical deployment constraints.

Extension Opportunities:

Scale the network beyond 4 users (e.g., 1:16 or 1:32) to test performance degradation and identify practical scaling limits for metropolitan deployments
Integrate the adaptable CV-QN protocol with wavelength-division multiplexing (WDM) on deployed telecom fiber to demonstrate coexistence with classical traffic
Develop software-defined networking (SDN) control planes that dynamically allocate security levels and key rates across users based on real-time demand

Replicability: No code or data availability is mentioned in the abstract. Reproduction would require a CV-QKD optical testbed with coherent state preparation, homodyne/heterodyne detection, an active 1:4 optical switch, and ~11 km of fiber per channel — a specialized photonics lab setup. The finite-size security analysis could be reproduced computationally given the protocol parameters.

Research Gaps:

Performance under asymmetric channel conditions (varying distances/losses per user) and larger user counts remains unexplored
Long-term stability, key management, and integration with higher-layer network protocols in real deployed telecom environments are not addressed

🏥 HEALTHCARE

1. Crab: A Semantics-Aware Checkpoint/Restore Runtime for Agent Sandboxes

Authors: Tianyuan Wu, Chaokun Chang, Lunxi Cao... Published: 2026-04-30 | Citations: 0 arXiv | PDF

Research Question: How to efficiently checkpoint and restore autonomous agent sandboxes (containers/microVMs) without the extremes of either losing OS-side state (chat-only recovery) or paying the full cost of per-turn checkpointing, by bridging the semantic gap between agent frameworks (which see tool calls but not OS effects) and the OS (which sees state changes but lacks turn-level context).

Summary: Crab is a transparent host-side runtime that bridges the semantic gap between agent frameworks and OS-level checkpoint/restore by using eBPF to classify which agent turns actually produce recovery-relevant state changes. By exploiting the sparsity that 75%+ of turns need no checkpoint, it achieves full recovery correctness at a fraction of the I/O cost, with negligible overhead on fault-free execution.

Key Results: On shell-intensive and code-repair workloads: recovery correctness raised from 8% (chat-only baseline) to 100% (matching full C/R); checkpoint traffic reduced by up to 87% via sparsity exploitation (75%+ of agent turns produce no recovery-relevant state); overhead stays within 1.9% of fault-free execution time.

Key Findings:

Over 75% of agent turns produce no recovery-relevant OS state changes, meaning most per-turn checkpoints in naive full-C/R approaches are entirely wasted work
Chat-history-only recovery achieves just 8% correctness on shell-intensive workloads because it misses filesystem, process, and runtime artifact state — quantifying the severity of the agent-OS semantic gap
Overlapping checkpoint I/O with LLM wait time (the period the sandbox is idle waiting for model inference) effectively hides C/R latency, keeping total overhead to 1.9% of fault-free execution time

Technical Novelty: Three-layer architecture that is novel in combination: (1) an eBPF-based inspector that classifies each agent turn's OS-visible side effects to determine whether a checkpoint is actually needed (exploiting the 75%+ sparsity), (2) a coordinator that aligns checkpoint boundaries with agent turn boundaries and overlaps C/R I/O with LLM inference latency to hide costs, and (3) a host-scoped scheduling engine that manages checkpoint traffic across co-located sandboxes. All of this runs transparently on the host without modifying the agent framework or the C/R backend.

What's New: First system to formally identify and bridge the agent-OS semantic gap for checkpoint/restore. Prior work either operated purely at the application layer (losing OS state) or at the OS layer (lacking agent-turn semantics). Crab's eBPF-based turn classification is a new mechanism that brings agent-level intent into OS-level C/R decisions without coupling to any specific agent framework.

Extension Opportunities:

Extend the eBPF turn classifier to multi-agent orchestration scenarios where multiple agents share a sandbox and turns interleave, requiring causal dependency tracking across agent boundaries
Apply the sparsity-aware checkpoint scheduling to GPU-attached sandboxes (e.g., agents running local inference or CUDA workloads), where device memory snapshotting adds a major cost dimension Crab doesn't currently address
Build an RL training harness on top of Crab's branching primitives to enable efficient tree-of-thought or Monte Carlo Tree Search rollouts over real execution environments rather than simulated ones

Replicability: No code or data availability is mentioned in the abstract. Reproduction would require: a container/microVM hosting platform with eBPF support, an agent framework generating shell-intensive or code-repair workloads, a CRIU-class C/R backend, and a multi-sandbox co-location setup. Moderate-to-high systems engineering effort; compute needs are modest (standard server hardware) but the integration surface is broad.

Research Gaps:

Evaluation is limited to shell-intensive and code-repair workloads — unclear how the 75% sparsity ratio holds for agents with heavy network I/O, database mutations, or GPU workloads
No discussion of multi-agent or multi-sandbox causal dependencies — if agent A's sandbox state depends on agent B's actions, per-sandbox checkpointing may produce inconsistent global snapshots

🔥 GitHub Trending

1. appergb/openless

⭐ 509 stars | HTML

Hold a key, speak, release — AI-polished text appears at your cursor in any app. Open-source voice input for macOS & Windows. (按住快捷键说话，松开即得润色后的文字)

ai-prompt asr dictation llm macos open-source

2. endless-sky-team/ai-trading-agent

⭐ 206 stars | TypeScript

llm trading bot ai agent crypto bitcoin ethereum perpetuals openrouter lighter algorithmic automated quant typescript react postgres bun fintech defi machine-learning open-source crypto-bot signals

ai-trading algorithmic-trading automated-trading bitcoin bun crypto-bot

3. lukiIabs/trading-agents

⭐ 204 stars | Python

TradingAgents LLM multi-agent finance trading stocks crypto fintech quantitative algo trading sentiment analysis OpenAI JavaScript Node.js research OSS

agents algorithmic-trading crypto finance fintech javascript

4. warpdot-dev/composio

⭐ 202 stars | TypeScript

typescript python sdk ai-agents anthropic openapi langchain openai-agents llamaindex mastra vercel-ai mcp oauth saas llm integrations agent-tools automation cloudflare google-gemini tooling rag multi-

ai-agents anthropic automation cloudflare devtools google-gemini

5. peter123023/awesome-claude-api

⭐ 172 stars | Unknown

收集和整理的可靠Claude中转API资源列表，帮助开发者快速找到稳定、高性价比的Claude API服务

api awesome awesome-list claude claude-code llm

6. HKUDS-AI/polymarket-ai-trading

⭐ 141 stars | HTML

AI-assisted paper trading for Polymarket: OpenAI market insight, vector search, Kelly sizing, SQLite, polymarket ai trading bot, ai trading bot, trading bot,polymarket ai trading bot, ai trading bot,

dashboard defi express fintech hkuds-ai javascript

7. warpdot-dev/craft-agents-oss

⭐ 111 stars | TypeScript

electron anthropic claude-agent-sdk mcp bun websocket thin-client multi-llm openapi skills automations oauth desktop-ai github-copilot google-ai chatgpt devtools apache-2 headless-server pi-sdk vscode

ai-agents anthropic automations bun chatgpt claude

8. maria-rcks/clawd.rip

⭐ 102 stars | TypeScript

Everything that went wrong with Claude.

agent ai anthropic claude claudecode llm

9. mebrown47/CUDA_spectrum

⭐ 69 stars | Python

A simple CUDA FFT spectrum tool

cuda dsp fft sdr spectrum-analyzer

10. PKU-YuanGroup/TIDE

⭐ 64 stars | Python

Turning the TIDE: Cross-Architecture Distillation for Diffusion Large Language Models

discrete-diffusion distillation llm nlp

11. 0xhimanshu/governor

⭐ 61 stars | Python

Claude Code usage governor: compact professional output, context slimming, tool-output filtering, telemetry, and drift guardrails.

ai-tools claude-code cli context-window developer-tools llm

12. elisaterumi-ai/agent-skills-in-practice

⭐ 54 stars | Unknown

Learn what AI skills are and how to design, structure, and use them in real-world agent systems.

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

13. LiangSu8899/FlashRT

⭐ 19 stars | C++

FlashRT is a high-performance realtime inference engine for small-batch, latency-sensitive AI workloads. The flagship integration is production VLA control for Pi0, Pi0.5, GROOT N1.6, and Pi0-FAST

cuda cuda-kernels gr00t gr00t-n1-6-3b pi pi05

14. YutoTerashima/hms-harmful-brain-activity-classification

⭐ 14 stars | Jupyter Notebook

Kaggle Silver Medal solution archive for HMS harmful brain activity EEG classification.

deep-learning eeg kaggle medical-ai pytorch spectrogram

15. jmerelnyc/ai-practice-examples

⭐ 10 stars | TypeScript

Course examples demonstrating AI implementation with industry best practices

ai-examples artificial-intelligence best-practices course-materials machine-learning typescript

Generated by Research Pulse on 2026-05-03 08:11