2026-05-25·Daily digest

🔬 Research Pulse

Daily Digest

May 25, 2026

🤖 AI

🧠 LLMs

1. SkillOpt: Executive Strategy for Self-Evolving Agent Skills

Authors: Yifan Yang, Ziyang Gong, Weiquan Huang... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How can agent skills (text-based instructions/context for frozen LLM agents) be systematically optimized with the same discipline and reliability as weight-space optimization, rather than being hand-crafted, one-shot generated, or loosely self-revised without guaranteed improvement over the starting point?

Summary: SkillOpt introduces a controllable text-space optimizer that trains agent skill documents as the external state of a frozen agent, using bounded edits accepted only on held-out validation improvement, with learning-rate budgets and rejected-edit buffers for stability. It dominates 52/52 evaluation cells across benchmarks, models, and harnesses, lifting GPT-5.5 accuracy by up to +24.8 points, and the optimized skills transfer across models, harnesses, and tasks without re-training.

Key Results: SkillOpt was best or tied on all 52 evaluated (model, benchmark, harness) cells across 6 benchmarks, 7 target models, and 3 execution harnesses (direct chat, Codex, Claude Code), beating per-cell competitors including human-written, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill skills. On GPT-5.5, it lifted average no-skill accuracy by +23.5 points in direct chat, +24.8 in Codex, and +19.1 in Claude Code. Optimized skills transferred across model scales, between Codex/Claude Code harnesses, and to a nearby math benchmark without re-optimization.

Key Findings:

Strict validation-gated acceptance of bounded edits reliably produces monotonic improvement over the starting skill, unlike loosely-controlled self-revision baselines
Optimized skill artifacts transfer across model scales, between Codex and Claude Code harnesses, and to nearby math benchmarks without further optimization
Adds zero inference-time model calls at deployment while delivering +19–25 point gains across three harnesses on GPT-5.5

Technical Novelty: First systematic controllable text-space optimizer that mirrors deep-learning optimizer discipline for skills: a separate optimizer model produces bounded add/delete/replace edits on a single skill document from scored rollouts, edits are only accepted on strict held-out validation improvement, and stability comes from a textual learning-rate budget, rejected-edit buffer, and epoch-wise slow/meta updates — distinct from prior loosely-controlled self-revision approaches (TextGrad, GEPA, EvoSkill, Trace2Skill).

What's New: Reframes skill creation as a disciplined text-space optimization problem analogous to weight-space training — with a separate optimizer model, bounded edit operations, validation gating, textual learning rates, and meta-updates — rather than treating skills as artifacts to be authored or freely rewritten by the agent itself.

Extension Opportunities:

Apply SkillOpt's discipline (textual learning rate, rejected-edit buffer, validation-gated acceptance) to multi-skill libraries with skill composition/dependency management rather than a single skill document
Extend the optimizer to jointly optimize skill text and tool definitions/MCP server configs as coupled external state for agentic harnesses
Investigate online/continual SkillOpt that incorporates production user feedback signals as the scoring function, with safety constraints on edits

Replicability: Abstract does not mention code/data release. Reproduction would require access to 7 target models (including GPT-5.5), 3 execution harnesses (Codex, Claude Code, direct chat), and substantial inference budget for repeated rollout-score-edit loops across 6 benchmarks; compute is dominated by validation rollouts rather than training, but the iterative optimizer loop on frontier models suggests meaningful API/inference cost.

Research Gaps:

Prior skill-generation methods (hand-crafted, one-shot LLM, self-revision) had no mechanism to guarantee improvement over the starting point under feedback
Lack of optimization-theoretic discipline (learning rates, rejection of bad updates, stability controls) in text-space prompt/skill optimization

2. From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills

Authors: Zisu Huang, Jingwen Xu, Yifan Yang... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: Do model-generated, domain-level agent skills actually improve language agents, when do they help vs. hurt, and what makes them succeed or fail across the full lifecycle (experience generation → skill extraction → skill consumption)?

Summary: The paper presents the first end-to-end empirical study of model-generated, domain-level agent skills across experience generation, extraction, and consumption. It shows skills help on average but cause meaningful negative transfer, that extractor and consumer abilities are independent, and proposes a 'meta-skill' that steers extraction toward utility-correlated features, consistently improving quality and reducing negative transfer.

Key Results: Built a utility-grounded evaluation framework spanning 5 diverse agentic task domains, testing multiple extractor/target model pairs. Demonstrated that model-generated skills are beneficial on average but exhibit non-trivial negative transfer; extractor strength and consumer strength are decoupled (a model can be strong as one and weak as the other); skill utility is independent of model scale or baseline task strength. A derived 'meta-skill' guiding extraction toward utility-tied features consistently improved skill quality across domains and substantially reduced negative transfer.

Key Findings:

Model-generated skills are net-beneficial but exhibit non-trivial negative transfer across tasks
Extractor strength and consumer strength are decoupled — a strong extractor may be a weak consumer of skills and vice versa; utility is uncorrelated with model scale or baseline task performance
A meta-skill that encodes the features tied to actual downstream utility consistently improves extracted-skill quality and reduces negative transfer across all studied domains

Technical Novelty: First systematic lifecycle-spanning evaluation of model-generated agent skills (vs. prior work focused on isolated extraction methods); introduces a utility-grounded evaluation framework decoupling extractor vs. consumer roles; introduces a 'meta-skill' artifact that operationalizes extraction guidance based on empirically identified utility features.

What's New: Prior work proliferates new extraction methods in isolation; this paper is the first to systematically interrogate the whole skill lifecycle, separate extractor vs. consumer roles as independent axes, and derive an actionable meta-skill grounded in measured utility rather than intuition.

Extension Opportunities:

Build an automatic skill-quality classifier that predicts negative transfer before deployment, using the utility-tied features identified in the paper
Extend the meta-skill approach to cross-domain transfer — train a meta-extractor that generalizes utility heuristics across unseen domains
Develop a consumer-aware skill extraction pipeline that tailors skill format/abstraction level to the target agent's known consumption profile

Replicability: Abstract does not mention code/data release. Reproduction would require multiple LLM-based extractor and target agent pairs across 5 agentic task domains — moderate-to-high API/compute cost depending on the model tiers used; experimental scaffolding (rollout collection, skill extraction, consumption eval) would need to be re-implemented if not released.

Research Gaps:

No characterization of which task/domain types are most prone to negative transfer, limiting safe deployment guidance
Meta-skill is derived empirically from the five studied domains; generalization to unseen domains and to multi-agent or tool-rich settings remains untested

3. MemAudit: Post-hoc Auditing of Poisoned Agent Memory via Causal Attribution and Structural Anomaly Detection

Authors: Zhewen Tan, Yilun Yao, Huiyan Jin... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: After an LLM agent exhibits harmful behavior due to poisoned persistent memory, which specific stored memory records are causally responsible? Existing defenses focus on online prevention (prompt filtering, output blocking) but cannot answer this post-hoc forensic question once an attack has already succeeded.

Summary: MemAudit is a post-hoc forensic framework for memory-poisoned LLM agents that identifies which stored memories caused harmful behavior by combining counterfactual influence scoring with structural anomaly detection over a memory consistency graph. Against the MINJA query-only injection attack, it drives attack success rates from 70%→0% (QA) and 83.3%→0% (reasoning agents).

Key Results: MemAudit was evaluated against MINJA, a query-only memory injection attack where malicious records are planted via normal agent interactions. In QA settings, attack success rate dropped from 70% to 0%; in reasoning-agent (RAP) settings, success dropped from 83.3% to 0%. Evaluation spans both QA and reasoning-agent configurations under realistic post-hoc auditing scenarios.

Key Findings:

Counterfactual ablation of individual memories yields a reliable causal-attribution signal for identifying poisoned records after harm has occurred.
Structural anomaly detection via a memory consistency graph complements influence scoring and catches malicious records that look locally plausible but are globally out-of-distribution.
Combining both signals fully neutralizes MINJA across QA (70%→0%) and reasoning-agent (83.3%→0%) settings under realistic auditing conditions.

Technical Novelty: The combination of two orthogonal signals is the key novelty: (1) a counterfactual memory influence score that ablates individual memories and measures the causal delta on harmful outputs, and (2) a memory consistency graph that flags records that are structurally anomalous relative to the rest of the memory store. Prior defenses are online interventions (filter inputs, block outputs); MemAudit is the first framework explicitly framed as post-hoc causal attribution over agent memory.

What's New: Reframes agent memory security from prevention to forensic attribution. Prior work intercepts attacks at input or output time; MemAudit is the first to ask 'which stored memory is to blame?' and answer it by fusing causal counterfactuals with graph-based structural anomaly detection over the memory store itself.

Extension Opportunities:

Extend MemAudit from post-hoc to streaming/online auditing — incrementally update influence scores and the consistency graph as new memories are written, enabling near-real-time quarantine of suspicious records before they steer behavior.
Apply the counterfactual influence score and consistency-graph signals to multi-agent memory pools (shared RAG stores, team-of-agents systems) where one compromised agent can poison memories consumed by peers — current evaluation appears single-agent.
Benchmark MemAudit against a broader attack zoo beyond MINJA — e.g., gradient-guided poisoning, semantic-drift attacks, or backdoors triggered only by rare query patterns — to test whether the structural anomaly signal generalizes beyond query-only injection.

Replicability: The abstract does not mention a public code release or dataset. Reproduction would require a MINJA-compatible memory-augmented agent stack (a RAG memory store plus an LLM backbone) and the ability to run counterfactual ablations — each memory removed and the agent re-queried — which is O(N) LLM calls per audit but tractable on a single GPU node or via API for modest memory banks (hundreds to low thousands of records).

Research Gaps:

Scalability of per-memory counterfactual ablation to very large persistent memory stores (millions of records) is not addressed by the abstract.
Evaluation is limited to a single attack family (MINJA); robustness against adaptive adversaries who craft memories to evade both the influence score and the consistency graph remains open.

🦾 ROBOTICS

1. ChainFlow-VLA: Causal Flow Planning with Vision-Language Models

Authors: Xiyang Wang, Xinlin Wang, Tingguang Zhou... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How can autonomous driving planners unify the causal temporal reasoning of autoregressive models with the global trajectory optimization of diffusion models, instead of treating them as separate paradigms with complementary weaknesses (AR error accumulation vs. diffusion's lack of causal constraints)?

Summary: ChainFlow-VLA unifies autoregressive causal trajectory generation with diffusion-based global refinement under a single probabilistic framework, using VLM hidden states as semantic priors for mode-conditioned residual correction. It achieves 94.85 on NAVSIM v1, matching human-level driving performance (94.8) and outperforming approaches that treat causal modeling and global optimization separately.

Key Results: ChainFlow-VLA achieves a state-of-the-art score of 94.85 on the NAVSIM v1 leaderboard, matching human-level performance (94.8). The paper demonstrates robust planning in ambiguous and long-tail driving scenarios by combining an AR 'Chain' generator producing discrete causal trajectory modes with a diffusion 'Flow' refiner that uses VLM hidden states as semantic priors for mode-conditioned residual correction.

Key Findings:

Combining AR mode generation (Chain) with diffusion residual refinement (Flow) outperforms either paradigm alone, resolving the trade-off between temporal causality and global trajectory consistency
VLM hidden states serve as effective semantic priors for fine-grained trajectory correction, especially in ambiguous and long-tail driving scenarios
The system reaches 94.85 on NAVSIM v1, effectively matching the 94.8 human-level performance benchmark

Technical Novelty: The novel contribution is formulating planning as a mixture over AR-induced modes with VLM-conditioned residual diffusion refinement. Unlike prior work that picks either AR or diffusion, ChainFlow-VLA composes them: AR produces a discrete set of causal trajectory hypotheses, then a diffusion model performs mode-conditioned correction in residual space (preserving causal structure) using VLM hidden states as semantic priors — a principled probabilistic unification rather than ensembling.

What's New: Prior end-to-end driving stacks either use AR (suffering error accumulation) or diffusion (lacking causal constraints). ChainFlow-VLA is the first to formulate planning as a mixture over AR-induced modes refined by mode-conditioned residual diffusion, with VLM semantic conditioning injected at the refinement stage rather than the generation stage.

Extension Opportunities:

Apply the Chain+Flow mixture-of-modes framework to other sequential decision domains beyond driving (e.g., robotic manipulation, drone navigation) where causal interaction and global smoothness both matter
Swap the VLM conditioning module to test whether larger or domain-specialized VLMs (e.g., driving-finetuned VLMs) further improve long-tail scenario handling, and ablate which VLM layers provide the most useful hidden-state priors
Extend the residual diffusion refiner to operate over multi-agent joint trajectory distributions rather than ego-only, enabling interaction-aware global refinement for dense traffic

Replicability: Code is promised at https://github.com/AFARI-Research/ChainFlow-VLA (not confirmed released at abstract time). Reproducing would require NAVSIM v1 benchmark data, a pretrained VLM backbone, and substantial GPU compute typical for VLM+diffusion training (likely multi-GPU A100-class). Exact compute is not stated in the abstract.

Research Gaps:

The abstract does not report performance under real-world deployment, closed-loop simulation beyond NAVSIM, or robustness to adversarial/sensor-corrupted inputs
It does not quantify inference latency or computational overhead of running both AR and diffusion modules plus a VLM — a practical concern for onboard autonomous driving

2. TactileReflex: Noise-Statistics-Driven Vision-Tactile Reflex Control for Force-Sensitive Manipulation

Authors: Ziyan Feng, Yulong Fu, Zheng Li... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How can robots manipulate fragile deformable containers (e.g., thin plastic cups with liquid) within an extremely narrow force margin where insufficient grip causes slip and excessive grip causes irreversible deformation, without requiring external force calibration, trial-and-error tuning, or material-specific models?

Summary: TactileReflex is a three-channel closed-loop tactile controller for force-sensitive manipulation of fragile deformable containers, where all controller thresholds are derived from the tactile sensor's intrinsic noise statistics rather than external calibration. It operates at ~12 Hz on image-level proxies from dual visuo-tactile sensors and is designed to act as a plug-and-play safety layer beneath higher-level policies including VLA models and VR teleoperation.

Key Results: TactileReflex demonstrated 5/5 success in preventing irreversible container deformation vs. at most 1/5 for partial (ablated) configurations. In a dynamic pouring task, it achieved 9/10 success across two water volumes, while fixed-effort baselines failed 0/10 due to pose drift. The controller operates at ~12 Hz using three image-level proxies (shear intensity Sy, contact intensity Fn, center of pressure C) from dual visuo-tactile sensors.

Key Findings:

Noise-statistics-based threshold derivation eliminates the need for external force calibration and material-specific physical models, enabling a calibration-free pipeline
All three reflex channels (slip suppression, weight-adaptive release, force protection) are necessary: the full system achieves 5/5 success at preventing deformation while any ablated configuration achieves at most 1/5
In dynamic pouring, the reflex controller compensates for pose drift that defeats fixed-effort baselines, achieving 9/10 vs. 0/10 success

Technical Novelty: Deriving controller thresholds directly from the sensor's intrinsic noise statistics via a brief static-hold-and-unload protocol — eliminating the standard requirements of external force calibration, manual threshold tuning, or material-specific contact/deformation models. Combined with a three-channel prioritized reflex architecture (slip suppression, weight-adaptive release, force protection) operating on image-level proxies rather than reconstructed forces.

What's New: Prior approaches to fragile object manipulation typically require force-calibrated sensors, material-specific models, or learned policies trained per-object. TactileReflex replaces all of these with a sensor-intrinsic noise calibration and image-level proxies, and explicitly targets the role of a model-agnostic safety layer beneath higher-level policies rather than a standalone manipulation policy.

Extension Opportunities:

Integrate TactileReflex as a safety layer beneath a VLA (vision-language-action) policy for general household manipulation, evaluating whether it broadens the operational envelope of pretrained policies on fragile-object tasks
Extend the noise-statistics calibration paradigm to other vision-based tactile sensors (GelSight, DIGIT, 9DTact) and benchmark whether the calibration-free property generalizes across sensor designs
Add a temperature/humidity-aware recalibration loop so the static-hold-and-unload protocol re-runs periodically, enabling long-horizon deployment where sensor noise statistics drift

Replicability: The abstract does not mention a code or data release. Reproduction would require dual vision-based tactile sensors (e.g., GelSight-style) mounted on a parallel gripper, a robot arm capable of ~12 Hz closed-loop control, and modest compute (image-level proxies suggest lightweight CV, likely a single GPU or even CPU sufficient). The static-hold-and-unload calibration protocol appears to be the only data requirement, making physical reproduction relatively low-overhead if the hardware stack is available.

Research Gaps:

Evaluation is limited to disposable plastic cups with liquid; generalization to other fragile deformables (eggs, thin-walled fruit, soft food) is untested
The 12 Hz control rate may bound performance for faster dynamic tasks like catching or rapid pouring, and the integration with downstream VLA/teleoperation policies is proposed but not empirically validated in the abstract

3. Direct Dynamic Retargeting for Humanoid Imitation Learning from Videos

Authors: Constant Roux, Ludovic De Matteïs, Armand Jordana... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How can we translate human motion from monocular videos to humanoid robots without the geometric bias introduced by intermediate kinematic projections in standard retargeting pipelines?

Summary: The paper introduces Direct Dynamic Retargeting (DDR), a single-stage framework that converts monocular human videos into dynamically feasible humanoid trajectories without intermediate kinematic projections. By combining task-space formulation with sampling-based MPC inside a physics simulator, DDR avoids the geometric bias of prior pipelines and produces references that improve both tracking accuracy and downstream RL training.

Key Results: DDR outperforms state-of-the-art Geometric Retargeting and Indirect Dynamic Retargeting baselines in demonstration tracking accuracy. Physically viable references generated by DDR accelerate RL training convergence and improve execution of agile/balancing behaviors. Note: abstract lacks specific numerical benchmarks, datasets, or quantitative metrics.

Key Findings:

Intermediate kinematic projections in standard retargeting introduce a geometric bias that restricts search space and yields suboptimal dynamics
Task-space formulation with sampling-based MPC natively handles complex contact sequences and mitigates input drift
Physically viable references from DDR accelerate RL convergence and improve final policy execution for agile and balancing skills

Technical Novelty: Single-stage framework that bypasses intermediate kinematic retargeting entirely — formulates retargeting in task space and uses sampling-based MPC within a physics simulator to natively handle contact sequences and input drift, rather than the standard two-stage geometric-then-dynamic pipeline.

What's New: Unlike prior Geometric Retargeting and Indirect Dynamic Retargeting that decompose the problem into sequential kinematic-then-dynamic stages, DDR is a unified single-stage approach operating directly in task space with physics-aware optimization, eliminating bias from intermediate representations.

Extension Opportunities:

Extend DDR to multi-agent or human-robot interaction scenarios where contact sequences involve external entities
Integrate learned dynamics models with the sampling-based MPC to reduce reliance on physics simulator fidelity and enable real-time on-robot retargeting
Apply DDR framework to non-humanoid morphologies (quadrupeds, soft robots) to test generality of direct dynamic retargeting beyond bipedal cases

Replicability: Source code promised to be publicly released. Reproduction requires a physics simulator (likely MuJoCo or similar), sampling-based MPC implementation, and humanoid robot model. Compute needs are moderate for MPC sampling but no specific GPU/CPU requirements stated in abstract.

Research Gaps:

No quantitative benchmarks or sim-to-real transfer results disclosed in the abstract
Reliance on sampling-based MPC inside a simulator may limit real-time deployment and generalization across robot platforms

💻 COMPUTE

1. Noise and Configuration Recovery Impact on Quantum Selected Configuration Interaction

Authors: Nonia Vaquero-Sabater, Abel Carreras, Lukas Broers... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How do ansatz expressivity, sampling noise, and configuration recovery interact in Quantum-Selected Configuration Interaction (QSCI) with the LUCJ ansatz, and can noise/recovery overcome the bias of compact ideal sampling distributions?

Summary: The paper dissects QSCI performance with the LUCJ ansatz on N2 dissociation, showing that ideal noiseless sampling yields biased, overly compact configuration spaces. Sampling noise and—more strikingly—classical configuration recovery (even from random starts) dramatically improve CI accuracy, suggesting recovery is the true engine of QSCI rather than ansatz quality.

Key Results: Using N2 dissociation in a large active space as benchmark, the authors demonstrate: (1) noiseless LUCJ sampling produces compact, biased configurational spaces that limit CI energy accuracy in strongly correlated regimes; (2) a simple noise model expands Hilbert-space exploration beyond the ideal ansatz's support; (3) combining noise with configuration recovery yields systematically improved energies; (4) configuration recovery starting from randomly generated configurations alone can construct accurate CI spaces.

Key Findings:

Noiseless LUCJ sampling produces biased, compact configurational spaces that cap CI accuracy in strongly correlated regimes
Sampling noise paradoxically improves results by expanding configurational coverage beyond the ansatz's natural support
Configuration recovery starting from random configurations alone constructs accurate CI spaces, making it the central mechanism of QSCI success

Technical Novelty: Reframes hardware noise from a liability into a useful resource for QSCI by showing it acts as an implicit Hilbert-space exploration mechanism, and isolates configuration recovery as the dominant contributor to QSCI accuracy — challenging the assumption that ansatz quality is the central bottleneck.

What's New: Prior QSCI work focused on ansatz design and viewed noise as detrimental. This paper inverts that framing — showing noise is beneficial for exploration and that configuration recovery, not the quantum ansatz, does most of the heavy lifting.

Extension Opportunities:

Test the noise-assisted QSCI framework on other strongly correlated benchmarks (Cr2, FeMoco, transition metal complexes) to assess generality beyond N2
Develop tailored/structured noise injection protocols that optimize Hilbert-space exploration rather than relying on generic device noise
Investigate purely classical configuration-recovery-based CI solvers (since random-start recovery alone works), potentially eliminating the quantum sampling step entirely for some regimes

Replicability: Abstract does not mention code/data availability. Reproduction would require a quantum chemistry stack supporting LUCJ ansatz simulation (e.g., Qiskit/ffsim), classical CI diagonalization tools, and access to noiseless + noisy quantum circuit simulators. Compute is modest for N2 active spaces on classical simulators; no real quantum hardware strictly required.

Research Gaps:

Lack of theoretical understanding of why LUCJ sampling distributions are systematically biased and how to design ansatzes that natively cover relevant configurations
Unclear how the noise-assisted advantage scales to larger active spaces and other molecular systems beyond N2

2. Twirled Perfect Tensor Networks: Computationally covariant holographic tensor networks

Authors: Gurbir Arora, Matthew Headrick, Albion Lawrence... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: Do random tensor network models of holography correctly realize the Python's Lunch Conjecture (PLC) bound on the complexity of reconstructing operators behind non-minimal extremal surfaces, and if not, what tensor network class does?

Summary: The paper identifies a mismatch between the Python's Lunch Conjecture's complexity bound and the actual reconstruction complexity in random tensor network models of holography, attributing it to a missing 'computational covariance' property. To fix this, it introduces twirled perfect tensor networks (TPTNs), which provably saturate the PLC bound, satisfy a lattice Ryu–Takayanagi formula for arbitrary boundary subregions, and unify desirable features of perfect and random tensor networks.

Key Results: The authors show explicitly in random tensor network (RTN) models of the python's lunch that the exponential reconstruction complexity is NOT generically controlled by the PLC exponent — RTNs lack 'computational covariance,' meaning space cannot be arbitrarily decomposed into low-complexity units. They construct an alternative class, 'twirled perfect tensor networks' (TPTNs), and prove these (i) satisfy computational covariance, (ii) have complexity bounded by the PLC value, and (iii) obey a lattice Ryu–Takayanagi formula for arbitrary boundary subregions. A residual discrete limitation from local postselection persists that is absent in gravity.

Key Findings:

Random tensor networks fail to realize the PLC complexity bound because they lack computational covariance
Twirled perfect tensor networks satisfy computational covariance and have complexity bounded by the PLC exponent
TPTNs obey a lattice Ryu–Takayanagi formula for arbitrary boundary subregions, combining strengths of perfect and random tensor networks
A residual discrete gap with gravity persists due to local postselection limitations in TPTNs

Technical Novelty: Introduction of 'twirled perfect tensors' — perfect tensors averaged (twirled) over a group action — yielding a tensor class that simultaneously enjoys perfect-tensor structural rigidity (RT for arbitrary subregions) and random-tensor flexibility (typical entanglement), while satisfying a new 'computational covariance' property that prior RTN and perfect-tensor (e.g., HaPPY) constructions do not.

What's New: Prior holographic tensor network constructions (HaPPY perfect tensors, random tensor networks) each capture some bulk-boundary features but neither satisfies the computational covariance property implicit in the Python's Lunch Conjecture. The twirled-perfect-tensor construction is a new hybrid that fills this gap and provides the first tensor network class explicitly engineered to respect the PLC's complexity scaling.

Extension Opportunities:

Implement TPTNs numerically (e.g., as PEPS/MERA variants) and benchmark entanglement-spectrum and postselection-complexity scaling against RTN baselines on small lattices
Apply TPTNs as a holographic-inspired ansatz for quantum error-correcting codes or fault-tolerant architectures where covariance under local decomposition is desired
Extend the construction to dynamical/time-evolving holographic settings (e.g., model evaporating black holes or shockwave geometries) and test whether the residual local-postselection gap with gravity closes

Replicability: The abstract does not mention released code or datasets; the work appears primarily analytic/theoretical. Reproducing the arguments requires expertise in holography, random tensor networks, and complexity theory but no significant compute. Numerical verification of TPTN properties on small networks would be tractable on a workstation.

Research Gaps:

A residual discrete complexity penalty from local postselection remains in TPTNs that is absent in semiclassical gravity, suggesting the construction is still not fully holographic
The framework is static — extension to dynamical bulk geometries and time-evolving holographic codes is not addressed

3. The Floquet-Magnus expansion of unbounded operators

Authors: Daniel Burgarth, Robin Hillier, Davide Lonigro... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How can the Floquet-Magnus expansion, a standard tool for approximating time-periodic quantum systems by effective time-independent Hamiltonians, be rigorously extended beyond bounded Hamiltonians to the broad class of unbounded Hamiltonians that arise in realistic quantum models?

Summary: The paper rigorously extends the Floquet-Magnus expansion to time-periodic unbounded Hamiltonians via a new nonperturbative framework, proving that the resulting effective dynamics approximates the true propagator to arbitrary order in the high-frequency limit without needing the Magnus series to converge. The approach is demonstrated on the Rabi model and the periodically driven harmonic oscillator.

Key Results: The authors construct a nonperturbative framework for effective Hamiltonians of time-periodic unbounded systems and prove it reproduces the Floquet-Magnus expansion. Crucially, they prove the effective dynamics approximates the original propagators to arbitrary order in the high-frequency limit without requiring convergence of the Floquet-Magnus series. Validated on two representative models: the quantum Rabi Hamiltonian in the interaction picture and the periodically driven quantum harmonic oscillator.

Key Findings:

A nonperturbative construction of effective Hamiltonians can be defined for a broad class of unbounded time-periodic Hamiltonians and matches the formal Floquet-Magnus expansion
Arbitrary-order high-frequency approximation of propagators is provable without requiring Magnus-series convergence, removing a restrictive condition
The framework applies to physically relevant unbounded models: the quantum Rabi Hamiltonian (interaction picture) and the periodically driven harmonic oscillator

Technical Novelty: Prior Floquet-Magnus theory required bounded Hamiltonians for convergence arguments. This work introduces a distinct nonperturbative construction of effective Hamiltonians valid for unbounded operators, then shows it formally agrees with Floquet-Magnus while bypassing the convergence requirement — yielding arbitrary-order high-frequency approximation guarantees that were previously inaccessible.

What's New: Existing Floquet-Magnus analyses are restricted to bounded operators and rely on series convergence; this work provides the first rigorous extension to unbounded Hamiltonians and decouples approximation guarantees from convergence of the expansion itself.

Extension Opportunities:

Apply the framework to many-body unbounded driven systems (e.g., Bose-Hubbard with periodic modulation) where standard Magnus expansions fail
Develop numerical schemes for computing the effective Hamiltonian terms order-by-order for specific unbounded models and benchmark against exact Floquet diagonalization
Extend to dissipative time-periodic Lindbladians with unbounded generators, relevant for driven open quantum systems

Replicability: No code or data is mentioned in the abstract; the contribution is mathematical/analytical. Reproduction requires no compute — only functional-analytic verification of the proofs and symbolic computation for the worked examples (Rabi, driven harmonic oscillator).

Research Gaps:

Convergence of the Floquet-Magnus expansion itself remains restrictive even after this extension — only asymptotic high-frequency control is established
Many-body and dissipative unbounded driven systems are not directly addressed

🔬 MATERIALS

1. Democratising Optical Orbital Angular Momentum: a Set of Cost-Effective Tools

Authors: Natasha Bierrum, Lyuxuan Chen, Ananya Kudaloor... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How can optical orbital angular momentum (OAM) and quantum mechanics concepts be taught affordably and interactively in classrooms without expensive spatial light modulators (SLMs)?

Summary: The paper introduces a 'poor man's' spatial light modulator built from a fork diffraction grating printed on photographic slide film and illuminated with a laser pointer, successfully generating vortex beams that carry orbital angular momentum. It positions this kit as an engaging, affordable hands-on tool for teaching quantum mechanics and optical OAM, bypassing the cost and abstraction barriers of conventional SLM-based instruction.

Key Results: Demonstrated that fork diffraction gratings printed cheaply on photographic slide film (outsourced to an external company) can successfully generate vortex beams carrying orbital angular momentum when paired with a basic laser pointer, replicating the core function of a costly SLM in a classroom-friendly setup. No quantitative benchmarks (efficiency, mode purity, topological charge fidelity) are reported in the abstract.

Key Findings:

Fork diffraction gratings outsourced for printing on photographic slide film are cheap, easy to produce, and optically functional.
A simple laser-pointer + printed grating setup reliably produces observable OAM-carrying vortex beams.
The toolkit supports both demonstrative teaching and open-ended investigative student inquiry, broadening pedagogical use cases.

Technical Novelty: The novelty is pedagogical/practical rather than physical: outsourcing fork-grating fabrication onto photographic slide film yields a disposable, low-cost ($-scale) alternative to spatial light modulators (typically thousands of dollars), making OAM experiments accessible at the high-school/undergraduate level.

What's New: Prior OAM teaching demonstrations rely on expensive SLMs or laboratory-grade holograms; this work shows a sub-$10-class consumable alternative produced via commercial photographic printing, making OAM experimentation feasible in resource-limited education settings.

Extension Opportunities:

Quantify the mode purity and diffraction efficiency of the slide-film gratings versus commercial SLMs across multiple topological charges (l = 1, 2, 3...) to characterize the trade-off.
Develop a companion open-source toolkit (grating design files, lab worksheets, interferometric detection guides) so educators worldwide can self-replicate, including a Mach-Zehnder add-on for measuring OAM via spiral interference patterns.
Extend the kit to demonstrate quantum-relevant phenomena such as OAM entanglement analogs, mode superposition, or single-photon OAM detection using inexpensive avalanche photodiodes.

Replicability: No code or dataset is mentioned. Reproduction requires only a laser pointer, a printed slide-film fork grating (outsourced printing), and an optical bench/screen — negligible compute, very low cost, and accessible to any teaching lab.

Research Gaps:

No quantitative characterization of the printed gratings (efficiency, mode purity, achievable topological charges, durability) is provided.
No assessment of learning outcomes — the pedagogical effectiveness claim is not evaluated against control instruction or student-performance metrics.

2. Graphene-based Photodetector with Engineered Hot Carrier Cooling Dynamics

Authors: Yishu Huang, Anand Nivedan, Florian Ludwig... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How can hot-carrier cooling dynamics in graphene photodetectors be engineered to enhance photoresponsivity in waveguide-integrated silicon photonics platforms without requiring an external bias?

Summary: The authors fabricate a WSe2/graphene/WSe2 waveguide-integrated photodetector and add a nearby graphite layer that screens Coulomb interactions, slowing hot-carrier cooling. Using a CW photomixing technique they measure up to a 4x increase in cooling time, which translates into ~50% higher internal photoresponsivity — showing proximity screening as a new lever for tuning graphene photodetector performance on silicon photonics platforms.

Key Results: Demonstrated that introducing proximity screening via a nearby graphite layer in a WSe2-graphene-WSe2 waveguide-integrated photodetector prolongs hot-carrier cooling time by up to 4x (measured via CW photomixing technique), translating to ~50% improvement in internal photoresponsivity over the unscreened baseline.

Key Findings:

Proximity screening by a nearby graphite layer prolongs hot-carrier cooling time by up to a factor of 4
Internal photoresponsivity of the waveguide-integrated WSe2/graphene/WSe2 detector improves by ~50% with screening
CW photomixing is a viable, device-relevant method to extract cooling dynamics without ultrafast pulsed lasers

Technical Novelty: Prior work on graphene photodetectors focused on optimizing the active heterostructure (e.g., hBN encapsulation, contact engineering). This paper introduces a passive proximity-screening graphite layer adjacent to the device that modifies Coulomb interactions and suppresses electron-phonon cooling pathways, decoupling cooling-time engineering from the active stack. Combined with CW photomixing as the characterization method (rather than ultrafast pump-probe), this provides a device-relevant measurement of cooling dynamics.

What's New: Introduces proximity dielectric/Coulomb screening from a separate graphite layer as a design knob for graphene photodetector dynamics — distinct from prior approaches that tuned the heterostructure stack itself or relied on bias-driven mechanisms.

Extension Opportunities:

Vary graphite-graphene separation distance and dielectric environment systematically to derive a quantitative scaling law between screening strength and cooling time enhancement
Replace graphite screening layer with tunable gated graphene to enable electrically controllable cooling dynamics and dynamic responsivity modulation
Extend the WSe2/graphene/WSe2 + screening stack to mid-IR or THz wavelengths to test whether proximity-screening benefits transfer to longer-wavelength on-chip detectors

Replicability: Abstract does not mention code/data availability. Reproduction requires a cleanroom with 2D-material dry-transfer capability (graphene, WSe2, graphite, hBN flakes), silicon photonics waveguide fabrication, cryogenic or RT photomixing setup with two tunable CW lasers, and microwave/RF readout — substantial experimental infrastructure rather than compute.

Research Gaps:

No quantitative model linking screening geometry (graphite thickness, separation) to cooling-time enhancement is described in the abstract
Bandwidth/speed trade-off from prolonged cooling time is not addressed — slower cooling typically reduces detector bandwidth

3. C-band 160 Gbs-1 Zero-bias Graphene Photodetectors: Breaking the Responsivity-Bandwidth Trade-off by Heterostructure Engineering

Authors: Karuppasamy Pandian Soundarapandian, Alberto Montanaro, Ioannis Vangelidis... Published: 2026-05-22 | Citations: 0 arXiv | PDF

Research Question: How can the fundamental responsivity-bandwidth trade-off in graphene photodetectors—caused by rapid hot-carrier cooling—be broken to enable simultaneously sensitive and ultrafast zero-bias operation at C-band telecom wavelengths?

Summary: The authors break the long-standing responsivity-bandwidth trade-off in graphene photodetectors by encapsulating graphene in WSe2, which suppresses out-of-plane hot-carrier cooling and extends the cooling length to ~2.68 um. The resulting zero-bias devices achieve ~0.12 A/W responsivity while supporting 160 Gb/s PAM-4 data transmission at C-band, establishing dielectric engineering as a new design lever for ultrafast optoelectronics.

Key Results: Demonstrated WSe2-encapsulated graphene photodetectors achieving responsivity ~0.12 A/W (potentially ~0.4 A/W), cooling length ~2.68 um, bandwidth beyond setup-limited 110 GHz, and direct detection at 120 Gb/s NRZ and 160 Gb/s PAM-4 with minimal DSP, all at zero bias.

Key Findings:

WSe2 encapsulation extends graphene hot-carrier cooling length to ~2.68 um by reducing the heat-exchange coefficient
Zero-bias responsivity reaches ~0.12 A/W (with ~0.4 A/W projected upper bound) while bandwidth exceeds the 110 GHz measurement limit
Direct detection at 120 Gb/s NRZ and 160 Gb/s PAM-4 is achieved with minimal DSP, demonstrating practical viability for next-gen interconnects

Technical Novelty: Using dielectric environment engineering (WSe2 encapsulation) as a deliberate design axis to suppress out-of-plane phonon-mediated energy dissipation from hot carriers, rather than relying on geometry, doping, or biasing schemes used in prior graphene PD work.

What's New: Prior graphene PDs forced a choice between sensitivity and speed; this work reframes the dielectric environment itself as a tunable knob for hot-carrier dynamics, achieving both simultaneously at zero bias—removing the need for power-hungry biasing typical of fast III-V or Ge detectors.

Extension Opportunities:

Explore alternative 2D dielectric encapsulants (e.g., MoS2, hBN/WSe2 stacks, or higher-k materials) to push cooling length further and approach the theoretical 0.4 A/W responsivity ceiling
Integrate the WSe2/graphene heterostructure with silicon photonics waveguides for wafer-scale co-packaged optics targeting 200G+ per-lane transceivers in AI datacenter interconnects
Characterize device performance beyond 110 GHz with higher-bandwidth measurement setups to identify the true intrinsic bandwidth limit and extend to O-band/2-micron wavelengths

Replicability: No code/data availability mentioned in the abstract. Reproduction requires a 2D-materials fab (CVD/exfoliated graphene + WSe2, dry transfer, e-beam lithography), a >110 GHz lightwave component analyzer, and a 160 Gb/s PAM-4 BERT—accessible only to well-equipped photonics labs.

Research Gaps:

The intrinsic device bandwidth ceiling remains unknown because measurements are setup-limited at 110 GHz
Path to scalable, wafer-level integration of WSe2/graphene heterostructures with silicon photonics platforms is not addressed

🔥 GitHub Trending

1. AndrewDzzz/browser-harness-bidi

⭐ 5 stars | Python

WebDriver BiDi-first browser harness for AI agents, inspired by browser-use/browser-harness.

agent-tools ai-agent bidi browser-automation browser-harness codex

2. maximilian32541-spec/npc-dialogue-starter

⭐ 2 stars | JavaScript

Starter template for building LLM-powered NPC dialogue systems in games

game-dev javascript llm npc starter template

3. pguso/mcp-from-scratch

⭐ 2 stars | JavaScript

Learn Model Context Protocol (MCP) from scratch in plain Node.js. Build a working MCP server step by step with zero black boxes, from JSON-RPC and stdio transport to tools, prompts, resources, samplin

ai ai-agents json-rpc learning llm mcp

4. datallmhub/agentflow4j-cookbook

⭐ 2 stars | Java

Runnable multi-agent LLM workflow examples in Java — RAG, ticket triage, web research, Slack bot, batch processing — with AgentFlow4J and Spring AI.

agent-orchestration agentflow4j ai-agents generative-ai java langchain4j

5. dizena/open-code-api

⭐ 2 stars | Go

支持自定义配置任意主流大模型，完美适配 Codex、Claude Code、Gemini CLI 等开发工具;Supports custom configuration of any mainstream large language model, with perfect compatibility for development tools such as Codex, Claude Code

claude-code cli code codex gemini-cli llm

6. maximilian32541-spec/game-ai-utils

⭐ 2 stars | JavaScript

CLI utilities for testing and debugging game AI pipelines

cli debugging game-dev javascript llm npc

7. martsokha/bentoml

⭐ 1 stars | Rust

Unofficial async Rust client for BentoML services: generic calls, task queues, streaming, and file I/O over reqwest.

api-client bentoml client http-client inference machine-learning

8. 3123595861-byte/Helmet_Compliance

⭐ 1 stars | Python

基于 YOLOv8 + Streamlit 构建的工业级实时安全帽检测面板，采用边缘计算架构，支持多线程画面流采集与自动化违规留痕。以及相关基础知识的一些学习路径。

helmet-detection machine-learning opencv python streamlit yolov8

9. ghalaxqahtani/AI-Agriculture-Stress-Detection

⭐ 1 stars | JavaScript

AI-powered crop stress detection system using Deep Learning and Computer Vision.

agriculture artificial-intelligence computer-vision deep-learning express machine-learning

10. FreeAutomation-Tech/fastapi-ai-kit

⭐ 1 stars | Python

Production-ready FastAPI template with OpenAI, Anthropic, and Ollama AI integration. Streaming, rate limiting, Docker, and more.

ai anthropic api backend docker fastapi

11. crystalknife/NeuroDrive-RL

⭐ 1 stars | Python

Autonomous driving system using PPO-based Reinforcement Learning and CARLA Simulator for lane following and navigation.

ai artificial-intelligence autonomous-driving autonomous-driving-vehicles carla carla-driving-simulator

12. praveena-0407/RouteNavigator

⭐ 1 stars | Python

AI - powered AP Smart Route Finder using Python, Gradio, NetworkX & Folium

ai folium gradio machine-learning maps networkx

13. ruiyli/coeus

⭐ 1 stars | Python

A research-assistant pipeline that takes a clear ML research thesis as input and produces a reviewer-grade paper.tex as output, with a human in the loop at five decision gates.

ai-research apple-silicon latex llm-agents machine-learning mps

14. Siddharth-k7/Image-classification

⭐ 1 stars | Jupyter Notebook

Computer vision practice with CNNs, MNIST, RGB image basics, AlexNet, VGG, ResNet, Inception, CIFAR-10, flowers, and Rock Paper Scissors classification.

cifar10 cnn computer-vision deep-learning image-classification keras

15. garyqlin/gbase

⭐ 1 stars | Python

GBase — Recursive Self-Improvement Agent Framework. Memory, evolution, quality gates, identity system, and 40+ auto-registered tools.

agent-framework ai-agents cognitive-architecture ebbinghaus llm mirror-memory

Generated by Research Pulse on 2026-05-25 06:10