the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain), with configuration-level gains ranging from +4.1 to +25.7 pp. Focused Skills with at most three modules outperform larger or exhaustive bundles, and smaller models with Skills can match

larger models without them. SkillsBench establishes paired evaluation as the foundation for rigorous measurement of Skill efficacy on agentic, expertise-heavy work.

Fig. 2 | Agent architecture stack and resolution rates across 18 model–harness configurations on 87 SkillsBench tasks. Each bar stacks the no-Skills baseline and curated-Skills lift; open/filled interval markers denote 95% CIs for the

baseline/total. Bars are ordered by Curated-Skills pass rate and colored by model family. Harness abbreviations: OH = OpenHands, CC = Claude Code, GCLI = Gemini CLI. Figure 1 of arXiv:2602.12670 (CC BY 4.0).

Li, X.*, Choe, K. W.*, Liu, Y., Chen, X., Tao, C., You, B. et al. In COLM (2026); preprint at arXiv:2604.05172.

Large language model (LLM) agents are increasingly deployed to automate productivity tasks (e.g., email, scheduling, document management), but evaluating them on live services is risky due to potentially irreversible changes. Existing benchmarks rely on simplified environments and fail to capture realistic, stateful, multi-service workflows. We introduce ClawsBench, a benchmark for evaluating and improving

LLM agents in realistic productivity settings. It includes five high-fidelity mock services (Gmail, Slack, Calendar, Docs, Drive) with full state management and deterministic snapshot/restore, along with 44 structured tasks covering single-service, cross-service, and safety-critical scenarios. We decompose agent scaffolding into two independent levers (domain skills that inject API knowledge via progressive

evidence that a concrete run remained within its intended evaluation boundary. This paper presents BenchShield, a model-backed instrumentation layer for reward integrity in LLM-agent evaluation. BenchShield grounds detection in a finite lifecycle model of an evaluation’s reward-relevant events. Within the benchmark infrastructure, two complementary analyses operate over this model. A static, phase-aware taint analysis exposes reward-hacking paths before a run. Its runtime counterpart uses infrastructure-side evidence to attribute concrete agent use and emit evidence-backed claims.

We construct BenchShield Trajectories, a human-labeled corpus of 456 adjudicated trajectories from more than 31,000 public agent runs across three benchmarks. Compared with an agentic hackability scanner baseline on the same tasks and model, BenchShield improves full-chain recall from 23–94% to 77–100%, same-vector coverage from 16–56% to 43–78%, and reduces per-task cost by up to 65%. Its runtime analysis achieves 96% accuracy in detecting reward hacking from infrastructure-side evidence.

Fig. 4 | Two reward-hacking failures from Terminal-Bench 3 in which an isolated verifier is insufficient. Left: agent-controlled executable content disables a kernel check through the declared patch (I2 → I7). Right: the agent

downloads hidden labels and submits answer-derived clusters (I1 → I7). Both runs retain a native reward of 1.00; BenchShield reports AgentViolation. Figure 1 of arXiv:2609.11028 (CC BY 4.0).

Experimental AMO Physics

Xu, S.*, Xia, X.*, Yu, Q., Parakh, A., Khan, S., Megidish, E., You, B. et al. Nature 645, 362–368 (2025). arXiv:2310.00595.

Trapped-ion applications, such as in quantum information, precision measurements, optical clocks, and mass spectrometry, rely on specialized high-performance ion traps. The latter applications typically employ traditional machining to customize macroscopic 3D Paul traps, while quantum information processing experiments usually rely on photo-lithographic techniques to miniaturize the traps and meet scalability requirements. Using photolithography, however, it is challenging to fabricate the complex three-dimensional electrode structures required for optimal confinement. Here we address these limitations by adopting a high-resolution 3D printing technology based on two-photon polymerization supporting fabrication of large

arrays of high-performance miniaturized 3D traps. We show that 3D-printed ion traps combine the advantages of traditionally machined 3D traps with the miniaturization provided by photolithography by confining single calcium ions in a small 3D-printed ion trap with radial trap frequencies ranging from 2 MHz to 24 MHz. The tight confinement eases ion cooling requirements and allows us to demonstrate high-fidelity coherent operations on an optical qubit after only Doppler cooling. With 3D printing technology, the design freedom is drastically expanded without sacrificing scalability and precision so that ion trap geometries can be optimized for higher performance and better functionality.

Temporally multiplexed ion-photon quantum interface via fast ion-chain transport

You, B.*, Wu, Q.*, Miron, D., Ke, W., Monga, I., Saglamyurek, E. & Haeffner, H. Phys. Rev. Appl. 26, 014101 (2026). arXiv:2405.10501.

High-rate remote entanglement between photon and matter-based qubits is essential for distributed quantum information processing. A key technique to increase the modest entangling rates of existing long-distance quantum networking approaches is multiplexing. Here, we demonstrate a temporally multiplexed ion-photon interface via rapid transport of a chain of nine calcium ions across 74 µm within 86 µs. The non-classical nature of the multiplexed photons is verified by measuring the second-order correlation function with an average value of g⁽²⁾(0) = 0.060(13). This indicates low crosstalk of about 1% between the multiplexed modes, which can be reduced further once coupling of the photons to a single-mode fiber is incorporated. In addition, we characterize the motional degree-of-freedom of the ion crystal after transport and find that it is coherently excited to n̄_α ≈ 110 for the center-of-mass mode. Our proof-of-principle implementation paves the way for large-scale quantum networking with trapped ions, but highlights some challenges that must be overcome.

Fig. 7 | Schematics of multiplexed ion-photon interface. (a) A nine-ion chain is confined in an RF Paul trap. Controlling DC endcap voltages allows for ion transport. A beam of 397 nm and 866 nm light illuminating all ions is used for Doppler cooling. An objective collects the 397 nm single photons and guides them to a 50/50 beamsplitter, followed by a photomultiplier tube on each exit port for photon detection. An 866 nm beam counter-propagates with the single photons and addresses individual ions. (b), (c) Excitation scheme and pulse sequence for the 397 nm single-photon generation. First, a global 397 nm beam prepares the ions in the 3²D_3/2 state. Then, the 866 nm addressing beam (resonant with 3²D_3/2 ↔ 4²P_1/2) is stroboscopically switched on when the target ion is in the interaction zone to extract photons from the target ion. Figure 1 of arXiv:2405.10501 (CC BY 4.0).

Momenzadeh, M., Sun, K., Wu, Q., You, B., Tang, Y.-L., Häffner, H. & Shcherbakov, M. R. npj Nanophotonics 3, 3 (2026). arXiv:2505.08997.

Trapped-ion quantum computing requires precise optical control for individual qubit manipulation. However, conventional free-space optics face challenges in alignment stability and scalability as the number of qubits increases. Integrated photonics offers a promising alternative, providing miniaturized optical systems on a chip. Here, we propose a design for a multimode photonic circuit integrated with a surface-electrode ion trap capable of targeted and reconfigurable light delivery. Three closely positioned ions can be addressed using a focusing grating coupler that emits multimode light through electrode openings to ions trapped 80 µm above the chip. Simulations show that the couplers achieve diffraction-limited spot with a 4.3 µm beam waist

along the trap axis and 2.2 µm perpendicular to the trap axis. Controlled interference of the TE₁₀ and TE₂₀ modes results in crosstalk of –20 dB to –30 dB at ion separations of 5-8 µm when addressing ions individually, and down to –60 dB when two of the three ions are addressed simultaneously. Additionally, the higher-order TE modes can offer a novel mechanism for driving spin-motion coupling transitions, potentially enabling alternative approaches to quantum gates and simulations. The proposed integrated platform offers a viable path for constructing large-scale trapped-ion systems, leveraging the benefits of nanophotonic design for precise and reliable ion manipulation.

Trapped ions provide a high-fidelity platform for quantum information processing, yet delivery of multiple, distinct wavelengths across large networks of interaction zones remains a bottleneck. Conventional free-space light delivery lacks scalability, while on-chip grating couplers suffer from narrow operational bandwidth that increases circuit footprint and optical interfacing complexity. Here we show a broadband photonic integrated circuit capable of addressing individual ions. The circuit combines a planar waveguide lens with a micromirror fabricated using two-photon polymerization at wafer

scale. This implementation can address three individual ions from λ = 405 – 880 nm with -27 dB average intensity crosstalk at 5 μm pitch. We trap Ca and Ba ions above such devices, characterize optical crosstalk with barium ions, and demonstrate individual repumping of calcium ions. This monolithic photonic architecture brings broadband addressing in an on-chip modality to trapped-ion technology. More generally, integrating additive manufacturing into quantum devices is poised to unlock expanded design space for implementing novel quantum architectures.

References

1. Li, X.*, Liu, Y.*, Chen, W.*, You, B.* et al. SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks. In NeurIPS (2026); preprint at arXiv:2602.12670.

2. Li, X.*, Choe, K. W.*, Liu, Y., Chen, X., Tao, C., You, B. et al. ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces. In COLM (2026); preprint at arXiv:2604.05172.

3. Zheng, S., Di, Z., Liu, Y., …, You, B. et al. BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure. Preprint at arXiv:2609.11028 (2026).

4. Xu, S.*, Xia, X.*, Yu, Q., Parakh, A., Khan, S., Megidish, E., You, B. et al. 3D-printed micro ion trap technology for quantum information applications. Nature 645, 362–368 (2025).

5. Broz, J., You, B., Khan, S., Häffner, H., Kaplan, D. E. & Rajendran, S. Test of Causal Nonlinear Quantum Mechanics by Ramsey Interferometry with a Trapped Ion. Phys. Rev. Lett. 130, 200201 (2023).

6. You, B.*, Wu, Q.*, Miron, D., Ke, W., Monga, I., Saglamyurek, E. & Haeffner, H. Temporally multiplexed ion-photon quantum interface via fast ion-chain transport. Phys. Rev. Appl. 26, 014101 (2026).

7. Momenzadeh, M., Sun, K., Wu, Q., You, B., Tang, Y.-L., Häffner, H. & Shcherbakov, M. R. Individual trapped-ion addressing with adjoint-optimized multimode photonic circuits. npj Nanophotonics 3, 3 (2026).

8. Klawson, D.*, Zhi, Y.*, You, B.*, Bareian, M.*, Mossman, E.* et al. A broadband, individually addressing two- and three-dimensional photonic integrated circuit for trapped-ion qubit control. Preprint at arXiv:2607.25062 (2026).