Search arXiv⌕ Search

arXiv subjects

Chenxu Liu

Publications and source records attributed to Chenxu Liu.

At least 19 recordsLinked to original sources

Satellite Streaks in HETDEX: Spectra, Orbital Identification, and Survey Impact

We present a spectroscopic database and analysis of satellite streaks in Public Data Release 1 (PDR1) from the Hobby-Eberly Telescope Dark Energy Experiment (HETDEX), covering 6778 observations and $\sim$90 deg$^2$ of sky from January 2017 to July 2024. We detect 527 streaks across 492 observations, a contamination fraction of 7.3% that rises to monthly values above 20% by 2023--2024, although the streaks mask only 0.09% of the survey area. Of the 527 streaks, 475 (90.1%) are identified with a cataloged object using archival two-line elements: 258 (54%) lie in low or Earth orbit and 217 (46%) are geostationary or highly elliptical. Only 35% were launched since 2010; the rest are spent rocket bodies, debris, and older spacecraft. This reflects HETDEX's sensitivity to faint objects and its 2017--2024 observing baseline. The Hobby-Eberly Telescope's large 10\,m aperture and HETDEX's $\sim$6\,minute exposures result in similar observed magnitudes for low Earth orbit (LEO) and geostationary Earth orbit (GEO) objects (median $g_{\rm AB}=16.2\pm1.1$ for LEO, $16.7\pm1.1$ for GEO). However, once transit time is corrected for, instantaneous magnitudes differ widely: $g_{\rm inst}=7.4$ for LEO versus $g_{\rm inst}=13.0$ for GEO objects, a difference of 5.6\,mag, while the exposure-averaged magnitudes differ by only 0.5 mag. Normalized to 550 km, the 34 Starlink streaks give $g(550 {\rm km})=5.95$, consistent with the mean $V(550\,{\rm km})=5.92\pm0.04$ reported for VisorSat-design spacecraft. The spectra are consistent with reflected sunlight, with modest, signal-to-noise-driven differences in slope. The catalog, spectra, and code are archived at https://doi.org/10.5281/zenodo.22962479.

astro-ph.IM↗

Early Fault-Tolerant Magic State Synthesis and Spacetime Efficient Analog Rotations with Color Code on a Superconducting Quantum Processor

With recent improvements to syndrome extraction scheduling and decoding, and the availability of complete transversal Clifford operations and lattice surgery, color code is an especially promising platform for quantum processors with local connectivity. The key ingredient for universal computation is thus generating useful magic states with limited resources. Space-Time Efficient Analog Rotations (STAR) and Fault-tolerant magic state synthesis (MSS) are two such methods for magic state generation, targeted for near-term and long-term quantum computing, respectively. In this work, we perform the first implementations of both magic state paradigms on a superconducting quantum processor. We present a STAR-type injection protocol on the color code, with favorable error rate and spacetime costs versus prior protocols on both the color and surface codes. We then implement a MSS factory by performing multi-patch magic state cultivations, interleaved with transversal Clifford operations, and lattice-surgery magic state teleportation. Under post-selection, we achieve a kickback tomography correct sign probability of $99.9859\%$ on cultivated $T_L$ states, $96\%$ tomographic fidelity on logical analog rotation states, and $91\%$ fidelity on synthesized logical magic states consuming 3 cultivated $T_L$ states, 3 transversal Hadamard operations, and real-time feed-forward $S_L$ corrections during magic state teleportation. We compare the efficacy of post-selection, Most-Likely-Error (MLE) decoding, and concatenated-matching decoding throughout, highlighting the potential of end-to-end color code computation.

quant-ph↗

Simultaneous Real-Color Time-Domain Survey of the Coma Field with the Multi-channel Photometric Survey Telescope (Mephisto)

We present the results from a high-cadence, multi-channel time-domain survey targeting the Coma galaxy cluster, utilizing the newly commissioned 1.6-m Multi-channel Photometric Survey Telescope (Mephisto). Mephisto's pioneering mosaic focal-plane design enables strictly simultaneous imaging in three distinct optical bands ($u/v, g/r, i/z$). This unique capability allows for the extraction of true, instantaneous spectral energy distributions (SEDs) of fast-evolving transients. Our survey continuously monitors a 20-square-degree footprint centered on the Coma cluster. By strategically targeting this massive local structure, we exploit cosmic anisotropy to significantly amplify the volumetric discovery rate of stellar explosions. During the initial six-month pilot operation, we successfully detected and tracked nine new supernovae (SNe). Remarkably, our high-cadence, multi-band strategy allowed us to capture ultra-early rising light curves and pre-discovery signals (e.g., SN 2026dyr) hours to days prior to formal discovery alerts. These instantaneous early-phase colors allow us to reconstruct intrinsic SEDs without temporal interpolation, providing reliable constraints on the physical parameters of early supernova (SN) evolution. Furthermore, our observed supernova yield shows good agreement with theoretical forecasts based on local stellar mass and star-formation rate overdensities. These pilot results validate the efficiency of the Mephisto Coma survey, demonstrating its objective capability to systematically explore the local transient population and provide rapid optical localizations for future multi-messenger events.

astro-ph.HE↗

Atmospheric Parameter Determination of RR Lyrae Stars With Mephisto Multi-band Photometry

Phase-resolved atmospheric parameters of pulsating stars, including RR Lyrae stars and Cepheids, are essential for studying stellar structure and pulsation physics and for establishing accurate stellar benchmarks for Galactic structure and assembly history. We present a photometric framework that simultaneously infers Teff, log g, and [Fe/H] from Mephisto multi-epoch, multi-band (u, v, g, r, i, z) photometry. A supervised neural-network forward model is built by training on synthetic colors derived from theoretical stellar atmosphere models. It reproduces the synthetic colors to a precision of 0.01 per cent. The model is applied to observed multi-epoch colors to derive phase-dependent atmospheric parameters. We apply the method to six RRab and four RRc stars in three globular clusters observed with Mephisto. We find typical per-phase uncertainties of about 62 K in effective temperature (Teff), 0.10 dex in surface gravity (log g), and 0.16 dex in metallicity ([Fe/H]). Repeated measurements across the pulsation cycle reduce the typical formal uncertainty of the final [Fe/H] estimates to roughly 0.03 dex. A sensitivity analysis indicates that the five Mephisto colors provide distinct and complementary constraints on the three stellar atmospheric parameters: (u-v) provides robust constraints on log g, (v-g) is informative for estimating [Fe/H], and (u-v), (v-g), (g-r), (r-i), and (i-z) all provide information on Teff. The results demonstrate that Mephisto multi-epoch, multi-band photometry can deliver precise, cost-effective phase-resolved atmospheric parameters for RR Lyrae stars, offering an efficient way to construct large samples of stellar benchmarks for studies of stellar pulsation physics and Galactic structure and evolution.

astro-ph.SR↗

Demonstration of Logical Error Suppression in a Surface Code Memory on a Cloud-Accessible IBM Superconducting Processor

Quantum error correction can suppress logical errors exponentially with code distance when physical noise is below threshold. Realizing this on hardware requires tight coordination among quantum gates, mid-circuit measurements, and qubit coherence. Device variation, noise asymmetry, and correlations across space and time complicate the translation of these capabilities into logical error suppression. Here, we study rotated surface-code memory on an IBM Nighthawk r2 processor. Performance varies across tested processor placements, which we attribute to differences in qubit quality. We optimize the code by combining short readout, fused measurement and reset, and an XX echo on data qubits during ancilla readout. Compared with standard readout without reset or echo, the optimized syndrome extraction circuit reduces the median per-round logical error rate by factors of 3.8 in the X basis and 1.4 in the Z basis. In memory experiments of up to 150 rounds, the distance advantage weakens, alongside long-lived correlations that may reflect a persistent error component. Under our optimization recipe, increasing the protected distance from 3 to 6 while holding the other distance fixed reduces the per-round logical error rate from 0.53% to 0.24% in Z memory and from 0.52% to 0.22% in X memory. The corresponding error-suppression factors under maximum-likelihood decoding are 1.68 and 1.85. These results connect circuit compilation and optimization to surface-code memory performance and limitations, and motivate QPU and QEC circuit codesign for fault-tolerant quantum computing primitives on current cloud hardware.

quant-ph↗

Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents

Agents driven by large language models are increasingly deployed on long-horizon tasks, where they use tools and modify real workspaces. How well they perform depends as much on the harness around the model as on the model itself. Existing benchmarks measure downstream task success under a fixed harness, leaving harness engineering itself unexplored. We propose Doc2Harness, a two-stage benchmark that evaluates long-horizon harness engineering directly. In Stage 1, a build model operating inside a fixed builder harness receives an interface-only contract skeleton and a development document that consolidates behavior across twelve frontier harnesses, and then constructs a complete candidate harness in an isolated container. In Stage 2, the frozen candidate harness runs in containers separate from the construction environment, one per test case, and is driven end to end by a scripted mock model. We score each test case using four observable channels, namely the requests sent to the model, the files on disk, the process exit code, and the command output; positive controls rule out vacuous passes. Across nineteen frontier models, each evaluated with three independent builds on 328 test cases, GPT-6 Astra achieves the highest pass rate at 53.05%, while the weakest model reaches 27.44%, a gap of roughly twenty-six percentage points. Even the strongest model fails nearly half of the test cases, and failures cluster around how cross-cutting functionality is split across modules. To our knowledge, Doc2Harness provides the first deterministic benchmark for measuring long-horizon harness engineering.

cs.SE↗

Star formation powers optical line emission from the CGM

Using integral field spectroscopy, we explore the disk-halo interface, or the inner circumgalactic medium (CGM), of individual galaxies by constructing and analyzing emission-line maps for a large sample (72) of normal, low-redshift galaxies spanning three orders of magnitude in stellar mass and four orders in star formation rate (SFR). We find a steep turnover occurring at $(1-2) R_e$ in the H$α$, [O {\small II}], and [O {\small III}] line emission radial profiles. Beyond this radius, the slope of the line emission radial profiles becomes shallower as the SFR of the central galaxy decreases, which might reflect the strength of the feedback processes. The line emission fluxes at large radius ($(5-10) R_e$ or $\sim (0.1-0.25)r_{\rm vir}$) correlate with the galaxy's SFR, but not with its stellar mass. These findings suggest that ionizing photons escaping from star-forming regions in the central galaxy account for the observed emission line fluxes from the inner CGM, with escape fractions inferred from the [O {\small III}] and [O {\small II}] ratio. Different state-of-the-art theoretical models do not agree on the predicted dependence of cool gas on the SFR of the central galaxies, highlighting the importance of CGM emission line measurements to distinguish between different subgrid models for star formation and feedback processes.

astro-ph.GA↗

REALHOP: Rethinking Multi-Hop Reasoning Evaluation via Behavioral Auditing

Complex questions often require multi-hop reasoning that connects facts distributed across sources or distant regions of a long context through intermediate steps. Benchmarks commonly evaluate this ability with questions built around predefined reasoning chains, treating a correct answer as evidence that the intended composition was used. Yet answer correctness alone leaves open whether success depends on the evidence associated with each intended step: models may instead rely on memorized associations, shorter paths, or partial evidence. We examine this dependence using the Behavioral Necessity Rate (BNR), which measures how often targeted evidence removal prevents answer recovery on initially correct instances. Across five existing benchmarks, panel-mean BNR ranges from 16.6% to 48.9%, exposing a substantial gap between annotated structure and observed dependence. Guided by this diagnosis, we introduce REALHOP, a diagnose-construct-verify framework that rebinds entities, factorizes selected relations, adds complete competing paths, and places evidence at traceable locations. Structural and semantic checks precede freezing; behavioral interventions follow. On 790 paired MuSiQue questions, REALHOP raises panel-mean BNR from 27.4% to 94.4% while retaining high Full accuracy. It also yields high BNR on REALHOP-FRAMES and REALHOP-LONGBENCH. On 216 long-context questions, the matched multiple-choice spread across 16 models grows from 13.9 to 59.2 points and persists under repeated open-ended evaluation. Together, these results show that a conceptually coherent chain and a correct final answer do not by themselves establish multi-hop reasoning. Verifying that success depends on every intended hop is therefore as fundamental to multi-hop evaluation as measuring answer accuracy itself.

cs.AI↗

Adaptive Dissipative State Preparation through Reinforcement Learning

Dissipative algorithms approach the problem of ground state preparation by mimicking the natural thermalization of a quantum system in contact with a large, low-temperature thermal environment. The environment can be efficiently simulated by a single ancilla qubit with a variable energy gap that is repeatedly coupled to the system qubits to generate a dissipative channel, and then reset after each interaction. Here we present an adaptive implementation of the dissipative algorithm, RL-Adapt, that uses single-shot reinforcement learning to optimize the selection of ancilla frequency and system-bath interaction operator to maximize energy dissipation without relying on a priori knowledge of the system spectrum. The adaptive implementation results in significantly reduced convergence times and can successfully find the ground state even for non-ideal operator pools that fail to converge using non-adaptive, uniform random operator and ancilla frequency selection.

quant-ph↗

Decoder Model Compatibility Provides Information beyond the Logical Gap under Drifting and Correlated Quantum Noise

Quantum error-correcting decoders provide both corrections and confidence estimates for the logical-error risk of individual measurement records. Such confidence can guide postselection, but it depends on the accuracy of the decoder noise model. When the physical device drifts away from that model, confidence measures such as the logical gap can become miscalibrated. We show that this mismatch can itself provide useful information. On IBM quantum hardware, syndrome statistics drift across consecutive acquisition windows, and the resulting detector histories differ substantially from those predicted by the stationary detector error model used for decoding. We observe a similar hardware-model mismatch in independent Google surface-code memory data. We exploit this effect by assigning additional rejection weight to detector histories that are unlikely under the assumed model. For minimum-weight perfect matching (MWPM), the minimum weight $W$ provides an inexpensive proxy for syndrome surprisal. On held-out IBM repetition-code memories, likelihood-aware postselection reduces the retained logical-error probability (LEP) by $10.6\%$ at $d=7$ and $28.8\%$ at $d=9$, both at $15\%$ rejection. On Google ten-round $d=5$ XZZX surface-code memory data, it reduces retained LEP by $75.4\%$ at $90\%$ rejection. Applied to Google's magic-state cultivation experiment with the Tesseract decoder, it reduces tomography infidelity by $34.7\%$ with $30\%$ additional grafting-stage rejection, or increases accepted yield by $30.4\%$ at a target fidelity of $97.0\%$. Burst-noise simulations show that the advantage persists as code distance and syndrome-extraction volume increase. The method reuses quantities already produced during logical-gap decoding and requires only an additional fitting step.

quant-ph↗

ResTest: Targeted Coverage of Residual Not-Covered Code Using Large Language Models for Web GUI Testing

Automated web GUI testing (AWGT) approaches explore web applications through GUI actions to achieve code coverage. However, existing approaches, whether random-based, model-based, or reinforcement-learning-based, often struggle to generate continuous and semantically meaningful action sequences for testing complex functionalities, limiting their achievable code coverage. Recent LLM-based approaches partially alleviate this problem but still fall short due to limited capability in inferring testable functionalities and low success rates in executing tasks on complex web applications. In this paper, we propose ResTest, a complementary approach that uses coverage-report-guided LLMs to target residual not-covered code left by existing AWGT approaches. ResTest first runs an existing AWGT approach to broadly explore the application while constructing a state transition graph with summarized information. Once coverage plateaus, ResTest employs an LLM to infer not-covered functionalities based on the state transition graph and coverage report, and then utilizes a specialized LLM-driven GUI agent to execute these functionalities in a targeted manner. Our evaluation on ten open-source web applications shows that ResTest improves three categories of AWGT approaches by 17.52\% to 24.36\% in average code coverage. Ablation studies further confirm the sustained coverage improvement capability and the contribution of each component.

cs.SE↗

Hy-MultiTurn: A Six-Dimensional Benchmark for Deep Multi-Turn Dialogue Understanding

Long-running multi-turn interactions with chatbots and agents are now common, and a correct response often depends on remembering earlier details, tracking later revisions, identifying intended objects or referents, and withholding action when required conditions are unmet. Existing multi-turn benchmarks typically cover short exchanges and do not fully evaluate these capabilities in long multi-turn interactions, particularly in Chinese, while offering limited insight into how and why models fail. To address these limitations, we analyze real chatbot failures to identify six recurring mechanisms and use them to define six controlled evaluation modes in Hy-MultiTurn, a Chinese benchmark for deep multi-turn dialogue understanding. The six modes evaluate constraint memory, precise execution, constraint synthesis, object localization, action suppression, and reference resolution. Across the six modes, we construct 209 controlled tasks spanning 12-76 turns, with dialogue length, irrelevant-topic distraction, and colloquial phrasing adding further difficulty. Evaluation of 22 frontier model configurations shows that Hy-MultiTurn is broadly challenging, as even GPT-5.5, the strongest overall configuration, satisfies all requirements in only 41.1 percent of responses and no model performs best in all six modes.

cs.CL↗

WebCraftBench: Evaluating Web Application Generation from a Software Testing Perspective

Human evaluation provides a direct measure of the quality of LLM-generated web applications. However, fitting human judgments through automated evaluation remains challenging. Static benchmarks can credit functionality that exists in source code but is unreachable at runtime. Interactive benchmarks exercise the application, yet incomplete exploration can cause them to miss implemented functionality and confound application defects with agent execution failures. To address these limitations, we propose WebCraftBench, an interactive benchmark for evaluating web application generation from a software testing perspective. WebCraftBench instruments each generated application and uses code coverage to guide an agent in exploring its functionality through user-simulated interactions. It then abstracts the interaction trace into a state-transition graph and evaluates the application along three dimensions: visual aesthetics, usability, and requirement alignment. By separating exploration from scoring, WebCraftBench collects runtime evidence without constraining exploration to predefined acceptance criteria. WebCraftBench comprises 369 real-world user requirements and 5,088 acceptance criteria. Evaluation of 17 frontier LLMs reveals distinct strengths across the three dimensions, with no model leading on every dimension. On 197 validated sessions sampled from an internal arena, WebCraftBench achieves 85.3\% agreement with human preferences, with agreement generally increasing as the score difference between paired applications grows. Further experiments show that coverage guidance improves exploration coverage and the model rankings remain stable when the judge model is replaced.

cs.SE↗

Population III Host Candidates at $z\sim2$: Strong He II $\lambda1640$ and Absent UV Metal Lines in HETDEX Ly$α$ Emitters

Population III (Pop III) stars are expected to produce extremely hard ionizing spectra, yet direct evidence for their signature--strong narrow He II $\lambda1640$ emission with weak or absent metal lines--remains elusive. Although Pop III formation peaks at $z\gtrsim10$, models predict that nearly pristine gas pockets can survive to later times, making intermediate-redshift searches a probe of metal mixing across cosmic time. We search 109,545 high-confidence Ly$α$-emitting galaxies at $1.9<z<2.3$ in the Hobby--Eberly Telescope Dark Energy Experiment's database for systems with strong He II $\lambda1640$ and no detected UV metal lines, recovering eight candidates and a first-order, tentative comoving number density of $\sim30$ Gpc$^{-3}$. The eight-object stack yields a rest-frame He II equivalent width of $28.7\pm6.7$ Angstrom, He II/Ly$α=0.354\pm0.094$, and a resolution-corrected He II FWHM of $449\pm105$ km s$^{-1}$. We find no significant N V, C IV, or O III] emission; using the He II width to set the scale, we derive $3σ$ upper limits of N V/He II $<0.167$, C IV/He II $<0.128$, and O III]/He II $<0.140$. This combination of strong He II, elevated He II/Ly$α$, moderate line width, and weak metal lines is difficult to reproduce with metal-enriched stellar populations, AGN narrow-line regions, shocks, or classical Wolf--Rayet features. We interpret the sample as candidates for very metal-poor or Pop III-like ionizing sources at $z\sim2$. Definitive confirmation requires deeper rest-UV spectroscopy to measure He II and metal-line limits in individual objects, followed by rest-optical spectroscopy of H$β$, [O III] $\lambda4959,\lambda5007$, [O II] $\lambda3727$, and H$α$ to constrain gas-phase metallicity, ionization conditions, and AGN activity.

astro-ph.GA↗

FTCircuitBench: A Benchmark Suite for Fault-Tolerant Quantum Compilation and Architecture

Realizing large-scale quantum advantage is expected to require quantum error correction (QEC), making the compilation and optimization of logical operations a critical area of research. Logical computation imposes distinct constraints and operational paradigms that differ from those of the Noisy Intermediate-Scale Quantum (NISQ) regime, motivating the continued evolution of compilation tools. Given the complexity of this emerging stack, where factors such as gate decomposition precision and computational models must be co-designed, standardized benchmarks and toolkits are valuable for evaluating progress. To support this need, we introduce FTCircuitBench, which serves as: (1) a benchmark suite of impactful quantum algorithms, featuring pre-compiled instances in both Clifford+T and Pauli Based Computation models; (2) a modular end-to-end pipeline allowing users to compile and decompose algorithms for various fault-tolerant architectures, supporting both prebuilt and custom optimization passes; and (3) a toolkit for evaluating the impact of algorithms and optimization across the full compilation stack, providing detailed numerical analysis at each stage. FTCircuitBench is fully open-sourced and maintained on Github.

quant-ph↗

Rapid Variability and Broadband Spectral Modeling in the Flaring Activity of BL Lacertae

We report a multi-wavelength study of two flaring episodes of the blazar BL Lacertae during MJD 60500-60800 (9 July 2024 - 5 May 2025). The source reached a daily-averaged $γ$-ray flux of $(1.03 \pm 0.05) \times 10^{-5} \, \mathrm{ph \, cm^{-2} \, s^{-1}}$ ($E > 100$ MeV) on MJD 60588 (5 October 2024). Using orbit-binned data from the Large Area Telescope (LAT) onboard the \textit{Fermi Gamma-ray Space Telescope}, we identify a minimum flux halving timescale of $τ= 1.33 \pm 0.29$ hr. This constrains the upper limit on the $γ$-ray emitting region size to $R \le 2.0 \times 10^{15}$ cm, as well as its distance from the central supermassive black hole to $R_\mathrm{H} \le 5.9 \times 10^{16}$ cm, assuming a Doppler factor of $δ= 14.8$ derived from the spectral energy distribution (SED) modeling. We find tentative evidence for sub-minute $γ$-ray variability with a minimum doubling time of $0.7 \pm 0.2$ min ($p$-value = 0.03). This may originate from an extremely compact region with a size of $R \le 1.8 \times 10^{13}$ cm, suggesting that the emission arises from magnetohydrodynamic substructures, such as plasmoids within a magnetic reconnection zone. Spectral analysis reveals a significant ``softer-when-brighter'' trend ($r = 0.96, p = 4.5 \times 10^{-4}$) during the minute-scale flare peaks, indicating a complex interplay between particle acceleration and radiative cooling. The SED is reproduced using a one-zone leptonic model, in which synchrotron self-Compton (SSC) and external Compton (EC) scattering effectively account for the high-energy emissions. The reduced magnetic field strengths and hard electron injection spectral indices observed during the flaring states suggest enhanced particle acceleration efficiency, possibly associated with relativistic magnetic reconnection.

astro-ph.HE↗

GameXpert-Bench: How Far Are Coding Agents from Expert Game Development?

Recent large language models (LLMs) can operate as coding agents that build complete games from natural language requests. Game development is especially demanding because program logic, visual and audio content, interfaces, interaction and playability must function together in one executable artifact. Measuring this capability therefore requires evaluation of both game product and the development process. Existing benchmarks often assess the game development capabilities of LLMs by evaluating the final artifact or an isolated development stage. Our analysis of complete human-agent development trajectories identifies three stages that together span the lifecycle of game development with a coding agent: initial game generation, bug diagnosis and repair, and optimization over multiple turns. Therefore, we introduce GameXpert-Bench, which operationalizes the three lifecycle stages as three complementary benchmark tracks. GameGen evaluates complete game creation from a single request in an empty workspace. GameFix evaluates diagnosis and repair when defects are reported or left for the agent to discover. GameOpt evaluates cumulative optimization through request chains seeded by real development trajectories between users and agents. We evaluate each track using live game interaction, deterministic behavioral tests, or final product criteria with regression checks. The suite contains 97 generation tasks across 11 genres; 100 repair tasks from 50 game levels verified by humans, each with 19-27 injected bugs; and 17 optimization chains with six turns and 102 requests. Across the three tracks, current agents are more reliable at producing playable foundations and implementing explicit requirements than at discovering defects, verifying runtime behavior, and preserving functionality across changes.

cs.AI↗

GDPevo: Evaluating Agent Self-Evolution on Real Business Tasks

Agent self-evolution updates an agent's persistent state from prior experience and reuses it to solve related tasks more effectively. Evaluating self-evolution is difficult: existing benchmarks provide limited coverage of economically valuable task domains, do not always design training and test tasks such that test-time gains can be attributed to training experience, and remain vulnerable to data contamination. We present GDPevo, an evolution-native benchmark grounded in GDP-related enterprise workflows, together with the fully automated data pipeline that generates it. Its core mechanism, rule hybridization, decomposes each enterprise workflow into atomic business rules, distributes subsets of these rules across training tasks, and recombines them in held-out test tasks so that test-time gains are attributable. GDPevo spans CRM, ERP, finance, healthcare, legal, and data-centric workflows. Its V1 release contains 120 tasks in 12 groups, with five training and five held-out test tasks per group. Full automation enables the pipeline to expand the suite to 240 tasks in 24 groups (V2) within two days, providing a practical response to contamination. Using GDPevo, we evaluate four agents, each comprising a harness and a model, under four supervision types. Self-evolution consistently improves held-out accuracy by up to 16.44 percentage points. But the best evolved agents remain far below the fully informed oracle ceiling of 91.6%, indicating that the self-evolution ability of current agents remains far from fully realized. We publicly release the pipeline, benchmark, and full evaluation results at https://github.com/Prism-Shadow/GDPevo.

cs.AI↗