Zhichao Chang | Searcharxiv

arXiv2025

Step-DeepResearch Technical Report

As LLMs shift toward autonomous agents, Deep Research has emerged as a pivotal metric. However, existing academic benchmarks like BrowseComp often fail to meet real-world demands for open-ended research, which requires robust skills in intent recognition, long-horizon decision-making, and cross-source verification. To address this, we introduce Step-DeepResearch, a cost-effective, end-to-end agent. We propose a Data Synthesis Strategy Based on Atomic Capabilities to reinforce planning and report writing, combined with a progressive training path from agentic mid-training to SFT and RL. Enhanced by a Checklist-style Judger, this approach significantly improves robustness. Furthermore, to bridge the evaluation gap in the Chinese domain, we establish ADR-Bench for realistic deep research scenarios. Experimental results show that Step-DeepResearch (32B) scores 61.4% on Scale AI Research Rubrics. On ADR-Bench, it significantly outperforms comparable models and rivals SOTA closed-source models like OpenAI and Gemini DeepResearch. These findings prove that refined training enables medium-sized models to achieve expert-level capabilities at industry-leading cost-efficiency.

cs.CL↗

arXiv2025

Probability distributions for kinetic roughening in the Kardar-Parisi-Zhang growth with long-range temporal and spatial correlations

We investigate numerically the effects of long-range temporal and spatial correlations based on the rescaled distributions of the squared interface width $W^2(L,t)$ and the interface height $h(x,t)$ in the (1+1)-dimensional Kardar-Parisi-Zhang (KPZ) growth system within the early growth regimes. Through extensive numerical simulations, we find that long-range temporally correlated noise could not significantly impact the distribution form of the interface width. Generally, $W^2(L,t)$ obeys approximately lognormal distribution when the temporal correlation exponent $θ\ge 0$. On the other hand, the effects of long-range spatially correlated noise are evidently different from the temporally correlated case. Our results show that, when the spatial correlation exponent $ρ\le 0.20$, the distribution forms of $W^2(L,t)$ approach the lognormal distribution, and when $ρ> 0.20$, the distribution becomes more asymmetric, steep, and fat-tailed, and tends to an unknown distribution form. As a comparison, probability distributions of the interface height are also provided in the temporally and spatially correlated KPZ system, exhibiting quite different characteristics from each other within the whole correlated regimes. For the temporal correlation, the height distributions satisfy Tracy-Widom Gaussian orthogonal ensemble (TW-GOE) when $θ\to 0$, and with increasing $θ$, the height distributions crossover continously to an unknown distribution. However, for the spatial correlation, the height distributions gradually transition from the TW-GOE distribution to the standard Gaussian form.

cond-mat.stat-mech↗