arXiv subjects

Jisu Bae

Publications and source records attributed to Jisu Bae.

4 recordsLinked to original sources

arXiv2026

Solar Open 2 Technical Report

We present Solar Open 2, a 250B-A15B Mixture-of-Experts language model built for long-horizon agentic tasks, scaled up from Solar Open 1 (Solar Open 100B). To hold entire agent trajectories in a single context, Solar Open 2 reaches a 1M-token window through a hybrid attention stack that interleaves one softmax layer among every three linear-attention layers, using no positional encoding and a gated delta rule extended to negative eigenvalues. To train at this scale under a fixed compute budget, we make training efficient in two ways: a stronger starting point, and higher-value data. For the starting point, we initialize Solar Open 2 from Solar Open 1, transferring the 5.69B-parameter shared skeleton that survives the architectural change and learning everything else through full pre-training. For the data, we curate for value per token: quality- and rarity-aware data curation and mixture-ratio optimization refine a 20T pool into a 10T mixture that, at equal token budget, outperforms the Solar Open 1 recipe. To build its agent skills, we train twelve domain specialists across purpose-built scenarios, then consolidate them into a single model by Multi-teacher On-Policy Distillation (MOPD). Against comparably sized open-weight models on English benchmarks, Solar Open 2 leads on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic suite, and stays competitive with the strongest (DeepSeek-V4-Flash and MiMo-V2.5) elsewhere. On Korean benchmarks, Solar Open 2 records the highest average of any model compared, including fast-tier closed APIs, and on Ko-GDPval, an in-house Korean officework-agent benchmark, it is competitive with DeepSeek-V4-Pro (1.6T) at less than a sixth of its size.

cs.CL↗

arXiv2025

Towards Fully-Automated Materials Discovery via Large-Scale Synthesis Dataset and Expert-Level LLM-as-a-Judge

Materials synthesis is vital for innovations such as energy storage, catalysis, electronics, and biomedical devices. Yet, the process relies heavily on empirical, trial-and-error methods guided by expert intuition. Our work aims to support the materials science community by providing a practical, data-driven resource. We have curated a comprehensive dataset of 17K expert-verified synthesis recipes from open-access literature, which forms the basis of our newly developed benchmark, AlchemyBench. AlchemyBench offers an end-to-end framework that supports research in large language models applied to synthesis prediction. It encompasses key tasks, including raw materials and equipment prediction, synthesis procedure generation, and characterization outcome forecasting. We propose an LLM-as-a-Judge framework that leverages large language models for automated evaluation, demonstrating strong statistical agreement with expert assessments. Overall, our contributions offer a supportive foundation for exploring the capabilities of LLMs in predicting and guiding materials synthesis, ultimately paving the way for more efficient experimental design and accelerated innovation in materials science.

cs.CL↗

arXiv2018

Efficient Beam Training and Channel Estimation for Millimeter Wave Communications Under Mobility

In this paper, we propose an efficient beam training technique for millimeter-wave (mmWave) communications. When some mobile users are under high mobility, the beam training should be performed frequently to ensure the accurate acquisition of the channel state information. In order to reduce the resource overhead caused by frequent beam training, we introduce a dedicated beam training strategy which sends the training beams separately to a specific high mobility user (called a target user) without changing the periodicity of the conventional beam training. The dedicated beam training requires small amount of resources since the training beams can be optimized for the target user. In order to satisfy the performance requirement with low training overhead, we propose the optimal training beam selection strategy which finds the best beamforming vectors yielding the lowest channel estimation error based on the target user's probabilistic channel information. Such dedicated beam training is combined with the greedy channel estimation algorithm that accounts for sparse characteristics and temporal dynamics of the target user's channel. Our numerical evaluation demonstrates that the proposed scheme can maintain good channel estimation performance with significantly less training overhead compared to the conventional beam training protocols.

cs.IT↗

arXiv2017

New Beam Tracking Technique for Millimeter Wave-band Communications

In this paper, we propose an efficient beam tracking method for mobility scenario in mmWave-band communications. When the position of the mobile changes in mobility scenario, the base-station needs to perform beam training frequently to track the time-varying channel, thereby spending significant resources for training beams. In order to reduce the training overhead, we propose a new beam training approach called "beam tracking" which exploits the continuous nature of time varying angle of departure (AoD) for beam selection. We show that transmission of only two training beams is enough to track the time-varying AoD at good accuracy. We derive the optimal selection of beam pair which minimizes Cramer-Rao Lower Bound (CRLB) for AoD estimation averaged over statistical distribution of the AoD. Our numerical results demonstrate that the proposed beam tracking scheme produces better AoD estimation than the conventional beam training protocol with less training overhead.

cs.IT↗