Yichuan Ma, Linyang Li, Yongkang Chen, Peiji Li, Xiaozhe Li, Qipeng Guo, Dahua Lin, Kai Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Synthetic high-quality multi-step reasoning data can significantly enhance the performance of large language models on various tasks. However, most existing methods rely on rejection sampling, which generates trajectories independently and suffers from inefficiency and imbalanced sampling across problems of varying difficulty. In this work, we introduce FastMCTS, an innovative data synthesis strategy inspired by Monte Carlo Tree Search. FastMCTS provides a more efficient sampling method for multi-step reasoning data, offering step-level evaluation signals and promoting balanced sampling across problems of different difficulty levels. Experiments on both English and Chinese reasoning datasets demonstrate that FastMCTS generates over 30% more correct reasoning paths compared to rejection sampling as the number of generated tokens scales up. Furthermore, under comparable synthetic data budgets, models trained on FastMCTS-generated data outperform those trained on rejection sampling data by 3.9% across multiple benchmarks. As a lightweight sampling strategy, FastMCTS offers a practical and efficient alternative for synthesizing high-quality reasoning data.

Large Language Models (LLMs) have shown outstanding breakthroughs in code generation. Recent work improves code LLMs by training on synthetic data generated by some powerful LLMs, which can be challenging to scale due to the dependence on a teacher model and high generation costs. In this paper, we focus on synthesizing code data at scale and propose a Case2Code task by exploiting the expressiveness and correctness of programs. Case2Code is an inductive inference task that aims to infer underlying code implementations by observing input-output examples or program behaviors, By incorporating LLMs to generate program inputs, and executing the program with these inputs to obtain the program outputs, we can synthesize diverse and high-quality Case2Code data at scale for training and evaluating code LLMs. Experimental results show that case-to-code induction is challenging for current representative LLMs if they are untrained. Models trained with Case2Code improve performance not only on distribution case-to-code induction but also various coding-generation tasks, demonstrating the great potential of large-scale synthetic data and inductive learning.

Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks, yet code generation remains a major challenge.Despite the abundant sources of code data, constructing high-quality training datasets at scale poses a significant challenge.Pre-training code data typically suffers from inconsistent data quality issues.Conversely, instruction-based methods which use a high-quality subset as seed samples suffer from limited task diversity.In this paper, we introduce UnitCoder, which directly supervises pre-training data quality through automatically generated unit tests, while ensuring the correctness via an iterative fix and refine flow.Code synthesized by Unit-Coder benefits from both the diversity of pretraining corpora and the high quality ensured by unit test supervision.Our experiments demonstrate that models fine-tuned on our synthetic dataset exhibit consistent performance improvements.Our work presents a scalable approach that leverages model-generated unit tests to guide the synthesis of high-quality code data from pre-training corpora, demonstrating the potential for producing diverse and high-quality post-training data at scale.All code and data will be released 1 .

Large language models (LLMs) have demonstrated exceptional performance in reasoning tasks such as mathematics and coding, matching or surpassing human capabilities. However, these impressive reasoning abilities face significant challenges in specialized domains. Taking Go as an example, although AlphaGo has established the high performance ceiling of AI systems in Go, mainstream LLMs still struggle to reach even beginner-level proficiency, let alone perform natural language reasoning. This performance gap between general-purpose LLMs and domain experts is significantly limiting the application of LLMs on a wider range of domain-specific tasks. In this work, we aim to bridge the divide between LLMs' general reasoning capabilities and expert knowledge in domain-specific tasks. We perform mixed fine-tuning with structured Go expertise and general long Chain-of-Thought (CoT) reasoning data as a cold start, followed by reinforcement learning to integrate expert knowledge in Go with general reasoning capabilities. Through this methodology, we present LoGos, a powerful LLM that not only maintains outstanding general reasoning abilities, but also conducts Go gameplay in natural language, demonstrating effective strategic reasoning and accurate next-move prediction. LoGos achieves performance comparable to human professional players, substantially surpassing all existing LLMs. Through this work, we aim to contribute insights on applying general LLM reasoning capabilities to specialized domains. We will release the first large-scale Go dataset for LLM training, the first LLM Go evaluation benchmark, and the first general LLM that reaches human expert-level performance in Go.

Sloping terrain distorts the sun-target-sensor geometry, resulting in biases of the optical reflectance measured by remote sensors relative to flat situations. Performing topographic correction (TC) is, therefore, deemed mandatory to foster the full exploitation of satellite images worldwide to support various applications in mountainous regions. Various TC models have already been proposed and developed, while most of them were previously evaluated at local or regional scales using a few images with various evaluation criteria. Therefore, a systematic and comprehensive assessment has yet to be done on these TC models in the global mountainous regions. In the present study, 10523 Landsat 8 OLI images filtered by land cover types and seasons sampled in the global mountains are corrected by ten popular TC models (SE, b correction, VECA, CC, SCS, DS, SCS+C, PLC, Minnaert, and Minnaert+SCS) with a unified evaluation criterion on the Google Earth Engine platform. The outcomes are that: (1) global TC effects on Landsat 8 OLI images generally increase with sun zenith angles and latitudes; (2) six models (SE, b correction, CC, VECA, Minnaert, and Minnaert+SCS) show good adaptability among the ten models for the global mountainous placing a disregard to land cover types and seasons; (3) considering permanent snow and ice, needle-leaved forests in winter, and null values might appear in b correction, SE is deemed to be with the most global adaptability. This study pioneers an evaluation of fashionable TC models concerning mountainous regions worldwide and will be useful for applying TC to Landsat images for the benefit of making global TC products in the future and a fair inter-comparison of OLI surface reflectance measured in various mountainous areas of the globe.

It is well known that complex topography can affect satellite observations, leading to substantial uncertainties in surface parameter estimation when topographic effects are ignored. However, most existing studies have focused on high-resolution satellite data (e.g., < 100 m resolution), while the impacts of topography on the moderate-resolution satellite data (e.g., Moderate Resolution Imaging Spectroradiometer (MODIS)) observation, product generation, and further applications have not been well explored. In this context, we investigated how topography-induced deviations propagate through moderate-resolution satellite observations, product generation, and downstream applications. We examined proxies such as top-of-atmosphere (TOA) reflectance, surface reflectance, land surface temperature (LST), leaf area index (LAI), and gross primary production (GPP), systematically analyzing their topographic effects across representative mountainous regions using multiple satellite datasets and radiative transfer models. Specifically, we conducted the following three tasks: (i) we utilized simultaneous observations from Geostationary Operational Environmental Satellite–16 (GOES-16) and GOES-17, which have differing viewing angles, to evaluate the topographic effects on geostationary satellite data; (ii) we analyzed MODIS-Terra and MODIS-Aqua data, with varying sun and viewing angles, to assess the impact of topography on polar-orbiting satellite products; and (iii) we employed radiative transfer models to gain theoretical insights into how topography influences satellite data across different terrain conditions. Our findings showed that topography induced an average deviation of 7.4% in near-infrared (NIR) band TOA reflectance in concurrent GOES-16 and GOES-17 observations. The surface reflectance and LST had similar deviation patterns as TOA reflectance. For NIR band surface reflectance, topographic effects lead to a maximum error of 0.37 in simulated data and an average of 16.7% deviation in MODIS-based evaluations. Furthermore, topographic impacts on LAI and GPP were found to average 36.0% and 10.4%, respectively, across four 1° × 1° mountainous regions globally. Long-term GPP trend analyses revealed uncertainties of 5.2% in the Alps and 3.8% in the Qinghai-Xizang Plateau, attributable to topographic effects. Our study demonstrates that topographic influences not only affect satellite observations but also propagate through to downstream applications for moderate-resolution data. By quantifying these effects, we underscore the importance of integrating topographic considerations into high-level satellite products over mountainous regions.

The bidirectional reflectance factor (BRF) is a key parameter for understanding the radiative transfer process within cities, influenced by the sun-target-sensor geometry. Simulating urban BRF poses significant challenges due to the complex three-dimension (3-D) structures of urban landscapes. Previous attempts have been facing significant hurdles, either inadequately addressing the complexities of real-world urban structures or consuming too many computing resources. To overcome these challenges, this study introduces an analytical model—urban bidirectional reflectance analytical model (UBRAM), which takes advantage of landscape-level geometric and radiometric parameters incorporating 3-D urban structures to simulate the BRF of neighborhood-scale urban landscapes ($100\times 100- 1000\times 1000$m). UBRAM underwent verification using simulations from the well-known radiative transfer model (discrete anisotropic radiative transfer, DART) under various solar-illumination geometries and diverse urban morphologies. Results demonstrated a good agreement between UBRAM and DART, with an overall${R} ^{2}$ranging from 0.829 to 0.983 and the shape similarity index exceeding 0.99 in most scenes, underscoring the effectiveness of UBRAM. Notably, leveraging a modular design, UBRAM enables independent processing of 3-D urban scenes and radiative transfer simulation, ensuring efficiency and scale applicability. This study establishes UBRAM as a viable model, offering a novel approach for quantifying BRF of neighborhood-scale urban surfaces. Moreover, UBRAM requires a limited number of easily obtainable parameters, facilitating its straightforward adaptation for application in other urban areas. UBRAM facilitates the simulation of radiative transfer processes based on real-world 3-D urban models, thereby enhancing the accuracy of scene rendering and downstream applications such as urban heat islands and environment monitoring.

Downward shortwave radiation (DSR) is critical to many surface processes, and many satellite-derived DSR products have been released. Few studies have validated DSR over mountains where it is highly heterogeneous and so the shortwave flux measured at ground stations does not match kilometer-scale DSR products. To tackle this challenge, we used a high spatial resolution (30 m) daily DSR over Sierra Nevada, Spain for 2008–2015, and a mountainous radiative transfer model to explore how topographic effects impacted the performances of DSR products. Four widely-used satellite products were selected as proxies for our evaluation: (i) MCD18A1 V6.1 (with a spatial resolution of 1 km); (ii) MSG DSR (~ 3.3 km); (iii) GLASS DSR V42 (0.05°); and (iv) BESS DSR (0.05°). There are three main findings under clear skies. Firstly, the product accuracies were slope-dependent, decreasing by 59.8–134.6% with slope ≥ 25° compared to areas with slope < 10°. Secondly, the product accuracies were aspect-dependent, exhibiting a higher degree of overestimation (i.e., average of 27.6 W/m²) on the north side and underestimation (i.e., average of -1.3 W/m²) on the south side. Thirdly, and finally, the product accuracies were time-dependent, exhibiting seasonal variations and pronounced overestimation in summer (i.e., 8.8 to 18.2 W/m²). Moreover, the impact of topography decreased with increasing cloud cover. Our findings can be applied to various mountainous areas due to the same mechanism of how topography influences the DSR estimation. This study corroborates the substantial uncertainties of the current DSR products in mountains and the necessity of incorporating topographic information into DSR estimations.

Surface albedo plays a significant role in Earth’s energy budget and global climate change. The spaceborne remote sensing technique is efficient for deriving and monitoring long-term surface albedo over large regions. Numerous satellite surface albedo products have been established and used for climate change research. However, the surface daily mean albedo is not regularly produced from satellite observations, even though it is more critical than instantaneous albedo for calculating daily shortwave radiation budget. Compared to polar-orbiting satellites, the geostationary satellites offer greater potential for mapping daily albedo with more diurnal observations. This study proposes an innovative multisensor combined direct estimation algorithm, which takes advantage of multiple clear-sky observations from new-generation geostationary satellite sensors FengYun-4 Advanced Geostationary Radiation Imager (AGRI) and Himawari-8 Advanced Himawari Imager (AHI) taken during the same day to improve the surface albedo estimation. Compared to albedo estimates derived from a single sensor, the root mean squared error (RMSE) decreases from 0.030 to 0.022 at OzFlux sites and from 0.040 to 0.031 at Heihe sites. Moreover, an information index of top-of-atmosphere (II_TOA) reflectance is proposed to quantify the amount of information that multiangular TOA observations carry in the surface albedo estimation. As a result, combining observations from two sensors increases such information by 20%, as compared to observations from a single sensor. This study demonstrates the combined multisensor direct estimation method has significant potential for improving surface albedo estimation.

Estimation of net shortwave radiation at the top-of-the atmosphere (Rns,TOA) at high spatial and temporal resolutions is essential for studying the Earth’s energy budget and its associated radiative forcing of natural or anthropogenic events on global or regional scales. Existing products typically use broadband sensors with coarse spatial resolution for the estimation. While narrowband sensors offer higher spatial resolution, they have a limited number of daily observations. Traditional estimation methods often necessitate atmospheric products as inputs, while inaccurate cloud and aerosol information can result in substantial estimation errors. Furthermore, geostationary satellites-based products are often developed for specific regions, with limited spatial coverage and varying accuracy due to the diverse range of satellites and algorithms used. To overcome these challenges and obtain globalRns,TOAwith improved spatiotemporal resolution and accuracy, a universal approach was proposed in this study to derive hourly 3-km globalRns,TOA, which takes the advantages from radiative transfer model, machine learning algorithm, and dense observations from five geostationary satellites. OurRns,TOAestimation shows reasonably good agreement with the Earth’s Radiant Energy System (CERES) product, with root mean square errors (RMSE) ranging from 53.43 W/m2to 75.67 W/m2and bias ranging from -12.78 W/m2to -2.01 W/m2on instantaneous scales, and the RMSE on daily scale improved by up to 6.7 W/m2compared to those of the sinusoidal-integrated values. Our generated 3-km dailyRns,TOAexhibits highly consistency of spatial pattern with 1° CERES product at multiple temporal conditions, while providing much more spatial details. Furthermore, we find the diurnal patterns at 3km resolution differ significantly from the sinusoidal patterns at 1°, exhibiting greater variability. The difference in daytimeRns,TOAestimation between the two reaches up to 86W/m2. This significant difference reflects the unreliability of relying solely on sinusoidal pattern and the necessity of high frequency observations to estimate daily values at high spatial resolution. However, increasing the observation frequency beyond a certain point (120-min) yields only limited improvements in the accuracy of daily radiation estimates. Overall, the algorithms proposed in this study are reliable, and can be easily applied to any satellite equipped with MSG (Meteosat Second Generation)-like or more bands. This study demonstrates the feasibility of jointly using multiple geostationary satellites for,Rns,TOAestimation with high spatial and temporal resolutions.

Surface albedo plays a key role in global climate modeling as a factor controlling the energy budget. Satellite observations were utilized to estimate surface albedo at global and regional scales with good precision over flat areas. However, because topography greatly complicates radiative transfer (RT) processes, estimating the albedo of rugged terrain with satellite data remains a challenge. In addition, albedo definitions over sloping terrain differ from that for flat areas. They include horizontal/horizontal sloped surface albedo (HHSA) and inclined/inclined sloped surface albedo (IISA). Methods for retrieving HHSA and IISA in mountains have not been well-explored. Here, we retrieved HHSA and IISA on sloping terrain from Landsat 8 using a direct estimation algorithm. We simulated a dataset of Landsat top-of-atmosphere (TOA) reflectance and surface albedo with discrete anisotropic radiative transfer (DART) model, for variable atmospheric, vegetation, soil, and topography properties. Then, we used artificial neural networks (ANNs) to derive an empirical relationship between TOA reflectance and surface albedo. The accuracy of our method was verified within situmeasurements: root mean squared error (RMSE) and bias equal to 0.029 and −0.010 for HHSA, and 0.023 and −0.001 for IISA, respectively. Several albedo results (HHSA, IISA, values without topographic consideration) were evaluated and compared. HHSA was found similar to albedo without topographic consideration, but IISA, considered as the “true albedo” for sloping terrain, showed large difference from them. This study demonstrated the feasibility of surface albedo estimation from Landsat TOA reflectance directly in rugged terrains and advanced our understanding of energy budget in mountains.

The objective of this study is to investigate the potential of multitemporal remote sensing images for crop classification. Multi-temporal Landsat 8 OLI/TIRS C1 Level-1 images were acquired. The surface reflectance of visible and near infrared bands was used to represent the characteristics of crops. A time series model of surface reflectance was constructed for crop classification. Cloud cover is critical for the accuracy of classification. In order to remove the influence of clouds, the cloud pixels were neglected by setting a constant. Pearson correlation coefficient was used in the time series model of surface reflectance to classify the crop type. Finally, the overall accuracy reaches 78.26% and Kappa reaches 71.33%. Therefore, the method has the operational potential for crop classification even in the special area with cloudy or foggy weather.

This reported study was intended to advance the retrieval of leaf area index (LAI) using synthetic aperture radar (SAR) data. A novel method was proposed by introducing the vegetation coverage into the Water Cloud Model (WCM) to improve the retrieval accuracy of the LAI. LAI is a strong indicator of crop productivity, and vegetation coverage has a strong relationship with the LAI (R2=0.9733), a function can be created to express their relation. Finally, the accuracy in this innovative LAI retrieval method were evaluated. The results showed that the accuracy of estimation was improved greatly (R2was increased to 0.6055 and 0.6422 from 0.3491 and 0.3561 in VH and HH polarization). Thus, the method has operational potential for the LAI retrieval of crop in agriculture regions.