Xinjie Chen, Minpeng Liao, Guoxin Chen, Chengxi Li, Biao Fu, Kai Fan, Xinggao Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

As oil and gas explorations progressively advance towards deeper and more complex geological formations, the imperative for precise characterization of subsurface structures has become increasingly prominent. The efficacy of noise suppression is a critical determinant for the quality of subsequent inversion and imaging processes. In recent years, deep learning methodologies have garnered significant attention and widespread application in seismic denoising, primarily due to their inherent data-driven advantages. While conventional deep learning implementations have achieved notable denoising performance, they are confronted with inherent limitations, including incomplete noise reduction and potential signal degradation. To address these challenges, this study proposes an innovative multi-round SCU-Net (MR-SCU) denoising approach. The MR-SCU methodology based on SCU-Net employs noise as labeled data to generate an initial denoised outcome in the first round. Denoising results are used as input while utilizing the residuals between the labeled and predicted data as labels for subsequent denoising round. Multiple rounds are iteratively repeated to achieve more thorough denoising effect while preserving effective signals from being compromised. The incorporation of SSIM (Structural Similarity Index Measure) as the loss function further enhances the method’s precision in detail-oriented denoising tasks. Numerical experiments conducted on synthetic data and field data acquired from a specific region in western China substantiate the efficacy of the MR-SCU, demonstrating its capability to deliver superior denoising performance while optimally preserve valuable seismic information.

Despite the remarkable capabilities of large language models (LLMs) in various reasoning tasks, they still struggle with table reasoning tasks, particularly in maintaining consistency throughout multi-step reasoning processes.While existing approaches have explored various decomposition strategies, they often lack effective mechanisms to identify and correct errors in intermediate reasoning steps, leading to cascading error propagation.To address these issues, we propose Table-Critic, a novel multi-agent framework that facilitates collaborative criticism and iterative refinement of the reasoning process until convergence to correct solutions.Our framework consists of four specialized agents: a Judge for error identification, a Critic for comprehensive critiques, a Refiner for process improvement, and a Curator for pattern distillation.To effectively deal with diverse and unpredictable error types, we introduce a self-evolving template tree that systematically accumulates critique knowledge through experience-driven learning and guides future reflections.Extensive experiments have demonstrated that Table-Critic achieves substantial improvements over existing methods, achieving superior accuracy and error correction rates while maintaining computational efficiency and lower solution degradation rate.The code is available at https: //github.com/

Retrieval-Augmented Generation (RAG) systems have emerged as a pivotal methodology for enhancing Large Language Models (LLMs) through the dynamic integration of external knowledge.To further improve RAG's flexibility, Agentic RAG introduces autonomous agents into the workflow.However, Agentic RAG faces several challenges: (1) the success of each step depends on both high-quality planning and accurate search, (2) the lack of supervision for intermediate reasoning steps, and (3) the exponentially large candidate space for planning and searching.To address these challenges, we propose DecoupleSearch, a novel framework that decouples planning and search processes using dual value models, enabling independent optimization of plan reasoning and search grounding.Our approach constructs a reasoning tree, where each node represents planning and search steps.We leverage Monte Carlo Tree Search to assess the quality of each step.During inference, Hierarchical Beam Search iteratively refines planning and search candidates with dual value models.Extensive experiments across policy models of varying parameter sizes, demonstrate the effectiveness of our method.

Tool learning enables large language models (LLMs) to interact with external tools and APIs, greatly expanding the application scope of LLMs. However, due to the dynamic nature of external environments, these tools and APIs may become outdated over time, preventing LLMs from correctly invoking tools. Existing research primarily focuses on static environments and overlooks this issue, limiting the adaptability of LLMs in real-world applications. In this paper, we propose ToolEVO, a novel framework designed to enhance the adaptive and reflective capabilities of LLMs against tool variability. By leveraging Monte Carlo Tree Search, ToolEVO facilitates active exploration and interaction of LLMs within dynamic environments, allowing for autonomous self-reflection and self-updating of tool usage based on environmental feedback. Additionally, we introduce ToolQA-D, a benchmark specifically designed to evaluate the impact of tool variability. Extensive experiments demonstrate the effectiveness and stability of our approach, highlighting the importance of adaptability to tool variability for effective tool learning.

Retrieval-augmented generation (RAG) systems face a fundamental challenge in aligning independently developed retrievers and large language models (LLMs). Existing approaches typically involve modifying either component or introducing simple intermediate modules, resulting in practical limitations and sub-optimal performance. Inspired by human search behavior—typically involving a back-and-forth process of proposing search queries and reviewing documents, we propose C-3PO, a proxy-centric framework that facilitates communication between retrievers and LLMs through a lightweight multi-agent system. Our framework implements three specialized agents that collaboratively optimize the entire RAG pipeline without altering the retriever and LLMs. These agents work together to assess the need for retrieval, generate effective queries, and select information suitable for the LLMs. To enable effective multi-agent coordination, we develop a tree-structured rollout approach for reward credit assignment in reinforcement learning. Extensive experiments in both in-domain and out-of-distribution scenarios demonstrate that C-3PO significantly enhances RAG performance while maintaining plug-and-play flexibility and superior generalization capabilities.

Elucidating the reasoning process with structured explanations from question to answer is crucial, as it significantly enhances the interpretability, traceability, and trustworthiness of question-answering (QA) systems.However, structured explanations demand models to perform intricately structured reasoning, which poses great challenges.Most existing methods focus on single-step reasoning through supervised learning, ignoring logical dependencies between steps.Moreover, existing reinforcement learning (RL) based methods overlook the structured relationships, underutilizing the potential of RL in structured reasoning.In this paper, we propose SEER, a novel method that maximizes a structure-based return to facilitate structured reasoning and explanation.Our proposed structure-based return precisely describes the hierarchical and branching structure inherent in structured reasoning, effectively capturing the intricate relationships between different reasoning steps.In addition, we introduce a fine-grained reward function to meticulously delineate diverse reasoning steps.Extensive experiments show that SEER significantly outperforms state-of-theart methods, achieving an absolute improvement of 6.9% over RL-based methods on En-tailmentBank, a 4.4% average improvement on STREET benchmark, and exhibiting outstanding efficiency and cross-dataset generalization performance.Our code is available at https://github.com/Chen-GX/SEER.

The goal of community search is to provide effective solutions for real-time, high-quality community searches within large networks. In many practical applications, such as event organization and friend recommendations, discovering various community structures within a network is crucial for users. However, existing community search algorithms rarely address issues within temporal graphs, and those that do often have two main limitations: (1) traditional community search methods become inefficient and experience significant increases in computation time when scaled to large graphs; (2) while GNN-based community search methods for temporal graphs offer generalizability, they often focus solely on community connectivity and lack cohesiveness. Therefore, we propose a new model PK-GCN, based on Graph Neural Networks (GNNs), to identify persistent k-core communities in temporal networks. This model can handle dynamic changes in temporal graphs and identify communities that persist over time. Compared to existing community search methods, our model not only finds communities with tighter structures but also allows for dynamic queries based on user input without needing retraining. Specifically, our model constructs features by integrating k-core information from core decomposition, graph features, and query features, resulting in more expressive node representations. Additionally, we designed a flexible dynamic query mechanism that allows users to input time information to query communities. Experiments on multiple datasets demonstrate that our model outperforms other GNN-based community search algorithms in F1-score.

Community search is a personalized community discovery problem designed to identify densely connected subgraphs containing the query node. Recently, community search in heterogeneous information networks (HINs) has received considerable attention. Existing methods typically focus on modeling relationships in HINs through predefined meta-paths or user-specified relational constraints. However, metapath-based methods are primarily designed to identify single-type communities with nodes of the same type rather than multi-type communities involving nodes of different types. Constraint-based methods require users to have a good understanding of community patterns to define a suitable set of relational constraints, which increases the burden on users. In this paper, we propose FCS-HGNN, a novel method for flexibly identifying both single-type and multi-type communities in HINs. Specifically, FCS-HGNN extracts complementary information from different views and dynamically considers the contribution of each relation instead of treating them equally, thereby capturing more fine-grained heterogeneous information. Furthermore, to improve efficiency on large-scale graphs, we further propose LS-FCS-HGNN, which incorporates i) the neighbor sampling strategy to improve training efficiency, and ii) the depth-based heuristic search strategy to improve query efficiency. We conducted extensive experiments to demonstrate the superiority of our proposed methods over state-of-the-art methods, achieving average improvements of 14.3% and 11.1% on single-type and multi-type communities, respectively.

Although recent advancements in large language models (LLMs) have significantly improved their performance on various tasks, they still face challenges with complex and symbolic multi-step reasoning, particularly in mathematical reasoning. To bolster the mathematical reasoning capabilities of LLMs, most existing efforts concentrate on seeking assistance from either domain experts or GPT-4 for high-quality process-supervised data, which is not only expensive but also labor-intensive. In our study, we propose an innovative framework, AlphaMath, that bypasses the need for process annotations (from humans or GPTs) by leveraging Monte Carlo Tree Search (MCTS). This framework focuses on unleashing the potential of a well-pretrained LLM to autonomously enhance its mathematical reasoning. Specifically, we integrate a value model with the LLM, automatically generating both process supervision and step-level evaluation signals in MCTS. Furthermore, we propose an efficient inference strategy—step-level beam search, where the value model is crafted to assist the policy model (i.e., LLM) in navigating more effective reasoning paths, rather than solely relying on prior probabilities. The experimental results on both in-domain and out-of-domain datasets demonstrate that even without GPT-4 or human-annotated process supervision, our AlphaMath framework achieves comparable or superior results to previous state-of-the-art methods.

Large language models (LLMs) have recently showcased remarkable capabilities, spanning a wide range of tasks and applications, including those in the medical domain. Models like GPT-4 excel in medical question answering but may face challenges in the lack of interpretability when handling complex tasks in real clinical settings. We thus introduce the diagnostic reasoning dataset for clinical notes (DiReCT), aiming at evaluating the reasoning ability and interpretability of LLMs compared to human doctors. It contains 511 clinical notes, each meticulously annotated by physicians, detailing the diagnostic reasoning process from observations in a clinical note to the final diagnosis. Additionally, a diagnostic knowledge graph is provided to offer essential knowledge for reasoning, which may not be covered in the training data of existing LLMs. Evaluations of leading LLMs on DiReCT bring out a significant gap between their reasoning ability and that of human doctors, highlighting the critical need for models that can reason effectively in real-world clinical scenarios.

In hydrocarbon exploration, seismic data collected in the field inevitably encounters noise interference, which subsequently affects the data processing and interpretation. Recently, deep learning methods have gained widespread popularity in seismic denoising. Among these methods, the U-Net has shown some potential, but its performance in complex noise suppression needs further improvement due to the limitations of the U-Net structure. Moreover, the majority of existing noise suppression methods primarily focus on synthetic noises with single characteristics, such as Gaussian random noise and linear interference. To devise methods that can effectively suppress more intricate field noise, this paper proposes a novel noise suppression method based on an encoder-decoder architecture called Multiscale Channel Attention Swin Conv UNet. Notably, it enhances the U-Net through the integration of the following two modules: (1) Swin-Conv Block, which replaces the convolution operation of U-Net and integrates the non-local modeling ability of Swin Transformer and the local modeling ability of residual connection convolution layers to achieve multi-dimensional feature extraction; (2) Multiscale Channel Attention Block, which replaces skip connection modules between the encoder and decoder in the U-Net with multi-channel feature fusion to capture more complex channel dependencies. This paper evaluates the proposed algorithm on both synthetic and field seismic data, and compares the results with several established denoising methods. Our algorithm enhances the network’s noise perception capabilities and improves signal-to-noise ratio and structural similarity index measure of seismic data. Finally, a concise discussion on the limitations of our method and potential avenues for enhancement is provided.

Most existing methods that address out-of-distribution (OOD) generalization for node classification on graphs primarily focus on a specific type of data biases, such as label selection bias or structural bias. However, anticipating the type of bias in advance is extremely challenging, and designing models solely for one specific type may not necessarily improve overall generalization performance. Moreover, limited research has focused on the impact of mixed biases, which are more prevalent and demanding in real-world scenarios. To address these limitations, we propose a novel Causality and Independence Enhancement (CIE) framework, applicable to various graph neural networks (GNNs). Our approach estimates causal and spurious features at the node representation level and mitigates the influence of spurious correlations through the backdoor adjustment. Meanwhile, independence constraint is introduced to improve the discriminability and stability of causal and spurious features in complex biased environments. Essentially, CIE eliminates different types of data biases from a unified perspective, without the need to design separate methods for each bias as before. To evaluate the performance under specific types of data biases, mixed biases, and low-resource scenarios, we conducted comprehensive experiments on five publicly available datasets. Experimental results demonstrate that our approach CIE not only significantly enhances the performance of GNNs but outperforms state-of-the-art debiased node classification methods.

Affected by the low-frequency seismic data missing and multiple parameters coupling, elastic full waveform inversion is easy to fall into local minima. This paper attempts to solve the local minima problem from two aspects: low-frequency seismic data reconstruction and wave mode decomposition. First, by introducing the envelope into sparse constrained deconvolution, an envelope-based sparse constrained deconvolution method is proposed to overcome the problem caused by the phase shift and side lobes. However, the resolution of the envelope is insufficient to identify overlapping seismic events, which are generated by velocity models rich in thin layers. Therefore, sparse constrained deconvolution and envelope-based sparse constrained deconvolution are combined, and a dual deconvolution method is proposed: sparse constrained deconvolution is used to improve the resolution of the original seismic data, then envelope-based sparse constrained deconvolution is used to reconstruct high-precision reflection sequence. Convolve the reconstructed reflection sequence with the full-band source wavelet to obtain the full-band seismic data. Secondly, for the multi-parameter coupling problem, we use wave mode decomposition to obtain separated P- and S-wave. Finally, a multi-scale elastic full waveform inversion method based on dual deconvolution and wave mode decomposition is proposed. Numerical experiment results demonstrate the algorithm proposed in the article.

Full waveform inversion (FWI) is, as an optimization problem, strongly nonconvex and is influenced intensely by the cycle skipping issue, especially for multiparameter inversions like the elastic case. When there are strong scattering heterogeneities in the target media, there will be more challenges for the inversion problem. The direct envelope inversion strategy uses the envelope Fréchet derivative to tackle the cycle skipping problem for strong scattering inversion and has been effectively used for the acoustic case. We extend the direct envelope inversion method to the elastic situation in this letter. We derive the elastic envelope Fréchet derivative and show how the strong scattering multiparameter elastic inversion is accomplished under the direct envelope inversion framework. Numerical tests with the SEG/EAGE salt velocity model proved the effectiveness of this method for the strong scattering elastic medium.

Full waveform inversion is often troubled by falling into a local minimum due to cycle-skipping problem when missing low-frequency seismic data. According to the dynamics of seismic waves, the travel-time information will not change due to the variation in the frequency band of seismic data, which is the working mechanism of travel-time inversion. However, the current travel-time inversion method often requires various assumptions for the convenience of calculation, resulting in limited inversion effects. We present a novel velocity building method based on travel-time information. First, we use the sparse-constrained deconvolution (SCD) to convert travel-time information of seismic data into reflection sequences, which greatly reduces the complexity of the travel-time inversion. Then, the phase-independent characteristic of the envelope is introduced into the SCD to deal with the phase shift of the seismic wave. The combination of SCD and envelope greatly improves the reconstruction accuracy of the reflection sequences. Finally, the reconstructed reflection sequences are convolved with the full-band source wavelet to obtain full-band seismic data, and thus, the envelope-based SCD (E-SCD) inversion method is proposed. The results of numerical experiments on the partial basic tracking (BP) model and SEG/EAGE overthrust model verify the performance of the E-SCD inversion method. The limitations of the method and the direction of future development are also briefly discussed.

Building high-fidelity velocity models for salt structures is a valuable and difficult problem in seismic exploration. Acoustic-based full-waveform inversion (FWI) methods usually produce velocity artifacts around high-contrast interfaces due to the generation of converted waves. Therefore, elastic FWI (EFWI) should be used in salt model velocity building. Two problems that restrict EFWI are: lack of low-frequency seismic data and multiparameter coupling. For the first problem, envelope is a good choice because of its ability to reconstruct low-frequency components independent of the frequency range of seismic data. However, envelope is instantaneous energy flow and lacks polarity information, while the elastic waves are vectors. Thus, the direct use of envelope to reconstruct the low-frequency components of elastic waves causes serious artificial artifacts in envelope inversion. Therefore, we introduce signed demodulation and window average function to obtain the multiscale (MS) envelope with polarity, defined as the MS signed envelope to reconstruct low-frequency elastic data. The reconstructed low-frequency data are then used in EFWI, and an elastic MS signed direct envelope inversion algorithm is proposed. For the second problem, wave mode decomposition and hierarchical inversion strategies are integrated into the inversion to eliminate the multiparameter coupling effect. A salt layer model and BP model are used to verify the effectiveness of the algorithm. Finally, the deficiencies in the research of this article and further improvement plans are also discussed.

The exploration capacity of full waveform inversion (FWI) for deep targets is affected by the acquisition geometry, complexity of overlying strata and dip angle of the target, etc. It is of great importance to analyze the relationship between the dip angle of the target reflector and the incident/scattered angle of the wavefields near the target, so as to achieve the illumination distribution and improve the inversion quality accordingly. We propose an angle domain illumination compensated FWI strategy, which utilizes the local resolution function as preconditioning to the gradient of FWI in the local angle domain, in order to improve the inversion capability for deep targets. The local resolution function describes the inversion capacity for the local target in the target dip coordinate, which can be generated from the local illumination matrix that contains the illumination information for the target from different incident and scattering directions. The Marmousi model and the SEG/EAGE salt model are used to show the validity of this method. Results from the numerical experiments prove that the proposed method can effectively improve the inversion performance as well as increase the convergence of FWI.

Strong scattering perturbations like large-scale salt structures in the model parameters make the task of full-waveform inversion more difficult than the weak scattering inversion. The problem becomes even tougher when both the velocity and density are taken into consideration because the tradeoff among the multiparameters further influences the inversion. In order to accomplish effective estimation for both the velocity and density with strong scatterings, we introduce an angle domain direct envelope inversion method with the new Fréchet derivative. The direct envelope inversion method works well on salt structure recovery for the velocity model. However, it may not work well if the density parameter is considered. By introducing angle information into the inversion, the tradeoff between velocity and density can be greatly reduced. Numerical examples using the SEG/EAGE salt model show that by accomplishing the direct envelope inversion in the angle domain, both the velocity and density estimation with strong scattering perturbations are greatly improved, which demonstrates the validity of the proposed method.

For full-waveform inversion (FWI), building the salt structure velocity model is a challenging problem. In the absence of a high-fidelity initial model, low-frequency seismic data are indispensable for accurate reconstruction of the long-wavelength components of salt domes, which are often missing in field data. The envelope isolates the amplitude information from the seismic data and has a stronger linear relationship with the velocity than the seismic data. It is used to enhance the convexity of the objective function in an envelope inversion (EI). However, the properties of the envelope are not effectively utilized in EI due to the lack of their proper understanding. In order to solve the problems that EI has in building the salt structure velocity, such as the cycle-skipping problem, the gradient calculation problem caused by the waveform Fréchet derivative, and the multisolution problem caused by the missing polarity information of the envelope, we propose a multiscale direct signed EI (MSDSEI) based on the direct envelope Fréchet derivative and multiscale signed envelope. The development process from EI to MSDSEI also inspired us to understand the physical mechanism involved in the low-frequency components of the envelope: the signed envelope can be regarded as a low-pass filter of the reflection sequences in the subsurface. Thus, we use the smoothness and phase-independent properties of the envelope to propose an envelope-based full-band seismic data reconstruction method for multiscale FWI. Finally, the performance of various EI methods is verified on the Sigsbee2A model.