Xin Cao

Large language model (LLM) agents choose which external tool to call from the tools' documentation, and this description matching fails in three measurable ways: similar tools create ambiguity, truncated or anonymized documentation creates degradation, and untrusted catalogs admit adversarial injection. We present Three-Axis Tool Selection (TATS), which answers each condition with a different signal: a contextual bandit that learns online to separate similar tools, schema features that survive the loss of descriptions, and a success-filtered anomaly penalty (ADAB) that downweights candidates unlike anything that has succeeded before. All three run inside one compact eight-feature selector with no LLM call at inference, evaluated axis by axis on BFCL and MetaTool. Under ambiguity, TATS's advantage over description matching widens from 12.6 to 21.4 points as similar candidates grow from 5 to 50; under name-only documentation, schema features keep 62.1% accuracy where description matching falls to 30.2%; under injection, ADAB cuts the selection rate of a malicious tool from 6.9% to 1.7% at a 3.4-point clean-catalog cost. We state when any feature-only defense cannot stop an attacker who knows the scoring rule, confirm it with a relaxed oracle, and show that an unforgeable registry of trusted tools reduces the same attack to about 1% of queries.

Financial question answering is often treated as document question answering, although financial evidence is both multimodal and time-dependent. Semantically equivalent facts expressed in narrative text, tables, page images, or Extensible Business Reporting Language (XBRL) should support consistent answers, whereas a disclosure may support a query only after becoming public. We formalise this combination as crossmodal evidence symmetry under a causal temporal boundary and introduce TempFinRAG, a multimodal temporal retrieval-augmented generation (RAG) framework for point-in-time financial question answering. Given a question, company, and as-of date, the framework enforces the information boundary defined by U.S. Securities and Exchange Commission (SEC) filing availability; aligns page layout, text, table structure, and XBRL facts; retrieves time-valid evidence; executes auditable financial calculations; and generates a cited answer. A verifier checks temporal validity, claim support, numerical consistency, and the need to abstain. We further introduce TempFinQA, a point-in-time evaluation protocol built from public filings and XBRL facts, and evaluate the framework on complementary evidence-grounded, numerical, conversational, and multi-table benchmarks. On TempFinQA, TempFinRAG improves answer accuracy from 66.7% to 78.9% over hybrid RAG while reducing temporal evidence leakage from 10.8% to 1.7% and hallucination from 17.3% to 7.9%. Reliable financial question answering therefore requires consistent treatment across evidence representations and deliberately asymmetric access across time.

Over the past two decades, the Amazon has experienced four severe large-scale droughts (i.e. 2005, 2010, 2015/16 and 2023), leading to drastically reduced water availability, slowed vegetation growth and higher forest mortality. As future droughts are expected to become more frequent and severe, accurately predicting the unprecedentedly low water storage levels and water shortages in advance is crucial. Herein, we developed a new approach to predict terrestrial water storage (TWS) during droughts, based on monthly changes in TWS (ΔTWS) and meteorological variables from 2003 to 2023. The model was trained during non-drought months and assessed during the four droughts when TWS values are well below the range of training data. The ΔTWS-based model excels in predicting drought-month TWS even only using precipitation and incoming solar radiation, with average correlation (R) over 0.9 and RMSE below 50 mm. The model also showed superior skills for predicting drought TWS months lead-time, with the 3-month prediction achieved high performance (R > 0.8, RMSE < 80 mm). We further examined TWS predictions during the large-scale 2023 drought and found that the predicted TWS showed high spatial agreement with observed TWS, with all 1-, 2-, and 3-month lead-times reaching average R values over 0.9. Then we evaluated water deficits in the driest months (September—December) in 2023. The model predicted the affected regions with reasonable accuracy, achieving an average of 72% even at 3-month lead-time. We also analyzed how uncertainty in meteorological inputs affects model performance, revealing higher input uncertainty reduced the model performance. This study presents a reliable approach for estimating and predicting low water storage during severe large-scale droughts, enabling early warnings of water deficits across the Amazon. This study could be generalized to other regions, supporting proactive water resource management, water security policies, ecosystem protection and climate adaptation strategies.

Network alignment (NA) is the task of discovering node correspondences across multiple networks. Although NA methods have achieved remarkable success in a myriad of scenarios, their effectiveness is not without additional information such as prior anchor links and/or node features, which may not always be available due to privacy concerns or access restrictions. To tackle this challenge, we propose Grad-Align+, a novel NA method built upon a recent state-of-the-art NA method, the so-called Grad-Align, that gradually discovers a part of node pairs until all node pairs are found. In designing Grad-Align+, we account for how to augment node features in the sense of performing the NA task and how to design our NA method by maximally exploiting the augmented node features. To achieve this goal, Grad-Align+ consists of three key components: 1) centrality-based node feature augmentation (CNFA), 2) graph neural network (GNN)-aided embedding similarity calculation alongside the augmented node features, and 3) gradual NA with similarity calculation using aligned cross-network neighbor-pairs (ACNs). Through comprehensive experiments, we demonstrate that Grad-Align+ exhibits (a) the superiority over benchmark NA methods, (b) empirical validations as well as our theoretical findings to see the effectiveness of CNFA, (c) the influence of each component, (d) the robustness to network noises, and (e) the computational efficiency.

Large language models (LLMs) are increasingly deployed in sensitive contexts where fairness and inclusivity are critical. Pronoun usage, especially concerning gender-neutral and neopronouns, remains a key challenge for responsible AI. Prior work, such as the MISGENDERED benchmark, revealed significant limitations in earlier LLMs' handling of inclusive pronouns, but was constrained to outdated models and limited evaluations. In this study, we introduce MISGENDERED+, an extended and updated benchmark for evaluating LLMs' pronoun fidelity. We benchmark five representative LLMs, GPT-4o, Claude 4, DeepSeek-V3, Qwen Turbo, and Qwen2.5, across zero-shot, few-shot, and gender identity inference. Our results show notable improvements compared with previous studies, especially in binary and gender-neutral pronoun accuracy. However, accuracy on neopronouns and reverse inference tasks remains inconsistent, underscoring persistent gaps in identity-sensitive reasoning. We discuss implications, model-specific observations, and avenues for future inclusive AI research.