arXiv subjects

Ruizhe Li

Publications and source records attributed to Ruizhe Li.

At least 37 records · Page 2Linked to original sources

arXiv2026

Spurious Rewards Paradox: Mechanistically Understanding How RLVR Activates Memorization Shortcuts in LLMs

Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen 2.5 achieve significant gains even with spurious or incorrect rewards. We investigate this phenomenon and identify a "Perplexity Paradox": spurious RLVR triggers a divergence where answer-token perplexity drops while prompt-side coherence degrades, suggesting the model is bypassing reasoning in favor of memorization. Using Path Patching, Logit Lens, JSD analysis, and Neural Differential Equations, we uncover a hidden Anchor-Adapter circuit that facilitates this shortcut. We localize a Functional Anchor in the middle layers (L18-20) that triggers the retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations to accommodate the shortcut signal. Finally, we demonstrate that scaling specific MLP keys within this circuit allows for bidirectional causal steering-artificially amplifying or suppressing contamination-driven performance. Our results provide a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models. Code is available at https://github.com/idwts/How-RLVR-Activates-Memorization-Shortcuts.

cs.LG↗

arXiv2026

Orthogonal Hierarchical Decomposition for Structure-Aware Table Understanding with Large Language Models

Complex tables with multi-level headers, merged cells and heterogeneous layouts pose persistent challenges for LLMs in both understanding and reasoning. Existing approaches typically rely on table linearization or normalized grid modeling. However, these representations struggle to explicitly capture hierarchical structures and cross-dimensional dependencies, which can lead to misalignment between structural semantics and textual representations for non-standard tables. To address this issue, we propose an Orthogonal Hierarchical Decomposition (OHD) framework that constructs structure-preserving input representations of complex tables for LLMs. OHD introduces an Orthogonal Tree Induction (OTI) method based on spatial--semantic co-constraints, which decomposes irregular tables into a column tree and a row tree to capture vertical and horizontal hierarchical dependencies, respectively. Building on this representation, we design a dual-pathway association protocol to symmetrically reconstruct semantic lineage of each cell, and incorporate an LLM as a semantic arbitrator to align multi-level semantic information. We evaluate OHD framework on two complex table question answering benchmarks, AITQA and HiTab. Experimental results show that OHD consistently outperforms existing representation paradigms across multiple evaluation metrics.

cs.CL↗

arXiv2026

Business as Rulesual: A Benchmark and Framework for Business Rule Flow Modeling with LLMs

Extracting structured procedural knowledge from unstructured business documents is a critical yet unresolved bottleneck in process automation. While prior work has focused on extracting linear action flows from instructional texts, such as recipes, it has insufficiently addressed the complex logical structures, including conditional branching and parallel execution, that are pervasive in real-world regulatory and administrative documents. Furthermore, existing benchmarks are limited by simplistic schemas and shallow logical dependencies, restricting progress toward logic-aware large language models.To bridge this Logic Gap, we introduce BREX, a carefully curated benchmark comprising 409 real-world business documents and 2,855 expert-annotated rules. Unlike prior datasets centered on narrow service scenarios, BREX spans over 30 vertical domains, covering scientific, industrial, administrative, and financial regulations. We further propose ExIde, a structure-aware reasoning framework that investigates five distinct prompting strategies, ranging from implicit semantic alignment to executable grounding via pseudo-code generation. This enables explicit modeling of rule dependencies and provides an out-of-the-box framework for different business customers without finetuning their own large language models. We benchmark ExIde using 13 state-of-the-art large language models. Our extensive evaluation reveals that executable grounding serves as a superior inductive bias, significantly outperforming standard prompts in rule extraction. In addition, reasoning-optimized models demonstrate a distinct advantage in tracing long-range and non-linear rule dependencies compared to standard instruction-tuned models.

cs.CL↗

arXiv2026

Visuals Lie, Consistency Speaks: Disentangling Spatial Attention from Reliability in Vision-Language Models

Multimodal Foundation Models are increasingly used as reasoning agents, making reliability, knowing when a model may hallucinate, critical. A common intuition, which we call the Attention-Confidence Assumption, holds that reliability follows from "structural" visual perception: tight attention on relevant regions should signal a trustworthy answer, while scattered attention signals confusion. We challenge this through the VLM Reliability Probe (VRP), a systematic cross-family study of reliability signals in contemporary Vision-Language Models (VLMs). We introduce structural-attention metrics, cluster counts (C_k) and spatial entropy (H_s), to quantify the visual encoder's gaze, and track its evolution (Delta H_s) across layers. This reveals a "Symbolic Detachment": models often "Early Lock" visual features only to diffuse attention later, severing early perception from final generation. Contrary to the grounding hypothesis, we find a "Cluster Failure": spatial attention has near-zero correlation (R approx 0.001) with accuracy. Instead, reliability is a phenomenon of generation dynamics and internal-state distributions. Self-Consistency, the agreement rate across sampled reasoning paths, is the dominant predictor of truth (R = 0.429). Scaling causal interventions exposes a sharp architectural divergence: LLaVA locks its prediction in a fragile late-stage bottleneck, whereas PaliGemma and Qwen2-VL distribute reliability globally, staying resilient even when ~50% or more of their most predictive layer is destroyed. For current VLMs, reliability signals are detached from visual grounding maps and are best inferred from generation-time dynamics and hidden-state probes.

cs.CV↗

arXiv2026

StoryVideoQA: Scaling Deep Video Understanding with a Large-Scale, Multi-Genre and Auto-Generated Dataset

Video question answering (VideoQA) aims to answer questions about given videos. While existing approaches excel on factoid VideoQA, they struggle with deep video understanding (DVU), which requires the comprehension of complex storylines. This challenge arises from the inherent long-range video content, multi-faceted question types, and instance-level story elements, all of which constrain the scale and diversity of manually constructed DVU datasets. These difficulties constrain the scale and diversity of manually-constructed DVU dataset. To address these, we previously introduced StoryMind to automatically construct DVU datasets with balanced fine-grained topics. Though it can generate high-quality question-answer pairs (QAs) for TV series, it suffers significant performance degradation when handling longer and more complex movies. In this paper, we further design StoryMindv2, an enhanced multi-agent collaboration framework to generate high-quality DVU datasets for both TV series and movies. By integrating a novel supervisor-guided generation mechanism and a refined multi-reviewer voting strategy, the framework is utilized to construct StoryVideoQA, the largest DVU dataset to date, featuring over 363K QAs on 393.2 hours diverse story videos including TV series (avg. 1,635 seconds) and movies (avg. 7,878 seconds). Comprehensive evaluations of 20 state-of-the-art VideoQA methods on this large-scale benchmark reveal that they cannot fully maintain long-range character associations or construct a coherent understanding of complex storylines. To bridge this gap, we propose PlotTree, a novel video understanding agent, re-organizing long-range video content into a hierarchical plot structure, enabling efficient storyline reasoning on StoryVideoQA. Project page: https://github.com/nercms-mmap/StoryVideoQA/

cs.CV↗

arXiv2026

Food Noise & False Safety: A Systematic Evaluation of How LLMs Fail to Adapt to Eating Disorder Queries with Clinician Feedback

Recent evidence shows that people with eating disorders (EDs) are increasingly seeking guidance, advice, and emotional support from Large Language Model (LLM)-based chat systems. Although these systems are not designed to provide clinical advice, their perceived expertise, neutrality and accessibility make them a frequent, albeit risky, source of support. This paper investigates potential patterns of interaction between users with EDs and LLMs, focusing on the potential harms arising from models that uncritically adapt to, and facilitate unsafe or self-harming user requests. We find, in consultation with clinical ED experts, that specific linguistic cues in prompts increase the likelihood of unsafe responses and, through systematically varying the degree of potential risk present in the user prompt, report the extent to which LLMs uncritically adapt to problematic, and potentially dangerous user inputs.

cs.AI↗

arXiv2026

Trust No Tool: Evaluating and Defending LLM Agents under Untrusted Tool Feedback

Tool-using LLM agents increasingly rely on external tools to make consequential decisions, yet most existing agent-security benchmarks and defenses implicitly assume that tool feedback is trustworthy once a tool has been selected. We study a different failure mode, cognitive poisoning, in which a malicious tool behaves plausibly during exploration, accumulates trust through benign-looking feedback, and becomes harmful only when hidden state conditions align with the final executable action. To study this setting, we construct TRUST-Bench, a task-conditioned benchmark of 1,970 hidden-trigger tool-compromise episodes with matched safe controls, introduce an asymmetric penalty metric, GuardedJoint, to better reflect real deployment risk, and present VISTA-Guard, a backbone-agnostic framework for final-action risk scoring. The core idea is to abstract multi-step tool interaction into structured environment variables that encode trust-formation dynamics and then score the risk of the final executable action from this trajectory-conditioned representation. Experiments show that prompt-centric heuristics, scalarized features, and zero-shot judges fail in this regime, whereas trajectory-aware final-action scoring yields strong in-domain discrimination and remains effective under balanced out-of-distribution transfer. Under GuardedJoint, VISTA-Guard reaches $84.2$ in-domain and $56.9$ on balanced out-of-distribution evaluation, while methods that optimize only one side of the safety--utility tradeoff collapse to zero. These findings support a broader view of agent security in black-box tool ecosystems: the decisive defense target is not local prompt text or tool descriptors alone, but the way trust is formed across the interaction trajectory and committed through the final action.

cs.CR↗

arXiv2026

Nationality encoding in language model hidden states: Probing culturally differentiated representations in persona-conditioned academic text

Large language models are increasingly used as writing tools and pedagogical resources in English for Academic Purposes, but it remains unclear whether they encode culturally differentiated representations when generating academic text. This study tests whether Gemma-3-4b-it encodes nationality-discriminative information in hidden states when generating research article introductions conditioned by British and Chinese academic personas. A corpus of 270 texts was generated from 45 prompt templates crossed with six persona conditions in a 2 x 3 design. Logistic regression probes were trained on hidden-state activations across all 35 layers, with shuffled-label baselines, a surface-text skyline classifier, cross-family tests, and sentence-level baselines used as controls. Probe-selected token positions were annotated for structural, lexical, and stance features using the Stanza NLP pipeline. The nationality probe reached 0.968 cross-validated accuracy at Layer 18, with perfect held-out classification. Nationality encoding followed a non-monotonic trajectory across layers, with structural effects strongest in the middle to upper network and lexical-domain effects peaking earlier. At high-signal token positions, British-associated patterns showed more postmodification, hedging, boosting, passive voice, and evaluative or process-oriented vocabulary, while Chinese-associated patterns showed more premodification, nominal predicates, and sociocultural or internationalisation vocabulary. However, sentence-level analysis found no significant nationality differences in the full generated surface text. The findings extend probing methodology to a sociolinguistic attribute and have practical implications for EAP and language pedagogy.

cs.CL↗

arXiv2026

Search-MIND: Training-Free Multi-Modal Medical Image Registration

Multi-modal image registration plays a critical role in precision medicine but faces challenges from non-linear intensity relationships and local optima. While deep learning models enable rapid inference, they often suffer from generalization collapse on unseen modalities. To address this, we propose Search-MIND, a training-free, iterative optimization framework for instance-specific registration. Our pipeline utilizes a coarse-to-fine strategy: a hierarchical coarse alignment stage followed by deformable refinement. We introduce two novel loss functions: Variance-Weighted Mutual Information (VWMI), which prioritizes informative tissue regions to shield global alignment from background noise and uniform regions, and Search-MIND (S-MIND), which broadens the convergence basin of structural descriptors by considering larger local search range. Evaluations on CARE Liver 2025 and CHAOS Challenge datasets show that Search-MIND consistently outperforms classical baselines like ANTs and foundation model-based approaches like DINO-reg, offering superior stability across diverse modalities.

eess.IV↗

arXiv2026

FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning

With the rapid advancement of large language models (LLMs), growing efforts have been made on LLM-based table retrieval. However, existing studies typically focus on single-table query, and implement it by similarity matching after encoding the entire table. These methods usually result in low accuracy due to their coarse-grained encoding which incorporates much query-irrelated data, and are also inefficient when dealing with large tables, failing to fully utilize the reasoning capabilities of LLM. Further, multi-table query is under-explored in retrieval tasks. To this end, we propose a hierarchical multi-table query method based on LLM: Fine-Grained Multi-Table Retrieval FGTR, a new retrieval paradigm that employs a human-like reasoning strategy. Through hierarchical reasoning, FGTR first identifies relevant schema elements and then retrieves the corresponding cell contents, ultimately constructing a concise and accurate sub-table that aligns with the given query. To comprehensively evaluate the performance of FGTR, we construct two new benchmark datasets based on Spider and BIRD . Experimental results show that FGTR outperforms previous state-of-the-art methods, improving the F_2 metric by 18% on Spider and 21% on BIRD, demonstrating its effectiveness in enhancing fine-grained retrieval and its potential to improve end-to-end performance on table-based downstream tasks.

cs.IR↗

arXiv2026

T$_2$* and Susceptibility Mapping as Indicators of Placental Health

Objective(s): T$_2$* and susceptibility ($χ$) MRI mapping provide complimentary measures of the haemodynamic environment in the placenta. The aims of this work were to use these simultaneously obtained measures to investigate the role of oxygen distribution on the well-established reduction of T$_2$* with gestational age found in healthy pregnancies and explore differences in both measures in compromised placentas. Methods: T$_2$* and $χ$ were measured simultaneously from a double echo, echo planar scan of the whole placenta, across a range of gestational ages and pregnancy complications. Regional variations across the placenta were investigated. Results: Whole placental mean T$_2$* was more correlated with standard deviation of $χ$ than mean $χ$ indicating it is more driven by increasing local inhomogeneities rather than bulk deoxygenation with healthy gestation. Compromised placentas also showed increased standard deviation of $χ$ as well as lower mean T$_2$* suggesting flow/uptake mismatch and reduced oxygenation. Regionally, the susceptibility was lowest (most oxygenated) and least variable in the central region of the placenta indicating good mixing and refreshment of blood in this area. The susceptibility was highest (most deoxygenated) and most variable at the fetal side, suggesting less effective perfusion in this region. Compromised cases showed the greatest difference on the fetal side for both mean and standard deviation of $χ$. T$_2$* was lowest at the fetal side for healthy and compromised cases but the maternal and central regions better distinguished between the two groups. Conclusion(s): T$_2$* and susceptibility can be mapped simultaneously from a single MRI scan and provide complimentary information about the function of the placenta across healthy gestational development, and as a potential indicator of placental compromise.

physics.med-ph↗

arXiv2026

Llettuce: An Open Source Natural Language Processing Tool for the Translation of Medical Terms into Uniform Clinical Encoding

This paper introduces Llettuce, an open-source tool designed to address the complexities of converting medical terms into OMOP standard concepts. Unlike existing solutions such as the Athena database search and Usagi, which struggle with semantic nuances and require substantial manual input, Llettuce leverages advanced natural language processing, including large language models and fuzzy matching, to automate and enhance the mapping process. Developed with a focus on GDPR compliance, Llettuce can be deployed locally, ensuring data protection while maintaining high performance in converting informal medical terms to standardised concepts.

cs.CL↗

arXiv2026

MANATEE: Inference-Time Lightweight Diffusion Based Safety Defense for LLMs

Defending LLMs against adversarial jailbreak attacks remains an open challenge. Existing defenses rely on binary classifiers that fail when adversarial input falls outside the learned decision boundary, and repeated fine-tuning is computationally expensive while potentially degrading model capabilities. We propose MANATEE, an inference-time defense that uses density estimation over a benign representation manifold. MANATEE learns the score function of benign hidden states and uses diffusion to project anomalous representations toward safe regions--requiring no harmful training data and no architectural modifications. Experiments across Mistral-7B-Instruct, Llama-3.1-8B-Instruct, and Gemma-2-9B-it demonstrate that MANATEE reduce Attack Success Rate by up to 100\% on certain datasets, while preserving model utility on benign inputs.

cs.CR↗

arXiv2026

Placental contractions in uncomplicated pregnancies

In 2020 we first described placental contractions, and we have now undertaken a study to characterise them and seek features that might automatically separate them from uterine contractions. We recruited 36 healthy pregnant women to undergo magnetic resonance imaging (MRI) between 29 and 42 weeks of pregnancy in a single-centre, prospective, observational study. Participants had fetal ultrasound to confirm normal growth. Dynamic MRI was acquired for between 15 and 32 minutes using respiratory triggered, multi-slice, single shot, gradient echo, echo planar imaging covering the whole uterus. All participants had a live birth of a healthy baby weighing over the 10th centile for gestational age and none developed any associated conditions of placental dysfunction e.g. pre-eclampsia, or severe maternal or fetal villous malperfusion on placental histopathology. Any visible contractions were recorded for all participants who completed their MRI scan and placental contractions occurred in at least 60% of our healthy pregnant population with a median frequency of approximately 2 per hour, and a median duration of 2.4 minutes. Contractions involving a decrease in placental volume of >10% were classified as either placental or uterine by visual observation. Placental contractions occurred more frequently than uterine contractions (p=0.0061), were associated with a larger increase in the surface area of the uterine wall not covered by the placenta (p=0.0015), placental sphericity (p<0.0001) and longer duration (p=0.0151). All contractions led to an increase in the MRI parameter R2* in the placenta. There was large variation both between participants and between contractions from the same individual, in terms of time course and contractions features, with no apparent change across the gestational age range studied, although the largest fractional volume changes were detected at early gestation.

physics.med-ph↗

arXiv2026

Attributing Response to Context: A Jensen-Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) leverages large language models (LLMs) combined with external contexts to enhance the accuracy and reliability of generated responses. However, reliably attributing generated content to specific context segments, context attribution, remains challenging due to the computationally intensive nature of current methods, which often require extensive fine-tuning or human annotation. In this work, we introduce a novel Jensen-Shannon Divergence driven method to Attribute Response to Context (ARC-JSD), enabling efficient and accurate identification of essential context sentences without additional fine-tuning, gradient-calculation or surrogate modelling. Evaluations on a wide range of RAG benchmarks, such as TyDi QA, Hotpot QA, and Musique, using instruction-tuned LLMs in different scales demonstrate superior accuracy and significant computational efficiency improvements compared to the previous surrogate-based method. Furthermore, our mechanistic analysis reveals specific attention heads and multilayer perceptron (MLP) layers responsible for context attribution, providing valuable insights into the internal workings of RAG models and how they affect RAG behaviours. Our code is available at https://github.com/ruizheliUOA/ARC_JSD.

cs.CL↗

arXiv2026

Semi-supervised Liver Segmentation and Patch-based Fibrosis Staging with Registration-aided Multi-parametric MRI

Liver fibrosis poses a substantial challenge in clinical practice, emphasizing the necessity for precise liver segmentation and accurate disease staging. Based on the CARE Liver 2025 Track 4 Challenge, this study introduces a multi-task deep learning framework developed for liver segmentation (LiSeg) and liver fibrosis staging (LiFS) using multiparametric MRI. The LiSeg phase addresses the challenge of limited annotated images and the complexities of multi-parametric MRI data by employing a semi-supervised learning model that integrates image segmentation and registration. By leveraging both labeled and unlabeled data, the model overcomes the difficulties introduced by domain shifts and variations across modalities. In the LiFS phase, we employed a patchbased method which allows the visualization of liver fibrosis stages based on the classification outputs. Our approach effectively handles multimodality imaging data, limited labels, and domain shifts. The proposed method has been tested by the challenge organizer on an independent test set that includes in-distribution (ID) and out-of-distribution (OOD) cases using three-channel MRIs (T1, T2, DWI) and seven-channel MRIs (T1, T2, DWI, GED1-GED4). The code is freely available. Github link: https://github.com/mileywang3061/Care-Liver

cs.CV↗

arXiv2026

ProMoral-Bench: Evaluating Prompting Strategies for Moral Reasoning and Safety in LLMs

Prompt design significantly impacts the moral competence and safety alignment of large language models (LLMs), yet empirical comparisons remain fragmented across datasets and models.We introduce ProMoral-Bench, a unified benchmark evaluating 11 prompting paradigms across four LLM families. Using ETHICS, Scruples, WildJailbreak, and our new robustness test, ETHICS-Contrast, we measure performance via our proposed Unified Moral Safety Score (UMSS), a metric balancing accuracy and safety. Our results show that compact, exemplar-guided scaffolds outperform complex multi-stage reasoning, providing higher UMSS scores and greater robustness at a lower token cost. While multi-turn reasoning proves fragile under perturbations, few-shot exemplars consistently enhance moral stability and jailbreak resistance. ProMoral-Bench establishes a standardized framework for principled, cost-effective prompt engineering.

cs.AI↗

arXiv2026

Benchmarking Text-to-Python against Text-to-SQL: The Impact of Explicit Logic and Ambiguity

While Text-to-SQL remains the dominant approach for database interaction, real-world analytics increasingly require the flexibility of general-purpose programming languages such as Python or Pandas to manage file-based data and complex analytical workflows. Despite this growing need, the reliability of Text-to-Python in core data retrieval remains underexplored relative to the mature SQL ecosystem. To address this gap, we introduce BIRD-Python, a benchmark designed for cross-paradigm evaluation. We systematically refined the original dataset to reduce annotation noise and align execution semantics, thereby establishing a consistent and standardized baseline for comparison. Our analysis reveals a fundamental paradigmatic divergence: whereas SQL leverages implicit DBMS behaviors through its declarative structure, Python requires explicit procedural logic, making it highly sensitive to underspecified user intent. To mitigate this challenge, we propose the Logic Completion Framework (LCF), which resolves ambiguity by incorporating latent domain knowledge into the generation process. Experimental results show that (1) performance differences primarily stem from missing domain context rather than inherent limitations in code generation, and (2) when these gaps are addressed, Text-to-Python achieves performance parity with Text-to-SQL. These findings establish Python as a viable foundation for analytical agents-provided that systems effectively ground ambiguous natural language inputs in executable logical specifications. Resources are available at https://anonymous.4open.science/r/Bird-Python-43B7/.

cs.AI↗