In this paper, we describe our system under the team name BLEU Monday for the English-to-Indic Multimodal Translation Task at W AT 2025. We participate in the text-only translation tasks for English-Hindi, English-Bengali, English-Malayalam, and English-Odia language pairs. We present a two-stage approach that addresses quality issues in the training data through automated error detection and correction, followed by parameter-efficient model fine-tuning. Our methodology introduces a vision-augmented judge-corrector pipeline that leverages multimodal language models to systematically identify and correct translation errors in the training data. The judge component classifies translations into three categories: correct, visually ambiguous (requiring image context), or mistranslated (poor translation quality). Identified errors are routed to specialized correctors: GPT-4o-mini regenerates captions requiring visual disambiguation, while IndicTrans2 retranslates cases with pure translation quality issues. This automated pipeline processes 28,928 training examples across four languages, correcting an average of 17.1% of captions per language. We then apply Low-Rank Adaptation (LoRA) to fine-tune the IndicTrans2 en-indic 200M distilled model on both original and corrected datasets.

Conventional road-situation detection methods achieve strong performance in predefined scenarios but fail in unseen cases and lack semantic interpretation, which is crucial for reliable traffic recommendations. This work introduces a multi-agent AI framework that combines multimodal large language models (MLLMs) with vision-based perception for road-situation monitoring. The framework processes camera feeds and coordinates dedicated agents for situation detection, distance estimation, decision-making, and Cooperative Intelligent Transport System (C-ITS) message generation. Evaluation is conducted on a custom dataset of 103 images extracted from 20 videos of the TAD dataset. Both Gemini-2.0-Flash and Gemini-2.5-Flash were evaluated. The results show 100\% recall in situation detection and perfect message schema correctness; however, both models suffer from false-positive detections and have reduced performance in terms of number of lanes, driving lane status and cause code. Surprisingly, Gemini-2.5-Flash, though more capable in general tasks, underperforms Gemini-2.0-Flash in detection accuracy and semantic understanding and incurs higher latency (Table II). These findings motivate further work on fine-tuning specialized LLMs or MLLMs tailored for intelligent transportation applications.

Large Language Models for Simulating Professions (SP-LLMs), particularly as teachers, are pivotal for personalized education. However, ensuring their professional competence and ethical safety is a critical challenge, as existing benchmarks fail to measure role-playing fidelity or address the unique teaching harms inherent in educational scenarios. To address this, we propose EduGuardBench, a dual-component benchmark. It assesses professional fidelity using a Role-playing Fidelity Score (RFS) while diagnosing harms specific to the teaching profession. It also probes safety vulnerabilities using persona-based adversarial prompts targeting both general harms and, particularly, academic misconduct, evaluated with metrics including Attack Success Rate (ASR) and a three-tier Refusal Quality assessment. Our extensive experiments on 14 leading models reveal a stark polarization in performance. While reasoning-oriented models generally show superior fidelity, incompetence remains the dominant failure mode across most models. The adversarial tests uncovered a counterintuitive scaling paradox, where mid-sized models can be the most vulnerable, challenging monotonic safety assumptions. Critically, we identified a powerful Educational Transformation Effect: the safest models excel at converting harmful requests into teachable moments by providing ideal Educational Refusals. This capacity is strongly negatively correlated with ASR, revealing a new dimension of advanced AI safety. EduGuardBench thus provides a reproducible framework that moves beyond siloed knowledge tests toward a holistic assessment of professional, ethical, and pedagogical alignment, uncovering complex dynamics essential for deploying trustworthy AI in education. See https://github.com/YL1N/EduGuardBench for Materials.

Efficiently fine-tuning pre-trained models for downstream tasks is a key challenge in the era of foundation models. Parameter-efficient fine-tuning (PEFT) presents a promising solution, achieving performance comparable to full fine-tuning by updating only a small number of adaptation weights per layer. Traditional PEFT methods typically rely on a single expert, where the adaptation weight is a low-rank matrix. However, for complex tasks, the data's inherent diversity poses a significant challenge for such models, as a single adaptation weight cannot adequately capture the features of all samples. To address this limitation, we explore how to integrate multiple small adaptation experts into a compact structure to defeat a large adapter. Specifically, we propose Tucker Adaptation (TuckA), a method with four key properties: (i) We use Tucker decomposition to create a compact 3D tensor where each slice naturally serves as an expert. The low-rank nature of this decomposition ensures that the number of parameters scales efficiently as more experts are added. (ii) We introduce a hierarchical strategy that organizes these experts into groups at different granularities, allowing the model to capture both local and global data patterns. (iii) We develop an efficient batch-level routing mechanism, which reduces the router's parameter size by a factor of $L$ compared to routing at every adapted layer (where $L$ is the number of adapted layers) (iv) We propose data-aware initialization to achieve loss-free expert load balancing based on theoretical analysis. Extensive experiments on benchmarks in natural language understanding, image classification, and mathematical reasoning speak to the efficacy of TuckA, offering a new and effective solution to the PEFT problem.

Zero-shot object navigation (ZSON) in unseen environments remains a challenging problem for household robots, requiring strong perceptual understanding and decision-making capabilities. While recent methods leverage metric maps and Large Language Models (LLMs), they often depend on depth sensors or prebuilt maps, limiting the spatial reasoning ability of Multimodal Large Language Models (MLLMs). Map-less ZSON approaches have emerged to address this, but they typically make short-sighted decisions, leading to local deadlocks due to a lack of historical context. We propose PanoNav, a fully RGB-only, mapless ZSON framework that integrates a Panoramic Scene Parsing module to unlock the spatial parsing potential of MLLMs from panoramic RGB inputs, and a Memory-guided Decision-Making mechanism enhanced by a Dynamic Bounded Memory Queue to incorporate exploration history and avoid local deadlocks. Experiments on the public navigation benchmark show that PanoNav significantly outperforms representative baselines in both SR and SPL metrics.

Detecting Alzheimer's disease (AD) from narrative transcripts challenges large language models (LLMs): pre-training rarely covers this out-of-distribution task, and all transcript demos describe the same scene, producing highly homogeneous contexts. These factors cripple both the model's built-in task knowledge (\textbf{task cognition}) and its ability to surface subtle, class-discriminative cues (\textbf{contextual perception}). Because cognition is fixed after pre-training, improving in-context learning (ICL) for AD detection hinges on enriching perception through better demonstration (demo) sets. We demonstrate that standard ICL quickly saturates, its demos lack diversity (context width) and fail to convey fine-grained signals (context depth), and that recent task vector (TV) approaches improve broad task adaptation by injecting TV into the LLMs' hidden states (HSs), they are ill-suited for AD detection due to the mismatch of injection granularity, strength and position. To address these bottlenecks, we introduce \textbf{DA4ICL}, a demo-centric anchoring framework that jointly expands context width via \emph{\textbf{Diverse and Contrastive Retrieval}} (DCR) and deepens each demo's signal via \emph{\textbf{Projected Vector Anchoring}} (PVA) at every Transformer layer. Across three AD benchmarks, DA4ICL achieves large, stable gains over both ICL and TV baselines, charting a new paradigm for fine-grained, OOD and low-resource LLM adaptation.

Attention is a core component of transformer architecture, whether encoder-only, decoder-only, or encoder-decoder model. However, the standard softmax attention often produces noisy probability distribution, which can impair effective feature selection at every layer of these models, particularly for long contexts. We propose Focal Attention, a simple yet effective modification that sharpens the attention distribution by controlling the softmax temperature, either as a fixed hyperparameter or as a learnable parameter during training. This sharpening enables the model to concentrate on the most relevant tokens while suppressing irrelevant ones. Empirically, Focal Attention scales more favorably than standard transformer with respect to model size, training data, and context length. Across diverse benchmarks, it achieves the same accuracy with up to 42% fewer parameters or 33% less training data. On long-context tasks, it delivers substantial relative improvements ranging from 17% to 82%, demonstrating its effectiveness in real world applications.

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in vision-language answering tasks. Despite their strengths, these models often encounter challenges in achieving complex reasoning tasks such as mathematical problem-solving. Previous works have focused on fine-tuning on specialized mathematical datasets. However, these datasets are typically distilled directly from teacher models, which capture only static reasoning patterns and leaving substantial gaps compared to student models. This reliance on fixed teacher-derived datasets not only restricts the model's ability to adapt to novel or more intricate questions that extend beyond the confines of the training data, but also lacks the iterative depth needed for robust generalization. To overcome these limitations, we propose \textbf{\method}, a \textbf{Math}ematical \textbf{S}elf-\textbf{E}volving framework for MLLMs. In contrast to traditional one-shot fine-tuning paradigms, \method iteratively refines the model through cycles of inference, reflection, and reward-based feedback. Specifically, we leverage iterative fine-tuning by incorporating correct reasoning paths derived from previous-stage inference and integrating reflections from a specialized Outcome Reward Model (ORM). To verify the effectiveness of \method, we evaluate it on a suite of challenging benchmarks, demonstrating significant performance gains over backbone models. Notably, our experimental results on MathVL-test surpass the leading open-source multimodal mathematical reasoning model QVQ. Our code and models are available at \texttt{https://zheny2751\allowbreak-dotcom.github.io/\allowbreak MathSE.github.io/}.

The growing complexity of urban mobility systems has made traffic simulation indispensable for evidence-based transportation planning and policy evaluation. However, despite the analytical capabilities of platforms such as the Simulation of Urban MObility (SUMO), their application remains largely confined to domain experts. Developing realistic simulation scenarios requires expertise in network construction, origin-destination modeling, and parameter configuration for policy experimentation, creating substantial barriers for non-expert users such as policymakers, urban planners, and city officials. Moreover, the requests expressed by these users are often incomplete and abstract-typically articulated as high-level objectives, which are not well aligned with the imperative, sequential workflows employed in existing language-model-based simulation frameworks. To address these challenges, this study proposes AgentSUMO, an agentic framework for interactive simulation scenario generation via large language models. AgentSUMO departs from imperative, command-driven execution by introducing an adaptive reasoning layer that interprets user intents, assesses task complexity, infers missing parameters, and formulates executable simulation plans. The framework is structured around two complementary components, the Interactive Planning Protocol, which governs reasoning and user interaction, and the Model Context Protocol, which manages standardized communication and orchestration among simulation tools. Through this design, AgentSUMO converts abstract policy objectives into executable simulation scenarios. Experiments on urban networks in Seoul and Manhattan demonstrate that the agentic workflow achieves substantial improvements in traffic flow metrics while maintaining accessibility for non-expert users, successfully bridging the gap between policy goals and executable simulation workflows.

Multimodal Large Language Models (MLLMs) extend foundation models to real-world applications by integrating inputs such as text and vision. However, their broad knowledge capacity raises growing concerns about privacy leakage, toxicity mitigation, and intellectual property violations. Machine Unlearning (MU) offers a practical solution by selectively forgetting targeted knowledge while preserving overall model utility. When applied to MLLMs, existing neuron-editing-based MU approaches face two fundamental challenges: (1) forgetting becomes inconsistent across modalities because existing point-wise attribution methods fail to capture the structured, layer-by-layer information flow that connects different modalities; and (2) general knowledge performance declines when sensitive neurons that also support important reasoning paths are pruned, as this disrupts the model's ability to generalize. To alleviate these limitations, we propose a multimodal influential neuron path editor (MIP-Editor) for MU. Our approach introduces modality-specific attribution scores to identify influential neuron paths responsible for encoding forget-set knowledge and applies influential-path-aware neuron-editing via representation misdirection. This strategy also enables effective and coordinated forgetting across modalities while preserving the model's general capabilities. Experimental results demonstrate that MIP-Editor achieves a superior unlearning performance on multimodal tasks, with a maximum forgetting rate of 87.75% and up to 54.26% improvement in general knowledge retention. On textual tasks, MIP-Editor achieves up to 80.65% forgetting and preserves 77.9% of general performance.