Large language models (LLMs) often achieve high performance in native language identification (NLI) benchmarks by leveraging superficial contextual clues such as names, locations, and cultural stereotypes, rather than the underlying linguistic patterns indicative of native language (L1) influence. To improve robustness, previous work has instructed LLMs to disregard such clues. In this work, we demonstrate that such a strategy is unreliable and model predictions can be easily altered by misleading hints. To address this problem, we introduce an agentic NLI pipeline inspired by forensic linguistics, where specialized agents accumulate and categorize diverse linguistic evidence before an independent final overall assessment. In this final assessment, a goal-aware coordinating agent synthesizes all evidence to make the NLI prediction. On two benchmark datasets, our approach significantly enhances NLI robustness against misleading contextual clues and performance consistency compared to standard prompting methods.

Large Language Models have demonstrated impressive fluency across diverse tasks, yet their tendency to produce toxic content remains a critical challenge for AI safety and public trust. Existing toxicity mitigation approaches primarily manipulate individual neuron activations, but these methods suffer from instability, context dependence, and often compromise the model's core language abilities. To address these shortcomings, we investigate three key questions: the stability of neuron-level toxicity indicators, the advantages of structural (layer-wise) representations, and the interpretability of mechanisms driving toxic generation. Through extensive experiments on Jigsaw and ToxiCN datasets, we show that aggregated layer-wise features provide more robust signals than single neurons. Moreover, we observe conceptual limitations in prior works that conflate toxicity detection experts and generation experts within neuron-based interventions. To mitigate this, we propose a novel principled intervention technique, EigenShift, based on eigen-decomposition of the language model's final output layer. This method selectively targets generation-aligned components, enabling precise toxicity suppression without impairing linguistic competence. Our method requires no additional training or fine-tuning, incurs minimal computational cost, and is grounded in rigorous theoretical analysis.

ABSTRACT This report presents the AIST A T team's submission to the lan guage-based audio retrieval task in DCASE 2025 Task 6. Our proposed system employs dual encoder architecture, where audi o and text modalities are encoded separately, and their repre senta-tions are aligned using contrastive learning. Additionally, we incorporat ed clustering to introduce an auxiliary classification task for fur ther fine-tuning. Our best single system achieved a mAP@16 of 46.62, wh ile an ensem-ble of four systems reached a mAP@16 of 48.83 on the Clotho development test split. Index T erms -- Audio-text retrieval, contrastive learning, knowledge distillation, topic modeling 1. INTRODUCTION DCASE 2025 Task 6 challenge [1] focuses on language-based au - dio retrieval, a task that requires retrieving audio record ings from a database that best matches a given textual query, and vice v ersa.

Large Vision-Language Models (L-VLMs) have demonstrated remarkable performance in various vision and language tasks, including visual question answering (VQA). However, their high computational cost makes them impractical for resource-constrained settings and inference-heavy applications. In contrast, Small Vision-Language Models (S-VLMs) offer efficiency but suffer from a significant performance gap compared to their larger counterparts. In this work, we introduce the Model Parity Aligner (MPA), a novel framework designed to systematically improve S-VLMs by leveraging unlabeled images and effective knowledge transfer from L-VLMs. Instead of traditional knowledge distillation methods that rely on labeled training data, MPA employs a strategic parity-based approach that precisely identifies the knowledge disparities between S-VLMs and L-VLMs, and optimizes training by targeting only these disparities. We conduct extensive experiments on four diverse VQA benchmarks, namely TextVQA, ST-VQA, ChartQA, and OKVQA, each of which requires specialized reasoning capabilities such as text recognition, chart interpretation, and commonsense and factual understanding. Our results demonstrate that MPA consistently enhances the performance of S-VLMs on all benchmarks, reducing the performance gap while maintaining computational efficiency. We make our code publicly available.

In recent years, Visual Question Localized-Answering in robotic surgery (Surgical-VQLA) has gained significant attention for its potential to assist medical students and junior doctors in understanding surgical scenes. Recently, the rapid development of Large Language Models (LLMs) has provided more promising solutions for this task. However, current methods struggle to establish complex dependencies between text and visual details, and have difficulty perceiving the spatial information of surgical scenes. To address these challenges, we propose a novel method, Surgical-MambaLLM, which is the first to combine Mamba2 with LLM in the surgical domain, that leverages Mamba2's ability to effectively capture cross-modal dependencies and perceive spatial information in surgical scenes, thereby enhancing the LLMs' understanding of surgical images. Specifically, we propose the Cross-modal Bidirectional Mamba2 Integration (CBMI) module to leverage Mamba2 for effective multimodal fusion, with its cross-modal integration capabilities. Additionally, tailored to the geometric characteristics of surgical scenes, we design the Surgical Instrument Perception (SIP) scanning mode for Mamba2 to scan the surgical images, enhancing the model's spatial understanding of the surgical scene. Extensive experiments demonstrate that our Surgical-MambaLLM model outperforms the state-of-the-art methods on the EndoVis17-VQLA and EndoVis18-VQLA datasets, significantly improving the performance of the Surgical-VQLA task.

Reinforcement learning (RL) is a promising approach for robotic manipulation, but it can suffer from low sample efficiency and requires extensive exploration of large state-action spaces. Recent methods leverage the commonsense knowledge and reasoning abilities of large language models (LLMs) to guide exploration toward more meaningful states. However, LLMs can produce plans that are semantically plausible yet physically infeasible, yielding unreliable behavior. We introduce LLM-TALE, a framework that uses LLMs' planning to directly steer RL exploration. LLM-TALE integrates planning at both the task level and the affordance level, improving learning efficiency by directing agents toward semantically meaningful actions. Unlike prior approaches that assume optimal LLM-generated plans or rewards, LLM-TALE corrects suboptimality online and explores multimodal affordance-level plans without human supervision. We evaluate LLM-TALE on pick-and-place tasks in standard RL benchmarks, observing improvements in both sample efficiency and success rates over strong baselines. Real-robot experiments indicate promising zero-shot sim-to-real transfer. Code and supplementary material are available at https://llm-tale.github.io.

Aiming at the problems of computational inefficiency and insufficient interpretability faced by large models in complex tasks such as multi-round reasoning and multi-modal collaboration, this study proposes a three-layer collaboration framework based on model-controller-task adaptation (MCP). By decoupling large model functions into reasoning, generation and retrieval modules, and combining reinforcement learning-driven dynamic routing algorithms and task adaptation mechanisms, the systematic integration of control theory and large model dynamic reasoning is achieved for the first time. Experiments show that the MCP framework improves the performance of cross-modal benchmarking tasks, such as GLUE, COCO, ScienceQA, etc., by 15-30% compared with the baseline model, improves the reasoning efficiency by 40%, and generates the interpretable intermediate results through the Presenter layer, obtaining 90% of the manual interpretability scores, which provides a brand-new technological path to solve the bottleneck of the practical application of the large model.

Language-conditioned robotic manipulation in open-world settings requires not only accurate task execution but also the ability to detect failures for robust deployment in real-world environments. Although recent advances in vision-language models (VLMs) have significantly improved the spatial reasoning and task-planning capabilities of robots, they remain limited in their ability to recognize their own failures. In particular, a critical yet underexplored challenge lies in detecting semantic misalignment errors, where the robot executes a task that is semantically meaningful but inconsistent with the given instruction. To address this, we propose a method for building datasets targeting Semantic Misalignment Failures detection, from existing language-conditioned manipulation datasets. We also present I-FailSense, an open-source VLM framework with grounded arbitration designed specifically for failure detection. Our approach relies on post-training a base VLM, followed by training lightweight classification heads, called FS blocks, attached to different internal layers of the VLM and whose predictions are aggregated using an ensembling mechanism. Experiments show that I-FailSense outperforms state-of-the-art VLMs, both comparable in size and larger, in detecting semantic misalignment errors. Notably, despite being trained only on semantic misalignment detection, I-FailSense generalizes to broader robotic failure categories and effectively transfers to other simulation environments and real-world with zero-shot or minimal post-training. The datasets and models are publicly released on HuggingFace (Webpage: https://clemgris.github.io/I-FailSense/).

Embodied agents operating in the physical world must make decisions that are not only effective but also safe, spatially coherent, and grounded in context. While recent advances in large multimodal models (LMMs) have shown promising capabilities in visual understanding and language generation, their ability to perform structured reasoning for real-world embodied tasks remains underexplored. In this work, we aim to understand how well foundation models can perform step-by-step reasoning in embodied environments. To this end, we propose the Foundation Model Embodied Reasoning (FoMER) benchmark, designed to evaluate the reasoning capabilities of LMMs in complex embodied decision-making scenarios. Our benchmark spans a diverse set of tasks that require agents to interpret multimodal observations, reason about physical constraints and safety, and generate valid next actions in natural language. We present (i) a large-scale, curated suite of embodied reasoning tasks, (ii) a novel evaluation framework that disentangles perceptual grounding from action reasoning, and (iii) empirical analysis of several leading LMMs under this setting. Our benchmark includes over 1.1k samples with detailed step-by-step reasoning across 10 tasks and 8 embodiments, covering three different robot types. Our results highlight both the potential and current limitations of LMMs in embodied reasoning, pointing towards key challenges and opportunities for future research in robot intelligence. Our data and code will be made publicly available.

Large models achieve strong performance on Vision-and-Language Navigation (VLN) tasks, but are costly to run in resource-limited environments. Token pruning offers appealing tradeoffs for efficiency with minimal performance loss by reducing model input size, but prior work overlooks VLN-specific challenges. For example, information loss from pruning can effectively increase computational cost due to longer walks. Thus, the inability to identify uninformative tokens undermines the supposed efficiency gains from pruning. To address this, we propose Navigation-Aware Pruning (NAP), which uses navigation-specific traits to simplify the pruning process by pre-filtering tokens into foreground and background. For example, image views are filtered based on whether the agent can navigate in that direction. We also extract navigation-relevant instructions using a Large Language Model. After filtering, we focus pruning on background tokens, minimizing information loss. To further help avoid increases in navigation length, we discourage backtracking by removing low-importance navigation nodes. Experiments on standard VLN benchmarks show NAP significantly outperforms prior work, preserving higher success rates while saving more than 50% FLOPS.