This letter proposes UniToCom, a unified token communication paradigm that treats tokens as the fundamental units for both processing and wireless transmission. Specifically, to enable efficient token representations, we propose a generative information bottleneck (GenIB) principle, which facilitates the learning of tokens that preserve essential information while supporting reliable generation across multiple modalities. By doing this, GenIB-based tokenization is conducive to improving the communication efficiency and reducing computational complexity. Additionally, we develop $σ$-GenIB to address the challenges of variance collapse in autoregressive modeling, maintaining representational diversity and stability. Moreover, we employ a causal Transformer-based multimodal large language model (MLLM) at the receiver to unify the processing of both discrete and continuous tokens under the next-token prediction paradigm. Simulation results validate the effectiveness and superiority of the proposed UniToCom compared to baselines under dynamic channel conditions. By integrating token processing with MLLMs, UniToCom enables scalable and generalizable communication in favor of multimodal understanding and generation, providing a potential solution for next-generation intelligent communications.

We present Depth Anything at Any Condition (DepthAnything-AC), a foundation monocular depth estimation (MDE) model capable of handling diverse environmental conditions. Previous foundation MDE models achieve impressive performance across general scenes but not perform well in complex open-world environments that involve challenging conditions, such as illumination variations, adverse weather, and sensor-induced distortions. To overcome the challenges of data scarcity and the inability of generating high-quality pseudo-labels from corrupted images, we propose an unsupervised consistency regularization finetuning paradigm that requires only a relatively small amount of unlabeled data. Furthermore, we propose the Spatial Distance Constraint to explicitly enforce the model to learn patch-level relative relationships, resulting in clearer semantic boundaries and more accurate details. Experimental results demonstrate the zero-shot capabilities of DepthAnything-AC across diverse benchmarks, including real-world adverse weather benchmarks, synthetic corruption benchmarks, and general benchmarks. Project Page: https://ghost233lism.github.io/depthanything-AC-page Code: https://github.com/HVision-NKU/DepthAnythingAC

Bias and stereotypes in language models can cause harm, especially in sensitive areas like content moderation and decision-making. This paper addresses bias and stereotype detection by exploring how jointly learning these tasks enhances model performance. We introduce StereoBias, a unique dataset labeled for bias and stereotype detection across five categories: religion, gender, socio-economic status, race, profession, and others, enabling a deeper study of their relationship. Our experiments compare encoder-only models and fine-tuned decoder-only models using QLoRA. While encoder-only models perform well, decoder-only models also show competitive results. Crucially, joint training on bias and stereotype detection significantly improves bias detection compared to training them separately. Additional experiments with sentiment analysis confirm that the improvements stem from the connection between bias and stereotypes, not multi-task learning alone. These findings highlight the value of leveraging stereotype information to build fairer and more effective AI systems.

--In this paper, we investigate the transferability of pre-trained language models to low-resource Indonesian local languages through the task of sentiment analysis. We evaluate both zero-shot performance and adapter-based transfer on ten local languages using models of different types: a monolingual Indonesian BERT, multilingual models such as mBERT and XLM-R, and a modular adapter-based approach called MAD-X. T o better understand model behavior, we group the target languages into three categories: seen (included during pre-training), partially seen (not included but linguistically related to seen languages), and unseen (absent and unrelated in pre-training data). Our results reveal clear performance disparities across these groups: multilingual models perform best on seen languages, moderately on partially seen ones, and poorly on unseen languages. We find that MAD-X significantly improves performance, especially for seen and partially seen languages, without requiring labeled data in the target language. Additionally, we conduct a further analysis on tokenization and show that while subword fragmentation and vocabulary overlap with Indonesian correlate weakly with prediction quality, they do not fully explain the observed performance. Instead, the most consistent predictor of transfer success is the model's prior exposure to the language, either directly or through a related language.

Mixture-of-Experts (MoE) approaches have recently gained traction in robotics applications due to their ability to dynamically allocate computational resources and specialize sub-networks for distinct tasks or environmental contexts, enabling more efficient decision-making. Such systems often comprise sparsely activated experts combined under a single monolithic architecture and require a well-configured internal routing mechanism, which does not allow for selective low-level expert and router customization and requires additional training. We propose MoIRA, an architecture-agnostic modular MoE framework designed to coordinate existing experts with an external text-based router. MoIRA incorporates two zero-shot routing options: embedding-based similarity and prompt-driven language model inference. In our experiments, we choose large Vision-Language-Action models, gr00t-N1 and $π_0$, as the underlying experts, and train low-rank adapters for low-overhead inference. We evaluate MoIRA on various GR1 Humanoid tasks and LIBERO Spatial and Goal benchmarks, where it consistently outperforms generalist models and competes with other MoE pipelines. Additionally, we analyse the robustness of the proposed approach to the variations of the instructions. While relying solely on textual descriptions of tasks and experts, MoIRA demonstrates the practical viability of modular deployment with precise, low-effort routing and provides an alternative, scalable foundation for future multi-expert robotic systems.

Time series forecasting models have diverse real world applications (e.g., from electricity metrics to software workload). Latest foundational models trained for time series forecasting show strengths (e.g., for long sequences and in zero-shot settings). However, foundational model was not yet used for forecasting rare, spiky events, i.e., a challenging target because those are a corner case of extreme events. In this paper, we optimize a state-of-the-art foundational model to forecast sporadic or spiky production outages of high-performance machine learning services powering billions of client devices. We evaluate the forecasting errors of the foundational model compared with classical stochastic forecasting models (e.g., moving average and autoregressive). The analysis helps us understand how each of the evaluated models performs for the sporadic or spiky events. For example, it identifies the key patterns in the target data that are well tracked by the foundational model vs. each of the stochastic models. We use the models with optimal parameters to estimate a year-long outage statistics of a particular root cause with less than 6% value errors.

The proliferation of multimodal memes in the social media era demands that multimodal Large Language Models (mLLMs) effectively understand meme harmfulness. Existing benchmarks for assessing mLLMs on harmful meme understanding rely on accuracy-based, model-agnostic evaluations using static datasets. These benchmarks are limited in their ability to provide up-to-date and thorough assessments, as online memes evolve dynamically. To address this, we propose AdamMeme, a flexible, agent-based evaluation framework that adaptively probes the reasoning capabilities of mLLMs in deciphering meme harmfulness. Through multi-agent collaboration, AdamMeme provides comprehensive evaluations by iteratively updating the meme data with challenging samples, thereby exposing specific limitations in how mLLMs interpret harmfulness. Extensive experiments show that our framework systematically reveals the varying performance of different target mLLMs, offering in-depth, fine-grained analyses of model-specific weaknesses. Our code is available at https://github.com/Lbotirx/AdamMeme.

Legal work, characterized by its text-heavy and resource-intensive nature, presents unique challenges and opportunities for NLP research. While data-driven approaches have advanced the field, their lack of interpretability and trustworthiness limits their applicability in dynamic legal environments. To address these issues, we collaborated with legal experts to extend an existing dataset and explored the use of Large Language Models (LLMs) and in-context learning to evaluate the legality of clauses in German employment contracts. Our work evaluates the ability of different LLMs to classify clauses as "valid," "unfair," or "void" under three legal context variants: no legal context, full-text sources of laws and court rulings, and distilled versions of these (referred to as examination guidelines). Results show that full-text sources moderately improve performance, while examination guidelines significantly enhance recall for void clauses and weighted F1-Score, reaching 80\%. Despite these advancements, LLMs' performance when using full-text sources remains substantially below that of human lawyers. We contribute an extended dataset, including examination guidelines, referenced legal sources, and corresponding annotations, alongside our code and all log files. Our findings highlight the potential of LLMs to assist lawyers in contract legality review while also underscoring the limitations of the methods presented.

Scientific applications of machine learning often rely on small, specialized models tuned to particular domains. Such models often achieve excellent performance, but lack flexibility. Foundation models offer versatility, but typically underperform specialized approaches, especially on non-traditional modalities and long-tail domains. We propose MARVIS (Modality Adaptive Reasoning over VISualizations), a training-free method that enables even small vision-language models to predict any data modality with high accuracy. MARVIS transforms latent embedding spaces into visual representations and then leverages the spatial and fine-grained reasoning skills of VLMs to successfully interpret and utilize them. MARVIS achieves competitive performance on vision, audio, biological, and tabular domains using a single 3B parameter model, achieving results that beat Gemini by 16\% on average and approach specialized methods, without exposing personally identifiable information (P.I.I.) or requiring any domain-specific training. We open source our code and datasets at https://github.com/penfever/marvis

Large language models (LLMs) excel at reasoning tasks requiring long thought sequences for planning, reflection, and refinement. However, their substantial model size and high computational demands are impractical for widespread deployment. Yet, small language models (SLMs) often struggle to learn long-form CoT reasoning due to their limited capacity, a phenomenon we refer to as the "SLMs Learnability Gap". To address this, we introduce \textbf{Mi}d-\textbf{Co}T \textbf{T}eacher \textbf{A}ssistant Distillation (MiCoTAl), a framework for improving long CoT distillation for SLMs. MiCoTA employs intermediate-sized models as teacher assistants and utilizes intermediate-length CoT sequences to bridge both the capacity and reasoning length gaps. Our experiments on downstream tasks demonstrate that although SLMs distilled from large teachers can perform poorly, by applying MiCoTA, they achieve significant improvements in reasoning performance. Specifically, Qwen2.5-7B-Instruct and Qwen2.5-3B-Instruct achieve an improvement of 3.47 and 3.93 respectively on average score on AIME2024, AMC, Olympiad, MATH-500 and GSM8K benchmarks. To better understand the mechanism behind MiCoTA, we perform a quantitative experiment demonstrating that our method produces data more closely aligned with base SLM distributions. Our insights pave the way for future research into long-CoT data distillation for SLMs.