Nearly all human work is collaborative; thus, the evaluation of real-world NLP applications often requires multiple dimensions that align with diverse human perspectives. As real human evaluator resources are often scarce and costly, the emerging "LLM-as-a-judge" paradigm sheds light on a promising approach to leverage LLM agents to believably simulate human evaluators. Yet, to date, existing LLM-as-a-judge approaches face two limitations: persona descriptions of agents are often arbitrarily designed, and the frameworks are not generalizable to other tasks. To address these challenges, we propose MAJ-EVAL, a Multi-Agent-as-Judge evaluation framework that can automatically construct multiple evaluator personas with distinct dimensions from relevant text documents (e.g., research papers), instantiate LLM agents with the personas, and engage in-group debates with multi-agents to Generate multi-dimensional feedback. Our evaluation experiments in both the educational and medical domains demonstrate that MAJ-EVAL can generate evaluation results that better align with human experts' ratings compared with conventional automated evaluation metrics and existing LLM-as-a-judge methods.

Given the vastness of chemical space and the ongoing emergence of previously uncharacterized proteins, zero-shot compound-protein interaction (CPI) prediction better reflects the practical challenges and requirements of real-world drug development. Although existing methods perform adequately during certain CPI tasks, they still face the following challenges: (1) Representation learning from local or complete protein sequences often overlooks the complex interdependencies between subsequences, which are essential for predicting spatial structures and binding properties. (2) Dependence on large-scale or scarce multimodal protein datasets demands significant training data and computational resources, limiting scalability and efficiency. To address these challenges, we propose a novel approach that pretrains protein representations for CPI prediction tasks using subsequence reordering, explicitly capturing the dependencies between protein subsequences. Furthermore, we apply length-variable protein augmentation to ensure excellent pretraining performance on small training datasets. To evaluate the model's effectiveness and zero-shot learning ability, we combine it with various baseline methods. The results demonstrate that our approach can improve the baseline model's performance on the CPI task, especially in the challenging zero-shot scenario. Compared to existing pre-training models, our model demonstrates superior performance, particularly in data-scarce scenarios where training samples are limited. Our implementation is available at https://github.com/Hoch-Zhang/PSRP-CPI.

Modeling user behavior is critical across many industries where understanding preferences, intent, or decisions informs personalization, targeting, and strategic outcomes. Surveys have long served as a classical mechanism for collecting such behavioral data due to their interpretability, structure, and ease of deployment. However, surveys alone are inherently limited by user fatigue, incomplete responses, and practical constraints on their length making them insufficient for capturing user behavior. In this work, we present LANTERN (Late-Attentive Network for Enriched Response Modeling), a modular architecture for modeling user behavior by fusing adaptive survey responses with supplemental contextual signals. We demonstrate the architectural value of maintaining survey primacy through selective gating, residual connections and late fusion via cross-attention, treating survey data as the primary signal while incorporating external modalities only when relevant. LANTERN outperforms strong survey-only baselines in multi-label prediction of survey responses. We further investigate threshold sensitivity and the benefits of selective modality reliance through ablation and rare/frequent attribute analysis. LANTERN's modularity supports scalable integration of new encoders and evolving datasets. This work provides a practical and extensible blueprint for behavior modeling in survey-centric applications.

To facilitate the wider adoption of robotics, accessible programming tools are required for non-experts. Observational learning enables intuitive human skills transfer through hands-on demonstrations, but relying solely on visual input can be inefficient in terms of scalability and failure mitigation, especially when based on a single demonstration. This paper presents a human-in-the-loop method for enhancing the robot execution plan, automatically generated based on a single RGB video, with natural language input to a Large Language Model (LLM). By including user-specified goals or critical task aspects and exploiting the LLM common-sense reasoning, the system adjusts the vision-based plan to prevent potential failures and adapts it based on the received instructions. Experiments demonstrated the framework intuitiveness and effectiveness in correcting vision-derived errors and adapting plans without requiring additional demonstrations. Moreover, interactive plan refinement and hallucination corrections promoted system robustness.

Visual imitation learning (VIL) provides an efficient and intuitive strategy for robotic systems to acquire novel skills. Recent advancements in foundation models, particularly Vision Language Models (VLMs), have demonstrated remarkable capabilities in visual and linguistic reasoning for VIL tasks. Despite this progress, existing approaches primarily utilize these models for learning high-level plans from human demonstrations, relying on pre-defined motion primitives for executing physical interactions, which remains a major bottleneck for robotic systems. In this work, we present FMimic, a novel paradigm that harnesses foundation models to directly learn generalizable skills at even fine-grained action levels, using only a limited number of human videos. Extensive experiments demonstrate that our FMimic delivers strong performance with a single human video, and significantly outperforms all other methods with five videos. Furthermore, our method exhibits significant improvements of over 39% and 29% in RLBench multi-task experiments and real-world manipulation tasks, respectively, and exceeds baselines by more than 34% in high-precision tasks and 47% in long-horizon tasks.

Psychology research has shown that humans are poor at estimating their performance on tasks, tending towards underconfidence on easy tasks and overconfidence on difficult tasks. We examine three LLMs, Llama-3-70B-instruct, Claude-3-Sonnet, and GPT-4o, on a range of QA tasks of varying difficulty, and show that models exhibit subtle differences from human patterns of overconfidence: less sensitive to task difficulty, and when prompted to answer based on different personas -- e.g., expert vs layman, or different race, gender, and ages -- the models will respond with stereotypically biased confidence estimations even though their underlying answer accuracy remains the same. Based on these observations, we propose Answer-Free Confidence Estimation (AFCE) to improve confidence calibration and LLM interpretability in these settings. AFCE is a self-assessment method that employs two stages of prompting, first eliciting only confidence scores on questions, then asking separately for the answer. Experiments on the MMLU and GPQA datasets spanning subjects and difficulty show that this separation of tasks significantly reduces overconfidence and delivers more human-like sensitivity to task difficulty.

We introduce Kimi K2, a Mixture-of-Experts (MoE) large language model with 32 billion activated parameters and 1 trillion total parameters. We propose the MuonClip optimizer, which improves upon Muon with a novel QK-clip technique to address training instability while enjoying the advanced token efficiency of Muon. Based on MuonClip, K2 was pre-trained on 15.5 trillion tokens with zero loss spike. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage, where the model improves its capabilities through interactions with real and synthetic environments. Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, with strengths in agentic capabilities. Notably, K2 obtains 66.1 on Tau2-Bench, 76.5 on ACEBench (En), 65.8 on SWE-Bench Verified, and 47.3 on SWE-Bench Multilingual -- surpassing most open and closed-sourced baselines in non-thinking settings. It also exhibits strong capabilities in coding, mathematics, and reasoning tasks, with a score of 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, 75.1 on GPQA-Diamond, and 27.1 on OJBench, all without extended thinking. These results position Kimi K2 as one of the most capable open-source large language models to date, particularly in software engineering and agentic tasks. We release our base and post-trained model checkpoints to facilitate future research and applications of agentic intelligence.

In the context of the rapidly evolving information technology landscape, marked by the advent of 6G communication networks, we face an increased data volume and complexity in network environments. This paper addresses these challenges by focusing on Quality of Service (QoS) in edge computing frameworks. We propose a novel approach to enhance QoS through the development of General Artificial Intelligence Lifelong Learning Systems, with a special emphasis on Federated Layering Techniques (FLT). Our work introduces a federated layering-based small model collaborative mechanism aimed at improving AI models' operational efficiency and response time in environments where resources are limited. This innovative method leverages the strengths of cloud and edge computing, incorporating a negotiation and debate mechanism among small AI models to enhance reasoning and decision-making processes. By integrating model layering techniques with privacy protection measures, our approach ensures the secure transmission of model parameters while maintaining high efficiency in learning and reasoning capabilities. The experimental results demonstrate that our strategy not only enhances learning efficiency and reasoning accuracy but also effectively protects the privacy of edge nodes. This presents a viable solution for achieving resilient large model lifelong learning systems, with a significant improvement in QoS for edge computing environments.

--Transfer Learning (TL) offers the potential to accelerate learning by transferring knowledge across tasks. However, it faces critical challenges such as negative transfer, domain adaptation and inefficiency in selecting solid source policies. These issues often represent critical problems in evolving domains, i.e. game development, where scenarios transform and agents must adapt. The continuous release of new agents is costly and inefficient. In this work we challenge the key issues in TL to improve knowledge transfer, agents performance across tasks and reduce computational costs. The proposed methodology, called F AST - Framework for Adaptive Similarity-based Transfer, leverages visual frames and textual descriptions to create a latent representation of tasks dynamics, that is exploited to estimate similarity between environments. The similarity scores guides our method in choosing candidate policies from which transfer abilities to simplify learning of novel tasks. Experimental results, over multiple racing tracks, demonstrate that F AST achieves competitive final performance compared to learning-from-scratch methods while requiring significantly less training steps. Learning is often thought of as a process rooted in interactions with the environment. Reinforcement Learning (RL) expands on this core concept by viewing learning as a trial-and error process, in which agents engage with the environment, make choices, and receive feedback in the form of reward or penalties. Traditionally, agents are trained from scratch to accomplish a single task, requiring extensive interactions with the environment to achieve proficiency far more than a human would need for comparable tasks. One primary challenge in RL is the substantial computational demands imposed by simulation, where training time and data requirements scale up for complex tasks. In game development and other evolving environments it is expensive and sub-optimal to start at each iteration from zero.

Causal discovery from observational data is challenging, especially with large datasets and complex relationships. Traditional methods often struggle with scalability and capturing global structural information. To overcome these limitations, we introduce a novel graph neural network (GNN)-based probabilistic framework that learns a probability distribution over the entire space of causal graphs, unlike methods that output a single deterministic graph. Our framework leverages a GNN that encodes both node and edge attributes into a unified graph representation, enabling the model to learn complex causal structures directly from data. The GNN model is trained on a diverse set of synthetic datasets augmented with statistical and information-theoretic measures, such as mutual information and conditional entropy, capturing both local and global data properties. We frame causal discovery as a supervised learning problem, directly predicting the entire graph structure. Our approach demonstrates superior performance, outperforming both traditional and recent non-GNN-based methods, as well as a GNN-based approach, in terms of accuracy and scalability on synthetic and real-world datasets without further training. This probabilistic framework significantly improves causal structure learning, with broad implications for decision-making and scientific discovery across various fields.