Plasticity, or the ability of an agent to adapt to new tasks, environments, or distributions, is crucial for continual learning. In this paper, we study the loss of plasticity in deep continual RL from the lens of churn: network output variability for out-of-batch data induced by mini-batch training. We demonstrate that (1) the loss of plasticity is accompanied by the exacerbation of churn due to the gradual rank decrease of the Neural Tangent Kernel (NTK) matrix; (2) reducing churn helps prevent rank collapse and adjusts the step size of regular RL gradients adaptively. Moreover, we introduce Continual Churn Approximated Reduction (C-CHAIN) and demonstrate it improves learning performance and outperforms baselines in a diverse range of continual learning environments on OpenAI Gym Control, ProcGen, DeepMind Control Suite, and MinAtar benchmarks.

A retriever, which retrieves relevant knowledge pieces from a knowledge base given a context, is an important component in many natural language processing (NLP) tasks. Retrievers have been introduced in knowledge-grounded dialog systems to improve knowledge acquisition. In knowledge-grounded dialog systems, when conditioning on a given context, there may be multiple relevant and correlated knowledge pieces. However, knowledge pieces are usually assumed to be conditionally independent in current retriever models. To address this issue, we propose Entriever, an energy-based retriever. Entriever directly models the candidate retrieval results as a whole instead of modeling the knowledge pieces separately, with the relevance score defined by an energy function. We explore various architectures of energy functions and different training methods for Entriever, and show that Entriever substantially outperforms the strong cross-encoder baseline in knowledge retrieval tasks. Furthermore, we show that in semi-supervised training of knowledge-grounded dialog systems, Entriever enables effective scoring of retrieved knowledge pieces and significantly improves end-to-end performance of dialog systems.

Social media platforms have become central to modern communication, yet they also harbor offensive content that challenges platform safety and inclusivity. While prior research has primarily focused on textual indicators of offense, the role of emojis, ubiquitous visual elements in online discourse, remains underexplored. Emojis, despite being rarely offensive in isolation, can acquire harmful meanings through symbolic associations, sarcasm, and contextual misuse. In this work, we systematically examine emoji contributions to offensive Twitter messages, analyzing their distribution across offense categories and how users exploit emoji ambiguity. To address this, we propose an LLM-powered, multi-step moderation pipeline that selectively replaces harmful emojis while preserving the tweet's semantic intent. Human evaluations confirm our approach effectively reduces perceived offensiveness without sacrificing meaning. Our analysis also reveals heterogeneous effects across offense types, offering nuanced insights for online communication and emoji moderation.

Directly training Large Language Models (LLMs) for Multi-Agent Systems (MAS) remains challenging due to intricate reward modeling, dynamic agent interactions, and demanding generalization requirements. This paper explores whether post-training techniques, specifically Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR), can effectively $\textit{generalize}$ to multi-agent scenarios. We use economic reasoning as a testbed, leveraging its strong foundations in mathematics and game theory, its demand for structured analytical reasoning, and its relevance to real-world applications such as market design, resource allocation, and policy analysis. We introduce $\textbf{Recon}$ ($\textbf{R}$easoning like an $\textbf{ECON}$omist), a 7B-parameter open-source LLM post-trained on a hand-curated dataset of 2,100 high-quality economic reasoning problems. Comprehensive evaluation on economic reasoning benchmarks and multi-agent games reveals clear improvements in structured reasoning and economic rationality. These results underscore the promise of domain-aligned post-training for enhancing reasoning and agent alignment, shedding light on the roles of SFT and RL in shaping model behavior. Code is available at https://github.com/MasterZhou1/Recon .

Modern wireless networks must adapt to dynamic conditions while efficiently managing diverse service demands. Traditional deep reinforcement learning (DRL) struggles in these environments, as scattered and evolving feedback makes optimal decision-making challenging. Large Language Models (LLMs) offer a solution by structuring unorganized network feedback into meaningful latent representations, helping RL agents recognize patterns more effectively. For example, in O-RAN slicing, concepts like SNR, power levels and throughput are semantically related, and LLMs can naturally cluster them, providing a more interpretable state representation. To leverage this capability, we introduce a contextualization-based adaptation method that integrates learnable prompts into an LLM-augmented DRL framework. Instead of relying on full model fine-tuning, we refine state representations through task-specific prompts that dynamically adjust to network conditions. Utilizing ORANSight, an LLM trained on O-RAN knowledge, we develop Prompt-Augmented Multi agent RL (PA-MRL) framework. Learnable prompts optimize both semantic clustering and RL objectives, allowing RL agents to achieve higher rewards in fewer iterations and adapt more efficiently. By incorporating prompt-augmented learning, our approach enables faster, more scalable, and adaptive resource allocation in O-RAN slicing. Experimental results show that it accelerates convergence and outperforms other baselines.

When aligning large language models (LLMs), their performance on various tasks (such as being helpful, harmless, and honest) depends heavily on the composition of their training data. However, selecting a data mixture that achieves strong performance across all tasks is challenging. Existing approaches rely on large ablation studies, heuristics, or human intuition, but these can be prohibitively expensive and suboptimal. We study this problem in the setting of preference optimization via DPO and introduce AutoMixAlign (AMA), a theoretically-grounded algorithm that adaptively mixes datasets during training to balance performance across tasks. AMA first trains \textit{specialist models} for each task to determine losses that correspond to strong task performance. Then, it trains a generalist model using a novel minimax optimization that prioritizes tasks for which generalist model losses deviate most from specialist model losses. To optimize this problem, we propose two algorithms: (1) AMA-R, which adaptively reweights the objective to prioritize tasks, and (2) AMA-S, which adaptively adjusts how much data is sampled from each task to prioritize tasks. Both algorithms achieve a convergence rate of $O(1/\sqrt{T})$ in the convex case. AMA-R's convergence result follows from Sagawa et al. (2019), and we provide a convergence proof for AMA-S using online learning techniques such as EXP3. We evaluate AMA on several multitask alignment setups and find that AMA outperforms the standard alignment approach -- which simply optimizes the total loss across all tasks -- and also outperforms model merging methods.

Evaluation frameworks for text summarization have evolved in terms of both domain coverage and metrics. However, existing benchmarks still lack domain-specific assessment criteria, remain predominantly English-centric, and face challenges with human annotation due to the complexity of reasoning. To address these, we introduce MSumBench, which provides a multi-dimensional, multi-domain evaluation of summarization in English and Chinese. It also incorporates specialized assessment criteria for each domain and leverages a multi-agent debate system to enhance annotation quality. By evaluating eight modern summarization models, we discover distinct performance patterns across domains and languages. We further examine large language models as summary evaluators, analyzing the correlation between their evaluation and summarization capabilities, and uncovering systematic bias in their assessment of self-generated summaries. Our benchmark dataset is publicly available at https://github.com/DISL-Lab/MSumBench.

Existing attacks against multimodal language models (MLLMs) primarily communicate instructions through text accompanied by adversarial images. In contrast, we exploit the capabilities of MLLMs to interpret non-textual instructions, specifically, adversarial images or audio generated by our novel method, Con Instruction. We optimize these adversarial examples to align closely with target instructions in the embedding space, revealing the detrimental implications of MLLMs' sophisticated understanding. Unlike prior work, our method does not require training data or preprocessing of textual instructions. While these non-textual adversarial examples can effectively bypass MLLM safety mechanisms, their combination with various text inputs substantially amplifies attack success. We further introduce a new Attack Response Categorization (ARC) framework, which evaluates both the quality of the model's response and its relevance to the malicious instructions. Experimental results demonstrate that Con Instruction effectively bypasses safety mechanisms in multiple vision- and audio-language models, including LLaVA-v1.5, InternVL, Qwen-VL, and Qwen-Audio, evaluated on two standard benchmarks: AdvBench and SafeBench. Specifically, our method achieves the highest attack success rates, reaching 81.3% and 86.6% on LLaVA-v1.5 (13B). On the defense side, we explore various countermeasures against our attacks and uncover a substantial performance gap among existing techniques. Our implementation is made publicly available.

Knowledge editing aims to efficiently update Large Language Models (LLMs) by modifying specific knowledge without retraining the entire model. Among knowledge editing approaches, in-context editing (ICE) offers a lightweight solution by injecting new knowledge directly into the input context, leaving model parameters unchanged. However, existing ICE approaches do not explicitly separate the newly injected knowledge from the model's original reasoning process. This entanglement often results in conflicts between external updates and internal parametric knowledge, undermining the consistency and accuracy of the reasoning path.In this work, we conduct preliminary experiments to examine how parametric knowledge influences reasoning path planning. We find that the model's reasoning is tightly coupled with its internal knowledge, and that naively injecting new information without adapting the reasoning path often leads to performance degradation, particularly in multi-hop tasks. To this end, we propose DecKER, a novel ICE framework that decouples reasoning from knowledge editing by generating a masked reasoning path and then resolving knowledge edits via hybrid retrieval and model-based validation. Experiments on multi-hop QA benchmarks show that DecKER significantly outperforms existing ICE methods by mitigating knowledge conflicts and preserving reasoning consistency. Our code is available at: https://github.com/bebr2/DecKER .

The integration of wind energy into power grids necessitates accurate ultra-short-term wind power forecasting to ensure grid stability and optimize resource allocation. This study introduces M2WLLM, an innovative model that leverages the capabilities of Large Language Models (LLMs) for predicting wind power output at granular time intervals. M2WLLM overcomes the limitations of traditional and deep learning methods by seamlessly integrating textual information and temporal numerical data, significantly improving wind power forecasting accuracy through multi-modal data. Its architecture features a Prompt Embedder and a Data Embedder, enabling an effective fusion of textual prompts and numerical inputs within the LLMs framework. The Semantic Augmenter within the Data Embedder translates temporal data into a format that the LLMs can comprehend, enabling it to extract latent features and improve prediction accuracy. The empirical evaluations conducted on wind farm data from three Chinese provinces demonstrate that M2WLLM consistently outperforms existing methods, such as GPT4TS, across various datasets and prediction horizons. The results highlight LLMs' ability to enhance accuracy and robustness in ultra-short-term forecasting and showcase their strong few-shot learning capabilities.