Chinese essay writing and its evaluation are critical in educational contexts, yet the capabilities of Large Language Models (LLMs) in this domain remain largely underexplored. Existing benchmarks often rely on coarse-grained text quality metrics, largely overlooking the structural and rhetorical complexities of Chinese essays, particularly across diverse genres. To address this gap, we propose \benchName, a multi-genre benchmark specifically designed for Chinese essay writing across four major genres: Argumentative, Narrative, Descriptive, and Expository. We curate and refine a total of 728 real-world prompts to ensure authenticity and meticulously categorize them into the \textit{Open-Ended} and \textit{Constrained} sets to capture diverse writing scenarios. To reliably evaluate generated essays, we develop a fine-grained, genre-specific scoring framework that hierarchically aggregates scores. We further validate our evaluation protocol through a comprehensive human agreement study. Finally, we benchmark 15 large-sized LLMs, analyzing their strengths and limitations across genres and instruction types. With \benchName, we aim to advance LLM-based Chinese essay evaluation and inspire future research on improving essay generation in educational settings.

Measurement systems (e.g., currencies) differ across cultures, but the conversions between them are well defined so that humans can state facts using any measurement system of their choice. Being available to users from diverse cultural backgrounds, large language models (LLMs) should also be able to provide accurate information irrespective of the measurement system at hand. Using newly compiled datasets we test if this is the case for seven open-source LLMs, addressing three key research questions: (RQ1) What is the default system used by LLMs for each type of measurement? (RQ2) Do LLMs' answers and their accuracy vary across different measurement systems? (RQ3) Can LLMs mitigate potential challenges w.r.t. underrepresented systems via reasoning? Our findings show that LLMs default to the measurement system predominantly used in the data. Additionally, we observe considerable instability and variance in performance across different measurement systems. While this instability can in part be mitigated by employing reasoning methods such as chain-of-thought (CoT), this implies longer responses and thereby significantly increases test-time compute (and inference costs), marginalizing users from cultural backgrounds that use underrepresented measurement systems.

Although region-specific large language models (LLMs) are increasingly developed, their safety remains underexplored, particularly in culturally diverse settings like Indonesia, where sensitivity to local norms is essential and highly valued by the community. In this work, we present IndoSafety, the first high-quality, human-verified safety evaluation dataset tailored for the Indonesian context, covering five language varieties: formal and colloquial Indonesian, along with three major local languages: Javanese, Sundanese, and Minangkabau. IndoSafety is constructed by extending prior safety frameworks to develop a taxonomy that captures Indonesia's sociocultural context. We find that existing Indonesian-centric LLMs often generate unsafe outputs, particularly in colloquial and local language settings, while fine-tuning on IndoSafety significantly improves safety while preserving task performance. Our work highlights the critical need for culturally grounded safety evaluation and provides a concrete step toward responsible LLM deployment in multilingual settings. Warning: This paper contains example data that may be offensive, harmful, or biased.

Large language models (LLMs) have revolutionized natural language processing, yet their substantial model sizes often require substantial computational resources. To preserve computing resources and accelerate inference speed, it is crucial to prune redundant parameters, especially for experienced users who often need compact expert models tailored to specific downstream scenarios. However, most existing pruning methods focus on preserving the model's general capabilities, often requiring extensive post-training or suffering from degraded performance due to coarse-grained pruning. In this work, we design a $\underline{Cus}$tom $\underline{Prun}$ing method ($\texttt{Cus-Prun}$) to prune a large general model into a smaller lightweight expert model, which is positioned along the "language", "domain" and "task" dimensions. By identifying and pruning irrelevant neurons of each dimension, $\texttt{Cus-Prun}$ creates expert models without any post-training. Our experiments demonstrate that $\texttt{Cus-Prun}$ consistently outperforms other methods, achieving minimal loss in both expert and general capabilities across various models from different model families and sizes.

We study a common challenge in reinforcement learning for large language models (LLMs): the Zero-Reward Assumption, where non-terminal actions (i.e., intermediate token generations) receive zero task-specific immediate reward, while only the final token receives a reward for the entire response. This assumption arises frequently in practice, as precise token-level rewards are often difficult or infeasible to obtain in LLM applications. In this work, we provide a unifying theoretical perspective. We introduce the Trajectory Policy Gradient Theorem, which shows that the policy gradient based on true, unknown token-level rewards can be unbiasedly estimated using only a response-level reward model, regardless of whether the Zero-Reward Assumption holds or not, for algorithms in the REINFORCE and Actor-Critic families. This result reveals that widely used methods such as PPO, GRPO, ReMax, and RLOO inherently possess the capacity to model token-level reward signals, offering a theoretical justification for response-level reward approaches. Our findings pave the way for more practical, efficient LLM fine-tuning, allowing developers to treat training algorithms as black boxes and focus on improving the response-level reward model with auxiliary sub-models. We also offer a detailed analysis of popular RL and non-RL methods, comparing their theoretical foundations and practical advantages across common LLM tasks. Finally, we propose a new algorithm: Token-Reinforced Policy Optimization (TRePO), a theoretically grounded method that is simpler than PPO, matches GRPO in memory efficiency, and holds promise for broad applicability.

Machine unlearning is used to mitigate the privacy risks of Large Vision-Language Models (LVLMs) arising from training on large-scale web data. However, existing unlearning methods often fail to carefully select substitute outputs for forget targets, resulting in Unlearning Aftermaths-undesirable behaviors such as degenerate, hallucinated, or excessively refused responses. We highlight that, especially for generative LVLMs, it is crucial to consider the quality and informativeness of post-unlearning responses rather than relying solely on naive suppression. To address this, we introduce a new unlearning task for LVLMs that requires models to provide privacy-preserving yet informative and visually grounded responses. We also propose PUBG, a novel unlearning method that explicitly guides post-unlearning behavior toward a desirable output distribution. Experiments show that, while existing methods suffer from Unlearning Aftermaths despite successfully preventing privacy violations, PUBG effectively mitigates these issues, generating visually grounded and informative responses without privacy leakage for forgotten targets.

Recent advancements in Large Language Models (LLMs) and Reinforcement Learning (RL) have shown significant promise in decision-making tasks. Nevertheless, for large-scale industrial decision problems, both approaches face distinct challenges: LLMs lack real-time long-sequence decision-making capabilities, while RL struggles with sample efficiency in vast action spaces. To bridge this gap, we propose Agents Co-Evolution (ACE), a synergistic framework between LLMs and RL agents for large-scale decision-making scenarios. ACE introduces a dual-role trajectory refinement mechanism where LLMs act as both Policy Actor and Value Critic during RL's training: the Actor refines suboptimal actions via multi-step reasoning and environment validation, while the Critic performs temporal credit assignment through trajectory-level reward shaping. Concurrently, RL agent enhances LLMs' task-specific decision-making with high-quality fine-tuning datasets generated via prioritized experience replay. Through extensive experiments across multiple power grid operation challenges with action spaces exceeding 60K discrete actions, ACE demonstrates superior performance over existing RL methods and LLM-based methods.

LLMssuch as GPT-4 have shown a remarkable ability to solve complex questions by generating step-by-step rationales. Prior works have utilized this capability to improve smaller and cheaper LMs (say, with 7B parameters). However, various practical constraints, such as copyright and legal issues, owing to lack of transparency in the pre-training data of large (often closed) models, prevent their use in commercial settings. Little focus has been given to improving the innate reasoning ability of smaller models without distilling information from larger LLMs. To address this, we propose COLLATE, a trainable framework that tunes a (small) LLM to generate those outputs from a pool of diverse rationales that selectively improves the downstream task. COLLATE enforces multiple instances of the same LLM to exhibit distinct behavior and employs them to generate rationales to obtain diverse outputs. The LLM is then tuned via preference optimization to choose the candidate rationale which maximizes the likelihood of ground-truth answer. COLLATE outperforms several trainable and prompting baselines on 5 datasets across 3 domains: maths problem solving, natural language inference, and commonsense reasoning. We show the eff icacy of COLLATE on LLMs from different model families across varying parameter scales (1B to 8B) and demonstrate the benefit of multiple rationale providers guided by the end task through ablations. Code is released here (https://github.com/Sohanpatnaik106/collate).

Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to access broader knowledge sources, yet factual inconsistencies persist due to noise in retrieved documents-even with advanced retrieval methods. We demonstrate that enhancing generative models' capacity to process noisy content is equally critical for robust performance. In this paper, we present KARE-RAG (Knowledge-Aware Refinement and Enhancement for RAG), which improves knowledge utilization through three key innovations: (1) structured knowledge representations that facilitate error detection during training, (2) Dense Direct Preference Optimization (DDPO)-a refined training objective that prioritizes correction of critical errors, and (3) a contrastive data generation pipeline that maintains semantic consistency while rectifying factual inaccuracies. Experiments show our method significantly enhances standard RAG pipelines across model scales, improving both in-domain and out-of-domain task performance without compromising general capabilities. Notably, these gains are achieved with modest training data, suggesting data-efficient optimization is possible through targeted learning strategies. Our findings establish a new direction for RAG improvement: by improving how models learn to process retrieved content, we can enhance performance across diverse inference paradigms. All data and code will be publicly available on Github.

Evaluation is important for multimodal generation tasks. With the rapid progress of MLLMs, there is growing interest in applying MLLMs to build general evaluation systems. However, existing work overlooks two aspects: (1) the development of evaluation capabilities for text-to-image (T2I) generation task, and (2) the incorporation of large-scale human evaluation data. In this paper, we introduce Minos-Corpus, a large-scale multimodal evaluation dataset that combines evaluation data from both human and GPT. The corpus contains evaluation data across both image-to-text(I2T) and T2I generation tasks. Based on this corpus, we propose Data Selection and Balance, Mix-SFT training methods, and apply DPO to develop Minos, a multimodal evaluation model built upon a 7B backbone. Minos achieves state-of-the-art (SoTA) performance among all open-source evaluation models of similar scale on the average of evaluation performance on all tasks, and outperforms all open-source and closed-source models on evaluation of T2I generation task. Extensive experiments demonstrate the importance of leveraging high-quality human evaluation data and jointly training on evaluation data from both I2T and T2I generation tasks.