Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ”Deceitful” and ”Manipulative”, often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

Large language models (LLMs) are increasingly entrusted with high-stakes decisions that affect human welfare. However, the principles and values that guide these models when distributing scarce societal resources remain largely unexamined. To address this, we introduce the Social Welfare Function (SWF) Benchmark, a dynamic simulation environment in which an LLM acts as a dictator, distributing tasks to heterogeneous recipients with different returns on investment (ROI). The benchmark is designed to create a dilemma between maximizing collective efficiency (i.e., overall ROI) and ensuring distributive fairness (measured by the Gini coefficient). We evaluate 20 state-of-the-art LLMs. Our findings reveal several key insights, including: (i) LLMs’ general ability, as measured by popular Arena leaderboards, misaligns with their allocation skills; (ii) Most LLMs exhibit a strong default utilitarian orientation, prioritizing overall productivity at the expense of inequality. (iii) Allocation behaviors are highly manipulated, easily perturbed by common persuasion strategies. These results highlight the risks of deploying current LLMs as societal decision-makers and underscore the need for specialized benchmarks and alignment for AI governance.

Video captioning directly reflects a Video Large Language Model’s perception, making caption fidelity a strong proxy for audio-visual grounding. Existing approaches often rely on costly annotations or stronger external teachers, limiting scalability. In this paper, we instead explore whether a VideoLLM can improve its own captions by exploiting structure already present within the model through self-supervision. Our exploration builds on two asymmetries we observe in current VideoLLMs. The first is a task asymmetry: captioning requires broad open-ended perception, whereas question answering narrows reasoning to specific dimensions. The second is an information asymmetry: the same model reasons more reliably with richer context or higher-fidelity audio-visual inputs. Motivated by these observations, we propose SODA (Self-Supervised On-Policy Distillation via Dual-Stream Auditing), a framework that enables a VideoLLM to critique and refine its own captions. A student first produces an initial caption from a low-budget video view, followed by a dual-stream audit: a Caption QA Stream identifies factual errors, and a Video QA Stream detects omitted content. Conditioned on this Revision QA Bank, a teacher using a larger input budget rewrites the caption, and the refined outputs are distilled back into the student. Experiments suggest that, despite training only on the captioning objective, SODA improves caption fidelity, reduces hallucination, and remains competitive on general video understanding.

Classical Chinese poetry generation and evaluation with Large Language Models (LLMs) face two intertwined challenges: existing generation benchmarks rely on simplistic keyword inputs devoid of literary context, and no ground-truth benchmark exists for objectively validating poetry evaluation methods. We present a comprehensive framework addressing both challenges. On the **generation** side, we construct **Honglou-Poem**, the first narrative-driven poetry generation benchmark derived from Dream of the Red Chamber, where each entry provides rich author personas and narrative backgrounds—moving far beyond traditional keyword-to-poem paradigms. A ”Digital Poetry Party” pipeline evaluates 12 state-of-the-art LLMs in this challenging setting. On the **evaluation** side, we introduce **Poetry-Tree**, a benchmark with axiomatically ordered quality levels created through systematic merit degradation, providing objective ground truth for evaluator validation. Leveraging Poetry-Tree as a training signal, we employ **GRPO** reinforcement learning with Spearman’s correlation as the reward to train a dedicated poetry evaluation model, **Poet-4B**. Our Poet-4B achieves a Spearman’s 𝜌 of **0.8253**, dramatically improving from the base model’s 0.2929 and closely approaching the best closed-source model GLM-4.6 (0.8764) with over 100× fewer parameters—a gap of only 0.051 in Spearman’s 𝜌. We release all benchmarks, trained models, and a leaderboard to facilitate future research.

Long-term memory systems allow LLM agents to preserve information beyond a single context window, but most systems focus on storing and retrieving facts after extraction, leaving the write decision under-specified. What deserves memory can depend on the user’s current task, topic, activity, or interaction partner, while uniform extraction applies one notion of importance across these different situations. We formulate this challenge as preference-conditioned write control and introduce AdaMem, which uses adaptive natural-language Memory Policies to personalize what an agent writes to memory. Each policy represents the user’s memory preference for a particular interaction context, is updated from periodic feedback, and controls subsequent memory writing. We evaluate this loop in AdaMem-Bench, which assigns different memory preferences to six concurrent interaction personas across five ten-week stories. Across two extraction models and two feedback modes, AdaMem improves average QA accuracy over Mem0 from 80.0% to 84.35% while reducing persistent memory by 9.27%. Our analyses show that explicit feedback helps models learn better memory policies, but current models still struggle to translate those policies into reliably selective writing behavior. AdaMem thus demonstrates the promise of adaptive write control while exposing policy execution as a central limitation of current memory agents. Our code is publicly available.[<https://github.com/galaxyChen/AdaMem>]

Fixed benchmarks are costly to renew and cannot adapt their questions to model-specific failures. We ask whether LLMs can instead discover one another’s weaknesses and turn those observations into an evaluation process. To study this question, we introduce LivingArena, an automated peer-probing framework in which models take turns testing one another. Using the interaction history, each questioner identifies potential weaknesses of its opponent and constructs targeted, verifiable questions to probe them. A 3,600-round tournament of ten models reveals a clear role asymmetry: strong answerers are not always reliable questioners, because they may generate internally inconsistent tests or fail to verify their own reference answers. After a questioner exposes an answerer’s failure, it is more likely to pursue the same capability domain, while the answerer’s weakness recurs on independently generated questions, including questions written by different models. These findings show that peer probing can reveal persistent model-specific weaknesses while separately evaluating answering and reliable test construction. By automating this process and allowing test difficulty to evolve with model capabilities, LivingArena provides a “living” benchmark for model development, red-teaming, and capability-aware multi-agent coordination. We publicly release our code.[<https://github.com/galaxyChen/LivingArena>]

The rise of Large Language Models (LLMs) is reshaping multimodel models, with speech synthesis being a prominent application. However, existing approaches often underutilize the linguistic intelligence of these models, typically failing to leverage their powerful instruction-following capabilities. This limitation hinders the model’s ability to follow text instructions for controllable Text-to-Speech (TTS). To address this, we propose a new paradigm inspired by operationalism that decouples instruction understanding from speech generation. We introduce BatonVoice, a framework where an LLM acts as a conductor, understanding user instructions and generating a textual plan – explicit vocal features (e.g., pitch, energy). A separate TTS model, the orchestra, then generates the speech from these features. To realize this component, we develop BatonTTS, a TTS model trained specifically for this task. Our experiments demonstrate that BatonVoice achieves strong performance in controllable and emotional speech synthesis, outperforming strong open- and closed-source baselines. Notably, our approach enables remarkable zero-shot cross-lingual generalization, accurately applying feature control abilities to languages unseen during post-training. This demonstrates that objectifying speech into textual vocal features can more effectively unlock the linguistic intelligence of LLMs.