With the growing popularity of virtual character platforms like Character.AI, users are increasingly turning to role-playing agents for emotional support in daily life. Yet existing research mainly focuses on character consistency in fictional or game-based scenarios, overlooking user-centered interactions such as companionship and psychological support. To bridge this gap, we propose Emotionally Supportive Role-Playing (ESRP), a framework designed to align role-playing with real-world user scenarios and emotional needs. We focus on typical users of these platforms, i.e., anime enthusiasts—including students, office workers, freelancers, and self-employed individuals—and design scenario-based questions that reflect their everyday struggles such as work stress and social loneliness. Through a two-round data collection involving 40 anime fans and 10 Large Language Models (LLMs), we build ChatAnime: the first ESRP dataset with 2,400 human-written and 24,000 LLM-generated responses, supported by over 132,000 fine-grained human annotations. We also provide the ESRP evaluation framework featuring 9 fine-grained metrics across three dimensions: basic dialogue, role-playing and emotional support, along with an overall metric for diversity. Experimental results under our evaluation setting show that top-performing LLMs surpass anime fans in role-playing and emotional support, while humans still lead in diversity.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Revisiting Non-Verbatim Memorization in Large Language Models: The Role of Entity Surface Forms
PDF ↗Understanding what kinds of factual knowledge large language models (LLMs) memorize is essential for evaluating their reliability and limitations.Entity-based QA is a common framework for analyzing non-verbatim memorization, but typical evaluations query each entity using a single canonical surface form, making it difficult to disentangle fact memorization from access through a particular name.We introduce RedirectQA, an entity-based QA dataset that uses Wikipedia redirect information to associate Wikidata factual triples with categorized surface forms for each entity, including alternative names, abbreviations, spelling variants, and common erroneous forms.Across 13 LLMs, we examine surface-conditioned factual memorization and find that prediction outcomes often change when only the entity surface form changes.This inconsistency is category-dependent: models are more robust to minor orthographic variations than to larger lexical variations such as aliases and abbreviations.Frequency analyses further suggest that both entity- and surface-level frequencies are associated with accuracy, and that entity frequency often contributes beyond surface frequency.Overall, factual memorization appears neither purely surface-specific nor fully surface-invariant, highlighting the importance of surface-form diversity in evaluating non-verbatim memorization.
DREAM-S: Speculative Decoding with Searchable Drafting and Target-Aware Refinement for Multimodal Generation
PDF ↗Speculative decoding (SD) has proven to be an effective technique for accelerating autoregressive generation in large language models (LLMs), however its application to vision-language models (VLMs) remains relatively unexplored. We propose DREAM-S, a novel SD framework designed specifically for fast and efficient decoding in VLMs. DREAM-S leverages a neural architecture search (NAS) framework with target-aware supernet training to automatically identify both the optimal interaction strategy between the draft and target models, and the most suitable draft model architecture for the underlying hardware implementation platform. DREAM-S additionally incorporates adaptive intermediate feature distillation, guided by attention entropy, to enable efficient draft training. Experiments on a range of well-established VLMs show that DREAM-S achieves up to a 3.85\times speedup compared to standard decoding approaches and significantly outperforms existing SD baselines.
Evaluating the quality of tables generated by large language models (LLMs) remains an open challenge: existing metrics either flatten tables into text, ignoring structure, or rely on fixed references that limit generalization. We present TabReX, a reference-less, property-driven framework for evaluating tabular generation via graph-based reasoning. TabReX converts both source text and generated tables into canonical knowledge graphs, aligns them through an LLM-guided matching process, and computes interpretable, rubric-aware scores that quantify structural and factual fidelity. The resulting metric provides controllable trade-offs between sensitivity and specificity, yielding human-aligned judgments and cell-level error traces. To systematically asses metric robustness, we introduce TabReX-Bench, a large-scale benchmark spanning six domains and twelve planner-driven perturbation types across three difficulty tiers. Empirical results show that TabReX achieves the highest correlation with expert rankings, remains stable under harder perturbations, and enables fine-grained model-vs-prompt analysis establishing a new paradigm for trustworthy, explainable evaluation of structured generation systems.
Persona-assigned Large Language Models can adopt diverse roles, enabling personalized and context-aware reasoning. However, even minor demographic perturbations in personas, such as simple pronoun swaps, can alter reasoning trajectories, leading to divergent sets of correct answers on reasoning benchmarks. We explore the potential of these variations as a constructive resource to improve LLM reasoning performance. We propose CHOIR (Collaborative Harmonization fOr Inference Robustness), a test-time framework that harmonizes a set of demographically perturbed, persona-conditioned reasoning signals into a unified prediction. CHOIR orchestrates a collaborative decoding process among counterfactual personas perturbed across dimensions of gender, race, religion, disability, and age, dynamically balancing agreement and divergence in their reasoning paths to improve performance. Experiments demonstrate that CHOIR consistently enhances LLM reasoning across model architectures, scales, and tasks. Improvements reach up to 20.1% for individual groups and 15.1% on average, and we show that CHOIR remains effective even when base personas are suboptimal.
Many approaches to LLM red-teaming leverage an attacker LLM to discover jailbreaks against a target. Several of them task the attacker with identifying effective strategies through trial and error, resulting in a semantically limited range of successes. Another approach discovers diverse attacks by combining crowdsourced harmful queries and tactics into instructions for the attacker, but does so at random, limiting effectiveness. This article introduces a novel framework, Adaptive Instruction Composition, that combines crowdsourced texts according to an adaptive mechanism trained to jointly optimize effectiveness with diversity. We use reinforcement learning to balance exploration with exploitation in a combinatorial space of instructions to guide the attacker toward diverse generations tailored to target vulnerabilities. We demonstrate that our approach substantially outperforms random combination on a set of effectiveness and diversity metrics, even under model transfer. Further, we show that it surpasses a host of recent adaptive approaches on Harmbench. We employ a lightweight neural contextual bandit that adapts to contrastive embedding inputs, and provide ablations suggesting that the contrastive pretraining enables the network to rapidly generalize and scale to the massive space as it learns.
Causal explanations in political narratives are often framed and contested. Different sources may explain the same event by assigning responsibility to different actors and expressing varying levels of certainty. Standard Event Causality Identification (ECI) focuses on detecting causal links and does not capture these distinctions. We introduce Framing-Aware Event Causality Identification (FrECI), a framing-aware extension of ECI that models causal explanations as structured claims including responsibility targets, evaluative framing, source type, and epistemic modality grounded in established framing theories. We construct a multilingual dataset aligned across English, Chinese, and Arabic narratives using shared event anchors. We evaluate FrECI using prompt-based large language model baselines and supervised neural models. Results show that prompt-based baselines struggle to recover complete framed causal claims, while joint supervised models perform substantially better. Finally, we demonstrate that FrECI enables quantitative analysis of divergent causal attribution across narratives.
Open LLMs enable AI practitioners to control development costs by building on an existing foundation for downstream applications. While offering substantial promise, current models often fail to meet the needs of users needing open solutions aligned with responsible AI principles, including data compliance, transparency, and inclusivity. In this work, we present Apertus, a fully open suite of large language models (LLMs) designed to address responsibility shortcomings in today’s open model ecosystem, namely data responsibility and global representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting robots.txt exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of data memorization, we also adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. Apertus also drastically expands multilingual coverage, training on 15T tokens from over approximately 1800 languages, with about 40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivaling or surpassing open-weight counterparts.
LLMs increasingly excel on AI benchmarks, but doing so does not guarantee validity for downstream tasks. This study contrasts LLM alignment on benchmarks, downstream tasks, and, importantly the intended impact of those tasks. We evaluate the performance of leading LLMs (i.e., generative pre-trained base models) on difficult-to-verify tasks of the teaching and learning of schoolchildren. Across all LLMs, inter-model behaviors on disparate tasks correlate higher than they do with expert human behaviors on target tasks. These biases shared across LLMs are poorly aligned with downstream measures of teaching quality and often negatively aligned with the intended impact of student learning outcomes. Further, we find multi-model ensembles, both unanimous model voting and expert-weighting by benchmark performance, further exacerbate misalignment with learning. We measure that selection of LLM and/or prompting strategy only reliably accounts for 15% of all measured misalignment error and that variation in misalignment error is shared across LLMs, suggesting that common pretraining accounts for much of the misalignment in these tasks. We demonstrate methods for robustly measuring alignment of complex tasks and provide unique insights into practical applications of LLMs in high-noise contexts.
GitChameleon 2.0: Evaluating AI Code Generation Against Python Library Version Incompatibilities
PDF ↗The rapid evolution of software libraries poses a considerable hurdle for code generation, necessitating continuous adaptation to frequent version updates while preserving backward compatibility. While existing code evolution benchmarks provide valuable insights, they typically lack execution-based evaluation for generating code compliant with specific library versions. To address this, we introduce GitChameleon 2.0, a novel, meticulously curated dataset comprising 328 Python code completion problems, each conditioned on specific library versions and accompanied by executable unit tests. GitChameleon 2.0 rigorously evaluates the capacity of contemporary large language models (LLMs), LLM-powered agents, code assistants, and RAG systems to perform version-conditioned code generation that demonstrates functional accuracy through execution. Our extensive evaluations indicate that state-of-the-art systems encounter significant challenges with this task; enterprise models achieving baseline success rates in the 48-51% range, underscoring the intricacy of the problem. By offering an execution-based benchmark emphasizing the dynamic nature of code libraries, GitChameleon 2.0 enables a clearer understanding of this challenge and helps guide the development of more adaptable and dependable AI code generation methods.
What About the Scene With the Hitler Reference? HAUNT: A Framework to Probe LLMs’ Self-consistency in Closed Domains Via Adversarial Nudge
PDF ↗Hallucinations pose a critical challenge to the real-world deployment of large language models (LLMs) in high-stakes domains. In this paper, we present a framework for stress testing factual fidelity in LLMs in the presence of adversarial nudge. Our framework consists of three steps. First, we instruct the LLM to produce sets of truths and lies consistent with the closed domain in question. Next, we instruct the LLM to verify the same set of assertions as truths and lies consistent with the same closed domain. Finally, we test the robustness of the LLM against the lies generated (and verified) by itself. Our extensive evaluation, conducted using five widely known proprietary and six open LLMs across two closed domains of popular movies and novels, reveals a wide range of susceptibility to adversarial nudges: even among the strongest proprietary LLMs, Claude exhibits strong resilience, GPT and Grok demonstrate moderate resilience, while Gemini and DeepSeek show weak resilience and open models fall short significantly.
Large language models (LLMs) store and recall factual knowledge, yet the precise mechanism of how entity representations are transformed to enable specific attribute retrieval remains underexplored. In this work, we investigate this mechanism through the lens of an “attribute-computation path”—a sequence of computational steps over the entity representation required to elicit a target attribute. We then propose an iterative patching protocol to identify a minimal subset of layers necessary for this computation. Applying our method to LLaMA 3.1 8B and Qwen 3 8B, we find that these paths are non-contiguous, often skipping layers, and that models possess multiple, functionally-equivalent paths for the same entity and fact, highlighting a high degree of redundancy in attribute computation. This implies that knowledge computation is highly distributed, potentially explaining the localization-editing mismatch and suggesting that knowledge storage and retrieval in LLMs is far from being well understood.
This paper presents a novel Automated Red Teaming (ART) framework that shifts from example-based to policy-based evaluation, addressing critical limitations in scalability and validity. We define harmful content through abstract safety policies rather than specific static examples. We also introduce multiple evaluation objectives: risk coverage, semantic diversity, and fidelity, and discover Pareto trade-offs between them. We propose Jailbreak-Zero, a black-box method capable of both zero-shot generation and fine-tuned exploitation of a victim’s vulnerabilities to achieve Pareto optimality. Unlike prior approaches, it does not require expert-designed strategies/prompts, but still achieves superior, human-readable attacks against open-source and proprietary models (attack success rates of 99.5% against GPT-4o and 96.0% against Claude 3.5), even for unseen safety policies. It retains efficacy even after victim models undergo safety alignment, and exposes controls to navigate Pareto trade-offs without retraining. Lastly, we show that Jailbreak-Zero is the best-performing ART method at a given compute budget. Code is available at: https://github.com/hukkai/jailbreak-zero/ .
Modern evaluation of Legal QA systems is shifting from terminal accuracy toward process-aware analyses of model reasoning. We propose a diagnostic framework grounded in monotonic pedagogical scaffolding, where language models receive gold-standard, case-relevant information across stages aligned with the canonical legal framework FIRAC — Facts, Issue, Rules, Application, Conclusion. By strictly adding solution-relevant content at each step, we introduce a controlled monotonic intervention that allows for the evaluation of reasoning trajectories rather than isolated outcomes.This longitudinal design enables the introduction of two transition-based diagnostics: Errors-to-Success (E2S) quantifies the guidance required to reach correctness, while Success-to-Errors (S2E) measures the fragility of that correctness under additional structure. These local patterns define a global robustness criterion termed Stable Accuracy, which credits a response only if the model maintains correctness throughout all scaffolding stages and enforces a higher bar for correctness by distinguishing sustained reasoning from transient patterns.We instantiate the framework on 3,123 Brazilian Bar Exam questions paired with expert-annotated explanations. Our findings reveal model instability patterns hidden from accuracy-only metrics and demonstrate that terminal accuracy systematically overestimates legal reasoning competence. To test the robustness of our diagnostics, we also evaluate a majority-vote aggregation across multiple reasoning samples, finding that the observed instability patterns persist under this stronger inference setting. Furthermore, principal component analysis indicates that legal domains cluster into distinct regions, suggesting systematic differences in reasoning demands across domains. While focused on the legal domain, our evaluation protocol is generalizable to any task with a staged reasoning structure.
BatonVoice: An Operationalist Framework for Enhancing Controllable Speech Synthesis with Linguistic Intelligence from LLMs
PDF ↗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.
Limited access to mental healthcare, extended wait times, and increasing capabilities of Large Language Models (LLMs) has led individuals to turn to LLMs for fulfilling their mental health needs. However, examining the multi-turn mental health conversation capabilities of LLMs remains under-explored. Existing evaluation frameworks typically focus on diagnostic accuracy and win-rates and often overlook alignment with patient-specific goals, values, and personalities required for meaningful conversations. To address this, we introduce MedAgent, a novel framework for synthetically generating realistic, multi-turn mental health sensemaking conversations and use it to create the Mental Health Sensemaking Dialogue (MHSD) dataset, comprising over 2,200 patient–LLM conversations. Additionally, we present MultiSenseEval, a holistic framework to evaluate the multi-turn conversation abilities of LLMs in healthcare settings using human-centric criteria. Our findings reveal that frontier reasoning models yield below-par performance for patient-centric communication and struggle at precise ("hard") diagnostic capabilities with average accuracy of ~31%. Additionally, we observed variation in model performance based on patient’s persona and performance drop with increasing turns in the conversation. Our work provides a comprehensive synthetic data generation framework, a dataset and evaluation framework for assessing LLMs in multi-turn mental health conversations.
As large language models continue to advance, ensuring their trustworthiness is critical. However, inaccessible real-world ground truth labels pose a significant challenge in high-stakes domains. Recent studies have highlighted weak-to-strong generalization, where a strong model trained only on a weak model’s labels surpasses the weak model in task performance. Yet, whether critical trustworthiness properties such as robustness, fairness, and privacy can generalize similarly remains an open question. This is the first work to study this question by examining if a stronger model can enhance trustworthiness when fine-tuned on a weaker model’s labels, a paradigm we term weak-to-strong trustworthiness. To address this, we introduce two fundamental fine-tuning strategies that leverage trustworthiness regularization during the fine-tuning of the weak model and the weak-to-strong transfer. Our experimental evaluation on real-world datasets reveals that while some trustworthiness properties, such as fairness, adversarial robustness, and OOD robustness, show significant improvement in trustworthiness generalization when both models were regularized, others, like privacy, do not exhibit signs of weak-to-strong trustworthiness. Our results highlight the potential of weak-to-strong trustworthiness as a practical pathway for enhancing the trustworthiness of increasingly capable AI systems, even under imperfect real-world conditions.
Code-switching, alternating between languages within a conversation, is natural for multilingual users, yet poses fundamental challenges for large language models (LLMs). When a user code-switches in their prompt to an LLM, they typically do not specify the expected language of the LLM response, and thus LLMs must infer the output language from contextual and pragmatic cues. We find that current LLMs systematically fail to align with this expectation, responding in undesired languages even when cues are clear to humans. We introduce OLA, a benchmark to evaluate LLMs’ Output Language Alignment in code-switched interactions. OLA focuses on Korean–English code-switching and spans simple intra-sentential mixing to instruction–content mismatches. Even frontier models frequently misinterpret implicit language expectation, exhibiting a systematic bias toward non-English responses. We further show this bias generalizes beyond Korean to Chinese and Indonesian pairs. Models also show instability through mid-response switching and language intrusions. Chain-of-Thought prompting fails to resolve these errors, indicating weak pragmatic reasoning about output language. However, Code-Switching Aware DPO with minimal data (~1K examples) substantially reduces misalignment, suggesting these failures stem from insufficient alignment rather than fundamental limitations. Our results highlight the need to align multilingual LLMs with users’ implicit expectations in real-world code-switched interactions.
The advent of always-on personal AI assistants, enabled by all-day wearable devices such as smart glasses, demands a new level of contextual understanding—one that goes beyond short, isolated events to encompass the continuous, longitudinal stream of egocentric video. Achieving this vision requires advances in long-horizon video understanding, where systems must interpret and recall visual and audio information spanning days or even weeks. Existing methods, including large language models and retrieval-augmented generation, are constrained by limited context windows and lack the ability to perform compositional, multi-hop reasoning over very long video streams. In this work, we address these challenges through EGAgent, an enhanced agentic framework centered on entity scene graphs, which represent people, places, objects, and their relationships over time. Our system equips a planning agent with tools for structured search and reasoning over these graphs, as well as hybrid visual and audio search capabilities, enabling detailed, cross-modal, and temporally coherent reasoning. Experiments on the EgoLifeQA and Video-MME-long datasets show that our method achieves state-of-the-art performance on EgoLifeQA (57.5%) and competitive performance on Video-MME-long (74.1%) for complex longitudinal video understanding tasks.
Disinformation is an escalating global threat, making it essential to understand its content, dissemination, and evolution. To confront this challenge, researchers have begun grouping related false claims into broader disinformation narratives, which can be tracked across cultures, time periods, and media sources. Analyzing these narratives provides critical insights for developing more effective countermeasures. To this end, we introduce DiNO: Disinformation Narrative Observer, a novel method designed to extract disinformation narratives from news articles. We applied DiNO to news articles on the Ukraine War, COVID-19 and Migration, sourced from disinformation-prone outlets as well as a reputable source. We evaluated the narratives extracted by DiNO by measuring how well their topics and stances aligned with a recognized disinformation narratives dataset. DiNO outperforms competitive narrative mining approaches, including Relatio and CaNarEx, achieving a 41%–44% improvement in topical alignment and a 30%–41% improvment in stance alignment.