As large language models are increasingly adopted and trusted in real-world applications, understanding their behavior beyond single-turn prompting has become critical. Existing safety evaluations primarily focus on refusal-based methods that test whether models avoid responding to inappropriate or violent requests, leaving open questions about how models behave in interactive social settings. In this paper, we observe the adversarial behavior of LLM models through a multi-agent simulation across five diverse social deduction conversational games, acting as testbeds that provide social pressures and survival stress based on game design without explicit prompt injections. From these interactions, we construct a closed behavioral taxonomy derived through open card sorting, applied uniformly across models using a meta-LLM for behavior labeling. This approach displays that models exhibit distinct behavioral profiles and that models’ different ways of structured deliberation influence both social stability and competitive success.
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
Conjecture and Inquiry: Quantifying Software Performance Requirements via Interactive Retrieval-Augmented Preference Elicitation
PDF ↗Since software performance requirements are documented in natural language, quantifying them into mathematical forms is essential for software engineering. Yet, the vagueness in performance requirements and uncertainty of human cognition have caused highly uncertain ambiguity in the interpretations, rendering their automated quantification an unaddressed and challenging problem. In this paper, we formalize the problem and propose IRAP, an approach that quantifies performance requirements into mathematical functions via interactive retrieval-augmented preference elicitation. IRAP differs from the others in that it explicitly derives from problem-specific knowledge to retrieve and reason the preferences, which also guides the progressive interaction with stakeholders, while reducing the cognitive overhead. Experiment results against 10 state-of-the-art methods on four real-world datasets demonstrate the superiority of IRAP on all cases with up to 40x improvements under as few as five rounds of interactions.
While it is widely agreed that large language models (LLMs) store concepts from multiple semantic hierarchies, much remains unknown regarding the structure of this storage. The correspondence between the functional roles of LLM components and the semantic hierarchies of knowledge remains underexplored in the current literature. For example, is information organized hierarchically within sections of an LLM? We take an initial step towards causally examining the correspondence between hierarchical concepts and the multi-granular structures (layers and attention heads) of various models. Specifically, we generate a dataset of semantic hierarchies and investigate their storage locations in six LLMs using activation patching, a causal intervention technique. At the layer level, our findings show a moderate indication that concepts at finer levels of granularity are stored around 61-78% of the time (p < 0.01) before those at coarser granularity. There is evidence for this trend at the attention level; however, the high variability in attention level results suggests that concepts are stored across attention heads rather than within. Our results offer insight into semantic organization within LLMs.
Optimizing Conversational Quality in Spoken Dialogue Systems with Reinforcement Learning from AI Feedback
PDF ↗Reinforcement learning from human or AI feedback (RLHF/RLAIF) for speech-in/speech-out dialogue systems (SDS) remains underexplored, with prior work largely limited to single semantic rewards applied at the utterance level. Such setups overlook the multi-dimensional and multi-modal nature of conversational quality, which encompasses semantic coherence, audio naturalness, speaker consistency, emotion alignment, and turn-taking behavior. Moreover, they are fundamentally mismatched with duplex spoken dialogue systems that generate responses incrementally, where agents must make decisions based on partial utterances. We address these limitations with the first multi-reward RLAIF framework for SDS, combining semantic, audio-quality, and emotion-consistency rewards. To align utterance-level preferences with incremental, blockwise decoding in duplex models, we apply turn-level preference sampling and aggregate per-block log-probabilities within a single DPO objective. We present the first systematic study of preference learning for improving SDS quality in both multi-turn Chain-of-Thought and blockwise duplex models, and release a multi-reward DPO dataset to support reproducible research. Experiments show that single-reward RLAIF selectively improves its targeted metric, while joint multi-reward training yields consistent gains across semantic quality and audio naturalness. These results highlight the importance of holistic, multi-reward alignment for practical conversational SDS.
Style representation learning is a powerful tool for authorship analysis and modeling writing style, yet the latent nature of learned representations makes them difficult to interpret. Recent work has attempted to explain these representations by generating natural language descriptions with large language models (LLMs) conditioned on input text. However, such descriptions are often prone to the LLM’s biases and hallucinations, and they lack an explicit objective and practical utility. In this work, we propose a novel framework for interpreting style representations through style-eliciting prompts: natural language instructions designed to steer LLMs to generate text that reflects specific stylistic attributes. We curate 1,010 distinct style features spanning 26 stylistic categories and construct a dataset by prompting an LLM to generate text conditioned on these features. Using this data, we train a decoder to generate a style prompt from the style representation of the generated text. We evaluate our approach on three tasks: (1) recovering original style prompts from generated text, (2) generating text in the same style using the recovered prompts, and (3) steering LLM outputs to match the style of human-written texts. Experiments demonstrate that our method consistently outperforms strong baselines that directly prompt LLMs with target text, achieving superior performance in both style description and style imitation. These results highlight style-eliciting prompts can provide a practical and interpretable interface to stylistic information encoded in style representations.
While word embeddings derive meaning from co-occurrence patterns, human language understanding is grounded in sensory and motor experience. We present \mathit{SENSE} (\textbf{S}ensorimotor \textbf{E}mbedding \textbf{N}orm \textbf{S}coring \textbf{E}ngine), a learned projection model that predicts Lancaster sensorimotor norms from word lexical embeddings. We also conducted a behavioral study where 281 participants selected which among candidate nonce words evoked specific sensorimotor associations, finding statistically significant correlations between human selection rates and \mathit{SENSE} ratings across 6 of the 11 modalities. Sublexical analysis of these nonce word selection rates revealed systematic phonesthemic patterns for the interoceptive norm, suggesting a path towards computationally proposing candidate phonesthemes from text data.
We revisit continual pre-training for large language models and argue that progress now depends less on scaling parameters than on scaling the right structure. We introduce SCALE, a width upscaling architecture that inserts lightweight expansions into linear modules while freezing all pre-trained parameters, preserving residual and attention topologies and increasing capacity without perturbing the base model’s original functionality. SCALE follows two principles: Persistent Preservation, which maintains the base model’s behavior via preservation-oriented initialization and freezing of the pre-trained weights, and Collaborative Adaptation, which trains only selected expansion components to acquire new knowledge with minimal interference. We instantiate these ideas as SCALE-Preserve (preservation-first), SCALE-Adapt (adaptation-first), and SCALE-Route, an optional routing extension that performs token-level routing between preservation and adaptation heads. On a controlled synthetic biography benchmark, SCALE reduces the severe forgetting seen in depth expansion while still learning new knowledge. In continual pre-training on a Korean corpus, SCALE variants forget less on English evaluations and achieve competitive gains on Korean benchmarks, yielding the best overall stability-plasticity trade-off. We further analyze when preservation holds provably and why combining preservation and adaptation stabilizes optimization relative to standard continual learning.
Pretrained biomedical vision–language models (VLMs) such as BioMedCLIP perform well on average but often degrade on challenging modalities where inter-class margins are small and acquisition-specific variations are pronounced, especially under few-shot supervision and when modality priors differ from pretraining corpora substantially. We propose BioVLM, a prompt-learning framework that improves cross-domain generalization without extensive backbone fine-tuning. BioVLM learns a diverse prompt bank and introduces dynamic prompt selection: for each input, it selects the most discriminative prompts via a low-entropy criterion on the predictive distribution, effectively coupling sparse few-shot evidence with rich LLM semantic priors. To strengthen this coupling, we distill high-confidence LLM-derived attributes and enforce robust knowledge transfer through strong/weak augmentation consistency. At test time, BioVLM adapts by choosing modality-appropriate prompts, enabling transfer to unseen categories and domains, while keeping training lightweight and inference efficient. On 11 MedMNIST+ 2D datasets, BioVLM achieves new state of the art across three distinct generalization settings. Codes are available at https://github.com/mainaksingha01/BioVLM.
MEAV: Model Editing with Alignment Vectors for inference time LLM alignment in single and multidomain preference spectrum
PDF ↗Aligning Large Language Models (LLM) to address subjectivity and nuanced preference levels requires adequate flexibility and control, which can be a resource-intensive and time-consuming procedure. Existing training-time alignment methods require full re-training when a change is needed and inference-time ones typically require access to the reward model at each inference step. We introduce **MEAV**, an inference-time model-editing-based LLM alignment method that learns encoded representations of preference dimensions, called *Alignment Vectors* (AV). These representations enable dynamic adjusting of the model behavior during inference through simple linear operations. Here, we focus on three gradual response levels across three specialized domains: medical, legal, and financial, exemplifying its practical potential. This new alignment paradigm introduces adjustable preference knobs during inference, allowing users to tailor their LLM outputs while reducing the inference cost by half compared to the prompt engineering approach. Additionally, we find that AVs are transferable across different fine-tuning stages of the same model, demonstrating their flexibility. AVs also facilitate multidomain, diverse preference alignment, making the process 12x faster than the retraining approach.
Stance detection aims to ascertain whether an author’s text is in favor, against, or neutral toward specific targets like public policies or social issues. While pretrained language models (PLMs) have greatly enhanced stance detection, they remain vulnerable to adversarial attacks—manipulations that maintain textual semantics but lead to incorrect predictions. Such vulnerabilities remain underexplored for stance detection. In this study, we introduce StanceAttack, an innovative adversarial attack method leveraging ChatGPT to create adversarial examples that can mislead well-trained stance detection models. We conduct experiments to evaluate our attack model by attacking state-of-the-art PLMs on two benchmark datasets. Results demonstrate that StanceAttack outperforms traditional adversarial methods with higher success rates and fewer retries. Human evaluations confirm that our adversarial examples preserve the original semantic meanings and naturalness. We share our code and data in https://github.com/chenyez/StanceAttack.
The growing ubiquity of Extended Reality (XR) is driving Conversational Recommendation Systems (CRS) toward visually immersive experiences. We formalize this paradigm as Immersive CRS (ICRS), where recommended items are highlighted directly in the user’s scene-based visual environment and augmented with in-situ labels. While item recommendation has been widely studied, the problem of how to select and evaluate which information to present as immersive labels remains an open problem. To this end, we introduce a principled categorization of information needs into explicit intent satisfaction and proactive information needs and use these to define novel evaluation metrics for item label selection. We benchmark IR-, LLM-, and VLM-based methods across three datasets and ICRS scenarios: fashion, movie recommendation, and retail shopping. Our evaluation reveals three important limitations of existing methods: (1) they fail to leverage scenario-specific information modalities (e.g., visual cues for fashion, metadata for retail), (2) they present redundant information that is visually inferable, and (3) they poorly anticipate users’ proactive information needs from explicit dialogue alone. In summary, this work provides both a novel evaluation paradigm for in-situ item labeling in ICRS and highlights key challenges for future work.
Current AI-powered code assistance tools often struggle with ambiguous problem statements that lack sufficient task context and requirements specification. Recent analysis of software engineering agents reveals that failures on such ambiguous requests are highly correlated with longer trajectories involving either over-exploration or repeated attempts at applying the same fix without proper evolution or testing, leading to suboptimal outcomes across software development tasks. We introduce CodeScout, a breakthrough contextual query refinement approach that systematically converts ambiguous user requests into comprehensive, actionable problem statements through lightweight pre-exploration of the target codebase. Our key innovation is demonstrating that structured analysis before task execution can supplement existing agentic capabilities without requiring any modifications to their underlying scaffolds. CodeScout performs targeted context scoping, conducts multi-perspective analysis examining potential fixes and exploration opportunities, then synthesizes these insights into enhanced problem statements with reproduction steps, expected behaviors, and targeted exploration hints. This pre-exploration directly addresses the identified failure patterns by reducing non-converging agent trajectories while clarifying user intent in natural language space. We evaluate CodeScout using state-of-the-art agentic scaffolds and language models on SWEBench-Verified, demonstrating a 20% improvement in resolution rates with up to 27 additional issues resolved compared to the default baseline method. Our results suggest that systematic query refinement through contextual analysis represents a promising direction for enhancing AI code assistance capabilities.
Large language models (LLMs), despite strong performance on complex mathematical problems, exhibit systematic limitations in counting tasks. This issue arises from the architectural limits of transformers, where counting is performed across layers, leading to degraded precision for larger counting problems due to depth constraints. To address this limitation, we propose a simple test-time strategy inspired by System-2 cognitive processes that decomposes large counting tasks into smaller, independent sub-problems that the model can reliably solve. We evaluate this approach using observational and causal mediation analyses to understand the underlying mechanism of this System-2-like strategy. Our mechanistic analysis identifies key components: latent counts are computed and stored in the final item representations of each part, transferred to intermediate steps via dedicated attention heads, and aggregated in the final stage to produce the total count. Experimental results demonstrate that this strategy enables LLMs to surpass architectural limitations and achieve higher accuracy on large-scale counting tasks. This work provides mechanistic insight into System-2 counting in LLMs and presents a generalizable approach for improving and understanding their reasoning behavior.
Beyond Reproduction: A Paired-Task Framework for Assessing LLM Comprehension and Creativity in Literary Translation
PDF ↗Large language models (LLMs) are increasingly used for creative tasks such as literary translation. Yet translational creativity remains underexplored and is rarely evaluated at scale, while source-text comprehension is typically studied in isolation, despite the fact that, in professional translation, comprehension and creativity are tightly intertwined. We address these gaps with a paired-task framework applied to literary excerpts from 11 books. Task 1 assesses source-text comprehension, and Task 2 evaluates translational creativity through Units of Creative Potential (UCPs), such as metaphors and wordplay. Using a scalable evaluation setup that combines expert human annotations with UCP-based automatic scoring, we benchmark 23 models and four creativity-oriented prompts. Our findings show that strong comprehension does not translate into human-level creativity: models often produce literal or contextually inappropriate renderings, with particularly large gaps for the more distant English–Chinese language pair. Creativity-oriented prompts yield only modest gains, and only one model, Mistral-Large, comes close to human-level creativity (0.167 vs. 0.246). Across all model–prompt combinations, only three exceed a creativity score of 0.1, while the rest remain at or near zero.
LRBench and Judge-R1: Principled Evaluation and Training of LLM-Based Judges for Long-Context Reasoning
PDF ↗Large language models (LLMs) are increasingly used as judges to evaluate, rank, and supervise other models, yet their reliability in judging LLMs’ reasoning process under long-context settings remains underexplored. Existing benchmarks either overly rely on human annotators, who may miss subtle flaws in lengthy reasoning chains, or focus solely on final responses while ignoring the underlying context and reasoning process. We introduce Long-Reason Bench (LRBench), a large-scale benchmark for evaluating LLM-based judges. LRBench comprises over 100K annotated instances spanning medical, legal, and academic-review scenarios, with fine-grained labels indicating violations of six core principles: Logical Correctness, Factual Consistency, Bias and Fairness, Groundedness, Helpfulness, and Harmlessness. Experimental results reveal that state-of-the-art LLM judges struggle to identify nuanced reasoning errors in long contexts. To improve judge reliability, we further present Judge-R1, which combines reinforcement learning with multi-turn search to enable grounded and principle-aware evaluation. Across domains and principles, Judge-R1 consistently outperforms single-turn baselines, enabling scalable and trustworthy evaluation of LLM reasoning. Our dataset and code are available at https://github.com/Xinyi-0724/Judge-R1.
LLM agents with tool-calling capabilities often fail when user instructions are ambiguous or incomplete, leading to incorrect invocations and task failures. Existing approaches operate in unstructured language spaces, generating clarifying questions through prompting strategies that lack principled criteria for determining which questions to ask and when to stop. We introduce a principled formulation of structured uncertainty that operates directly over tool parameters and their domains, cleanly separating specification uncertainty (what the user wants) from model uncertainty (what the LLM predicts). Our formulation uses Expected Value of Perfect Information (EVPI) to quantify the disambiguation value of each potential question, balanced against aspect-based cost modeling that prevents redundant questioning. We demonstrate the versatility of this formulation through two applications. First, SAGE-Agent uses structured uncertainty for inference-time question selection, achieving 7–39% higher coverage on ambiguous tasks while reducing clarification questions by 1.5–2.7 x compared to strong prompting and uncertainty-based baselines. Second, we show that structured uncertainty provides effective training signals: uncertainty-guided reward modeling boosts When2Call accuracy from 36.5% to 65.2% (3B model) and 36.7% to 62.9% (7B model) through uncertainty-weighted GRPO training, demonstrating more sample-efficient reinforcement learning for tool-calling agents. To enable evaluation, we present ClarifyBench, the first multi-turn dynamic tool-calling disambiguation benchmark. Our results establish structured uncertainty as a principled framework that improves both inference-time interaction efficiency and training-time sample efficiency in tool-augmented agents.
Weight tying, i.e. sharing parameters between input and output embedding matrices, is common practice in language model design, yet its impact on the learned embedding space remains poorly understood. In this paper, we show that tied embedding matrices align more closely with output (unembedding) matrices than with input embeddings of comparable untied models, indicating that the shared matrix is shaped primarily for output prediction rather than input representation. This unembedding bias arises because output gradients dominate early in training. Using tuned lens analysis, we show this negatively affects early-layer computations, which contribute less effectively to the residual stream. Scaling input gradients during training reduces this bias, providing causal evidence for the role of gradient imbalance. This is mechanistic evidence that weight tying optimizes the embedding matrix for output prediction, compromising its role in input representation. These results help explain why weight tying can harm performance at scale and have implications for training smaller LLMs, where the embedding matrix contributes substantially to total parameter count.
Humans often hold different perspectives on the same issues. In many NLP tasks, annotation disagreement can reflect valid subjective perspectives. Modeling annotator perspectives and understanding their relationship with other human factors, such as socio-demographic attributes, have received increasing attention. Prior work typically focuses on single demographic factors or limited combinations. However, in real-world settings, annotator perspectives are shaped by complex social contexts, and finer-grained socio-demographic attributes can better explain human perspectives. In this work, we propose Socio-Contrastive Learning, a method that jointly models annotator perspectives while learning socio-demographic representations. Our method provides an effective approach for the fusion of socio-demographic features and textual representations to predict annotator perspectives, outperforming standard concatenation-based methods. The learned representations further enable analysis and visualization of how demographic factors relate to variation in annotator perspectives. Our code is available at GitHub: https://github.com/Leixin-Zhang/Socio_Contrastive_Learning
Multimodal LLMs can accurately perceive numerical content across modalities yet fail to perform exact multi-digit multiplication when the identical underlying arithmetic problem is presented as numerals, number words, images, or in audio form. Because existing benchmarks often lack systematically paired instances across modalities, it remains difficult to compare genuine arithmetic limits within and across model families. We therefore introduce a controlled multimodal multiplication benchmark that factorially varies digit length, digit sparsity, representation (e.g., numerals vs. number words), and modality (text, rendered images, and audio), with paired instances from a reproducible generator. We also define arithmetic load, C, as the product of the total and non-zero digit number as a compact, mechanistically motivated proxy for operation count. Across evaluations, accuracy falls sharply as C grows, often nearing zero by C > 100. Indeed, C remains predictive of performance across modalities and models, with R^2 > 0.5, nearing the value from more complex measures of arithmetic load that count the number of intermediate arithmetic steps. A separate perception-versus-computation decomposition shows that multimodal degradation is primarily computational rather than perceptual: on matched perception checks, models are near-perfect (>99\%) across modalities even when multiplication accuracy drops substantially. Beyond measuring when models fail, we ask which procedures they are predisposed to follow. We introduce a style-controlled forced-completion loss probe that scores heuristic-specific reasoning prefixes—including columnar multiplication, distributive decomposition, and rounding/compensation. Here, distributive decomposition is favored in both text and vision modalities; heuristic-specific LoRA adapters produce near-orthogonal updates yet degrade accuracy, indicating the base model maintains a well-tuned internal router.
Steering large language model (LLM) reasoning via high-level reasoning actions offers a promising approach to improve robustness and interpretability. However, existing action-based paradigms, ranging from training-free prompting to static plan retrieval or prediction, often fail to consistently outperform standard generation because their planners tend to degenerate into repetitive loops or fixed patterns. We propose PG-HAP (Policy-Guided High-Level Action Planning), a lightweight stepwise planner–executor framework that learns to select reasoning actions dynamically while keeping the executor LLM fully frozen. The planner is trained with reinforcement learning to optimize answer correctness. To prevent degeneration, we introduce two targeted mechanisms: (i) an Action-Dependency Logit Mask that enforces valid transitions to avoid redundancy, and (ii) an Action Diversity Reward that discourages mode collapse by promoting varied action sequences. Across mathematical and commonsense reasoning benchmarks, PG-HAP improves accuracy over strong baselines while producing less redundant, more adaptive trajectories. This demonstrates that learning high-level planning alone can substantially strengthen reasoning without expensive end-to-end model tuning.