Prompt sensitivity, which refers to how strongly the output of a large language model (LLM) depends on the exact wording of its input prompt, raises concerns among users about the LLM’s stability and reliability. In this work, we consider LLMs as multivariate functions and perform a first-order Taylor expansion, thereby analyzing the relationship between meaning-preserving prompts, their gradients, and the log probabilities of the model’s next token. We derive an upper bound on the difference between log probabilities using the Cauchy-Schwarz inequality. We show that LLMs do not internally cluster similar inputs like smaller neural networks do, but instead disperse them. This dispersing behavior leads to an excessively high upper bound on the difference of log probabilities between two meaning-preserving prompts, making it difficult to effectively reduce to 0. In our analysis, we also show which types of meaning-preserving prompt variants are more likely to introduce prompt sensitivity risks in LLMs. In addition, we demonstrate that the upper bound is strongly correlated with an existing prompt sensitivity metric, PromptSensiScore. Moreover, by analyzing the logit variance, we find that prompt templates typically exert a greater influence on logits than the questions themselves. Overall, our results provide a general interpretation for why current LLMs can be highly sensitive to prompts with the same meaning, offering crucial evidence for understanding the prompt sensitivity of LLMs.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Automated interaction with graphical user interfaces (GUIs) is central to General Artificial Intelligence yet remains challenging within Super App ecosystems, characterized by non-standard rendering and absent accessibility metadata. While GUI agents often rely on explicit accessibility trees or static imitation, they are less explored for dynamic environments marked by sparse feedback and implicit visual cues. We present GUI0, a framework synergizing autonomous data synthesis with dual-agent co-evolution. GUI0 establishes a domain-aware foundation model via synthesized corpora and employs curriculum-driven reinforcement learning, where a curriculum agent generates boundary tasks to optimize an actor agent.Empirical results demonstrate three key advantages: (1) State-of-the-art performance on the SuperAPP benchmark, outperforming Gemini-2.5-Pro and Claude-4-Sonnet; (2) universal efficacy across diverse base models, consistently yielding substantial improvements on both Qwen2.5-VL and GUI-Owl variants; and (3) robust zero-shot generalization to standard GUIs (e.g., +62.7% on ScreenSpot Pro).
Route to Rome Attack: Directing LLM Routers to Expensive Models via Adversarial Suffix Optimization
PDF ↗Cost-aware routing dynamically dispatches user queries to models of varying capability to balance performance and inference cost. However, the routing strategy introduces new security concern that adversaries may manipulate router to consistently select expensive high-capability models. Existing routing attacks depend either on white-box access or heuristic prompts, rendering them ineffective in real-world black-box scenarios. In this work, we propose R^2A, which aims to mislead black-box LLM routers to expensive models via adversarial suffix optimization. Specifically, R^2A deploys a hybrid ensemble surrogate router to mimic the black-box router. A suffix optimization algorithm is further adapted for the ensemble-based surrogate. Extensive experiments on multiple open-source and commercial routing systems demonstrate that R^2A significantly increases the routing rate to expensive models on queries of different distributions. Code and examples: https://github.com/thcxiker/R2A-Attack.
Traditional task-oriented dialog systems are unable to evolve from ongoing interactions or adapt to new domains after deployment, that is a critical limitation in real-world dynamic environments. Continual learning approaches depend on episodic retraining with human-curated data, failing to achieve autonomy lifelong improvement. While evolutionary computation and LLM driven self-improvement offer promising mechanisms for dialog optimization, they lack a unified framework for holistic, iterative strategy refinement. To bridge this gap, we propose DarwinTOD, a lifelong self-evolving dialog framework that systematically integrates these two paradigms, enabling continuous strategy optimization from a zero-shot base without task-specific fine-tuning. DarwinTOD maintains an Evolvable Strategy Bank and operates through a dual-loop process: online multi-agent dialog execution with peer critique, and offline structured evolutionary operations that refine the strategy bank using accumulated feedback. This closed-loop design enables autonomous continuous improvement without human intervention. Extensive experiments show that DarwinTOD surpasses previous state-of-the-art methods and exhibits continuous performance gains throughout evolution. Our work provides a novel framework for building dialog systems with lifelong self-evolution capabilities.
Do MLLMs Capture How Interfaces Guide User Behavior? A Benchmark for Multimodal UI/UX Design Understanding
PDF ↗User interface (UI) design goes beyond visuals to shape user experience (UX), underscoring the shift toward UI/UX as a unified concept. While recent studies have explored UI evaluation using Multimodal Large Language Models (MLLMs), they largely focus on surface-level features, overlooking how design choices influence user behavior at scale. To fill this gap, we introduce WiserUI-Bench, a novel benchmark for multimodal understanding of how UI/UX design affects user behavior, built on 300 real-world UI image pairs from industry A/B tests, with empirically validated winners that induced more user actions. For future design progress in practice, post-hoc understanding of why such winners succeed with mass users is also required; we support this via expert-curated key interpretations for each instance. Experiments across multiple MLLMs on WiserUI-Bench for two main tasks, (1) predicting the more effective UI image between an A/B-tested pair, and (2) explaining it post-hoc in alignment with expert interpretations, show that models exhibit limited understanding of the behavioral impact of UI/UX design. We believe our work will foster research on leveraging MLLMs for visual design in user behavior contexts.
Large language models (LLMs) have achieved strong performance on standard benchmarks, yet their performance is not robust across different task manifestations. It remains unclear how performance changes under controlled task rewrites that preserve the original solution structure, while varying the rewrite type and level. To address this question, we introduce ReTRE (Rewrite-based Transfer Robustness Evaluation), an evaluation benchmark inspired by learning transfer theory that probes transfer robustness along two rewrite levels: Near Transfer and Far Transfer. ReTRE employs a multi-agent system to construct textual and visual variants while preserving the structure of the original solution. Evaluations on mathematical and science tasks across state-of-the-art multimodal LLMs reveal a consistent transfer gap: performance exhibits a general declining trend as transfer similarity drops and strong text performance can face performance decline under cross-modal transfer. Crucially, we identify a divergence between post-training paradigms: reinforcement learning preserves transfer robustness, whereas supervised fine-tuning tends to overfit the training distribution, leading to severe degradation in far-transfer performance despite strong in-distribution accuracy.
Logical reasoning is a key capability of large language models, yet current benchmarks focus almost entirely on tasks that just check basic logical consistency and overlook the reflective reasoning required for paradox detection and resolution. To fill the gap, we present ParaSuite, the first pipeline dedicated to paradox research that automates data synthesis, evaluation, and training. We introduce PARADOX, a synthetic, high-quality data spanning two difficulty tiers and three academic domains, accompanied by specialized evaluation metrics and solving algorithms. We propose ParadoxBreaker-7B, trained with Mutual-Information Guided Fine-Tuning and reinforcement learning step verify paradox reward(PAPO). Experiments demonstrate significant improvements in both paradoxical and general STEM reasoning.
As people increasingly turn to AI for personal deliberation beyond task-oriented assistance, concerns about sycophancy in these value-laden contexts have grown. Unlike human flattery, which is intentional and self-interested, AI sycophancy emerges as a byproduct of RLHF’s reward structure for user-preference alignment. Yet the observable behavior is similar: both produce responses that preserve what users want to hear. Focusing on this phenomenon through Goffman’s face-work framework, we operationalize AI sycophancy as excessive face-saving, either active (preserving positive face through agreement) or passive (preserving negative face by withholding challenge). In a mixed-methods study (N=31), participants engaged with AI across three moral dilemmas under these conditions and a non-sycophantic neutral baseline. Sycophantic responses increased decision confidence but reduced open-minded thinking; participants felt supported yet found the conversations unproductive. Neutral responses, though initially uncomfortable, promoted cognitive flexibility and meaningful deliberation. These findings reveal a confidence-competence trade-off in AI-mediated moral reasoning and suggest that effective AI for personal deliberation requires calibrated friction, not unconditional agreement.
Multi-turn, long-horizon tasks are increasingly common for large language models (LLMs), but solving them typically requires many sequential model invocations, accumulating substantial inference costs. Here, we study cost-aware multi-turn LLM routing: selecting which model to invoke at each turn from a model pool, given a fixed cost budget. We propose MTRouter, which encodes the interaction history and candidate models into joint history–model embeddings, and learns an outcome estimator from logged trajectories to predict turn-level model utility. Experiments show that MTRouter improves the performance–cost trade-off: on ScienceWorld, it surpasses GPT-5 while reducing total cost by 58.7%; on Humanity’s Last Exam (HLE), it achieves competitive accuracy while reducing total cost by 43.4% relative to GPT-5, and these gains even carry over to held-out tasks. Further analyses reveal several mechanisms underlying its effectiveness: relative to prior multi-turn routers, MTRouter makes fewer model switches, is more tolerant to transient errors, and exhibits emergent specialization across models.Code: https://github.com/ZhangYiqun018/MTRouter
Evolution, the engine behind the survival and growth of life on Earth, operates through the population-based process of reproduction. Inspired by this principle, this paper formally defines a newly emerging problem: the population-based evolution of large language models (LLMs). We introduce a novel framework that starts with a population of parent LLMs and allows this population to evolve through four key operations: (i) crossover, merging the weights of different parents to create offspring LLMs, (ii) mutation, introducing small, random changes to model weights to foster diversity, (iii) selection, prioritizing high-performing models, and (iv) succession, transferring the learned experience from parent to offspring LLMs. With only 200 samples per new task, the LLM population evolves rapidly to adapt to the task at hand, without any gradients. Experiments on 12 datasets show that our framework consistently outperforms existing multi-LLM merging and adaptation methods, achieving relative performance gains of up to 54.8 over the best LLM in the initial population. Moreover, our framework allows for (i) the evolution of LLMs across multiple new tasks simultaneously, (ii) scaling effectively with populations of up to 40 LLMs, and (iii) even zero-shot generalization to unseen held-out tasks. Code: https://github.com/ZhangYiqun018/GENOME
POLYCHARTQA: Benchmarking Large Vision-Language Models with Multilingual Chart Question Answering
PDF ↗Charts are a universally adopted medium for data communication, yet existing chart understanding benchmarks are overwhelmingly English-centric, limiting their accessibility and relevance to global audiences. To address this limitation, we introduce PolyChartQA, the first large-scale multilingual benchmark for chart question answering, comprising 22,606 charts and 26,151 QA pairs across 10 diverse languages. PolyChartQA is constructed through a scalable pipeline that enables efficient multilingual chart generation via data translation and code reuse, supported by LLM-based translation and rigorous quality control. We systematically evaluate multilingual chart understanding with PolyChartQA on state-of-the-art LVLMs and reveal a significant performance gap between English and other languages, particularly low-resource ones. Additionally, we introduce a companion multilingual chart question answering training set, PolyChartQA-Train, on which fine-tuning LVLMs yields substantial gains in multilingual chart understanding across diverse model sizes and architectures. Together, our benchmark provides a foundation for developing globally inclusive vision-language models capable of understanding charts across diverse linguistic contexts. Codes and datasets are available on https://github.com/Road2Redemption/PolyChartQA.
Proactive streaming video understanding requires Video-LLMs to decide when to respond as a video unfolds, a task where existing methods often fall short due to their implicit, query-agnostic modeling of visual evidence. We introduce Response-G1, a novel framework that establishes explicit, structured alignment between the accumulated video evidence and the query’s expected response conditions via scene graphs. The framework operates in three fine-tuning-free stages: (1) online query-guided scene graph generation from streaming clips; (2) memory-based retrieval of the most semantically relevant historical scene graphs; and (3) retrieval-augmented trigger prompting for per-frame "silence/response" decisions. By grounding both evidence and conditions in a shared graph representation, Response-G1 achieves more interpretable and accurate response timing decisions. Experimental results on established benchmarks demonstrate the superiority of our method in both proactive and reactive tasks, validating the advantage of explicit scene graph modeling and retrieval in streaming video understanding.
Humor holds up a mirror to social perception: what we find funny often reflects who we are and how we judge others. When language models engage with humor, their reactions expose the social assumptions they have internalized from training data. In this paper, we investigate counterfactual unfairness through humor by observing how the model’s responses change when we swap who speaks and who is addressed while holding other factors constant. Our framework spans three tasks: humor generation refusal, speaker intention inference, and relational/societal impact prediction, covering both identity-agnostic humor and identity-specific disparagement humor. We introduce interpretable bias metrics that capture asymmetric patterns under identity swaps. Experiments across state-of-the-art models reveal consistent relational disparities: jokes told by privileged speakers are refused up to 67.5% more often, judged as malicious 64.7% more frequently, and rated up to 1.5 points higher in social harm on a 5-point scale. These patterns highlight how sensitivity and stereotyping coexist in generative models, complicating efforts toward fairness and cultural alignment.
Personalized Large Language Models (PLLMs) aim to align model outputs with individual user preferences, a crucial capability for user-centric applications. However, the prevalent approach of fine-tuning a separate module for each user faces two major limitations: (1) storage costs scale linearly with the number of users, rendering the method unscalable; and (2) fine-tuning a static model from scratch often yields suboptimal performance for users with sparse data. To address these challenges, we propose MTA, a Merge-then-Adapt framework for PLLMs. MTA comprises three key stages. First, we construct a shared Meta-LoRA Bank by selecting anchor users and pre-training meta-personalization traits within meta-LoRA modules. Second, to ensure scalability and enable dynamic personalization combination beyond static models, we introduce an Adaptive LoRA Fusion stage. This stage retrieves and dynamically merges the most relevant anchor meta-LoRAs to synthesize a user-specific one, thereby eliminating the need for user-specific storage and supporting more flexible personalization. Third, we propose a LoRA Stacking for Few-Shot Personalization stage, which applies an additional ultra-low-rank, lightweight LoRA module on top of the merged LoRA. Fine-tuning this module enables effective personalization under few-shot settings. Extensive experiments on the LaMP benchmark demonstrate that our approach outperforms existing SOTA methods across multiple tasks. Our code is also available.
Large Language Models (LLMs) are increasingly deployed as agents that invoke external tools through structured function calls. While recent work reports strong tool-calling performance under standard English-centric evaluations, the robustness of tool calling under multilingual user interactions remains underexplored. In this work, we introduce MLCL, a diagnostic benchmark, and conduct a systematic evaluation of multilingual tool calling across Chinese, Hindi, and the low-resource language Igbo. Through fine-grained error analysis, we show that many failures occur despite correct intent understanding and tool selection. We identify parameter value language mismatch as a dominant failure mode, where models generate semantically appropriate parameter values in the user’s language, violating language-invariant execution conventions. We further evaluate several inference-time system strategies and find that while these strategies substantially reduce language-induced execution errors, none of them can fully recover English-level performance.
CT-FineBench: A Diagnostic Fidelity Benchmark for Fine-Grained Evaluation of CT Report Generation
PDF ↗The evaluation of generated reports remains a critical challenge in Computed Tomography (CT) report generation, due to the large volume of text, the diversity and complexity of findings, and the presence of fine-grained, disease-oriented attributes. Conventional evaluation metrics offer only coarse measures of lexical overlap or entity matching and fail to reflect the granular diagnostic accuracy required for clinical use. To address this gap, we propose CT-FineBench, a benchmark built from CT-RATE and Merlin to evaluate the fine-grained factual consistency of CT reports, constructed from CT-RATE and Merlin. Our benchmark is constructed through a meticulous, Question-Answering (QA) based process: first, we identify and structure key, finding-specific clinical attributes (e.g., location, size, margin). Second, we systematically transform these attributes into a QA dataset, where questions probe for specific clinical details grounded in gold-standard reports. The evaluation protocol for CT-FineBench involves using this QA dataset to query a machine-generated report and scoring the correctness of the answers. This allows for a comprehensive, interpretable, and clinically-relevant assessment, moving beyond superficial lexical overlap to pinpoint specific clinical errors. Experiments show that CT-FineBench correlates better with expert clinical assessment and is substantially more sensitive to fine-grained factual errors than prior metrics.
Trajectory2Task: Training Robust Tool-Calling Agents with Synthesized Yet Verifiable Data for Complex User Intents
PDF ↗Tool-calling agents are increasingly deployed in real-world customer-facing workflows. Yet most studies on tool-calling agents focus on idealized settings with general, fixed, and well-specified tasks.In real-world applications, user requests are often (1) ambiguous, (2) changing over time, or (3) infeasible due to policy constraints, and training and evaluation data that cover these diverse, complex interaction patterns remain under-represented.To bridge the gap, we present Trajectory2Task a verifiable data generation pipeline for studying tool use at scale under three realistic user scenarios: ambiguous intent, changing intent, and infeasible intents.The pipeline first conducts multi-turn exploration to produce valid tool-call trajectories. It then converts these trajectories into user-facing tasks with controlled intent adaptations. This process yields verifiable task that support closed-loop evaluation and training. We benchmark several state-of-the-art LLMs on the generated complex user scenario tasks and observe frequent failures.Finally, using successful trajectories obtained from task rollouts, we fine-tune lightweight LLMs and find consistent improvements across all three conditions, along with better generalization to unseen tool-use domains, indicating stronger tool-calling ability.
The Bitter Lesson of Diffusion Language Models for Agentic Workflows: A Comprehensive Reality Check
PDF ↗The pursuit of real-time agentic interaction has driven interest in Diffusion-based Large Language Models (dLLMs) as alternatives to auto-regressive backbones, promising to break the sequential latency bottleneck. **However, does such efficiency gains translate into effective agentic behavior?** In this work, we present a comprehensive evaluation of dLLMs (e.g., LLaDA, Dream) across two distinct agentic paradigms: Embodied Agents (requiring long-horizon planning) and Tool-Calling Agents (requiring precise formatting).Contrary to the efficiency hype, our results on Agentboard and BFCL reveal a "**bitter lesson**": current dLLMs fail to serve as reliable agentic backbones, frequently leading to systematically failure. **(1) In Embodied settings**, dLLMs suffer repeated attempts, failing to branch under temporal feedback. **(2) In Tool-Calling settings**, dLLMs fail to maintain symbolic precision (e.g. strict JSON schemas) under diffusion noise. To assess the potential of dLLMs in agentic workflows, we introduce **DiffuAgent**, a multi-agent evaluation framework that integrates dLLMs as plug-and-play cognitive cores. Our analysis shows that dLLMs are effective in non-causal roles (e.g., memory summarization and tool selection) but require the incorporation of causal, precise, and logically grounded reasoning mechanisms into the denoising process to be viable for agentic tasks.
RATE: Reviewer Profiling and Annotation-free Training for Expertise Ranking in Peer Review Systems
PDF ↗Reviewer assignment is increasingly critical yet challenging in the LLM era, where rapid topic shifts render many pre-2023 benchmarks outdated and where proxy signals poorly reflect true reviewer familiarity. We address this evaluation bottleneck by introducing LR-bench, a high-fidelity, up-to-date benchmark curated from 2024–2025 AI/NLP manuscripts with five-level self-assessed familiarity ratings collected via a large-scale email survey, yielding 1,055 expert-annotated paper–reviewer–score annotations. We further propose a reviewer-centric ranking framework that distills each reviewer’s recent publications into compact keyword-based profiles and fine-tunes an embedding model with weak preference supervision constructed from heuristic retrieval signals, enabling to match each manuscript against a reviewer profile directly. Across the LR-bench and the CMU gold-standard dataset, our approach consistently achieves state-of-the-art performance, outperforming strong embedding baselines by a clear margin. We release LR-bench at https://huggingface.co/datasets/Gnociew/LR-bench, and an github repository at https://github.com/Gnociew/RATE-Reviewer-Assignment.
Can LLM Agents Simulate Multi-Turn Human Behavior? Evidence from Real Online Customer Behavior Data
PDF ↗Recent research shows that LLM Agents can generate “believable” human behaviors via prompt-only methods, and such agents have been increasingly adopted in downstream applications. However, existing evaluation of these agents only focuses on qualitative believability (whether human raters think they are accurate), leaving open questions of whether LLM agents can accurately generate step-by-step actions mimicking a particular human’s behavior in a multi-turn interaction task. In this work, we take shopping as a case study and present the first large-scale quantitative evaluation of state-of-the-art LLMs’ ability to accurately simulate human behavior. Using real-world data from 31,865 online shopping sessions containing 230,965 user actions, our evaluation reveals that prompt-based LLMs (DeepSeek-R1, Llama, Claude) achieve only 11.86% accuracy in generating human actions, highlighting a substantial gap in actual behavioral accuracy. Through experiments, we also showcase that strategies as simple as fine-tuning LLMs on real human click-through data augmented with synthesized reasoning traces can greatly enhance models’ performance. The fine-tuned Qwen2.5-7B achieves 17.26% action generation accuracy and 33.86% F1 score on final purchase prediction, representing substantial improvements of 5.4% and 13.85% over prompt-only baselines. This work establishes the first rigorous benchmark and dataset for human behavior simulation and provides actionable insights for developing more accurate LLM agents for future downstream applications.