The effective detection and governance of Large Language Model (LLM) generated content has become increasingly critical due to the growing risk of misuse. Despite the impressive performance of existing detectors, their reliability and potential in multilingual, real-world scenarios remain largely underexplored.In this study, we introduce DetectRL-X, a comprehensive multilingual benchmark designed to evaluate advanced detectors across 8 dimensions. The benchmark encompasses 8 languages commonly used in commercial contexts and collects human-written texts from 6 domains highly susceptible to LLM misuse. To better aligned with real-world applications, We create LLM-generated texts using 4 popular commercial LLMs, and include typical AI-assisted writing operations such as polishing, expanding, and condensing to capture authentic usage patterns. Furthermore, we develop a multilingual framework for paraphrasing and perturbation attacks to simulate diverse human modifications and writing noise, enabling stress testing of detectors across languages.Experimental results on DetectRL-X reveal the strengths and limitations of current state-of-the-art detectors when applied to diverse linguistic resources. We further analyze how domains, generators, attack strategies, text length, and refinement operations influence performance in different languages, underscoring DetectRL-X as an effective benchmark for strengthening multilingual and language-specific detectors.
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
Reasoning capabilities in large language models (LLMs) have generally advanced significantly. However, it is still challenging for existing reasoning-based LLMs to perform effective decision-making abilities in multi-agent environments, due to the absence of explicit foresight modeling. To this end, strategic reasoning, the most fundamental capability to anticipate the counterpart’s behaviors and foresee its possible future actions, has been introduced to alleviate the above issues. Strategic reasoning is fundamental to effective decision-making in multi-agent environments, yet existing reasoning enhancement methods for LLMs do not explicitly capture its foresight nature. In this work, we introduce **Fo**resight **P**olicy **O**ptimization (**FoPO**) to enhance strategic reasoning in LLMs, which integrates opponent modeling principles into policy optimization, thereby enabling explicit consideration of both self-interest and counterpart influence. Specifically, we construct two curated datasets, namely ***Cooperative RSA*** and ***Competitive Taboo***, equipped with well-designed rules and moderate difficulty to facilitate a systematic investigation of FoPO in a self-play framework. Our experiments demonstrate that FoPO significantly enhances strategic reasoning across LLMs of varying sizes and origins. Moreover, models trained with FoPO exhibit strong generalization to out-of-domain strategic scenarios, substantially outperforming standard LLM reasoning optimization baselines.
Complex agentic AI systems, powered by a coordinated ensemble of Large Language Models (LLMs), tool and memory modules, have demonstrated remarkable capabilities on intricate, multi-turn tasks. However, this success is shadowed by prohibitive economic costs and severe latency, exposing a critical, yet underexplored, trade-off. We formalize this challenge as the Agent System Trilemma: the inherent tension among achieving state-of-the-art performance, minimizing monetary cost, and ensuring rapid task completion. To dismantle this trilemma, we introduce EvoRoute, a self-evolving model routing paradigm that transcends static, pre-defined model assignments. Leveraging an ever-expanding knowledge base of prior experience, EvoRoute dynamically selects Pareto-optimal LLM backbones at each step, balancing accuracy, efficiency, and resource use, while continually refining its own selection policy through environment feedback. Experiments on challenging agentic benchmarks such as GAIA and BrowseComp+ demonstrate that EvoRoute, when integrated into off-the-shelf agentic systems, not only sustains or enhances system performance but also reduces execution cost by up to 80% and latency by over 70%.
Emergent Misalignment via In-Context Learning: Narrow in-context examples can produce broadly misaligned LLMs
PDF ↗Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across four model families (Gemini, Kimi-K2, Grok, and Qwen), narrow in-context examples cause models to produce misaligned responses to benign, unrelated queries. With 16 in-context examples, EM rates range from 1% to 24% depending on model and domain, appearing with as few as 2 examples. Neither larger model scale nor explicit reasoning provides reliable protection, and larger models are typically even more susceptible. Next, we formulate and test a hypothesis, which explains in-context EM as conflict between safety objectives and context-following behavior. Consistent with this, instructing models to prioritize safety reduces EM while prioritizing context-following increases it. These findings establish ICL as a previously underappreciated vector for emergent misalignment that resists simple scaling-based solutions.
Culture serves as a fundamental determinant of human affective processing and profoundly shapes how individuals perceive and interpret emotional stimuli. Despite this intrinsic link extant evaluations regarding cultural alignment within Large Language Models primarily prioritize declarative knowledge such as geographical facts or established societal customs. These benchmarks remain insufficient to capture the subjective interpretative variance inherent to diverse sociocultural lenses. To address this limitation, we introduce CEDAR, a multimodal benchmark constructed entirely from scenarios capturing **C**ulturally **E**licited **D**istinct **A**ffective **R**esponses. To construct CEDAR, we implement a novel pipeline that leverages LLM-generated provisional labels to isolate instances yielding cross-cultural emotional distinctions, and subsequently derives reliable ground-truth annotations through rigorous human evaluation. The resulting benchmark comprises 10,962 instances across seven languages and 14 fine-grained emotion categories, with each language including 400 multimodal and 1,166 text-only samples. Comprehensive evaluations of 17 representative multilingual models reveal a dissociation between language consistency and cultural alignment, demonstrating that culturally grounded affective understanding remains a significant challenge for current models.
Vision–Language Models (VLMs) are increasingly used in clinical diagnostics, yet their robustness to adversarial attacks remains largely unexplored, posing serious risks. Existing medical attacks focus on secondary objectives such as model stealing or adversarial fine-tuning, while transferable attacks from natural images introduce visible distortions that clinicians can easily detect. To address this, we propose MedFocusLeak, a highly transferable black-box multimodal attack that induces incorrect yet clinically plausible diagnoses while keeping perturbations imperceptible. The method injects coordinated perturbations into non-diagnostic background regions and employs an attention-distraction mechanism to shift the model’s focus away from pathological areas. Extensive evaluations across six medical imaging modalities show that MedFocusLeak achieves state-of-the-art performance, generating misleading yet realistic diagnostic outputs across diverse VLMs. We further introduce a unified evaluation framework with novel metrics that jointly capture attack success and image fidelity, revealing a critical weakness in the reasoning capabilities of modern clinical VLMs.
Post-training quantization has emerged as a widely adopted technique for compressing and accelerating the inference of Large Language Models (LLMs). The primary challenges in LLMs quantization stem from activation outliers, which significantly degrade model performance especially at lower bit precision. While recent approaches attempt to mitigate outliers through linear transformations across feature dimensions, our analysis reveals that the transformed weights and activations still exhibit persistent outlier patterns with concentrated magnitude distributions. In this paper, we first model the mathematical relationship between quantization error and outliers, and then introduce a new metric Flatness to quantify the distribution of outliers. Based on this, we derive the theoretical optimal solution with respect to Flatness. Building on these insights, we propose Bidirectional Diagonal Quantization (BDQ), a novel post-training quantization framework that effectively disperses outlier patterns through optimized matrix transformations. BDQ strategically distributes outlier magnitudes across matrix dimensions via learned diagonal operations. Extensive experiments demonstrate that BDQ establishes a new quantization benchmark. It achieves less than 1% accuracy drop in W4A4 quantization on the LLaMA-3-8B model. In the more challenging W2A4KV16 experiment, compared to state-of-the-art approaches, BDQ reduces the performance gap by 39.1% on the DeepSeek-R1-Distill-LLaMA-70B model.
Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition in Large Reasoning Models
PDF ↗Recent Large Reasoning Models (LRMs) excel at complex reasoning tasks but often suffer from overthinking, generating overly long and redundant reasoning trajectories. To explore its essence, our empirical analysis reveals that LRMs are primarily limited to recognizing task properties (i.e., difficulty levels) like humans before solving the problem, leading to a one-size-fits-all reasoning strategy. This observation motivates a fundamental question: Can we explicitly bootstrap such ability to alleviate overthinking in LRMs? To this end, we propose Think-How-to-Think (TH2T), a novel two-stage fine-tuning strategy that progressively inspires LRMs’ difficulty cognition and redundancy cognition of LRMs. Specifically, we first inject Difficulty Dypnosis into output prefixes as cues for global, prospective reasoning strategy selection, stimulating the model’s sharper sensitivity to task complexity and adaptive control of reasoning depth. Then, we incorporate Redundancy Hypnosis into in-progress reasoning steps, which serve as local, retrospective signals for behavior correction by identifying and eliminating superfluous reasoning detours. Experiments across 7B/14B/32B models demonstrate that TH2T significantly reduces inference costs by over 70% on easy tasks and 40% on complex ones without compromising performance. The resultant models exhibit a nascent ability for difficulty-aware reasoning, effectively mitigating behaviors like excessive reflection and looping, thereby paving the way for more cognitively efficient LRMs.
TaxPraBen: A Scalable Benchmark for Structured Evaluation of LLMs in Chinese Real-World Tax Practice
PDF ↗While Large Language Models (LLMs) excel in various general domains, they exhibit notable gaps in the highly specialized, knowledge-intensive, and legally regulated Chinese tax domain. Consequently, while tax-related benchmarks are gaining attention, many focus on isolated NLP tasks, neglecting real-world practical capabilities. To address this issue, we introduce TaxPraBen, the first dedicated benchmark for Chinese taxation practice. It combines 10 traditional application tasks, along with 3 pioneering real-world scenarios: tax risk prevention, tax inspection analysis, and tax strategy planning, sourced from 14 datasets totaling 7.3K instances. TaxPraBen features a scalable structured evaluation paradigm designed through process of "structured parsing—field alignment extraction—numerical and textual matching", enabling end-to-end tax practice assessment while being extensible to other domains. We evaluate 19 LLMs based on Bloom’s taxonomy. The results indicate significant performance disparities: all closed-source large-parameter LLMs excel, and Chinese LLMs like Qwen2.5 generally exceed multilingual LLMs, while the YaYi2 LLM, fine-tuned with some tax data, shows only limited improvement. TaxPraBen[<https://anonymous.4open.science/r/TaxPraBen/>] serves as a vital resource for advancing evaluations of LLMs in practical applications.
The efficiency of multi-agent systems driven by large language models (LLMs) largely hinges on their communication topology. However, designing an optimal topology is a non-trivial challenge, as it requires balancing competing objectives such as task performance, communication cost, and robustness. Existing frameworks often rely on static or hand-crafted topologies, which inherently fail to adapt to diverse task requirements, leading to either excessive token consumption for simple problems or performance bottlenecks for complex ones. To address this challenge, we introduce a novel generative framework called Guided Topology Diffusion (GTD). Inspired by conditional discrete graph diffusion models, GTD formulates topology synthesis as an iterative construction process. At each step, the generation is steered by a lightweight proxy model that predicts multi-objective rewards (e.g., accuracy, utility, cost), enabling real-time, gradient-free optimization towards task-adaptive topologies. This iterative, guided synthesis process distinguishes GTD from single-step generative frameworks, enabling it to better navigate complex design trade-offs. We validated GTD across multiple benchmarks, and experiments show that this framework can generate highly task-adaptive, sparse, and efficient communication topologies, significantly outperforming existing methods in LLM agent collaboration. Our code is available at https://anonymous.4open.science/r/diffusion_agent-953C.
Developing culturally grounded multilingual AI systems remains challenging, particularly for low-resource languages. While synthetic data offers promise, its effectiveness in multilingual and multicultural contexts is underexplored. We investigate bottom-up synthetic data generation using large open-source LLMs (>= 235B parameters) grounded in language-specific Wikipedia content, complementing dominant top-down translation-based approaches from English. We introduce Updesh, a high-quality large-scale synthetic instruction-following dataset comprising 9.5M data points across 13 Indian languages and English, encompassing diverse reasoning and generative tasks emphasizing on enhancing long-context and multi-turn capabilities while improving alignment with Indian cultural contexts. Comprehensive evaluation using automated metrics and 10K human assessments confirms high data quality. Downstream evaluations performed by fine-tuning models on various datasets and assessing performance across 13 diverse multilingual datasets and model comparative evaluations, demonstrate that models trained on Updesh consistently obtain significant improvements on NLG tasks and remain competitive on NLU tasks. Improvements are most pronounced for low and medium-resource languages, effectively narrowing performance gaps with high-resource languages. Our findings provide empirical evidence that effective multilingual AI development requires multi-faceted, culturally grounded data curation strategies beyond translation-based approaches.
Cultural alignment in Large Language Models (LLMs) is essential for producing contextually aware, respectful, and trustworthy outputs. Without it, models risk generating stereotyped, insensitive, or misleading responses that fail to reflect cultural diversity w.r.t Helpful, Harmless, and Honest (HHH) paradigm. Existing benchmarks represent early steps toward cultural alignment; yet, no benchmarks currently enables systematic evaluation of cultural alignment in line with UNESCO’s principles of cultural diversity w.r.t HHH paradigm. Therefore, to address this gap, we built Align-Cultura, two-stage pipeline for cultural alignment. Stage I constructs CULTURAX, the HHH-English dataset grounded in the UNESCO cultural taxonomy, through Query Construction, which reclassifies prompts, expands underrepresented domains (or labels), and prevents data leakage with SimHash. Then, Response Generation pairs prompts with culturally grounded responses via two-stage rejection sampling. The final dataset contains 1,500 samples spanning 30 subdomains of tangible and intangible cultural forms. Stage II benchmarks CULTURAX on general-purpose models, culturally fine-tuned models, and open-weight LLMs (Qwen3-8B and DeepSeek-R1-Distill-Qwen-7B). Empirically, culturally fine-tuned models improve joint HHH by 4%–6%, reduce cultural failures by 18%, achieve 10%–12% efficiency gains, and limit leakage to 0.3%.
Cross-Examination Framework: A Task-Agnostic Diagnostic for Information Fidelity in Text-to-Text Generation
PDF ↗Traditional metrics like BLEU and BERTScore fail to capture semantic fidelity in generative text-to-text tasks. We adapt the Cross-Examination Framework (CEF) for a reference-free, multi-dimensional evaluation by treating the source and candidate as independent knowledge bases. CEF generates verifiable questions from each text and performs a cross-examination to derive three interpretable scores: Coverage, Conformity, and Consistency. Validated across translation, summarization and clinical note-generation, our framework identifies critical errors, such as content omissions and factual contradictions, missed by standard metrics. A key contribution is a systematic robustness analysis to select a stable judge model. Crucially, the strong correlation between our reference-free and with-reference modes validates CEF’s reliability without gold references. Furthermore, human expert validation demonstrates that CEF mismatching questions align with meaning-altering semantic errors higher than with non-semantic errors, particularly excelling at identifying entity-based and relational distortions.
While Large Language Models (LLMs) demonstrates remarkable zero-shot annotation tasks, they often struggle with the specialized conventions of gold-standard benchmarks. We propose the systematic reuse and refinement of annotation guidelines as an alignment mechanism, introducing an iterative moderation framework that simulates the early phases of annotation projects. We evaluate three hypotheses: (1) the efficacy of guideline integration, (2) the advantage of reasoning-optimized models, and (3) the viability of moderation under minimal supervision. Testing across biomedical NER tasks (NCBI Disease, BC5CDR, BioRED) with three LLM families (GPT, Gemini, DeepSeek), our results empirically confirm all three hypotheses. While the iterative moderation framework shows a good potential in effectively refining guidelines, our analysis also reveals a significant room for improvement.
Safety alignment of large language models currently faces a central challenge: existing alignment techniques often prioritize mitigating responses to harmful prompts at the expense of overcautious behavior, leading models to incorrectly refuse benign requests. A key goal of safe alignment is therefore to improve safety while simultaneously minimizing false refusals. In this work, we introduce Energy Landscape Steering (ELS), a novel, fine-tuning free framework designed to resolve this challenge through dynamic, inference-time intervention. We trained a lightweight, external Energy-Based Model (EBM) to assign high energy to undesirable (false refusal or jailbreak) states and low energy to desirable (helpful response or safe reject) ones. During inference, the EBM maps the LLM’s internal activations to an energy landscape, and we use the gradient of the energy function to steer the hidden states toward low-energy regions in real time. This dynamically guides the model toward desirable behavior without modifying its parameters. By decoupling behavioral control from the model’s core knowledge, ELS provides a flexible and computationally efficient solution. Extensive experiments across diverse models demonstrate its effectiveness: raising compliance on the ORB-H benchmark from 57.3% to 82.6% while maintaining the baseline safety performance. Our work establishes a promising paradigm for building LLMs that simultaneously achieve high safety and low false refusal rates.
Large Language Models (LLMs) have demonstrated remarkable capabilities in machine translation. However, maintaining discourse coherence and terminological consistency remains a persistent challenge in document-level translation (DocMT). Existing solutions, such as memory-based agents, predominantly rely on explicit context concatenation. This paradigm treats historical context as a static external resource, which often leads to context dilution, high inference latency, and superficial knowledge integration. To address these limitations, we propose AdaDPI, an adaptive agentic framework that shifts the DocMT paradigm from static retrieval to dynamic parametric internalization. Specifically, we design a linguistic uncertainty monitor (LUM) to actively detect critical discourse discontinuities by the model’s epistemic uncertainty. Upon detection, a context-to-parameter integrator (CPI) compiles retrieved external constraints directly into the model’s intrinsic state via an online parameter adaptation mechanism. Through the online parameter adaptation on a lightweight adapter, AdaDPI internalizes document-specific norms into the model’s intrinsic representations, enabling a progressive evolution of the translation strategy as the discourse unfolds. Extensive experiments on the discourse-rich GuoFeng and IWSLT2017 datasets demonstrate that AdaDPI significantly outperforms the SoTA baselines by more than 5 points on the consistency metric.
Quality Estimation (QE) aims to assess machine translation quality without reference translations, but recent studies have shown that existing QE models exhibit systematic gender bias. In particular, they tend to favor masculine realizations in gender-ambiguous contexts and may assign higher scores to gender-misaligned translations even when gender is explicitly specified. To address these issues, we propose FairQE, a multi-agent-based, fairness-aware QE framework that mitigates gender bias in both gender-ambiguous and gender-explicit scenarios. FairQE detects gender cues, generates gender-flipped translation variants, and combines conventional QE scores with LLM-based unbiased reasoning through a dynamic bias-aware aggregation mechanism. This design preserves the strengths of existing QE models while calibrating their gender-related biases in a plug-and-play manner. Extensive experiments across multiple gender bias evaluation settings demonstrate that FairQE consistently improves gender fairness over strong QE baselines. Moreover, under MQM-based meta-evaluation following the WMT 2023 Metrics Shared Task, FairQE achieves competitive or improved general QE performance. These results show that gender bias in QE can be effectively mitigated without sacrificing evaluation accuracy, enabling fairer and more reliable translation evaluation.
Social Dynamics as Critical Vulnerabilities that Undermine Objective Decision-Making in LLM Collectives
PDF ↗Large language model (LLM) agents are increasingly acting as human delegates in multi-agent environments, where a representative agent integrates diverse peer perspectives to make a final decision. Drawing inspiration from social psychology, we investigate how the reliability of this representative agent is undermined by the social context of its network. We define four key phenomena—social conformity, perceived expertise, dominant speaker effect, and rhetorical persuasion—and systematically manipulate the number of adversaries, relative intelligence, argument length, and argumentative styles. Our experiments demonstrate that the representative agent’s accuracy consistently declines as social pressure increases: larger adversarial groups, more capable peers, and longer arguments all lead to significant performance degradation. Furthermore, rhetorical strategies emphasizing credibility or logic can further sway the agent’s judgment, depending on the context. These findings reveal that multi-agent systems are sensitive not only to individual reasoning but also to the social dynamics of their configuration, highlighting critical vulnerabilities in AI delegates that mirror the psychological biases observed in human group decision-making.
Reinforcement learning with verifiable rewards (RLVR) is a promising approach for improving the complex reasoning abilities of large language models (LLMs). However, current RLVR methods face two significant challenges: the near-miss reward problem, where a small mistake can invalidate an otherwise correct reasoning process, greatly hindering training efficiency; and exploration stagnation, where models tend to focus on solutions within their ”comfort zone”, lacking the motivation to explore potentially more effective alternatives. To address these challenges, we propose StepHint, a novel RLVR algorithm that utilizes multi-level stepwise hints to help models explore the solution space more effectively. StepHint partitions valid reasoning chains into reasoning steps using our proposed adaptive partitioning method. The initial few steps are used as hints, and simultaneously, multiple-level hints (each comprising a different number of steps) are provided to the model. This approach directs the model’s exploration toward a promising solution subspace while preserving its flexibility for independent exploration. By providing hints, StepHint mitigates the near-miss reward problem, thereby improving training efficiency. Additionally, the external reasoning pathways help the model develop better reasoning abilities, enabling it to move beyond its ”comfort zone” and mitigate exploration stagnation. StepHint outperforms competitive RLVR enhancement methods across six mathematical benchmarks and two out-of-domain benchmarks.
Revealing Procedural Reasoning Structures in Chain-of-Thought Training via Span-Level Gradient Organization
PDF ↗Chain-of-Thought (CoT) prompting enables large language models to produce multi-step reasoning, yet how such reasoning-related structure is expressed during training remains poorly understood. We present Gradient-based Structural Developer (GSD), an unsupervised framework with a principled gradient aggregation view that tracks span-level gradient during fine-tuning on reasoning benchmarks to understand how models develop structured, step-by-step reasoning capabilities. Our analysis shows that while gradients at the level of individual tokens are often noisy, aggregating gradients over contiguous reasoning-related spans reveals stable and recurring directional alignment across samples. We refer to these directionally aligned patterns as aligned sequential stresses, reflecting consistent gradient organization associated with similar reasoning procedures. Beyond capturing semantically similar reasoning instances, such gradient alignment also reveals structurally similar but semantically diverse cases that share common procedural organization. These findings position GSD as a diagnostic framework for analyzing how procedural reasoning structures emerge during training, with downstream selection results serving as auxiliary evidence correlating gradient alignment with adaptation efficiency.