Prompt injection attacks pose serious security risks across a wide range of real-world applications. While receiving increasing attention, the community faces a critical gap: the lack of a unified platform for prompt injection evaluation. This makes it challenging to reliably compare defenses, understand their true robustness under diverse attacks, or assess how well they generalize across tasks and benchmarks. For instance, many defenses initially reported as effective were later found to exhibit limited robustness on diverse datasets and attacks. To bridge this gap, we introduce PIArena, a unified and extensible platform for prompt injection evaluation that enables users to easily integrate state-of-the-art attacks and defenses and evaluate them across a variety of existing and new benchmarks. We also design a dynamic strategy-based attack that adaptively optimizes injected prompts based on defense feedback. Through comprehensive evaluation using PIArena, we uncover critical limitations of state-of-the-art defenses: limited generalizability across tasks, vulnerability to adaptive attacks, and fundamental challenges when an injected task aligns with the target task. The code and datasets are available at https://github.com/sleeepeer/PIArena.
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
Competitive programming has become a rigorous benchmark for evaluating the reasoning and problem-solving capabilities of large language models (LLMs). The International Olympiad in Informatics (IOI) stands out as one of the most prestigious annual competitions in competitive programming and has become a key benchmark for comparing human and AI-level programming ability. While several proprietary models have been claimed to achieve gold medal-level performance at the IOI, often with undisclosed methods, achieving comparable results with open-weight models remains a significant challenge. In this paper, we present GenCluster, a scalable and reproducible test-time compute framework that attains IOI gold-level performance using open-weight models. It combines large-scale generation, behavioral clustering, ranking, and a round-robin submission strategy to efficiently explore diverse solution spaces under limited validation budgets. Our experiments show that the performance of our proposed approach scales consistently with available compute, narrowing the gap between open and closed systems. Notably, we will show that GenCluster can achieve a gold medal at IOI 2025 for the first time with an open-weight model gpt-oss-120b, setting a new benchmark for transparent and reproducible evaluation of reasoning in LLMs.
Large Language Models (LLMs) for code generation (i.e., Code LLMs) have demonstrated impressive capabilities in AI-assisted software development and testing. However, recent studies have shown that these models are prone to generating vulnerable or even malicious code under adversarial settings. Existing red-teaming approaches rely on extensive human effort, limiting their scalability and practicality, and generally overlook the interactive nature of real-world AI-assisted programming, which often unfolds over multiple turns. To bridge these gaps, we present RedCoder, a red-teaming agent that engages victim models in multi-turn conversation to elicit vulnerable code. The pipeline to construct RedCoder begins with a multi-agent gaming process that simulates adversarial interactions, yielding a set of prototype conversations and an arsenal of reusable attack strategies. We then fine-tune an LLM on these prototype conversations to serve as the backbone of RedCoder. Once deployed, RedCoder autonomously engages Code LLMs in multi-turn conversations, dynamically retrieving relevant strategies from the arsenal to steer the dialogue toward vulnerability-inducing outputs. Experiments across multiple Code LLMs show that our approach outperforms prior single-turn and multi-turn red-team methods in inducing vulnerabilities in code generation, offering a scalable and effective tool for evaluating the security boundaries of modern code-generation systems.
Speculative Decoding has emerged as a popular technique for accelerating inference in Large Language Models. However, most existing approaches yield only modest improvements in production serving systems. Methods that achieve substantial speedups typically rely on an additional trained draft model or auxiliary model components, increasing deployment and maintenance complexity. This added complexity reduces flexibility, particularly when serving workloads shift to tasks, domains, or languages that are not well represented in the draft model’s training data.We introduce Simply-Scalable Speculative Decoding (SSSD), a training-free method that combines lightweight n-gram matching with hardware-aware speculation. Relative to standard autoregressive decoding, SSSD reduces latency by up to 2.9×. It achieves performance on par with leading training-based approaches across a broad range of benchmarks, while requiring substantially lower adoption effort—no data preparation, training or tuning are needed—and exhibiting superior robustness under language and domain shift, as well as in long-context settings.
LoPT: Lossless Parallel Tokenization Acceleration for Long Context Inference of Large Language Model
PDF ↗Long context inference scenarios have become increasingly important for large language models, yet they introduce significant computational latency. While prior research has optimized long-sequence inference through operators, model architectures, and system frameworks, tokenization remains an overlooked bottleneck. Existing parallel tokenization methods accelerate processing through text segmentation and multi-process tokenization, but they suffer from inconsistent results due to boundary artifacts that occur after merging. To address this, we propose LoPT, a novel Lossless Parallel Tokenization framework that ensures output identical to standard sequential tokenization. Our approach employs character-position-based matching and dynamic chunk length adjustment to align and merge tokenized segments accurately. Extensive experiments across diverse long-text datasets demonstrate that LoPT achieves significant speedup while guaranteeing lossless tokenization. We also provide theoretical proof of consistency and comprehensive analytical studies to validate the robustness of our method.
Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) dominate the post-training landscape for mathematical reasoning, yet differ fundamentally in their reliance on expert trajectories. To understand the optimal way to harness these trajectories for maximizing performance, we propose the Plasticity-Ceiling Framework. This framework empirically grounds the post-training landscape by decomposing the final performance ceiling into the foundational SFT performance and the subsequent RL plasticity (i.e., the maximum improvement via RL). Through extensive benchmarking, we establish the Sequential SFT-then-RL pipeline as the superior standard, overcoming the stability and premature convergence deficits inherent in synchronized approaches. Furthermore, we derive precise scaling guidelines: (1) Transitioning to RL at the Stable or Mild Overfitting Regime of SFT maximizes the final ceiling by securing a robust SFT foundation with substantial RL plasticity; (2) Refuting the “Less is More” hypothesis in SFT-then-RL scaling, we demonstrate that Data Scale determines the primary post-training potential, while Trajectory Difficulty acts as a performance multiplier; and (3) The Minimum Validation Loss of SFT serves as a reliable indicator for selecting the expert trajectories that maximize the ultimate performance ceiling. Our findings provide actionable guidelines for extracting maximum value from expert trajectories.
Language identification (LID) is a fundamental step in curating multilingual corpora. However, LID models still perform poorly for many languages, especially on the noisy and heterogeneous web data often used to train multilingual language models. In this paper, we introduce CommonLID, a community-driven, human-annotated LID benchmark for the web domain, covering 109 languages. Many of the included languages have been previously under-served, making CommonLID a key resource for developing more representative high-quality text corpora. We show CommonLID’s value by using it, alongside five other common evaluation sets, to test eight popular LID models. We analyse our results to situate our contribution and to provide an overview of the state of the art. In particular, we highlight that existing evaluations overestimate LID accuracy for many languages in the web domain. We make CommonLID and the code used to create it available under an open, permissive license.
Active Few-Shot Learning (AFSL) adapts LLMs to specialized domains by identifying the most valuable unlabeled samples for annotation and use as few-shot demonstrations, effectively reducing human annotation costs while promoting high performance. However, existing methods typically rely on output-level signals for the sample identification, such as predictive entropy or semantic similarities with test-time data based on external embeddings, which often overlook models’ internal dynamics which could pinpoint specific knowledge gaps. To bridge this gap, we propose NeuFS, a Neuron-Aware Active Few-Shot Learning framework that shifts the selection paradigm from output-level proxies to models’ internal dynamics. NeuFS utilizes neuron activation patterns to represent sample directly, and includes a dual-criteria selection strategy that: (1) ensures few-shot sample diversity with neuron patterns for broader example coverage, while (2) prioritizing on identifying informative and challenging few-shot samples LLMs tend to hallucinate by quantifying neuron consensus. Experiments on three datasets demonstrate that NeuFS excels in both reasoning and text classification tasks, outperforming existing AFSL baselines. Ablation studies further highlight that internal neuron activations provide a more principled and effective selection signal than external embeddings, validating the superiority of the proposed NeuFS.
Leveraging Outline-Optimized Generative Interactions and Critique for Self-Refining Outlines with Reinforcement Learning
PDF ↗Long-form outline generation requires satisfying multiple competing objectives simultaneously: outlines must be engaging, well-organized, topically relevant, and comprehensive while maintaining logical consistency across hierarchical structures. Current approaches either rely on expensive multi-turn interactions with large language models or employ procedural refinement pipelines that cannot systematically learn from critique. We present Logic-RL, a framework that transforms critique-guided outline refinement into a learnable policy through reinforcement learning. Our approach constructs refinement trajectories from teacher demonstrations, synthesizes explicit reasoning chains that decompose the critique-revision process, and optimizes a refinement policy using group relative policy optimization with structure-aware rewards. Experiments on FreshWiki and WikiOutline demonstrate that Logic-RL achieves substantial improvements over strong baselines, with the 0.6B model obtaining 79.17% relative gain and the 1.7B model achieving 8.67% improvement in average rubric scores compared to the best existing methods. Further analysis reveals that learned refinement policies generalize across domains and can be iteratively applied, with quality continuing to improve through three refinement rounds before diminishing returns.
Reinforcement learning (RL) is central to post-training, particularly for agentic models that require specialized reasoning behaviors. In this setting, model merging offers a practical mechanism for integrating multiple RL-trained agents from different tasks into a single generalist model. However, existing merging methods are designed for supervised fine-tuning (SFT), and they are suboptimal to preserve task-specific capabilities on RL-trained agentic models. The root is a task-vector mismatch between RL and SFT: on-policy RL induces task vectors that are highly sparse and heterogeneous, whereas SFT-style merging implicitly assumes dense and globally comparable task vectors. When standard global averaging is applied under this mismatch, RL’s non-overlapping task vectors that encode critical task-specific behaviors are reduced and parameter updates are diluted. To address this issue, we propose Reinforced Agent Merging (RAM), a distribution-aware merging framework explicitly designed for RL-trained agentic models. RAM disentangles shared and task-specific unique parameter updates, averaging shared components while selectively preserving and rescaling unique ones to counteract parameter update dilution. Experiments across multiple agent domains and model architectures demonstrate that RAM not only surpasses merging baselines, but also unlocks synergistic potential among agents to achieve performance superior to that of specialized agents in their domains.
The GDN-CC Dataset: Automatic Corpus Clarification for AI-enhanced Democratic Citizen Consultations
PDF ↗LLMs are ubiquitous in modern NLP, and while their applicability extends to texts produced for democratic activities such as online deliberations or large-scale citizen consultations, ethical questions have been raised for their usage as analysis tools. We continue this line of research with two main goals: (a) to develop resources that can help standardize citizen contributions in public forums at the pragmatic level, and make them easier to use in topic modeling and political analysis; (b) to study how well this standardization can reliably be performed by small, open-weights LLMs, i.e. models that can be run locally and transparently with limited resources. Accordingly, we introduce Corpus Clarification as a preprocessing framework for large-scale consultation data that transforms noisy, multi-topic contributions into structured, self-contained argumentative units ready for downstream analysis. We present GDN-CC, a manually-curated dataset of 1,231 contributions to the French Grand Débat National, comprising 2,285 argumentative units annotated for argumentative structure and manually clarified. We then show that finetuned Small Language Models match or outperform LLMs on reproducing these annotations, and measure their usability for an opinion clustering task. We finally release GDN-CC-large, an automatically annotated corpus of 240k contributions, the largest annotated democratic consultation dataset to date.
Cross-Tokenizer Knowledge Distillation (CTKD) enables knowledge transfer between a large language model and a smaller student, even when they employ different tokenizers. While existing approaches mainly focus on token-level alignment strategies, which are often brittle and sensitive to discrepancies between tokenizers, we argue that the method of aggregating tokens into more robust representations before distillation is of equal importance. In this paper, we introduce SRA (Span Representation Alignment for Large Language Model Distillation), a novel framework that reframes CTKD through the physical lens of Multi-Particle Dynamical Systems. SRA shifts the fundamental unit of alignment from tokens to robust, tokenizer-agnostic spans. We model each span as a cluster of particles and represent its state by its Center of Mass (CoM) - an attention-weighted average that captures rich semantic information. We leverage the concept of span centers of mass with attention-derived weighting to prioritize the most salient spans. In addition, we employ a geometric regularizer to preserve the structural integrity of the representation space and introduce aligned span logit distillation to enhance knowledge transfer across models. In challenging cross-architecture distillation experiments, SRA consistently and significantly outperforms state-of-the-art CTKD baselines, validating our physically-grounded approach.
Large Language Models (LLMs) often produce texts that appear coherent and credible, even when their factual reliability is uncertain. This paper investigates whether such perceived credibility correlates with the pervasive use of generics—generalizations without explicit quantification. We introduce a text-level genericity score derived from clause-level annotations and apply it to argumentative essays produced by humans and LLMs. To analyze how generics are realized in discourse, we employ Rhetorical Structure Theory to examine coherence relations across varying levels of genericity. Results show that according to our genericity metric, human texts are less generic than LLM-produced texts. As regards discourse, higher genericity correlates with less structured, paratactic structures, while for some models coherence is maintained through elaboration relations. Our findings suggest that some LLMs maintain well-structured coherence even in highly generic texts, which might enable them to "camouflage" argumentative texts as informative, enhancing their perceived credibility and persuasiveness.
Conversational agents are required to respond to their users not only with high quality (i.e. commonsense bearing) responses, but also considering multiple plausible alternative scenarios, reflecting the diversity in their responses. Despite the growing need to train diverse commonsense generators, the progress of this line of work has been significantly hindered by the lack of large-scale high-quality diverse commonsense training datasets. Due to the high annotation costs, existing Generative Commonsense Reasoning (GCR) datasets are created using a small number of human annotators, covering only a narrow set of commonsense scenarios. To address this training resource gap, we propose a two-stage method to create the first-ever synthetic dataset CommonSyn for diversified (GCR). The model fine-tuned on our synthetic data jointly increase both generation diversity and quality compared with vanilla models and the model fine-tuned on human-crafted dataset across different size Large Language Models (LLMs)
RSMeM: Knowledge-Enhanced Memory Evolution for Remote Sensing Agents with Systematic Evaluation
PDF ↗Geoscience research requires complex analysis and domain expertise, with remote sensing (RS) observations as a key foundation. However, existing RS agents built on general-purpose LLMs remain largely domain-agnostic, resulting in brittle and error-prone workflows. Moreover, these failures are seldom consolidated into a reusable experience for subsequent analyses. To address this issue, we introduce RSMeM, a knowledge-enhanced memory evolution mechanism that bootstraps RS agents with pre-distilled domain knowledge and iteratively integrates online experience for robust multi-step tool execution. RSMeM is composed of two components: (i) Hierarchical Knowledge Grounding, which performs taxonomy-aware retrieval over a hierarchical domain corpus to guide planning and tool selection; and (ii) Failure-Aware Experience Refinement, which distills failure-annotated tool-use traces into reusable constraints for next-round tool execution. By iteratively employing these two processes, RS agents can evolve to absorb task-level domain knowledge and effectively translate it into instance-level execution experience. Extensive experiments on EarthBench demonstrate that RSMeM consistently improves tool-use performance and end-to-end accuracy across a diverse set of LLM backbones. Notably, RSMeM achieves a 6% accuracy improvement on DeepSeek-V3.2 with less than 1% additional experience tokens, demonstrating the knowledge density of our distilled experience. All codes and models will be released to support reproducible research.
Recent advances in text-to-image (T2I) generation have enabled visually coherent image synthesis from descriptions, but generating images containing multiple given subjects remains challenging. As the number of reference identities increases, existing methods often suffer from subject missing and semantic drift. To address this problem, we propose MUSIC, the first MLLM specifically designed for MUlti-Subject In-Context image generation. To overcome the data scarcity, we introduce an automatic and scalable data generation pipeline that eliminates the need for manual annotation. Furthermore, we enhance the model’s understanding of multi-subject semantic relationships through a vision chain-of-thought (CoT) mechanism, guiding step-by-step reasoning from subject images to semantics and generation. To mitigate identity entanglement and manage visual complexity, we develop a novel semantics-driven spatial layout planning method and demonstrate its test-time scalability. By incorporating complex subject images during training, we improve the model’s capacity for chained reasoning. In addition, we curate MSIC, a new benchmark tailored for multi-subject in-context generation. Experimental results demonstrate that MUSIC significantly surpasses other methods in both multi- and single-subject scenarios.
Enhanced Reasoning for Biomedical Document-Level Relation Extraction via a Novel Cascade Language Model Framework
PDF ↗Biomedical document-level relation extraction poses significant challenges beyond sentence-level tasks, as it necessitates the integration of evidence from entire documents and the ability for coherent cross-sentence reasoning. While pretrained language models (PLMs) demonstrate efficiency in handling local contexts, they often struggle with global dependency modeling. Conversely, large language models (LLMs) exhibit strong reasoning capabilities but tend to generate hallucinations in knowledge-intensive biomedical domains. This paper introduces CoRE, a novel cascade framework that leverages the complementary strengths of PLMs and LLMs through a detect-then-rethink paradigm. The PLM serves as an efficient detector for high-confidence relations, while challenging cases are forwarded to an LLM enhanced with semantic retrieval and iterative reasoning mechanisms. Experimental results on BioRED and CDR datasets show that CoRE achieves substantial improvements over state-of-the-art baselines, validating the effectiveness of the proposed cascade paradigm for complex biomedical relation extraction.
Supervised Fine-Tuning (SFT) of large language models often suffers from task interference and catastrophic forgetting. Recent approaches alleviate this issue by isolating task-critical parameters during training. However, these methods represent a static solution to a dynamic problem, assuming that parameter importance remains fixed once identified. In this work, we empirically demonstrate that parameter importance exhibits temporal drift over the course of training. To address this, we propose Evolving Parameter Isolation (EPI), a fine-tuning framework that adapts isolation decisions based on online estimates of parameter importance. Instead of freezing a fixed subset of parameters, EPI periodically updates isolation masks using gradient-based signals, enabling the model to protect emerging task-critical parameters while releasing outdated ones to recover plasticity. Experiments on diverse multi-task benchmarks demonstrate that EPI consistently reduces interference and forgetting compared to static isolation and standard fine-tuning, while improving overall generalization. Our analysis highlights the necessity of synchronizing isolation mechanisms with the evolving dynamics of learning diverse abilities.
Terminal simulation, framed as a terminal command-level Turing test, is a long-standing problem of symbolic language generation in dialogue and interactive systems. Prior scripted simulators lack the flexibility needed for complex, multi-turn interactions, while LLM-based approaches often misinterpret commands, break output formats, drift from system state, and remain vulnerable to prompt injection. In this work, we propose MANTIS, a terminal simulation framework that improves realism, consistency, and robustness in command-language generation. MANTIS integrates a multi-agent architecture with a filter-based routing model that safely dispatches commands to external tools or an LLM-based agent, enabling support for interactive commands while defending against prompt injection attacks. In addition, we design an agentic file system with history pruning to preserve long-term state consistency. We release three datasets: 28,045 real terminal input-output pairs, a 1,000-session multi-turn interaction dataset, and a 25,849-instance labeled classification dataset. MANTIS outperforms state-of-the-art baselines by more than 9%, achieving over 95% accuracy on multi-turn terminal simulation. The dataset and source code are available at https://github.com/kaiwei666a/MANTIS_Terminal_Simulation
Diagnosing Spatial Consistency across Perspectives and Viewpoints in Large Vision-Language Models
PDF ↗Consistent reasoning about 3D spatial relations across changing viewpoints is fundamental for Embodied AI agents operating in dynamic environments. While Large Vision-Language Models (LVLMs) have advanced multimodal perception, their ability to maintain spatial consistency across diverse perspectives remains underexplored. Existing benchmarks primarily assess spatial capabilities from a static, single-view, and egocentric perspective, failing to capture the dynamic nature of real-world spatial cognition.To address this gap, we introduce SCOPE (Spatial COnsistency across PErspectives and Viewpoints), a comprehensive benchmark designed to rigorously diagnose spatial reasoning capabilities. Grounded in human cognitive theories of dual spatial representations, SCOPE discretizes the 360^\circ field into multiview scenarios to systematically evaluate both allocentric and egocentric reasoning capabilities. Our dataset comprises 20.1K spatial VQA pairs derived from high-quality 3D environments. Through an extensive evaluation of 26 state-of-the-art LVLMs, we identify two fundamental limitations that prevent consistent spatial understanding across viewpoints.We hope SCOPE facilitates the diagnosis of spatial reasoning, serving as a stepping stone toward reliable embodied action.