LLM-based agents are increasingly deployed to autonomously solve complex tasks, raising urgent needs for IP protection and regulatory provenance. While content watermarking effectively attributes LLM-generated outputs, it fails to directly identify the high-level planning behaviors (e.g., tool and subgoal choices) that govern multi-step execution. Critically, watermarking at the planning-behavior layer faces unique challenges: minor distributional deviations in decision-making can compound during long-term agent operation, degrading utility, and many agents operate as black boxes that are difficult to intervene in directly. To bridge this gap, we propose AgentMark, a behavioral watermarking framework that embeds multi-bit identifiers into planning decisions while preserving utility. It operates by eliciting an explicit behavior distribution from the agent and applying distribution-preserving conditional sampling, enabling deployment under black-box APIs while remaining compatible with action-layer content watermarking. Experiments across embodied, tool-use, and social environments demonstrate practical multi-bit capacity, robust recovery from partial logs, and utility preservation. Code is available at https://github.com/Tooooa/AgentMark.
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
Process Reward Models (PRMs), which assign fine-grained scores to intermediate reasoning steps within a solution trajectory, have emerged as a promising approach to enhance the reasoning quality of Large Language Models (LLMs).However, most existing PRMs rely on a unidirectional left-to-right (L2R) evaluation scheme, which restricts their utilization of global context.In light of this challenge, we propose a novel bidirectional evaluation paradigm, named \mathbf{Bi}directional \mathbf{P}rocess \mathbf{R}eward \mathbf{M}odel (\mathbf{BiPRM}).BiPRM incorporates a parallel right-to-left (R2L) evaluation stream, implemented via prompt reversal, alongside the conventional L2R flow.Then a gating mechanism is introduced to adaptively fuse the reward scores from both streams to yield a holistic quality assessment.Remarkably, compared to the original PRM, BiPRM introduces only a 0.3% parameter increase for the gating module, and the parallel execution of two streams incurs merely 5% inference time latency. Our extensive empirical evaluations spanning diverse benchmarks, LLM backbones, PRM objectives and sampling policies demonstrate that BiPRM consistently surpasses unidirectional baselines, achieving an average relative gain of 10.6% over 54 solution-level configurations and 37.7% in 12 step-level error detection scenarios.Generally, our results highlight the effectiveness, robustness and general applicability of BiPRM, offering a promising new direction for process-based reward modeling.
Vision-Language Models (VLMs) have demonstrated remarkable proficiency in general multi-modal understanding; yet they struggle to efficiently acquire continually evolving domain-specific skills. Conventional approaches to enhancing VLM capabilities, such as Supervised Fine-Tuning (SFT), require extensive dataset curation and substantial computational resources. Model merging has emerged as an efficient alternative that enables the transfer of domain-specific expertise from Large Language Models (LLMs) to VLMs without incurring additional training data requirements or significant computational overhead. Unlike conventional merging of homogeneous LLMs, which mainly aggregates existing capabilities, cross-modal skill injection aims to induce emergent cross-modal capabilities by integrating a domain-expert LLM into a VLM. However, existing research lacks a systematic analysis of the applicability and methodology of cross-modal skill injection. In this study, we investigate cross-modal skill injection across three main aspects: scenarios, methods, and hyperparameters. For scenarios, we find that cross-modal skill injection generally performs well in instruction-following and cross-lingual settings, yet struggles with mathematical reasoning. For methods, we find that classic approaches such as TA and DARE consistently achieve superior performance over alternative merging methods. We also provide a systematic and quantitative analysis of the hyperparameter tuning that these classic methods critically depend on.
Mock Worlds, Real Skills: Building Small Agentic Language Models with Synthetic Tasks, Simulated Environments, and Rubric-Based Rewards
PDF ↗Small LLMs often struggle to match the agentic capabilities of large, costly models. While reinforcement learning can help, progress has been limited by two structural bottlenecks: existing open-source agentic training data are narrow in task variety and easily solved; real-world APIs lack diversity and are unstable for large-scale reinforcement learning rollout processes. We address these challenges with SYNTHAGENT, a framework that jointly synthesizes diverse tool-use training data and simulates complete environments. Specifically, a strong teacher model creates novel tasks and tool ecosystems, then rewrites them into intentionally underspecified instructions. This compels agents to actively query users for missing details. When handling synthetic tasks, an LLM-based user simulator provides user-private information, while a mock tool system delivers stable tool responses. For rewards, task-level rubrics are constructed based on required subgoals, user-agent interactions, and forbidden behaviors. Across 14 challenging datasets in math, search, and tool use, models trained on our synthetic data achieve substantial gains, with small models showing performance comparable to some larger baselines in certain domains.
Large-scale social simulators are essential for studying complex social patterns. Prior work explores hybrid methods to scale up simulations, combining large language models (LLM)-based agents with numerical agent-based models (ABM). However, this incurs high latency due to expensive memory retrieval and sequential ABM execution. To address this challenge, we propose GASim, a graph-accelerated hybrid multi-agent framework for large-scale social simulations. For core agents driven by LLM, GASim introduces Graph-Optimized Memory (GOM) to replace intensive LLM-based retrieval pipelines with lightweight propagation over a sparse memory graph. For the majority of ordinary agents, GASim employs Graph Message Passing (GMP), substituting sequential ABM execution with parallel updates by fine-grained feature aggregation and Graph Attention Network. We further introduce Entropy-Driven Grouping (EDG) that coordinates this hybrid partitioning, leveraging information entropy to dynamically identify emergent core agents situated in information-diverse neighborhoods. Extensive experiments show that GASim not only delivers a substantial 9.94× end-to-end speedup over the traditional hybrid framework but also consumes less than 20% of baseline tokens, significantly reducing costs while preserving strong alignment with real-world public opinion trends.
Role-playing (RP) agents rely on behavioral profiles to act consistently across diverse narrative contexts, yet existing profiles are largely unstructured, non-executable, and weakly validated, leading to brittle agent behavior. We propose Codified Decision Trees (CDT), a data-driven framework that induces an executable and interpretable decision structure from large-scale narrative data. CDT represents behavioral profiles as a tree of conditional rules, where internal nodes correspond to validated scene conditions and leaves encode grounded behavioral statements, enabling deterministic retrieval of context-appropriate rules at execution time. The tree is learned by iteratively inducing candidate scene–action rules, validating them against data, and refining them through hierarchical specialization, yielding profiles that support transparent inspection and principled updates. Across multiple benchmarks, CDT substantially outperforms human-written profiles and prior profile induction methods, indicating that codified and validated behavioral representations lead to more reliable agent grounding.
Scientific writing is an expert-domain task that demands deep domain knowledge, task-specific requirements and reasoning capabilities that leverage the domain knowledge to satisfy the task specifications. While scientific text generation has been widely studied, its evaluation remains a challenging and open problem. It is critical to develop models that can be reliably deployed for evaluating diverse open-ended scientific writing tasks while adhering to their distinct requirements. However, existing LLM-based judges and reward models are primarily optimized for general-purpose benchmarks with fixed scoring rubrics and evaluation criteria. Consequently, they often fail to reason over sparse knowledge of scientific domains when interpreting task-dependent and multi-faceted criteria. Moreover, fine-tuning for each individual task is costly and impractical for low-resource settings. To bridge these gaps, we propose cost-efficient, open-source reward models tailored for scientific writing evaluation. We introduce a two-stage training framework that initially optimizes scientific evaluation preferences and then refines reasoning capabilities. Our multi-aspect evaluation design and joint training across diverse tasks enable fine-grained assessment and robustness to dynamic criteria and scoring rubrics. Experimental analysis shows that our training regime strongly improves LLM-based scientific writing evaluation. Our models generalize effectively across tasks and to previously unseen scientific writing evaluation settings, allowing a single trained evaluator to be reused without task-specific retraining.
The widespread utilization of language models in modern applications is inconceivable without Parameter-Efficient Fine-Tuning techniques, such as low-rank adaptation (LoRA), which adds trainable adapters to selected layers. Although LoRA may obtain accurate solutions, it requires significant memory to train large models and intuition on which layers to add adapters. In this paper, we propose a novel method, WeightLoRA, which overcomes this issue by adaptive selection of the most critical LoRA heads throughout the optimization process. As a result, we can significantly reduce the number of trainable parameters while maintaining the capability to obtain consistent or even superior metric values. We conduct experiments for a series of competitive benchmarks and DeBERTa, BART, Llama and Qwen models, comparing our method with different adaptive approaches. The experimental results demonstrate the efficacy of WeightLoRA and the superior performance of WeightLoRA+ in almost all cases. The source code is available at https://github.com/brain-lab-research/WLoRA
Author-in-the-Loop Response Generation and Evaluation: Integrating Author Expertise and Intent in Responses to Peer Review
PDF ↗Author response (rebuttal) writing is a critical stage of scientific peer review that demands substantial author effort. In practice, authors possess domain expertise, author-only information, and response strategies – concrete forms of author expertise and intent – and seek NLP assistance that integrates these signals into author response generation (ARG). Yet this author-in-the-loop paradigm lacks formal NLP formulation and systematic study: no dataset provides fine-grained author signals, existing ARG work lacks author inputs and controls, and no evaluation measures response reflection of author signals and effectiveness in addressing reviewer concerns. To fill these gaps, we introduce (i) Re³Align, the first large-scale dataset of aligned review–response–revision triplets, where revisions proxy author signals; (ii) REspGen, an author-in-the-loop ARG framework supporting flexible author input, multi-attribute control, and evaluation-guided refinement; and (iii) REspEval, a comprehensive evaluation suite with 20+ metrics spanning input utilization, controllability, response quality, and discourse. Experiments with SOTA LLMs demonstrate the benefits of author input and evaluation-guided refinement, the impact of input specificity on response quality, and controllability–quality trade-offs. We release our dataset, generation and evaluation tools.
Multilingual Idioms in Sentences and Conversations Across High-, Medium-, and Low-Resource Languages
PDF ↗Idiomatic expressions pose a major challenge for multilingual NLP because their meanings shift between figurative and literal usage, often requiring context for accurate interpretation. Prior work has focused on high-resource languages typically evaluates isolated idiom-meaning questions, overlooking realistic discourse. We introduce MIDI, a multilingual idiom dataset spanning 3 high-, 3 medium-, and 12 low-resource languages, curated by native speakers. Unlike previous datasets, MIDI provides idioms embedded in both sentence-level and conversational contexts, capturing both literal and figurative readings. Benchmarking state-of-the-art models shows that idiom comprehension degrades in low-resource languages and that, in all resource tiers, literal interpretations are substantially harder than figurative ones. Conversational context improves performance but does not eliminate these disparities. Through controlled tests and interventions on hidden representations, we further separate memorization from reasoning, exposing core limitations of current models.
Standard evaluators, such as reward models, compress diverse human judgments into a single scalar, conflating valid Subjective Preference with Cognitive Uncertainty. This structural mismatch often leads to brittle alignment and reward hacking. To address this, we propose PRISM which reinterprets reward evaluation as a conditional distribution parameterized by a Mixture of Gaussians. PRISM structurally disentangles these factors: distinct Gaussian experts emerge to capture conflicting preference dimensions, while their variance estimates quantify uncertainty, acting as a dynamic reliability gate during optimization. We introduce a two-stage training strategy to learn these disentangled representations from scalable pairwise comparisons without requiring massive fine-grained annotations. Empirical results show that PRISM significantly outperforms scalar baselines in both accuracy and generalization. Furthermore, in downstream Reinforcement Learning, PRISM effectively mitigates reward hacking, yielding policies that are more robust and resilient to distribution shifts.
Detecting pre-training data in Large Language Models (LLMs) is crucial for auditing data privacy and copyright compliance, yet it remains challenging in black-box, zero-shot settings where computational resources and training data are scarce. While existing likelihood-based methods have shown promise, they typically aggregate token-level scores using uniform weights, thereby neglecting the inherent information-theoretic dynamics of autoregressive generation. In this paper, we hypothesize and empirically validate that memorization signals are heavily skewed towards the high-entropy initial tokens, where model uncertainty is highest, and decay as context accumulates. To leverage this linguistic property, we introduce Positional Decay Reweighting (PDR), a training-free and plug-and-play framework. PDR explicitly reweights token-level scores to amplify distinct signals from early positions while suppressing noise from later ones. Extensive experiments show that PDR acts as a robust prior and can usually enhance a wide range of advanced methods across multiple benchmarks.
Capability Decomposition for Unified Information Extraction via Hierarchical Mixture-of-Experts
PDF ↗Unified Information Extraction (UIE) aims to handle heterogeneous IE tasks within a single framework, but existing methods often suffer from inconsistent schema representation, implicitly intermediate reasoning and full-parameter adaptation, which limit generalization, interpretability and parameter efficiency. To address these issues, we propose UC-UIE (Universal Capabilities-based Unified Information Extractor), a unified framework based on Large Language Model (LLM), which introduces a unified frame-and-slots schema for IE tasks and explicitly decomposes IE reasoning into three universal capabilities: judging, locating, and associating. Furthermore, UC-UIE adopts a Low-Rank Adaptation (LoRA) based hierarchical Mixture-of-Experts (MoE) adapter to fine-tune LLMs for IE tasks, which explicitly models these three capabilities in a task-driven way while ensuring parameter efficiency. With only 1.24% trainable parameters, UC-UIE outperforms full-parameter tuning methods, showing excellent parameter efficiency. Zero-shot evaluation reveals its strong generalization ability to unseen domains and schemas, benefiting from unified schema representation and explicit capability decomposition. Further experiments validate that the hierarchical MoE adapter learns capability specialization and composition, which enhances both UIE performance and interpretability.
Large Language Models (LLMs) achieve strong performance on many reasoning benchmarks, yet these evaluations typically focus on isolated tasks that differ from real-world usage in task-oriented dialogue (TOD). In this setting, LLMs must perform reasoning inherently while generating text and adhering to instructions on role, format, and style. This mismatch raises concerns about whether benchmark performance accurately reflects models’ reasoning robustness in TOD setting.We investigate how framing reasoning tasks within TOD affects LLM performance by introducing a new dynamic benchmark covering eight travel-related tasks that require arithmetic, spatial, temporal, and commonsense reasoning. Each problem is presented in both isolated and dialogue-based variants, enabling controlled comparison while mitigating data contamination. Experiments on nine LLMs reveal a substantial and consistent performance gap between isolated and dialogue settings. Through ablations and qualitative analysis, we show that this gap is largely driven by the multi-turn nature of dialogue, with additional effects from role conditioning and tool-use requirements. Our results highlight the need to evaluate LLM reasoning in realistic interactive scenarios.
Despite the recent success of instruction-tuned language models and their ubiquitous usage, very little is known of how models process instructions internally. In this work, we address this gap from a mechanistic point of view by investigating how instruction-specific representations are constructed and utilized in different stages of post-training: Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Via causal mediation, we identify that instruction representation is fairly localized in models. These representations, which we call Instruction Vectors (IVs), demonstrate a curious juxtaposition of linear separability along with non-linear causal interaction, broadly questioning the scope of the linear representation hypothesis commonplace in mechanistic interpretability. To disentangle the non-linear causal interaction, we propose a novel method to localize information processing in language models that is free from the implicit linear assumptions of patching-based techniques. We find that, conditioned on the task representations formed in the early layers, different information pathways are selected in the later layers to solve that task, i.e., IVs act as circuit selectors.
Mixture-of-Agents (MoA) improves LLM performance through layered collaboration, but its dense topology raises costs and latency. Existing methods employ LLM judges to filter responses, yet still require all models to perform inference before judging, failing to cut costs effectively. They also lack model selection criteria and struggle with large model pools, where full inference is costly and can exceed context limits. To address this, we propose **RouteMoA**, an efficient mixture-of-agents framework with dynamic routing. It employs a lightweight *scorer* to perform initial screening by predicting coarse-grained performance from the query, narrowing candidates to a high-potential subset without inference. A *mixture of judges* then refines these scores through lightweight self- and cross-assessment based on existing model outputs, providing posterior correction without additional inference. Finally, a *model ranking* mechanism selects models by balancing performance, cost, and latency. RouteMoA outperforms MoA across varying tasks and model pool sizes, reducing cost by 89.8% and latency by 63.6% in the large-scale model pool. Code is available at https://github.com/Jize-W/RouteMoA.
Table Question Answering in the Era of Large Language Models: A Comprehensive Survey of Tasks, Methods, and Evaluation
PDF ↗Table Question Answering (TQA) aims to answer natural language questions using tabular data, often accompanied by additional contexts, such as passages. The task spans diverse settings, varying in table representation, question/answer complexity, modality involved, and domain. While recent advances in large language models (LLMs) have driven substantial progress for TQA, the field still lacks a systematic organization and understanding of task formulations, core challenges, and methodological trends, particularly in light of emerging research directions, such as reinforcement learning. This survey addresses this gap by providing a comprehensive and structured overview of TQA research using LLMs. We introduce various task setups and summarize available benchmarks based on task features. We categorize current modeling strategies according to the challenges they target, and analyze their strengths and limitations. Furthermore, we highlight underexplored but timely topics that have not been systematically covered in prior surveys. By unifying disparate research threads and identifying open problems, our survey offers a consolidated foundation for the TQA community, enabling a deeper understanding of the state of the art and guiding future developments in this rapidly evolving area.
Focus-dLLM: Accelerating Long-Context Diffusion LLM Inference via Confidence-Guided Context Focusing
PDF ↗Diffusion Large Language Models (dLLMs) deliver strong long-context processing capability in a non-autoregressive decoding paradigm. However, the considerable computational cost of bidirectional full attention limits the inference efficiency. Although sparse attention is promising, existing methods remain ineffective. This stems from the need to estimate attention importance for tokens yet to be decoded, while the unmasked token positions are unknown during diffusion. In this paper, we present **Focus-dLLM**, a novel training-free attention sparsification framework tailored for accurate and efficient long-context dLLM inference. Based on the finding that token confidence strongly correlates across adjacent steps, we first design a *past confidence-guided indicator* to predict unmasked regions. Built upon this, we propose a *sink-aware pruning strategy* to accurately estimate and remove redundant attention computation, while preserving highly influential attention sinks. To further reduce overhead, this strategy reuses identified sink locations across layers, leveraging the observed cross-layer consistency. Experimental results show that our method offers more than 29\times lossless speedup under 32K context length.
Idiomatic Expression Generation, which aims to produce idiomatic text from plain text, is a valuable yet challenging NLP task. However, existing methods suffer from the scarcity of parallel data and dependence on high-quality manual annotations. To address this, we propose an iterative LLM-SLM (Large Language Model-Small Language Model) collaborative framework — Auto-IDEA, that replaces human supervision for idiomatic expression data generation. In this self-improving cycle, the LLM constructs parallel corpora (idiomatic and plain text) via bidirectional semantic reconstruction, automatically generating "Locate-Then-Polish" (LTP) annotations; the SLM filters low-quality corpora while continuously enhancing its verification ability through incremental learning. We instantiate Auto-IDEA for Chinese Idiom Polishing (CIP), constructing CIP-200K, a large-scale dataset of 206K parallel sentences with LTP annotations. The Qwen3-8B fine-tuned on CIP-200K achieves a 25.2% absolute Idiom Polishing Accuracy (IPA) improvement over a supervised fine-tuning (SFT) baseline, outperforming DeepSeek-R1 by 6.2%. Extensive experiments (e.g., Chinese idiom cloze tests and English idiom generation tasks) and human evaluations verify the generalization and effectiveness of Auto-IDEA, demonstrating a new pathway for high-quality, annotation-free data generation through LLM-SLM collaboration.
In knowledge-intensive creative tasks, Large Language Models (LLMs) often generate outputs that extend beyond established knowledge, making direct verification against current evidence impractical. Unlike factual hallucinations checked against ground truth, such outputs arise naturally in creative generation, where extending beyond current knowledge is often the goal. Yet prior work debates whether hallucination should be suppressed or embraced without empirically analyzing this unverifiable subclass. On the ideation evaluation side, existing work focuses on individual outputs without characterizing the unverifiable space as a whole. To address this gap, we propose a novelty-verifiability characterization that distinguishes Creative Synthesis (Region A) from Groundless Fabrication (Region B), and study it through a conceptual creation task where LLMs synthesize novel scientific concepts. Through 32,400 generations across three technical domains and 1,080 human judgments, we find that Region A is non-negligible (4.7%) and robust, persisting across generation strategies, models, domains, and embedding choices. A retrospective recovery experiment further shows that LLMs can approximate post-cutoff scientific concepts in controlled combinatorial settings. Our findings suggest that the unverifiable space is not uniformly noise but exhibits empirically distinguishable internal structure, providing an empirical basis for more selective hallucination governance.[<https://github.com/YuLab1/llm-concept-creation>]