AAAI 2026 Papers

Large language models (LLMs) often generate hallucinated content lacking factual or contextual grounding, hindering their reliability in critical applications. Traditional methods like supervised fine-tuning and reinforcement learning from human feedback are data-intensive and computationally expensive, while static parameter editing struggles with context-dependent errors and catastrophic forgetting. To overcome these limitations, we introduce LLM-CAS, a framework that formulates real-time hallucination correction as a hierarchical reinforcement learning (HRL) problem. LLM-CAS trains an agent to learn a sophisticated policy, dynamically selecting optimal, temporary neuron perturbations during inference based on the immediate context. This learned, policy-driven approach provides greater adaptability than prior dynamic methods that rely on heuristic or pre-defined adjustments. As a result, LLM-CAS achieves significant performance gains across various LLMs, improving accuracy by 10.98 percentage points on StoryCloze, 2.71 points on TriviaQA, and 2.06 points on TruthfulQA's MC1 score, thereby outperforming static methods like ITI and CAA, as well as the dynamic SADI framework. This context-aware, efficient approach promises enhanced reliability for LLMs in high-stakes domains, with future potential for multimodal extensions.

Recent generative models have significantly advanced speech restoration tasks, yet their training objectives often misalign with human perceptual preferences, resulting in suboptimal quality. While post-training alignment has proven effective in other generative domains like text and image generation, its application to generative speech restoration remains largely under-explored. This work investigates the challenges of applying preference-based post-training to this task, focusing on how to define a robust preference signal and curate high-quality data to avoid reward hacking. To address these challenges, we propose a multi-metric preference alignment strategy. We construct a new dataset, GenSR-Pref, comprising 80K preference pairs, where each chosen sample is unanimously favored by a complementary suite of metrics covering perceptual quality, signal fidelity, content consistency, and timbre preservation. This principled approach ensures a holistic preference signal. Applying Direct Preference Optimization (DPO) with our dataset, we observe consistent and significant performance gains across three diverse generative paradigms: autoregressive models (AR), masked generative models (MGM), and flow-matching models (FM) on various restoration benchmarks, in both objective and subjective evaluations. Ablation studies confirm the superiority of our multi-metric strategy over single-metric approaches in mitigating reward hacking. Furthermore, we demonstrate that our aligned models can serve as powerful ''data annotators'', generating high-quality pseudo-labels to serve as a supervision signal for traditional discriminative models in data-scarce scenarios like singing voice restoration.

Retrieval-Augmented Generation (RAG) improves large language models by retrieving external knowledge, often truncated into smaller chunks due to the input context window, which leads to information loss, resulting in response hallucinations and broken reasoning chains. Moreover, traditional RAG retrieves unstructured knowledge, introducing irrelevant details that hinder accurate reasoning. To address these issues, we propose TAdaRAG, a novel RAG framework for on-the-fly task-adaptive knowledge graph construction from external sources. Specifically, we design an intent-driven routing mechanism to a domain-specific extraction template, followed by supervised fine-tuning and a reinforcement learning-based implicit extraction mechanism, ensuring concise, coherent, and non-redundant knowledge integration. Evaluations on six public benchmarks and a real-world business benchmark (NowNewsQA) across three backbone models demonstrate that TAdaRAG outperforms existing methods across diverse domains and long-text tasks, highlighting its strong generalization and practical effectiveness.

Privacy concerns have long been a critical issue in AI models. With the rapid advancement of generative AI, the privacy awareness of models has drawn attention, raising new challenges for privacy protection that is independent of data and tasks. This paper introduces a novel framework for enhancing privacy protection through directional steering in representation space, which seamlessly integrates with both language and vision-language models. Specifically, we first construct a comprehensive privacy-related dataset based on the Solove taxonomy of privacy. Then, we leverage this dataset to enhance model privacy awareness in the representation space, steering the model to protect privacy during inference. Experiments on 12 models validate the effectiveness and generalization of our method. Moreover, we demonstrate the transferability of privacy-enhanced representations between same-source large language models (LLMs) and vision-language models (VLMs), offering a scalable solution for privacy protection in frontier AI models.

Web automation uses intelligent agents to perform high-level tasks by mimicking human interactions with webpages. Despite recent advances in LLM-based web agents, efficiently navigating complex, real-world webpages remains challenging due to massive DOM structures (10,000 ~ 100,000 tokens). Current approaches either truncate DOMs—losing vital information—or use inefficient heuristics and separate ranking models, failing to balance precision and scalability. We introduce Prune4Web, a novel paradigm that transforms DOM processing from LLM-based filtering to programmatic pruning. Our key innovation is DOM Tree Pruning Programming, where an LLM generates executable Python scoring programs to dynamically filter DOM elements based on semantic clues from decomposed sub-tasks. This approach eliminates the need for LLMs to process full DOMs, instead delegating traversal and scoring to lightweight, interpretable programs. The result is a 25 ~ 50 times reduction in candidate elements for grounding, enabling precise action localization without attention dilution. Additionally, we propose a data annotation method and a two-turn dialogue training strategy that jointly optimizes Planner, Programmatic Filter, and Grounder in a unified framework. Experiments demonstrate state-of-the-art performance. On our low-level task grounding task, our approach dramatically increases grounding accuracy from 46.80% to 88.28%, highlighting its effectiveness.

LLM's code generation capabilities have yielded substantial improvements in the effectiveness of programming tasks. However, LLM-generated code still suffers from compilation and runtime errors. Existing offline preference optimization methods primarily focus on enhancing LLMs' coding abilities using pass/fail signals in the preference data, overlooking the deep-level error types in the failed codes. To address this, we propose Adaptively Progressive Preference Optimization (AP2O) for coding (i.e., AP2O-Coder), a method that guides LLMs adaptively and methodically to reduce code errors for code generation. Specifically, we construct an error notebook from failed codes and progressively optimize the LLM to correct errors type by type. Furthermore, we adaptively replay error types to tailor to the LLM's evolving weaknesses throughout training. Through extensive experiments on both code and general LLMs (Llama, Qwen, and DeepSeek series) with parameters ranging from 0.5B to 34B, our AP2O-Coder improves code generation performance by up to 3% in pass@k while using less preference data.

The proliferation of multi-modal data on the internet has intensified the need for structured event understanding across textual and visual modalities. However, existing multi-modal event extraction models suffer from three major limitations: the absence of explicit event schema guidance, coarse-grained multi-modal alignment strategies, and reliance on heterogeneous, misaligned multi-modal training datasets. To address these issues, we propose LLaVA-MS-PIT, a Multi-modal Schema-Guided Progressive Instruction Tuning Framework that explicitly injects structured multi-modal event schema knowledge into the model before event extraction. Specifically, we introduce the textual event schema to establish the model’s prior knowledge of event concepts and enhance its ability to reason about event structures, while the visual event schema is employed to bridge the representation gap between textual and visual modalities at the event level, enabling unified and semantically aligned event representations across modalities. Moreover, to alleviate data scarcity and modality misalignment inherent in current benchmarks, we construct imSitu-MEE, a high-quality multi-modal parallel dataset generated and annotated through schema-guided procedures. Extensive experiments demonstrate that LLaVA-MS-PIT achieves competitive performance on multi-modal event extraction benchmarks, underscoring the effectiveness and necessity of schema-guided progressive instruction tuning.

Recently, End-to-End Speech Translation (E2E-ST) methods leveraging large language models (LLMs) have demonstrated strong generalization capabilities and excellent scalability by integrating pre-trained speech encoders with LLMs, where Low-Rank Adaptation (LoRA) is commonly used for parameter-efficient fine-tuning to reduce training costs. However, LoRA's low-rank assumption often fails in multilingual tasks, as the inherent complexity of cross-lingual semantic relationships and syntactic variations exceeds the representational capacity of low-rank matrices. This leads to parameter conflicts across languages, resulting in suboptimal performance. To address this issue, we propose Mixture of Low-Rank Adaptations (MoLoRA), which integrates the Mixture of Experts (MoE) mechanism with LoRA. MoLoRA effectively enhances the model's expressive capacity while maintaining parameter efficiency during training. Specifically, we treat multiple LoRA modules as low-rank experts and introduce a routing mechanism to dynamically activate language-specific experts. Additionally, shared experts are incorporated and consistently activated to model cross-lingual general knowledge. Furthermore, to enhance the robustness and accuracy of speech representations, we propose a Multi-Granularity Representation Fusion module (MGRF). This module mitigates local distortions in frame-level speech representations caused by noise by fusing frame-level and sentence-level features, thereby providing the LLM with more accurate high-level semantic information. We conduct multilingual experiments on the MuST-C and CoVoST-2 datasets. Our method achieves an average BLEU score of 32.2 across eight language pairs on the MuST-C dataset and an average of 36.3 across three language pairs on the CoVoST-2 dataset, establishing a new state-of-the-art (SOTA) performance.

As a knowledge-intensive and challenging task, automatic generation of long-form wiki-style articles has garnered increasing attention from researchers due to its ability to efficiently integrate, organize and present vast amounts of both structured and unstructured knowledge. To the best of our knowledge, most of the existing mainstream state-of-the-art methods for automatic wiki-style article generation typically follow a "one-shot generation" paradigm: given a topic, (1) first generating a structured outline, (2) then independently and in parallel generating the content of each outline chapter in a one-shot using the chapter title and references. However, the core limitation of the paradigm lies in its disregards inter-chapter correlation and lacks post-generation revision and refinement, resulting in content redundancy, weak relevance and logical inconsistency. To address these issues, we propose WikiREVIEW, a novel multi-perspective review framework for automatic wiki-style article generation. Specifically, our proposed method introduces multi-perspective experts to review the content of each outline chapter at both chapter and paragraph levels following the initial generation, offering evaluation feedback and continuously refining the numerous deficiencies in the initial long-form article, ultimately achieving high-quality wiki-style article generation. Extensive experimental results on the public English dataset FreshWiki and our own constructed high-quality Chinese dataset ChineseWiki, demonstrate that our proposed WikiREVIEW significantly outperforms existing state-of-the-art automatic wiki-style article generation methods across all automatic evaluation metrics and human evaluation.

Document-level event argument extraction (DEAE) is essential for knowledge acquisition, aiming to extract participants of events from documents. In the zero-shot setting, existing methods employ LLMs to generate synthetic data to address the challenge posed by the scarcity of annotated data. However, relying solely on Event-type-only prompts makes it difficult for the generated content to accurately capture the contextual and structural relationships of unseen events. Moreover, ensuring the reliability and usability of synthetic data remains a significant challenge due to the absence of quality evaluation mechanisms. To this end, we introduce a multi-agent collaboration framework for zero-shot document-level event argument extraction (ZS-DEAE), which simulates the human collaborative cognitive process of “Propose–Evaluate–Revise.” Specifically, the framework comprises a generation agent and an evaluation agent. The generation agent synthesizes data for unseen events by leveraging knowledge from seen events, while the evaluation agent extracts arguments from the synthetic data and assesses their semantic consistency with the context. The evaluation results are subsequently converted into reward signals, with event structure constraints incorporated into the reward design to enable iterative optimization of both agents via reinforcement learning. In three zero-shot scenarios constructed from the RAMS and WikiEvents datasets, our method achieves improvements both in data generation quality and argument extraction performance, while the generated data also effectively enhances the zero-shot performance of other DEAE models.

Dysarthric speech reconstruction (DSR) aims to enhance the intelligibility of dysarthric speech. Compared with normal speech, the dysarthric speech is characterized by its pathological features, including discontinuous pronunciation, slow speech, hoarseness, and improper pauses. Significant disparities in the feature space between normal and dysarthric speech may result in suboptimal speech reconstruction, thereby degrading speech intelligibility. To enhance the reconstruction ability of speech feature spaces, this paper proposes a DSR model named the Encoding-Aligned Variational Autoencoder (EA-VAE). By incorporating alignment modules of frame-level embedding features, prior distributions, and duration into the encoder of the VAE, the model explicitly aligns the dysarthric speech encoding with a representation of the parallel normal speech. A shared decoder is then used to generate speech with improved intelligibility. Experimental results on the UASpeech benchmark confirm that EA-VAE achieves state-of-the-art performance, with a 31.7% relative word error rate reduction and the highest subjective MOS score (4.48), thoroughly validating the effectiveness and advancements of the proposed method in dysarthric speech reconstruction.

Large language models (LLMs) often generate hallucinations in knowledge-intensive QA due to parametric knowledge limitations. While existing methods like KG-CoT improve reliability by integrating knowledge graph (KG) paths, they suffer from rigid hop-count selection (solely question-driven) and underutilization of reasoning paths (lack of guidance). To address this, we propose RFKG-CoT: First, it replaces the rigid hop-count selector with a relation-driven adaptive hop-count selector that dynamically adjusts reasoning steps by activating KG relations (e.g., 1-hop for direct ''brother" relations, 2-hop for indirect ''father-son" chains), formalized via a relation mask. Second, it introduces a few-shot in-context learning path guidance mechanism with CoT (think) that constructs examples in a ''question-paths-answer" format to enhance LLMs' ability to understand reasoning paths. Experiments on four KGQA benchmarks show RFKG-CoT improves accuracy by up to 14.7 pp (Llama2-7B on WebQSP) over KG-CoT. Ablations confirm the hop-count selector and the path prompt are complementary, jointly transforming KG evidence into more faithful answers.

Zero-shot stance detection (ZSSD) seeks to determine the stance of text toward previously unseen targets, a task critical for analyzing dynamic and polarized online discourse with limited labeled data. While large language models (LLMs) offer zero-shot capabilities, prompting-based approaches often fall short in handling complex reasoning and lack robust generalization to novel targets. Meanwhile, LLM-enhanced methods still require substantial labeled data and struggle to move beyond instance-level patterns, limiting their interpretability and adaptability. Inspired by cognitive science, we propose the Cognitive Inductive Reasoning Framework (CIRF), a schema-driven method that bridges linguistic inputs and abstract reasoning via automatic induction and application of cognitive reasoning schemas. CIRF abstracts first-order logic patterns from raw text into multi-relational schema graphs in an unsupervised manner, and leverages a schema-enhanced graph kernel model to align input structures with schema templates for robust, interpretable zero-shot inference. Extensive experiments on SemEval-2016, VAST, and COVID-19-Stance benchmarks demonstrate that CIRF not only establishes new state-of-the-art results, but also achieves comparable performance with just 30% of the labeled data, demonstrating its strong generalization and efficiency in low-resource settings.

Tool-use capabilities fundamentally transform large language models (LLMs) from passive language generators into active agents with real-world utility, drawing intense research focus. Yet, their emergent nature renders traditional scaling laws ineffective for early-stage prediction, obstructing principled model design and efficient training. In this work, we propose a proxy-task perspective that predicts tool-use capabilities by measuring early model performance on selected non-emergent proxy tasks. Our method quantifies two properties of each proxy task: alignment, which reflects how well it captures tool-use trajectories, and stability, which indicates how consistently it behaves across training conditions. These properties are used to weight predictive signals. Theoretically, we formalize how these weighted signals approximate emergent tool use through bounded extrapolation under relaxed assumptions. Empirically, we validate our approach across training checkpoints, model scales, and data setups. Results show that a carefully weighted ensemble of proxy tasks can accurately rank downstream tool-use ability long before it arises. Our findings provide new theoretical foundations and practical tools for efficient training and capability planning, and advance the understanding of how complex abilities arise in LLMs.

Large Reasoning Models (LRMs) achieve promising results on complex reasoning tasks but remain susceptible to hallucinations. Existing hallucination detection methods based on Large Language Models (LLMs) often focus solely on final answers, overlooking inconsistencies between the answer and reasoning process. This limitation reduces their ability to detect hallucinations during inference. Moreover, training-free approaches lack mechanisms for confidence estimation, resulting in an unquantified detection output. In contrast, training-based methods can provide fine-grained assessments but often neglect the self-correction capability of LRMs, where earlier errors may be corrected in subsequent steps, leading to inaccurate hallucination detection. To address these challenges, we propose ConfFuse, a unified framework that fuses global and local confidence scores for hallucination detection. A Global Hallucination Detection Model (GHDM) is trained using Direct Preference Optimization (DPO) to assess hallucinations at the level of entire reasoning chains, yielding global confidence estimates. Simultaneously, a Process Reward Model (PRM) estimates step-wise confidence scores to capture local logical flaws. A weighted fusion strategy combines the global confidence score with the minimum local score to jointly reflect overall reasoning consistency and local soundness. Experimental evaluations demonstrate that ConfFuse surpasses Qwen3-1.7B and Qwen3-8B by up to 11.86% and 5.46% in F1 score on in-distribution datasets, and achieves average improvements of 4.65% and 2.80% on out-of-distribution datasets. These results verify the effectiveness and generalizability of the proposed framework.

Existing Grammatical Error Correction (GEC) systems suffer from limited reference diversity, leading to underestimated evaluation and restricted model generalization. To address this issue, we introduce the Judge of Edit-Level Validity (JELV), an automated framework to validate correction edits from grammaticality, faithfulness, and fluency. Using our proposed human-annotated Pair-wise Edit-level Validity Dataset (PEVData) as benchmark, JELV offers two implementations: a multi-turn LLM-as-Judges pipeline achieving 90% agreement with human annotators, and a distilled DeBERTa classifier with 85% precision on valid edits. We then apply JELV to reclassify misjudged false positives in evaluation and derive a comprehensive evaluation metric by integrating false positive decoupling and fluency scoring, resulting in state-of-the-art correlation with human judgments. We also apply JELV to filter LLM-generated correction candidates, expanding the BEA19's single-reference dataset containing 38,692 source sentences. Retraining top GEC systems on this expanded dataset yields measurable performance gains. JELV provides a scalable solution for enhancing reference diversity and strengthening both evaluation and model generalization.

Large language models have achieved substantial progress in mathematical reasoning, yet their advancement is limited by the scarcity of high-quality, high-difficulty training data. Existing synthesis methods largely rely on transforming human-written templates, limiting both diversity and scalability. We propose MathSmith, a novel framework for synthesizing challenging mathematical problems to enhance LLM reasoning. Rather than modifying existing problems, MathSmith constructs new ones from scratch by randomly sampling concept–explanation pairs from PlanetMath, ensuring data independence and avoiding contamination. To increase difficulty, we design nine predefined strategies as soft constraints during rationales. We further adopts reinforcement learning to jointly optimize structural validity, reasoning complexity, and answer consistency. The length of the reasoning trace generated under autoregressive prompting is used to reflect cognitive complexity, encouraging the creation of more demanding problems aligned with long-chain-of-thought reasoning. Experiments across five benchmarks, categorized as easy & medium (GSM8K, MATH-500) and hard (AIME2024, AIME2025, OlympiadBench), show that MathSmith consistently outperforms existing baselines under both short and long CoT settings. Additionally, a weakness-focused variant generation module enables targeted improvement on specific concepts. Overall, MathSmith exhibits strong scalability, generalization, and transferability, highlighting the promise of high-difficulty synthetic data in advancing LLM reasoning capabilities.

Tool learning, which allows Large Language Models (LLMs) to leverage external tools for solving complex user tasks, has emerged as a promising avenue for extending model capabilities. However, existing approaches primarily focus on data synthesis for fine-tuning LLMs to invoke tools effectively, largely ignoring how to fully stimulate the potential of the model. In this paper, we propose ToolACE-R, a novel framework that includes both model-aware iterative training and adaptive refinement for tool learning. ToolACE-R features a model-aware iterative training procedure that progressively adjust training samples based on the model’s evolving capabilities to maximize its potential. Additionally, it incorporates self-refinement training corpus which emphasizes LLM's ability to iteratively refine their tool calls, optimizing performance without requiring external feedback. Furthermore, we introduce adaptive self-refinement for efficient test-time scaling, where the trained model can autonomously determine when to stop the process based on iterative self-refinement. We conduct extensive experiments across several benchmark datasets, showing that ToolACE-R achieves competitive performance compared to advanced LLMs. The performance can be further improved efficiently through adaptive self-refinement. These results highlight the effectiveness and generalizability of ToolACE-R, offering a promising direction for more efficient and scalable tool learning.

Automatic presentation slide generation can greatly streamline content creation. However, since preferences of each user may vary, existing under-specified formulations often lead to suboptimal results that fail to align with individual user needs. We introduce a novel task that conditions paper-to-slides generation on user-specified preferences. We propose a human behavior-inspired agentic framework, SlideTailor, that progressively generates editable slides in a user-aligned manner. Instead of requiring users to write their preferences in detailed textual form, our system only asks for a paper-slides example pair and a visual template—natural and easy-to-provide artifacts that implicitly encode rich user preferences across content and visual style. Despite the implicit and unlabeled nature of these inputs, our framework effectively distills and generalizes the preferences to guide customized slide generation. We also introduce a novel chain-of-speech mechanism to align slide content with planned oral narration. Such a design significantly enhances the quality of generated slides and enables downstream applications like video presentations. To support this new task, we construct a benchmark dataset that captures diverse user preferences, with carefully designed interpretable metrics for robust evaluation. Extensive experiments demonstrate the effectiveness of our framework.