AAAI 2026 论文列表

Large Language Models (LLMs) show remarkable potential for few-shot information extraction (IE), yet their performance is highly sensitive to the choice of in-context examples. Conventional selection strategies often fail to provide informative guidance, as they overlook a key source of model fallibility: confusion stemming not just from semantic content, but also from the generation of well-structured formats required by IE tasks. To address this, we introduce Active Prompting for Information Extraction (APIE), a novel active prompting framework guided by a principle we term introspective confusion. Our method empowers an LLM to assess its own confusion through a dual-component uncertainty metric that uniquely quantifies both Format Uncertainty (difficulty in generating correct syntax) and Content Uncertainty (inconsistency in extracted semantics). By ranking unlabeled data with this comprehensive score, our framework actively selects the most challenging and informative samples to serve as few-shot exemplars. Extensive experiments on four benchmarks show that our approach consistently outperforms strong baselines, yielding significant improvements in both extraction accuracy and robustness. Our work highlights the critical importance of a fine-grained, dual-level view of model uncertainty when it comes to building effective and reliable structured generation systems.

While large language models (LLMs) have demonstrated remarkable performance on high-level semantic tasks, they often struggle with fine-grained, token-level understanding and structural reasoning—capabilities that are essential for applications requiring precision and control. We introduce TASE, a comprehensive benchmark designed to evaluate LLMs' ability to perceive and reason about token-level information across languages. TASE covers 10 tasks under two core categories: token awareness and structural understanding, spanning Chinese, English, and Korean, with a 35,927-instance evaluation set and a scalable synthetic data generation pipeline for training. Tasks include character counting, token alignment, syntactic structure parsing, and length constraint satisfaction. We evaluate over 30 leading commercial and open-source LLMs, including O3, Claude 4, Gemini 2.5 Pro, and DeepSeek-R1, and train a custom Qwen2.5-14B model using the GRPO training method. Results show that human performance significantly outpaces current LLMs, revealing persistent weaknesses in token-level reasoning. TASE sheds light on these limitations and provides a new diagnostic lens for future improvements in low-level language understanding and cross-lingual generalization.

Large Language Model (LLM) agents have demonstrated strong potential in complex, interactive decision-making tasks. However, when training LLM agents end-to-end with reinforcement learning (RL), efficiently optimizing agent policies in dynamic environments remains a significant challenge. Existing RL-based LLM agent paradigms commonly organize interactions in a cycle where reasoning is followed by action. In our work, we observe a phenomenon we call Exploration Contraction, where the explicit introduction of a reasoning stage reduces the diversity of actions—quantified by lower action entropy—which in turn limits exploration and leads to premature policy convergence. To address this limitation, we propose Act-before-Reasoning (ActRe), a two-stage RL training framework. In the first stage, we reverse the typical rollout order, prompting the agent to generate actions prior to reasoning, which encourages exploration driven by model intuition. In the second stage, we restore the standard reasoning-then-action order for training and evaluation, ensuring robust and interpretable decision-making. Experiments on the ALFWorld and WebShop benchmarks show that ActRe effectively mitigates exploration contraction, yielding consistently higher task success rates and improved training robustness compared to strong RL baselines. Our analysis underscores the importance of action entropy in the exploration-exploitation trade-off during LLM agent training and provides a practical approach to maintain the benefits of explicit reasoning while promoting sufficient exploration.

Large language models sometimes inadvertently reproduce passages that are copyrighted, exposing downstream applications to legal risk. Most existing studies for inference-time defences focus on surface-level token matching and rely on external blocklists or filters, which add deployment complexity and may overlook semantically paraphrased leakage. In this work, we reframe copyright infringement mitigation as intrinsic semantic-space control and introduce SCOPE, an inference-time method that requires no parameter updates or auxiliary filters. Specifically, the sparse autoencoder (SAE) projects hidden states into a high-dimensional, near-monosemantic space; benefiting from this representation, we identify a copyright-sensitive subspace and clamp its activations during decoding. Experiments on widely recognized benchmarks show that SCOPE mitigates copyright infringement without degrading general utility. Further interpretability analyses confirm that the isolated subspace captures high-level semantics.

Existing large language model (LLM)-based table question answering (TableQA) methods primarily involve decomposition reasoning and answer verification processes. However, decomposing questions solely at the semantic level, without considering the factual evidence in tables, fails to significantly reduce the difficulty for LLMs in understanding the key information in questions. Furthermore, reasoning and verification without supporting factual evidence are often arbitrary and unreliable. In light of these issues, this paper proposes a Syllogism-Inspired Reasoning and Verification method (SIRV), which performs reliable decomposition reasoning and answer verification based on the evidential concept of syllogism. Specifically, SIRV extracts question-relevant factual evidence from the table to construct the premises. Based on the constructed premises, SIRV plans reasoning paths and generates sub-questions that explicitly indicate relevant factual evidence, performing evidence-centered reasoning. Additionally, SIRV examines the consistency between the premises and the table to focus on factual evidence, thereby reliably identifying and correcting errors in the reasoning process. Compared to state-of-the-art methods, SIRV achieves performance improvements of up to 5.24% in single-mode and 2.89% in joint reasoning, while also demonstrating excellent generalization ability and efficiency.

LLM-based approaches have recently achieved impressive results in zero-shot stance detection. However, they still struggle in complex real-world scenarios, where stance understanding requires dynamic background knowledge, target definitions involve compound entities or events that must be explicitly linked to stance labels, and rhetorical devices such as irony often obscure the author’s actual intent. To address these challenges, we propose MSME, a Multi-Stage, Multi-Expert framework for zero-shot stance detection. MSME consists of three stages: (1) Knowledge Preparation, where relevant background knowledge is retrieved and stance labels are clarified; (2) Expert Reasoning, involving three specialized modules—Knowledge Expert distills salient facts and reasons from a knowledge perspective, Label Expert refines stance labels and reasons accordingly, and Pragmatic Expert detects rhetorical cues such as irony to infer intent from a pragmatic angle; (3) Decision Aggregation, where a Meta-Judge integrates all expert analyses to produce the final stance prediction. Experiments on three public datasets show that MSME achieves state-of-the-art performance across the board.

Proprietary models are increasingly dominating the race for ever-larger language models. Can open-source, smaller models remain competitive across a broad range of tasks? In this paper, we present the Avengers---a lightweight framework that leverages the collective intelligence of these smaller models. The Avengers builds upon four lightweight operations: (i) embedding: encode queries using a text embedding model; (ii) clustering: group queries based on their semantic similarity; (iii) scoring: scores each model's performance within each cluster; and (iv) voting: improve outputs via repeated sampling and voting. At inference time, each query is embedded and assigned to its nearest cluster. The top-performing model(s) within that cluster are selected to generate the response with repeated sampling. Remarkably, with 10 open-source models (~7B parameters each), the Avengers surpasses GPT-4o, 4.1, and 4.5 in average performance across 15 diverse datasets spanning mathematics, coding, logical reasoning, general knowledge, and affective tasks. In particular, it surpasses GPT-4.1 on mathematics tasks by 18.21% and on code tasks by 7.46%. Furthermore, the Avengers delivers superior out-of-distribution generalization, and remains robust across various embedding models, clustering algorithms, ensemble strategies, data efficiency, and values of its sole parameter---the number of clusters.

Retrieval-Augmented Generation (RAG) plays a crucial role in grounding Large Language Models by leveraging external knowledge, whereas the effectiveness is often compromised by the retrieval of contextually flawed or incomplete information. To address this, knowledge graph-based RAG methods have evolved towards hierarchical structures, organizing knowledge into multi-level summaries. However, these approaches still suffer from two critical, unaddressed challenges: high-level conceptual summaries exist as disconnected ``semantic islands'', lacking the explicit relations needed for cross-community reasoning; and the retrieval process itself remains structurally unaware, often degenerating into an inefficient flat search that fails to exploit the graph's rich topology. To overcome these limitations, we introduce LeanRAG, a framework that features a deeply collaborative design combining knowledge aggregation and retrieval strategies. LeanRAG first employs a novel semantic aggregation algorithm that forms entity clusters and constructs new explicit relations among aggregation-level summaries, creating a fully navigable semantic network. Then, a bottom-up, structure-guided retrieval strategy anchors queries to the most relevant fine-grained entities and then systematically traverses the graph's semantic pathways to gather concise yet contextually comprehensive evidence sets. The LeanRAG can mitigate the substantial overhead associated with path retrieval on graphs and minimize redundant information retrieval. Extensive experiments on four challenging QA benchmarks with different domains demonstrate that LeanRAG significantly outperforms existing methods in response quality while reducing 46% retrieval redundancy.

Pairwise evaluation of Large Language Models (LLMs) is a common paradigm, but it is prone to preference bias, where judges systematically favor certain outputs, such as their own. This bias leads to inconsistent and skewed rankings across different judges. To address this, we first empirically demonstrate significant and heterogeneous biases in cross-model evaluations. We then propose UDA (Unsupervised Debiasing Alignment), a framework that reduces inter-judge disagreement by dynamically adjusting the Elo rating system. For each pairwise comparison, a compact neural network learns to adaptively set the K-factor and refine win probabilities. Crucially, UDA operates in a fully unsupervised manner, guided solely by the objective of minimizing the dispersion among the Elo trajectories of all judges. This forces an alignment towards a collective consensus, which serves as an unsupervised proxy for a more stable and reproducible evaluation. In addition, we provide theoretical motivation demonstrating how alignment towards a consensus can reduce aggregate system bias. Experiments show that UDA significantly reduces the inter-judge rating standard deviation by up to 63.4% and improves the average correlation with human judgments by 24.7%. Notably, UDA elevates the performance of poorly performing judges to achieve parity with high-quality ones, fostering a more robust and reliable evaluation ecosystem.

Recent studies reveal that adversaries can manipulate the internal knowledge of large language models (LLMs) on selected topics through model editing, causing attacker-specified harmful or biased outputs when queried about the edited content. Once such tampered LLMs are distributed, they can mislead users on the targeted topics, thereby potentially propagating misinformation or reinforcing stereotypes. However, existing knowledge manipulation attacks rely on the ability to redistribute compromised models, which is infeasible in constrained settings like Federated Instruction Tuning (FedIT), where a central server controls LLM's training and distribution. In this work, we introduce ShadeEdit, the first attack framework that leverages strengthened model editing to enable knowledge manipulation in FedIT scenarios. ShadeEdit introduces two key components to address two challenges posed by the training process of FedIT: (1) a paraphrase-based editing dataset selection strategy to mitigate the dilution from benign updates on malicious ones by constructing a high-quality editing dataset, and (2) an adaptive manipulation mechanism to evade aggregation-based defenses via an adaptive clipping strategy. ShadeEdit achieves an average 99.5% attack success rate over eight robust aggregation algorithms while preserving instruction-following accuracy, demonstrating its strong attack effectiveness and model-utility preservation.

Despite the rapid progress in large language models (LLMs), even sub-billion-scale systems perform at chance level on challenging natural language inference (NLI) benchmarks such as Adversarial Natural Language Inference (ANLI), while training larger models is often impractical due to limited computational resources. We address this parameter-efficiency bottleneck in NLI with a Complex-Vector Token Representation that explicitly decouples each token from its context, and a Token-Context Attention mechanism that updates each token based on the most informative contextual semantics. On ANLI, a 0.8B-parameter Token-Context Attention model achieves higher parameter efficiency (accuracy per parameter) than all 1B and comparable 0.8B self-attention baselines; it also suffers smaller performance degradation under Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks and achieves the largest few-shot gains on SNLI and MNLI while exhibiting no significant degradation in ANLI accuracy after adaptation. These results suggest that explicitly disentangling token and context offers a viable alternative to standard self-attention for NLI tasks.

With the widespread application of Large Language Models (LLMs), it has become a significant concern to ensure their safety and prevent harmful responses. While current safe-alignment methods based on instruction fine-tuning and Reinforcement Learning from Human Feedback (RLHF) can effectively reduce harmful responses from LLMs, they often require high-quality datasets and heavy computational overhead during model training. Another way to align language models is to modify the logit of tokens in model outputs without heavy training. Recent studies have shown that contrastive decoding can enhance the performance of language models by reducing the likelihood of confused tokens. However, these methods require the manual selection of contrastive models or instruction templates, limiting the degree of contrast. To this end, we propose Adversarial Contrastive Decoding (ACD), an optimization-based framework to generate two opposite soft system prompts, the Safeguarding Prompt (SP) and the Adversarial Prompt (AP), for prompt-based contrastive decoding. The SP aims to promote safer outputs while the AP aims to exploit the harmful parts of the model, providing a strong contrast to align the model with safety. ACD only needs to apply a lightweight prompt tuning on a rather small anchor dataset without training the target model. Experiments conducted on extensive models and benchmarks demonstrate that the proposed method achieves much better safety performance than previous model training-free decoding methods without sacrificing its original generation ability.

Fine-tuning large language models (LLMs) in a parameter-efficient manner while preserving their pre-trained world knowledge remains a significant challenge. While Low-Rank Adaptation (LoRA) and its variants effectively mitigate catastrophic forgetting, they do not fully eliminate the loss of critical pre-trained knowledge. In this work, we first analyze the layer-wise distribution of domain-specific knowledge within LLMs through knowledge localization, and empirically identify a clear layer-specific pattern: pre-trained world knowledge predominantly resides in lower layers, whereas knowledge relevant to downstream tasks is more concentrated in higher layers. Motivated by this observation, we propose L2-LoRA, a simple yet effective variant of LoRA that applies layer-specific L2 regularization to the LoRA weights during fine-tuning. Specifically, L2-LoRA imposes stronger regularization on lower layers to preserve pre-trained world knowledge, while allowing greater adaptation in higher layers to better align with downstream tasks. Experiments across multiple benchmarks show that L2-LoRA not only consistently outperforms vanilla LoRA in downstream performance, but also effectively mitigates catastrophic forgetting by retaining more pre-trained knowledge.

Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-Enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models.

Multilingual Alignment is an effective and representative paradigm to enhance LLMs' multilingual capabilities, which transfers the capabilities from the high-resource languages to the low-resource languages. Meanwhile, some research on language-specific neurons provides a new perspective to analyze and understand LLMs' mechanisms. However, we find that there are many neurons that are shared by multiple but not all languages and cannot be correctly classified. In this work, we propose a ternary classification methodology that categorizes neurons into three types, including language-specific neurons, language-related neurons, and general neurons. And we propose a corresponding identification algorithm to distinguish these different types of neurons. Furthermore, based on the distributional characteristics of different types of neurons, we divide the LLMs' internal process for multilingual inference into four parts: (1) multilingual understanding, (2) shared semantic space reasoning, (3) multilingual output space transformation, and (4) vocabulary space outputting. Additionally, we systematically analyze the models before and after alignment with a focus on different types of neurons. We also analyze the phenomenon of ''Spontaneous Multilingual Alignment''. Overall, our work conducts a comprehensive investigation based on different types of neurons, providing empirical results and valuable insights to better understand multilingual alignment and multilingual capabilities of LLMs.

Goal-driven persuasive dialogue, exemplified by applications like telemarketing, requires sophisticated multi-turn planning and strict factual faithfulness, which remains a significant challenge for even state-of-the-art Large Language Models (LLMs). A lack of task-specific data often limits previous works, and direct LLM application suffers from strategic brittleness and factual hallucination. In this paper, we first construct and release TeleSalesCorpus, the first real-world-grounded dialogue dataset for this domain. We then propose AI-Salesman, a novel framework featuring a dual-stage architecture. For the training stage, we design a Bayesian-supervised reinforcement learning algorithm that learns robust sales strategies from noisy dialogues. For the inference stage, we introduce the Dynamic Outline-Guided Agent (DOGA), which leverages a pre-built script library to provide dynamic, turn-by-turn strategic guidance. Moreover, we design a comprehensive evaluation framework that combines fine-grained metrics for key sales skills with the LLM-as-a-Judge paradigm. Experimental results demonstrate that our proposed AI-Salesman significantly outperforms baseline models in both automatic metrics and comprehensive human evaluations, showcasing its effectiveness in complex persuasive scenarios.

Knowledge distillation (KD) is a widely adopted technique for transferring the capabilities of large teacher models to smaller student models, thereby significantly reducing inference costs and memory consumption. However, existing KD methods are all constrained by an inherent greedy optimization objective, rooted in the assumption of teacher superiority: "Trust all teacher-generated outputs (TGOs)" and "Distrust any student-generated outputs (SGOs) unsupported by the teacher". We propose ASKD, a novel KD method with adaptive skewness determined by sample quality, refining this objective to: "Learn TGOs proportionally to their quality, and distrust only low-quality unsupported SGOs". ASKD comprises three key components: (1) A reinforcement learning-style optimization formulation to mitigate the inherent approximation bias in sample-based Kullback-Leibler (KL) divergence approximations used by previous KD methods; (2) Well-designed quality supervision signals to map and achieve adaptive skewness in skewed KL loss, pioneering the usage of sample quality to adjust learning magnitudes; (3) A gradient-clip function on high-quality SGOs for findings that high-quality SGOs in KL loss fail to yield positive updates and even cause adverse effects on some samples. Extensive experiments indicate that ASKD builds high-performance student models across various tasks, including instruction following, mathematical reasoning, and code generation, outperforming state-of-the-art methods comprehensively and surpassing GRPO-like approaches that use advantages as multiplicative factors. We also provide detailed mathematical proofs demonstrating properties such as Lipschitz continuity of the update coefficient and uniform convergence of the loss function, ensuring theoretical rigor for key components of ASKD.

Numerical reasoning over documents, which demands both contextual understanding and logical inference, is challenging for low-capacity local models deployed on computation-constrained devices. Although such complex reasoning queries could be routed to powerful remote models like GPT-4, exposing local data raises significant data leakage concerns. Existing mitigation methods generate problem descriptions or examples for remote assistance. However, the inherent complexity of numerical reasoning hinders the local model from generating logically equivalent queries and accurately inferring answers with remote guidance. In this paper, we present a model collaboration framework with two key innovations: (1) a context-aware synthesis strategy that shifts the query topics while preserving reasoning patterns; and (2) a tool-based answer reconstruction approach that reuses the remote-generated plug-and-play solution with code snippets. Experimental results demonstrate that our method achieves better reasoning accuracy than solely using local models while providing stronger data protection than fully relying on remote models. Furthermore, our method improves accuracy by 16.2% - 43.6% while reducing data leakage by 2.3% - 44.6% compared to existing data protection approaches.

While Large Reasoning Models (LRMs) exhibit remarkable capabilities in complex tasks, they often suffer from excessive redundancy in their chain-of-thought reasoning. This significantly reduces inference efficiency and increases computational costs. We identify that LRM redundancy is not uniformly homogeneous but can be taxonomized according to whether it is destructive to the final answer: destructive redundancy (e.g., logical drift, hallucination amplification) versus non-destructive redundancy (e.g., repetition, over-elaboration). Moreover, LRM's redundant and concise responses exhibit a significant distinction in their hidden layer representation spaces. Based on these insights, we propose CATS (Category-Aware Token-level Steering), a training-free and lightweight method to reduce the redundancy phenomenon. CATS decomposes redundancy into six semantically interpretable characteristic dimensions. By flexibly weighting and combining the differential vectors corresponding to these dimensions, CATS synthesizes a composite intervention vector, enabling zero-parameter intervention in the hidden layers. Experiments across three LRM models and five mathematical reasoning datasets demonstrate that CATS reduces reasoning length by an average of 25% while maintaining or even slightly improving task accuracy. CATS offers a pluggable, training-free, and lightweight solution, making it particularly beneficial for users in low-resource environments.