AAAI 2026 论文列表

Emojis are globally used non-verbal cues in digital communication, and extensive research has examined how large language models (LLMs) understand and utilize emojis across contexts. While usually associated with friendliness or playfulness, it is observed that emojis may trigger toxic content generation in LLMs. Motivated by such a observation, we aim to investigate: (1) whether emojis can clearly enhance the toxicity generation in LLMs and (2) how to interpret this phenomenon.* We begin with a comprehensive exploration of emoji-triggered LLM toxicity generation by automating the construction of prompts with emojis to subtly express toxic intent. Experiments across 5 mainstream languages on 7 famous LLMs along with jailbreak tasks demonstrate that prompts with emojis could easily induce toxicity generation. To understand this phenomenon, we conduct model-level interpretations spanning semantic cognition, sequence generation and tokenization, suggesting that emojis can act as a heterogeneous semantic channel to bypass the safety mechanisms. To pursue deeper insights, we further probe the pre-training corpus and uncover potential correlation between the emoji-related data polution with the toxicity generation behaviors.

Backdoor attacks embed malicious behaviors into Large Language Models (LLMs), enabling adversaries to trigger harmful outputs or bypass safety controls. However, the persistence of the implanted backdoors under user-driven post-deployment continual fine-tuning has been rarely examined. Most prior works evaluate the effectiveness and generalization of implanted backdoors only at releasing and empirical evidence shows that naively injected backdoor persistence degrades after updates. In this work, we study whether and how implanted backdoors persist through a multi‑stage post-deployment fine‑tuning. We propose P‑Trojan, a trigger‑based attack algorithm that explicitly optimizes for backdoor persistence across repeated updates. By aligning poisoned gradients with those of clean tasks on token embeddings, the implanted backdoor mapping is less likely to be suppressed or forgotten during subsequent updates. Theoretical analysis shows the feasibility of such persistent backdoor attacks after continual fine-tuning. And experiments conducted on the Qwen2.5 and LLaMA3 families of LLMs, as well as diverse task sequences, demonstrate that P‑Trojan achieves over \textbf{99\%} persistence while preserving clean‑task accuracy. Our findings highlight the need for persistence-aware evaluation and stronger defenses in realistic model adaptation pipelines.

Large language models (LLMs) often show poor performance in low-resource languages like Korean, partly due to unique linguistic challenges such as homophonous Sino-Korean words that are indistinguishable in Hangul script. To address this semantic ambiguity, we propose HanjaBridge, a novel meaning-injection technique integrated into a continual pre-training (CPT) framework. Instead of deterministically mapping a word to a single Hanja (Chinese character), HanjaBridge presents the model with all possible Hanja candidates for a given homograph, encouraging the model to learn contextual disambiguation. This process is paired with token-level knowledge distillation to prevent catastrophic forgetting. Experimental results show that HanjaBridge significantly improves Korean language understanding, achieving a 21% relative improvement on the KoBALT benchmark. Notably, by reinforcing semantic alignment between Korean and Chinese through shared Hanja, we observe a strong positive cross-lingual transfer. Furthermore, these gains persist even when Hanja augmentation is omitted at inference time, ensuring practical efficiency with no additional run-time cost.

This paper provides an early effort to systematically examine nation-level biases exhibited by Large Language Models (LLMs) within the domain of International Relations (IR), a dimension that has remained largely unexplored in prior research. Leveraging historical records from the United Nations Security Council (UNSC), we developed a bias evaluation framework comprising three distinct tests to explore nation-level bias in various LLMs, with a particular focus on the five permanent members of the UNSC. Experimental results show that, even with the general bias patterns across models (e.g., favorable biases toward the western nations, and unfavorable biases toward Russia), these still vary based on the LLM. Notably, even within the same LLM, the direction and magnitude of bias for a nation change depending on the evaluation context. This observation suggests that LLM biases are fundamentally multidimensional, varying across models and tasks. We also observe that models with stronger reasoning abilities show reduced bias and better prediction performance. Building on this finding, we introduce a debiasing framework that improves LLMs’ factual reasoning combining Retrieval-Augmented Generation with Reflexion-based self-reflection techniques. Experiments show it effectively reduces nation-level bias, and improves performance, particularly in GPT-4o-mini and LLama-3.3-70B. Our findings emphasize the need to assess nation-level bias alongside prediction performance when applying LLMs in the IR domain.

Automated essay scoring (AES) is a challenging task in cross-prompt settings due to the diversity of scoring criteria. While previous studies have focused on the output of large language models (LLMs) to improve scoring accuracy, we believe activations from intermediate layers may also provide valuable information. To explore this possibility, we evaluated the discriminative power of LLMs’ activations in cross-prompt essay scoring task. Specifically, we used activation to fit probes and further analyzed the effects of different models and input content of LLMs on this discriminative power. By computing the directions of essays across various trait dimensions under different prompts, we analyzed the variation in evaluation perspectives of large language models concerning essay types and traits. Results show that the activations possess strong discriminative power in evaluating essay quality and that LLMs can adapt their evaluation perspectives to different traits and essay types, effectively handling the diversity of scoring criteria in cross-prompt settings.

Large Language Models (LLMs) have demonstrated significant potential across various domains. However, they often struggle with integrating external knowledge and performing complex reasoning, leading to hallucinations and unreliable outputs. Retrieval Augmented Generation (RAG) has emerged as a promising paradigm to mitigate these issues by incorporating external knowledge. Yet, conventional RAG approaches, especially those based on vector similarity, fail to effectively capture relational dependencies and support multi-step reasoning. In this work, we propose CogGRAG, a human cognition-inspired, graph-based RAG framework designed for Knowledge Graph Question Answering (KGQA). CogGRAG models the reasoning process as a tree-structured mind map that decomposes the original problem into interrelated subproblems and explicitly encodes their semantic relationships. This structure not only provides a global view to guide subsequent retrieval and reasoning but also enables self-consistent verification across reasoning paths. The framework operates in three stages: (1) top-down problem decomposition via mind map construction, (2) structured retrieval of both local and global knowledge from external Knowledge Graphs (KGs), and (3) bottom-up reasoning with dual-process self-verification. Unlike previous tree-based decomposition methods such as MindMap or Graph-CoT, CogGRAG unifies problem decomposition, knowledge retrieval, and reasoning under a single graph-structured cognitive framework, allowing early integration of relational knowledge and adaptive verification. Extensive experiments demonstrate that CogGRAG achieves superior accuracy and reliability compared to existing methods.

Balancing exploration and exploitation is a central goal in reinforcement learning (RL). Despite recent advances in enhancing language model (LM) reasoning, most methods lean toward exploitation, and increasingly encounter performance plateaus. In this work, we revisit entropy -- a signal of exploration in RL -- and examine its relationship to exploratory reasoning in LMs. Through empirical analysis, we uncover positive correlations between high-entropy regions and three types of exploratory reasoning actions: (1) pivotal tokens that determine or connect logical steps, (2) reflective actions such as self-verification and correction, and (3) rare behaviors under-explored by the base LMs. Motivated by this, we introduce a minimal modification to standard RL with only one line of code: augmenting the advantage function with an entropy-based term. Unlike traditional maximum-entropy methods which encourage exploration by promoting uncertainty, we encourage exploration by promoting deeper and longer reasoning chains. Notably, our method achieves significant gains on the Pass@K metric -- an upper-bound estimator of LM reasoning capabilities -- even when evaluated with extremely large K values, pushing the boundaries of LM reasoning.

Recent advancements in improving the reasoning capabilities of Large Language Models have underscored the efficacy of Process Reward Models (PRMs) in addressing intermediate errors through structured feedback mechanisms. This study analyzes PRMs from multiple perspectives, including training methodologies, scalability, and generalization capabilities. We investigate the interplay between pre-training and reward model training FLOPs to assess their influence on PRM efficiency and accuracy in complex reasoning tasks. Our analysis reveals a pattern of diminishing returns in performance with increasing PRM scale, highlighting the importance of balancing model size and computational cost. Furthermore, the diversity of training datasets significantly impacts PRM performance, emphasizing the importance of diverse data to enhance both accuracy and efficiency. We further examine test-time scaling strategies, identifying Monte Carlo Tree Search as the most effective method when computational resources are abundant, while Best-of-N Sampling serves as a practical alternative under resource-limited conditions. Notably, our findings indicate that PRMs trained on mathematical datasets exhibit performance comparable to those tailored for code generation, suggesting robust cross-domain generalization. Employing a gradient-based metric, we observe that PRMs exhibit a preference for selecting responses with similar underlying patterns, further informing their optimization.

Human writers often begin their stories with an overarching mental scene, where they envision the interactions between characters and their environment. Inspired by this creative process, we propose a novel approach to long-form story generation, termed hybrid bottom-up long-form story generation, using multi-agent simulations. In our method, agents interact within a dynamic sandbox environment, where their behaviors and interactions with one another and the environment generate emergent events. These events form the foundation for the story, enabling organic character development and plot progression. Unlike traditional top-down approaches that impose rigid structures, our hybrid bottom-up approach allows for the natural unfolding of events, fostering more spontaneous and engaging storytelling. The system is capable of generating stories exceeding 10,000 words while maintaining coherence and consistency, addressing some of the key challenges faced by current story generation models. We achieve state-of-the-art performance across several metrics. This approach offers a scalable and innovative solution for creating dynamic, immersive long-form stories that evolve organically from agent-driven interactions.

Counterfactual reasoning is widely recognized as one of the most challenging and intricate aspects of causality in artificial intelligence. In this paper, we evaluate the performance of large language models (LLMs) in counterfactual reasoning. In contrast to previous studies that primarily focus on commonsense causal reasoning, where LLMs often rely on prior knowledge for inference, we specifically assess their ability to perform counterfactual inference using a set of formal rules. To support this evaluation, we introduce a new benchmark dataset, CounterBench, comprising 1.2K counterfactual reasoning questions. The dataset is designed with varying levels of difficulty, diverse causal graph structures, distinct types of counterfactual questions, and multiple nonsensical name variants. Our experiments demonstrate that counterfactual reasoning poses a significant challenge for LLMs, with most models performing at levels comparable to random guessing. To enhance LLM's counterfactual reasoning ability, we propose a novel reasoning paradigm, CoIn, which guides LLMs through iterative reasoning and backtracking to systematically explore counterfactual solutions. Experimental results show that our method significantly improves LLM performance on counterfactual reasoning tasks and consistently enhances performance across different LLMs.

The automation of diagram generation has gained significant attention in recent years. Previous studies mainly focused on generating diagrams from natural language, but often lacked support for user-friendly editing like drag-and-drop. This paper proposes a novel task: generating editable, high-fidelity diagrams from either text or raster images. It is also among the first to introduce diagram restoration and style transfer in this setting.To tackle these tasks, we constructed the Diagram-mxGraph dataset, covering restoration, text-to-diagram generation, and style transfer. We propose two core innovations: Fine-grained Adaptive Background Suppression (FABS) and Component-Aware Adaptive Loss (CAAL). Leveraging pre-trained Vision Transformers (ViTs) and the Diagram Adapter module, our method aligns diagram features with a Large Language Model (LLM) to output diagrams in editable mxGraph format.

Scene graphs have emerged as a structured and serializable environment representation for grounded spatial reasoning with Large Language Models (LLMs). In this work, we propose SG2, an iterative Schema-Guided Scene-Graph reasoning framework based on multi-agent LLMs. The agents are grouped into two modules: a (1) Reasoner module for abstract task planning and graph information queries generation, and a (2) Retriever module for extracting corresponding graph information based on code-writing following the queries. Two modules collaborate iteratively, enabling sequential reasoning and adaptive attention to graph information. The scene graph schema, prompted to both modules, serves to not only streamline both reasoning and retrieval process, but also guide the cooperation between two modules. This eliminates the need to prompt LLMs with full graph data, reducing the chance of hallucination due to irrelevant information. Through experiments in multiple simulation environments, we show that our framework surpasses existing LLM-based approaches and baseline single-agent, tool-based Reason-while-Retrieve strategy in numerical Q&A and planning tasks.

As an essential component of fine-tuning, warm-up plays a crucial role in promoting stability and generalization. Many studies have examined its underlying mechanisms from different aspects. However, most of the studies focus on incorporating these insights into optimizers to reduce the reliance on warm-up. Little attention has been paid to addressing the inherent limitations of the warm-up itself, which restricts its effectiveness. In this work, we revisit warm-up from a loss landscape perspective and identify several limitations with existing warm-up, including: (1) susceptibility to nearby suboptimal traps, (2) sensitivity to hyperparameters and random seeds, and (3) inefficiency during the early stages of training. To overcome these limitations, we propose Sensitivity-Aware Warm-Up (SAWU), a lightweight and adaptive strategy that dynamically leverages learning sensitivity during warm-up to guide updates toward better and more stable basins. In addition, SAWU also introduces an adaptive scheduling mechanism and phase transition strategy across warm-up, stable, and decay phases to further enhance robustness and efficiency. Extensive experiments on various downstream tasks show that SAWU significantly outperforms the vanilla method (e.g., average 3.43% improvement on RoBerta). Moreover, SAWU can be easily combined with various optimizers and remains effective even when warm-up-based methods fail (e.g, it lifts RAdam from 49.46% to 91.78% on qnli. Thanks to its lightweight nature, SAWU introduces minimal overhead and even reduces training time by over 5% compared to other methods.

Retrieval-Augmented Generation (RAG) is an effective solution to overcome the limitations of Large Language Models (LLMs) in terms of specific-domain knowledge and timely information updates. However, current RAG methods typically respond to queries based on isolated segments, lacking the ability to integrate information within the same document. This undermines performance in real-world tasks requiring coherent understanding across an entire document. Notably, the human brain naturally integrates and summarizes prior knowledge upon reading a given text, progressively formulating a comprehensive understanding. Motivated by this cognitive process, we propose the Hierarchical Two-Stage Summarization-based Information Retrieval (HTSIR) method, which preprocesses the corpus prior to retrieval, summarizes continuous texts to obtain integrated information, and constructs a retrieval tree with varying summary granularities. The retrieved information is then processed by a Reranker based on the current question to serve as a context for LLMs. Additionally, as single-step summarization is often imprecise in query-based summarization tasks, we further apply a Refinement module, allowing LLMs to reflect and revise their output to achieve the final result. By combining HTSIR with GPT-4o mini, we achieve state-of-the-art results on complex question tasks across four long-text datasets (NarrativeQA, QASPER, QuALITY, and QMSum), achieving an improvement of about 6 points on the Question Answering (QA) task in QuALITY-HRAD.

Multimodal summarization with multimodal output (MSMO) aims to generate coherent textual summaries while selecting the most semantically relevant images to enhance expressiveness. Despite the advancements of large multimodal models like GPT-4o, LLaMA-3, and Grok-3, these models often exhibit hallucination and weak visual-text alignment when applied to MSMO tasks. To address these challenges, we propose ModalSyncSum, a unified framework that enhances semantic consistency and visual faithfulness. It incorporates image-aware information extraction to mitigate visual-text misalignment, QA-based description verification to detect and correct hallucinated image descriptions, and named entity-guided refinement to ensure factual accuracy and entity alignment across modalities. Furthermore, we introduce a new evaluation metric M3AS, which jointly considers image content coverage, text-image alignment, and summary consistency, filling the gap in evaluating multimodal summary quality. Experimental results show that our model outperforms prompt-based baselines across multiple datasets, achieving significant gains on ROUGE, BLEU, and BERTScore, with BLEU improving by 21.95%. In human evaluation, M3AS exhibits stronger correlation with human judgments in consistency, image-summary relevance, and focus, surpassing existing automatic metrics.

Chain‑of‑thought (CoT) prompting boosts Large Language Models accuracy on multi‑step tasks, yet whether the generated ``thoughts'' reflect the true internal reasoning process is unresolved. We present the first feature‑level causal study of CoT faithfulness. Combining sparse autoencoders with activation patching, we extract monosemantic features from Pythia‑70M and Pythia‑2.8B while they tackle GSM8K math problems under CoT and plain (noCoT) prompting. Swapping a small set of CoT‑reasoning features into a noCoT run raises answer log‑probabilities significantly in the 2.8B model, but has no reliable effect in 70M, revealing a clear contrast for these two scales. CoT also leads to significantly higher activation sparsity and feature interpretability scores in the larger model, signalling more modular internal computation. For example, the model's confidence in generating correct answers improves from 1.2 to 4.3. We introduce patch‑curves and random‑feature patching baselines, showing that useful CoT information is not only present in the top-K patches but widely distributed. Overall, our results indicate that CoT can induce more interpretable internal structures in high-capacity LLMs, validating its role as a structured prompting method.

Effectively handling long contexts is challenging for Large Language Models (LLMs) due to the rarity of long texts, high computational demands, and substantial forgetting of short-context abilities. Recent approaches have attempted to construct long contexts for instruction tuning, but these methods often require LLMs or human interventions, which are both costly and limited in length and diversity. Also, the drop in short-context performances of present long-context LLMs remains significant. In this paper, we introduce Flora, an effortless (human/LLM-free) long-context construction strategy. Flora can markedly enhance the long-context performance of LLMs by arbitrarily assembling short instructions based on categories and instructing LLMs to generate responses based on long-context meta-instructions. This enables Flora to produce contexts of arbitrary length and scale with rich diversity, while only slightly compromising short-context performance. Experiments on Llama3-8B-Instruct and QwQ-32B show that LLMs enhanced by Flora excel in three long-context benchmarks while maintaining strong performances in short-context tasks.

Recent advances in large language models (LLMs) have greatly improved their reasoning and decision-making abilities when deployed as agents. Richer reasoning, however, often comes at the cost of longer chain of thought (CoT), hampering interaction efficiency in real-world scenarios. Nevertheless, there still lacks systematic definition of LLM‑Agent efficiency, hindering targeted improvements. To this end, we introduce dual‑efficiency, comprising (i) step-level efficiency, which minimizes tokens per step, and (ii) trajectory-level efficiency, which minimizes the number of steps to complete a task. Building on this definition, we propose DEPO, a dual-efficiency preference‑based optimization method that jointly rewards succinct responses and fewer action steps. Experiments on WebShop and BabyAI show that DEPO cuts token usage by up to 60.9% and steps by up to 26.9%, while achieving up to a 29.3% improvement in task performance. DEPO also generalizes to three out-of-domain math benchmarks and retains its efficiency gains when trained on only 25% of the data.

Large language models (LLMs) often suffer from hallucination, generating factually incorrect statements when handling questions beyond their knowledge and perception. Retrieval-augmented generation (RAG) addresses this by retrieving query-relevant contexts from knowledge bases to support LLM reasoning. Recent advances leverage pre-constructed graphs to capture the relational connections among distributed documents, showing remarkable performance in complex tasks. However, existing Graph-based RAG (GraphRAG) methods rely on a costly process to transform the corpus into a graph, introducing overwhelming token cost and update latency. Moreover, real-world queries vary in type and complexity, requiring different logic structures for accurate reasoning. The pre-built graph may not align with these required structures, resulting in ineffective knowledge retrieval. To this end, we propose a Logic-aware Retrieval Augmented Generation framework (LogicRAG) that dynamically extracts reasoning structures at inference time to guide adaptive retrieval without any pre-built graph. LogicRAG begins by decomposing the input query into a set of subproblems and constructing a directed acyclic graph (DAG) to model the logical dependencies among them. To support coherent multi-step reasoning, LogicRAG then linearizes the graph using topological sort, so that subproblems can be addressed in a logically consistent order. Besides, LogicRAG applies graph pruning to reduce redundant retrieval and uses context pruning to filter irrelevant context, significantly reducing the overall token cost. Extensive experiments demonstrate that LogicRAG achieves both superior performance and efficiency compared to state-of-the-art baselines.