AAAI 2026 论文列表

The text-to-SQL task is an active challenge in Natural Language Processing. Many existing solutions focus on using black-box language models extended with specialized components within customized end-to-end text-to-SQL pipelines. While these solutions use both closed-source proprietary language models and coding-oriented open-source models, there is a lack of research regarding SQL-specific small generative models. At the same time, recent advancements in self-correcting generation strategies show promise for improving the capabilities of existing architectures. The application of these concepts to the text-to-SQL task remains unexplored. In this paper, we introduce RetrySQL, a new approach to training text-to-SQL generation models. We prepare reasoning steps for reference SQL queries and then corrupt them to create retry data that contains both incorrect and corrected steps, divided with a special token. We continuously pre-train open-source coding models with this data and demonstrate that retry steps yield an improvements of up to 4 and 9 percentage points for overall and challenging execution metrics, respectively, as compared to pre-training without retry data. We showcase that the self-correcting behavior is learned by the model and the increase in downstream accuracy metrics is a result of this additional skill. Finally, we incorporate RetrySQL-trained models into the full text-to-SQL pipeline and showcase that they are competitive in terms of execution accuracy with proprietary models that contain orders of magnitude more parameters. RetrySQL demonstrates that self-correction can be learned in the text-to-SQL task and provides a novel way of improving generation accuracy for small SQL-oriented language models.

This paper presents a systematic investigation into the constrained generation capabilities of large language models (LLMs) in producing Songci, a classical Chinese poetry form characterized by strict structural, tonal, and rhyme constraints defined by Cipai templates. We first develop a comprehensive, multi-faceted evaluation framework that includes: (i) a formal conformity score, (ii) automated quality assessment using LLMs, (iii) human evaluation, and (iv) classification-based probing tasks. Using this framework, we evaluate the generative performance of 18 LLMs, including 3 proprietary models and 15 open-source models across 4 families, under five prompting strategies: zero-shot, one-shot, completion-based, instruction-based, and chain-of-thought. Finally, we propose a Generate-Critic architecture in which the evaluation framework functions as an automated critic. Leveraging the critic’s feedback as a scoring function for best-of-N selection, we fine-tune 3 lightweight open-source LLMs via supervised fine-tuning (SFT), resulting in improvements of up to 5.88% in formal conformity. Our findings offer new insights into the generative strengths and limitations of LLMs in producing culturally significant and formally constrained literary texts.

Large Language Models (LLMs) have exhibited remarkable capabilities but remain vulnerable to jailbreaking attacks, which can elicit harmful content from the models by manipulating the input prompts. Existing black-box jailbreaking techniques primarily rely on static prompts crafted with a single, non-adaptive strategy, or employ rigid combinations of several underperforming attack methods, which limits their adaptability and generalization. To address these limitations, we propose MAJIC, a Markovian adaptive jailbreaking framework that attacks black-box LLMs by iteratively combining diverse innovative disguise strategies. MAJIC first establishes a ''Disguise Strategy Pool'' by refining existing strategies and introducing several innovative approaches. To further improve the attack performance and efficiency, MAJIC formulate the sequential selection and fusion of strategies in the pool as a Markov chain. Under this formulation, MAJIC initializes and employs a Markov matrix to guide the strategy composition, where transition probabilities between strategies are dynamically adapted based on attack outcomes, thereby enabling MAJIC to learn and discover effective attack pathways tailored to the target model. Our empirical results demonstrate that MAJIC significantly outperforms existing jailbreak methods on prominent models such as GPT-4o and Gemini-2.0-flash, achieving over 90\% attack success rate with fewer than 15 queries per attempt on average.

Defending large language models (LLMs) against jailbreak attacks is crucial for ensuring their safe deployment. Existing defense strategies typically rely on predefined static criteria to differentiate between harmful and benign prompts. However, such rigid rules fail to accommodate the inherent complexity and dynamic nature of real-world jailbreak attacks. In this paper, we focus on the novel challenge of adaptive defense against diverse jailbreaks. We propose a new concept "mirror'', which is a dynamically generated prompt that reflects the syntactic structure of the input while ensuring semantic safety. The discrepancies between input prompts and their corresponding mirrors serve as guiding principles for defense. A novel defense model, MirrorShield, is further proposed to detect and calibrate risky inputs based on the crafted mirrors. Evaluated on multiple benchmark datasets and compared against ten state-of-the-art attack methods, MirrorShield demonstrates superior defense performance and promising generalization capabilities.

In Psychotherapy, Early Maladaptive Schemas (EMS) are entrenched negative perceptions of self or others that perpetuate mental health challenges, contribute to treatment resistance and relapse, and obstruct therapeutic progress. Addressing EMS using appropriate psychotherapeutic support (PS) strategies helps resolve core emotional deficits, mitigate resistance, and improve client engagement. Moreover, adapting polite and empathetic communication based on clients’ emotional states fosters trust, emotional safety, and a conducive therapeutic environment, which is critical for addressing EMS and achieving positive outcomes. Motivated by these insights, we introduce MATE - a novel EMS-guided polite and empAthetic dialogue sysTem for psychothErapeutic support. MATE integrates a Large Language Model (LLM) with a Mixture of Experts-based Reinforcement Learning (MoE-RL) approach to overcome the limitations of traditional RL methods, such as large action spaces and generic responses. The LLM captures diverse semantic patterns from dialogue context. MoE-RL leverages dedicated psychotherapeutic, politeness, and empathy experts, along with a new reward function, comprising PS, politeness, empathy, contextual consistency, and diversity rewards to guide policy learning for effective response generation. Evaluations on the HOPE and PSYCON datasets demonstrate MATE’s efficacy in generating polite and empathetic psychotherapeutic responses based on clients’ EMS and emotional cues while ensuring contextual consistency and diversity.

Refusal refers to the functional behavior enabling safety-aligned language models to reject harmful or unethical prompts. Following the growing scientific interest in mechanistic interpretability, recent work encoded refusal behavior as a single direction in the model’s latent space; e.g., computed as the difference between the centroids of harmful and harmless prompt representations. However, emerging evidence suggests that concepts in LLMs often appear to be encoded as a low-dimensional manifold embedded in the high-dimensional latent space. Motivated by these findings, we propose a novel method leveraging Self-Organizing Maps (SOMs) to extract multiple refusal directions. To this end, we first prove that SOMs generalize the prior work's difference-in-means technique. We then train SOMs on harmful prompt representations to identify multiple neurons. By subtracting the centroid of harmless representations from each neuron, we derive a set of multiple directions expressing the refusal concept. We validate our method on an extensive experimental setup, demonstrating that ablating multiple directions from models' internals outperforms not only the single-direction baseline but also specialized jailbreak algorithms, leading to an effective suppression of refusal. Finally, we conclude by analyzing the mechanistic implications of our approach.

This work explores the consistency of LLMs in answering multiple times the same question. In particular, we study how known, open-source LLMs respond to 10 repetitions of questions from the multiple-choice benchmarks MMLU-Redux and MedQA, considering different inference temperatures, small (2B-10B parameters) vs. medium models (50B-80B), finetuned vs. base models, and other parameters. The paper also examines the effects of requiring answer consistency in repetitive inferences on accuracy and the trade-offs involved in deciding which model best provides both of them, for what we propose some new representations. Results show that the number of questions which can be answered consistently vary wildly among models but typically is in the 50%-85% range for small models and that accuracy among consistent answers correlates to overall accuracy at low inference temperatures. Results for medium-sized models seem to indicate much higher levels of answer consistency.

Cross-lingual topic modeling seeks to uncover coherent and semantically aligned topics across languages—a task central to multilingual understanding. Yet most existing models learn topics in disjoint, language-specific spaces and rely on alignment mechanisms (e.g., bilingual dictionaries) that often fail to capture deep cross-lingual semantics, resulting in loosely connected topic spaces. Moreover, these approaches often overlook the rich semantic signals embedded in multilingual pretrained representations, further limiting their ability to capture fine-grained alignment. We introduce **GloCTM** (**Glo**bal Context Space for **C**ross-Lingual **T**opic **M**odel), a novel framework that enforces cross-lingual topic alignment through a unified semantic space spanning the entire model pipeline. GloCTM constructs enriched input representations by expanding bag-of-words with cross-lingual lexical neighborhoods, and infers topic proportions using both local and global encoders, with their latent representations aligned through internal regularization. At the output level, the global topic-word distribution, defined over the combined vocabulary, structurally synchronizes topic meanings across languages. To further ground topics in deep semantic space, GloCTM incorporates a Centered Kernel Alignment (CKA) loss that aligns the latent topic space with multilingual contextual embeddings. Experiments across multiple benchmarks demonstrate that GloCTM significantly improves topic coherence and cross-lingual alignment, outperforming strong baselines.

Written Multi-Party Conversations (WMPCs) are widely studied across disciplines, with social media as a primary data source due to their accessibility. However, these datasets raise privacy concerns and often reflect platform-specific properties. For example, interactions between speakers may be limited due to rigid platform structures (e.g., threads, tree-like discussions), which yield overly simplistic interaction patterns (e.g., one-to-one ``reply-to'' links). This work explores the feasibility of generating synthetic WMPCs with instruction-tuned Large Language Models (LLMs) by providing deterministic constraints such as dialogue structure and participants’ stance. We investigate two complementary strategies of leveraging LLMs in this context: (i.) LLMs as WMPC generators, where we task the LLM to generate a whole WMPC at once and (ii.) LLMs as WMPC parties, where the LLM generates one turn of the conversation at a time (made of speaker, addressee and message), provided the conversation history. We next introduce an analytical framework to evaluate compliance with the constraints, content quality, and interaction complexity for both strategies. Finally, we assess the level of obtained WMPCs via human and LLM-as-a-judge evaluations. We find stark differences among LLMs, with only some being able to generate high-quality WMPCs. We also find that turn-by-turn generation yields better conformance to constraints and higher linguistic variability than generating WMPCs in one pass. Nonetheless, our structural and qualitative evaluation indicates that both generation strategies can yield high-quality WMPCs.

Generating with citations is crucial for trustworthy Large Language Models (LLMs), yet even advanced LLMs often produce mismatched or irrelevant citations. Existing methods over-optimize citation fidelity while overlooking relevance to the user query, which degrades answer quality and robustness in real-world settings with noisy or irrelevant retrieved content. Moreover, the prevailing single-pass paradigm struggles to deliver optimal answers in long-form generation that requiring multiple citations. To address these limitations, we propose FineRef, a framework based on Fine-grained error Reflection, which explicitly teaches the model to self-identify and correct two key citation errors—mismatch and irrelevance—on a per-citation basis. FineRef follows a two-stage training strategy. The first stage instills an “attempt–reflect–correct” behavioral pattern via supervised fine-tuning, using fine-grained and controllable reflection data constructed by specialized lightweight models. An online self-reflective bootstrapping strategy is designed to improve generalization by iteratively enriching training data with verified, self-improving examples. To further enhance the self-reflection and correction capability, the second stage applies process-level reinforcement learning with a multi-dimensional reward scheme that promotes reflection accuracy, answer quality, and correction gain. Experiments on the ALCE benchmark demonstrate that FineRef significantly improves both citation performance and answer accuracy. Our 7B model outperforms GPT-4 by up to 18% in Citation F1 and 4% in EM Recall, while also surpassing the state-of-the-art model across key evaluation metrics. FineRef also exhibits strong generalization and robustness in domain transfer settings and noisy retrieval scenarios.

Large language models now draft news, legal analyses, and software code with human-level fluency. At the same time, regulations such as the EU AI Act mandate that each synthetic passage carry an imperceptible, machine-verifiable mark for provenance. Conventional logit-based watermarks satisfy this requirement by selecting a pseudorandom green vocabulary at every decoding step and boosting its logits, yet the random split can exclude the highest-probability token and thus erode fluency. WaterMod mitigates this limitation through a probability-aware modular rule. The vocabulary is first sorted in descending model probability; the resulting ranks are then partitioned by the residue rank mod k, which distributes adjacent—and therefore semantically similar—tokens across different classes. A fixed bias of small magnitude is applied to one selected class. In the zero-bit setting (k=2), an entropy-adaptive gate selects either the even or the odd parity as the green list. Because the top two ranks fall into different parities, this choice embeds a detectable signal while guaranteeing that at least one high-probability token remains available for sampling. In the multi-bit regime (k>2), the current payload digit d selects the color class whose ranks satisfy rank mod k = d. Biasing the logits of that class embeds exactly one base-k digit—equivalently log2(k) bits—per decoding step, thereby enabling fine-grained provenance tracing. The same modular arithmetic therefore supports both binary attribution and rich payloads. Experimental results demonstrate that WaterMod consistently attains strong watermark detection performance while maintaining generation quality in both zero-bit and multi-bit settings. This robustness holds across a range of tasks, including natural language generation, mathematical reasoning, and code synthesis.

Large Language Models (LLMs) demonstrate strong reasoning capabilities but still struggle with hallucinations and limited transparency. Recently, KG-enhanced LLMs that integrate knowledge graphs (KGs) have been shown to improve reasoning performance, particularly for complex, knowledge-intensive tasks. However, these methods still face significant challenges, including inaccurate retrieval and reasoning failures, often exacerbated by long input contexts that obscure relevant information. Furthermore, many of these approaches rely on LLMs to directly retrieve evidence from KGs, and to self-assess the sufficiency of this evidence, which often results in premature or incorrect reasoning. To address the retrieval and reasoning failures, we propose ProgRAG, a multi-hop knowledge graph question answering (KGQA) framework that decomposes complex questions into sub-questions, and progressively extends partial reasoning paths by answering each sub-question. At each step, external retrievers gather candidate evidence, which is then refined through uncertainty-aware pruning by the LLM. Finally, the context for LLM reasoning is optimized by organizing and rearranging the partial reasoning paths obtained from the sub-question answers. Experiments on two well-known datasets, WebQSP and CWQ, demonstrate that ProgRAG outperforms existing baselines in multi-hop KGQA, offering improved reliability and reasoning quality.

With the rapid advancement of mathematical reasoning capabilities in Large Language Models (LLMs), AI systems are increasingly being adopted in educational settings to support students’ comprehension of problem-solving processes. However, a critical component remains underexplored in current LLM-generated explanations: multimodal explanation. In real-world instructional contexts, human tutors routinely employ visual aids, such as diagrams, markings, and highlights, to enhance conceptual clarity. To bridge this gap, we introduce the multimodal solution explanation task, designed to evaluate whether models can identify visual keypoints, such as auxiliary lines, points, angles, and generate explanations that incorporate these key elements essential for understanding. To evaluate model performance on this task, we propose ME2, a multimodal benchmark consisting of 1,000 math problems annotated with visual keypoints and corresponding explanatory text that references those elements. Our empirical results show that current models struggle to identify visual keypoints. In the task of generating keypoint-based explanations, open-source models also face notable difficulties. This highlights a significant gap in current LLMs’ ability to perform mathematical visual grounding, engage in visually grounded reasoning, and provide explanations in educational contexts. We expect that the multimodal solution explanation task and the ME2 dataset will catalyze further research on LLMs in education and promote their use as effective, explanation-oriented AI tutors.

Large language models (LLMs) may generate harmful outputs on malicious inputs. Existing safety methods, including prompt engineering and model editing, rely on hand-crafted templates or target-driven parameter modifications, limiting their generalizability in unseen harmful scenarios. Post-training aims to ensure LLM safety in general domains via supervised fine-tuning (SFT) or reinforcement learning (RL) on diverse malicious inputs. SFT needs annotated refusal samples while RL learns to refuse risk by exploring diverse harmful inputs. However, these methods tend to harshly refuse over any possible risks, sacrificing potentially useful information and degrading model utility. We argue that realistic malicious inputs often mix both harmful and helpful semantics (i.e., entities and relations), and LLMs should identify and remove only harmful relations while preserving useful ones. Thus, the original malicious user inputs can shift into safe queries, to which LLMs can respond safely and helpfully. In this paper, we propose WALKSAFE, a graph-based risk-aware training framework that enables LLMs to identify potential risks of key semantics (entities and relations) in user inputs via graph structure. By filtering harmful relations, LLMs can respond to safe input queries and then generate their corresponding safe and helpful responses. First, we model all entities and relations in the inputs with a graph structure. Second, we adopt a risk-aware random walk on the graph to quantify potential risk under multiple entities and relations. Then, we reconstruct safe queries by filtering harmful relations to promote the LLM to answer safely and helpfully rather than with direct refusals. Finally, we propose Bi-GRPO to post-train LLMs. As vanilla GRPO conducts only the intra-group comparison, Bi-GRPO performs both intra-group and inter-group comparisons between different response groups. The extra inter-group rewards encourage the model to distinguish harmful and safe semantics, and thus prefer safe and helpful responses. Experiments on three LLMs show that our models obtain SOTA results.

Direct Preference Optimization (DPO) simplifies reinforcement learning from human feedback (RLHF) for large language models (LLMs) by directly training on offline preference data to align with human preferences. During DPO training, the reference model serves as a data weight adjuster. However, the common practice of initializing the policy and reference models identically in DPO can lead to inefficient data utilization and impose a performance ceiling. Meanwhile, the absence of a reference model in Simple Preference Optimization (SimPO) reduces training robustness and requires stricter conditions to prevent catastrophic forgetting. In this work, we propose Pre-DPO, a simple yet effective DPO-based training paradigm that improves preference optimization by introducing a guiding reference model. This reference model provides foresight into the desired policy state achievable through the training preference data, serving as a guiding mechanism that adaptively assigns higher weights to samples more suitable for the model and lower weights to those less suitable. Extensive experiments on the AlpacaEval 2 and Arena-Hard v0.1 benchmarks demonstrate that Pre-DPO consistently improves the performance of both DPO and SimPO, without relying on external models or additional data.

Social biases embedded in Large Language Models (LLMs) raise critical concerns, resulting in representational harms -- unfair or distorted portrayals of demographic groups -- that may be expressed in subtle ways through generated language. Existing evaluation methods often depend on predefined identity-concept associations, limiting their ability to surface new or unexpected forms of bias. In this work, we present the Bias Association Discovery Framework (BADF), a systematic approach for extracting both known and previously unrecognized associations between demographic identities and descriptive concepts from open-ended LLM outputs. Through comprehensive experiments spanning multiple models and diverse real-world contexts, BADF enables robust mapping and analysis of the varied concepts that characterize demographic identities. Our findings advance the understanding of biases in open-ended generation and provide a scalable tool for identifying and analyzing bias associations in LLMs.

We introduce DSCodeBench, a new benchmark designed to evaluate large language models (LLMs) on complicated and realistic data science code generation tasks. DSCodeBench consists of 1,000 carefully constructed problems sourced from realistic problems from GitHub across ten widely used Python data science libraries. DSCodeBench offers a more challenging and representative testbed, more complex code solutions, more comprehensive data science libraries, clearer and better structured problem descriptions, and stronger test suites. To construct the DSCodeBench, we develop a robust pipeline that combines task scope selection, code construction, test case generation, and problem description synthesis. The process is paired with rigorous manual editing to ensure alignment and enhance the reliability of the evaluation. Experimental result shows that DSCodeBench exhibits robust scaling behavior, where larger models systematically outperform smaller ones, validating its ability to distinguish model capabilities. The best LLM we test, GPT-4o, has a pass@1 of 0.392, indicating that LLMs still have a large room to improve for realistic data science code generation tasks. We believe DSCodeBench will serve as a rigorous and trustworthy foundation for advancing LLM-based data science programming.

Reinforcement Learning (RL) has been shown to improve the capabilities of large language models (LLMs). However, applying RL to open-domain tasks faces two key challenges: (1) the inherent subjectivity of these tasks prevents the verifiable rewards as required by Reinforcement Learning with Verifiable Rewards (RLVR); (2) Reinforcement Learning from Human Feedback (RLHF) relies on external reward mechanisms. To overcome these limitations, we propose Self-Examining Reinforcement Learning (SERL), a novel self-improving framework where the LLM serves as both Actor and Judge. SERL introduces two synergistic reward mechanisms without any external signals. On the one hand, to improve the Actor's capability, we derive rewards from Copeland-style pairwise comparison judgments across a group of generated responses. On the other hand, a self-consistency reward that encourages coherent judgments is proposed to improve the Judge's reliability. This refinement strengthens the Judge, consequently generating a more robust training signal for the Actor. Experiments show that our method outperforms existing self-improvement training methods. SERL improves the LC win rate of Qwen3-8B on AlpacaEval 2.0 from 52.37% to 59.90%. To the best of our knowledge, our method achieves state-of-the-art performance among self-improving approaches. Furthermore, it achieves a performance comparable to significantly larger models like Qwen3-32B, demonstrating superior effectiveness and robustness on open-domain tasks.

Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, DoRA, and HiRA, enable lightweight adaptation of large pre-trained models via low-rank updates. However, existing PEFT approaches apply static, input-agnostic updates to all tokens, disregarding the varying importance and difficulty of different inputs. This uniform treatment can lead to overfitting on trivial content or under-adaptation on more informative regions, especially in autoregressive settings with distinct prefill and decoding dynamics. In this paper, we propose GateRA, a unified framework that introduces token-aware modulation to dynamically adjust the strength of PEFT updates. By incorporating adaptive gating into standard PEFT branches, GateRA enables selective, token-level adaptation—preserving pre-trained knowledge for well-modeled inputs while focusing capacity on challenging cases. Empirical visualizations reveal phase-sensitive behaviors, where GateRA automatically suppresses updates for redundant prefill tokens while emphasizing adaptation during decoding. To promote confident and efficient modulation, we further introduce an entropy-based regularization that encourages near-binary gating decisions. This regularization prevents diffuse update patterns and leads to interpretable, sparse adaptation without hard thresholding. Finally, we present a theoretical analysis showing that GateRA induces a soft gradient-masking effect over the PEFT path, enabling continuous and differentiable control over adaptation. Experiments on multiple commonsense reasoning benchmarks demonstrate that GateRA consistently outperforms or matches prior PEFT methods.