IJCAI 2026 Papers

Tool-augmented methods aim to enhance the reasoning capabilities of large language models (LLMs) by invoking external tools, which can be broadly categorized into training-free and training-based methods. Training-free methods can directly instruct LLMs to invoke external tools, but they exhibit limited generalization in dynamic environments. In contrast, training-based methods use a teacher model to generate tool-use trajectories and fine-tune a student model to distill the trajectories that can generalize across tasks. However, they still face two major challenges: (1) Knowledge boundary mismatch: teacher trajectories may contain steps that the student cannot execute without tools due to their mismatch of knowledge boundaries. (2) Tool overuse propagation: the student inherits the misaligned tool-use strategy from teacher model. To address these challenges, we propose RADAR, an automated framework for knowledge boundary discovery and tool overuse mitigation. First, RADAR performs semantic anchoring to select representative seeds. Second, simulated annealing probing is employed to actively explore the student model’s tool-use boundary and acquire sparse, verified student-aware labels. Third, it globally propagates these labels and rewrites conflicts, yielding deployment-aligned decisions with fewer unnecessary tool calls. We conduct experiments on multiple benchmarks, compared to SOTA method, RADAR reduces average tool-use by 52.7% while improving accuracy by 3.2% across three models.

Large language models have shown promising potential in chemistry, with prior work exploring molecular recognition, classification, and property prediction. Despite the achieved progress, LLMs are still far from satisfactory when dealing with complex chemical multi-step reaction pathway inference task due to the lack of domain knowledge and limited ability to maintain consistent multi-step reasoning. To address these challenges, we propose ChemKGL, a novel multi-step reasoning framework enhanced by knowledge graph retrieval. We reformulate the task as a knowledge graph retrieval problem to better control LLM generation, and leverage graph technologies to integrate textual descriptions, reaction conditions, and materials. Based on this reformulation, we design a novel multi-retrieval strategy to improve the performance of reaction pathway inference, which includes forward retrieval, reverse retrieval, and historical retrieval. An entropy-based pruning mechanism is further introduced to alleviate cold-start issues caused by retrieval failure. In addition, we construct a comprehensive chemical knowledge graph and corresponding dataset to support model training and analysis. Finally, we employ multiple evaluation methods to assess our proposed ChemKGL, and the experimental results demonstrate its superiority. Our code and dataset can be obtained at https://github.com/Double-Sail/ChemKGL.

Training a unified language model that adapts between intuitive System 1 and deliberative System 2 remains challenging due to interference between their cognitive modes. Recent studies have thus pursued making System 2 models more efficient. However, these approaches focused on output control, limiting what models produce. We argue that this paradigm is misaligned: output length is merely a symptom of the model's cognitive configuration, not the root cause. In this work, we shift the focus to capability control, which modulates how models think rather than what they produce. To realize this, we leverage existing Instruct and Thinking checkpoints through dynamic parameter interpolation, without additional training. Our pilot study establishes that linear interpolation yields a convex, monotonic Pareto frontier, underpinned by representation continuity and structural connectivity. Building on this, we propose DAMI (DynAmic Model Interpolation), a framework that estimates a query-specific Reasoning Intensity lambda(q) to configure cognitive depth. For training-based estimation, we develop a preference learning method encoding accuracy and efficiency criteria. For zero-shot deployment, we introduce a confidence-based method leveraging inter-model cognitive discrepancy. Experiments on five mathematical reasoning benchmarks demonstrate that DAMI achieves higher accuracy than the Thinking model while remaining efficient, effectively combining the efficiency of System 1 with the reasoning depth of System 2.

The rapid advancement of generative AI has made audio deepfakes increasingly indistinguishable from authentic human vocals, posing significant threats to persons-of-interest (POI) such as public figures. Current detection systems primarily rely on generic, black-box models that fail to capture speaker-specific idiosyncratic traits and lack interpretability. In this paper, we propose Phoneme-based Voice Profiling (PVP), a novel personalized defense framework. By shifting the detection paradigm from macro-utterance analysis to micro-phonetic modeling, PVP captures the unique acoustic distributions underlying a POI’s habitual articulatory patterns. Specifically, our framework models speaker-specific phonetic realizations using lightweight Gaussian Mixture Models (GMMs) estimated solely from bona fide reference speech. This design enables data-efficient profiling and robust generalization to previously unseen spoofing attacks without requiring heavy spoof-specific training. Furthermore, we introduce the first large-scale Chinese POI deepfake dataset to benchmark speaker-specific detection. Experimental results demonstrate that PVP significantly outperforms state-of-the-art generic detectors in POI spoofing scenarios, achieving substantial EER reductions while providing fine-grained, phoneme-level interpretability for forensic analysis. Code and data are available at: https://github.com/JunXue-tech/PVP

Training critique models to provide useful feedback for actor models is an effective approach in scalable oversight, especially for complex tasks like math reasoning. However, current research lacks suitable datasets for effectively training critique models and integrating them in a principled way at both test time and training time. To bridge this gap, we first propose AutoMathCritique, an automated and scalable framework for collecting critique data. Using it, we create MathCritique-76k, a dataset of $76,321$ instances paired with step-level feedback, which is then used to train critique models for generating natural-language feedback. We show that critique models consistently improve the actor’s performance—particularly on difficult queries at test time—with larger gains as inference compute scales. Building on the findings, we propose a critique-in-the-loop self-improvement method that incorporates critique-based supervision into the actor’s self-training process. Extensive experiments demonstrate that this method improves the actor’s exploration efficiency and solution diversity, especially on challenging queries, leading to a stronger actor model. Our code and datasets are at https://mathcritique.github.io/.

LLMs have shown immense potential for code translation, yet they often struggle to ensure both syntactic correctness and semantic consistency. While preference-based learning offers a promising alignment strategy, it is hindered by unreliable semantic rewards derived from sparse test cases or restrictive reference translations. We argue that a robust semantic reward for code translation must be derived directly from the source code. In this paper, we propose CTO to improve code translation with syntax-guided and semantic-aware preference optimization. Through contrastive learning, we train a cross-lingual semantic model to directly assess functional equivalence between source and translated code. By formulating code translation as a multi-objective optimization problem, this robust semantic signal is seamlessly unified with compiler-based syntactic feedback within the direct preference optimization framework. Extensive experiments on C++, Java, and Python translations demonstrate that CTO significantly outperforms existing baselines and alternative preference optimization strategies.

Retrieval-Augmented Generation (RAG) systems have demonstrated remarkable effectiveness in mitigating hallucinations by incorporating external knowledge. However, the retrieval process inevitably introduces noise, posing significant challenges to RAG robustness. Fundamentally, noise robustness is a composite competency encompassing noise identification, retrieval filtering, and robust reasoning. Unfortunately, existing methods often fail to capture this holistic nature, typically restricted to single-dimensional enhancements or hampered by noise-induced optimization interference. To address these limitations, we decompose RAG robustness into three sub-tasks: knowledge boundary perception, retrieval verification, and robust reasoning. Accordingly, we propose ARMOR, an Adaptive Curriculum Meta-Learning for Noise-Robust RAG Reasoning. Inspired by the teaching process, ARMOR introduces an adaptive task scheduler as a "Teacher" to curate training plans according to the absolute learning progress of the model ("student"). Distinct from GRPO, ARMOR utilizes inter-group advantage estimation during learning to effectively mitigate mode collapse caused by limited output spaces. Experiments on three noise-injected complex reasoning datasets (2WikiMultiHopQA, MusiQue, and HotPotQA) indicate that ARMOR achieves synergistic improvements in retrieval noise robustness, significantly outperforming baselines in output quality.

In-context learning (ICL) enables large language models (LLMs) to adapt to new tasks with considerable performance gains, yet its effectiveness is highly sensitive to the choice of demonstrations. Most existing selection methods rely on heuristic or proxy signals, such as similarity, diversity, or uncertainty, and select demonstrations independently, which may misalign with downstream performance and overlook set-level composition effects. Therefore, we propose Performance-Driven Demonstration Selection (PDDS), which directly aligns demonstration selection with ICL performance. PDDS formulates selection as predicting the target LLM’s downstream task performance for a given query–in-context pair, replacing proxy heuristics with a performance-aware objective. Unlike prior approaches that rank demonstrations independently before composing the prompt, PDDS evaluates the entire in-context set, capturing inter-demonstration interactions and composition effects. PDDS trains an end-to-end scorer using supervision from the target LLM’s actual task outcomes. At inference time, it selects high-scoring in-context sets without additional target LLM calls or validation-time feedback. Across 6 NLP tasks, 8 datasets, and 8 target LLMs ranging from 1B to 70B parameters, PDDS achieves state-of-the-art results and generalizes well across LLMs and datasets. PDDS remains effective with as few as 30 training instances and can be integrated as a plug-and-play component to enhance existing ICL methods.

Remote Sensing Image-Text Retrieval (RSITR) aims to achieve precise retrieval between remote sensing images and textual descriptions. However, existing methods neglect the multi-dimensional cognitive attributes inherent in remote sensing data and struggle to handle them simultaneously, leading to suboptimal retrieval performance. In this paper, we propose a novel Progressive Decoupling-Aggregation-Refinement framework for RSITR (PDAR-RSITR) to comprehensively capture multi-dimensional cognitive attributes. Specifically, we adapt Multi-dimensional Cognitive Decoupling to learn features across cognitive attributes of different dimensions via multi-process clustering. Subsequently, we utilize Salient Cognitive Aggregation to select and aggregate the most salient attributes via dynamic routing. Furthermore, we propose Expert Collaborative Refinement to enhance critical cross-modal relationships via three complementary expert perspectives. Extensive experimental results demonstrate that PDAR-RSITR significantly outperforms existing state-of-the-art methods across multiple metrics.

Cross-domain Named Entity Recognition (CD-NER) aims to transfer the rich knowledge in the source domain to the target domain. Recent studies adopting decomposition or generation paradigms have achieved significant performance improvements, demonstrating high accuracy in entity span detection. However, during entity type classification, models severely suffer from entity type confusion, the erroneous tendency that models classify entities of one type in the text as another similar but incorrect type. To address this issue, we first propose a Multidimensional Confusion Quantification Model (MCQM) that quantifies a model's confusion extent between entity types from three dimensions: source-target hierarchy analysis, semantic similarity analysis, and explicit data evaluation. Moreover, we propose the Progressive Bidirectional Reasoning Chain (PBRC). PBRC leverages the source-target hierarchy and confusion analysis from the MCQM to prompt the LLM to generate two-stage reasoning information. The two-stage reasoning information is utilized to augment the knowledge of the model, significantly mitigating entity type confusion and improving the model's generalization performance. Experimental results demonstrate that our method achieves new state-of-the-art results on all domains of the CrossNER dataset.

Evaluating retrieval-augmented generation (RAG) systems in legal domains is challenging due to the nuanced nature of legal reasoning and scarcity of domain-specific benchmarks. We investigate semantic similarity and legal correctness in Brazilian legal question answering, developing a synthetic dataset of 3,012 evaluation instances from the Brazilian Civil Procedure Code spanning seven query types, validated through human expert assessment. Our evaluation framework combines BERTScore with domain-adapted LLM-as-Judge (GPT-4o-mini), validated against expert legal assessment. Analysis reveals 50.8% disagreement between semantic similarity and legal correctness, with correlation varying substantially by query type (ρ = 0.464 to ρ = 0.732). Explicit article citation emerges as the strongest quality predictor (Cohen’s d = 1.099), with cited responses achieving 145% higher legal correctness scores. Concrete examples demonstrate bidirectional divergence: responses may achieve high semantic similarity yet cite incorrect articles, or provide legally perfect answers with minimal lexical overlap. Findings demonstrate that semantic metrics alone are insufficient for legal RAG evaluation, with disagreement patterns structured by query type.

Current approaches to evaluating personality in large language models (LLMs) typically prompt them to self-report on psychological questionnaires such as the Big Five Inventory. However, these methods assess introspective labels rather than observable behavior, despite the fact that LLMs are deployed to act in realistic contexts, not to reflect on their own traits. To bridge this gap, we introduce BehaviorBench, a new benchmark for evaluating LLM personality through concrete behaviors in everyday scenarios. Grounded in established personality psychology, BehaviorBench links each Big Five trait to validated behavioral manifestations and embeds them in contextually plausible situations that naturally elicit trait-relevant actions. We evaluate a range of models and personality shaping strategies using BehaviorBench and find a substantial mismatch between self-reported personalities and actual behaviors. By grounding evaluation in observable behavior rather than introspection, our work reveals critical gaps in current LLM personality modeling and control mechanisms. Our code and data are available at https://github.com/butra1n/BehaviorBench

Mixture-of-Experts (MoE) large language models improve inference efficiency through sparse expert activation, but deployment on resource-constrained devices remains challenging due to the large expert parameter footprint. Expert offloading mitigates this issue by loading experts on demand, yet its effectiveness critically depends on accurate and efficient expert prediction: inaccurate predictions incur redundant expert transfers, while overly expensive predictors negate latency benefits. Existing trajectory-based methods suffer from low accuracy, whereas semantic-based approaches incur prohibitive similarity-matching overhead. We propose DoMoE, a domain-aware MoE inference system that exploits domain locality in inference workloads. DoMoE organizes routing information into domain-specific expert routing tables, restricts semantic matching to domain-relevant tokens and explicitly balances prediction accuracy against prediction overhead. Experiments show that DoMoE achieves a 1.32 X average throughput improvement and a 1.22 X increase in expert hit ratio across multiple MoE models and workloads, enabling efficient and accurate expert routing for practical inference.

AI-driven survey generation has advanced rapidly, yet related work generation (RWG) remains relatively underexplored. Unlike surveys that provide broad literature overviews, RWG synthesizes prior studies for a single focal paper, requiring contextual fit, cross-paper comparison, and accurate attribution. To address this gap, we propose MA-RWG, a fully automated multi-agent framework that generates polished related work sections from only a title and abstract. MA-RWG first retrieves high-quality candidate papers through semantic retrieval, optionally enhanced with a diversity-aware term. It then coordinates four specialized agents for summarization, organization, integration, and fact checking, enabling DAG-based taxonomy construction, feedback-guided refinement, and dual-model verification. For evaluation, we introduce a dedicated benchmark for paper-specific related work generation, covering generation quality, citation quality, and claim-level semantic similarity. Experimental results show that MA-RWG outperforms RAG-based baselines and survey-oriented agentic methods on the RWG task. Further ablation and cross-domain experiments demonstrate the soundness and robustness of the proposed framework.

Knowledge editing is pivotal for efficiently updating the parametric memory of Large Language Models (LLMs), enabling them to function as evolving agents in dynamic environments. However, mainstream in-parameter knowledge editing approaches suffer from Subject-Dominant Memory Interference: modifying a specific fact inadvertently corrupts the broader structural knowledge associated with the same subject within LLMs. We diagnose the root cause as a shortcut learning pathology, where the optimization objective overfits subject representations while bypassing the essential relational context. To rectify this, we propose Causal Path Alignment (CPA), a principled framework designed to anchor the optimization trajectory to valid causal pathways. CPA enforces parameter updates to route through relation-aware intermediate states, thereby preventing the erasure of contextual dependencies. Experimental results across diverse LLM backbones demonstrate that CPA consistently eliminates the shortcut, significantly improving relation specificity while exhibiting minimal side-effects. Moreover, CPA serves as a model-agnostic plug-in for existing editors, paving the way for reliable and trustworthy in-parameter knowledge editing.

Movie dubbing is the task of synthesizing speech from scripts conditioned on video scenes, requiring accurate lip sync, faithful timbre transfer, and proper modeling of character identity and emotion. However, existing methods face two major limitations: (1) high-quality multimodal dubbing datasets are limited in scale, suffer from high word error rates, contain sparse annotations, rely on costly manual labeling, and are restricted to monologue scenes, all of which hinder effective model training; (2) existing dubbing models rely solely on the lip region to learn audio-visual alignment, which limits their applicability to complex live-action cinematic scenes, and exhibit suboptimal performance in lip sync, speech quality, and emotional expressiveness. To address these issues, we propose FunCineForge, which comprises an end-to-end production pipeline for large-scale dubbing datasets and an MLLM-based dubbing model designed for diverse cinematic scenes. The pipeline enables the construction of the first television dubbing dataset, CineDub, which serves as a high-quality foundation for training and evaluation. Building on this, our dubbing model effectively captures multimodal cues and supports complex dubbing scenarios, including monologue, narration, dialogue, and multi-speaker settings. Experiments demonstrate that our approach consistently outperforms state-of-the-art methods in audio quality, word error rate, lip sync, temporal alignment, timbre transfer, and instruction following. Code and demos are available at https://funcineforge.github.io/.

Reasoning chains in Large Language Models (LLMs) often exhibit heavy-tailed length distributions, yet existing efficiency methods rely on suboptimal linear penalties that suppress complex reasoning, limiting both accuracy and generalization. To address this, we first empirically observe that reasoning lengths are well approximated by a log-normal distribution, and provide an intuitive explanation for this phenomenon. Based on this insight, we propose the Powered Length Penalty (PLP), an adaptive regularizer that penalizes redundancy in short sequences while gradually reducing penalties for longer sequences, preserving deep reasoning. Trained solely on the elementary GSM8K dataset, PLP significantly improves reasoning efficiency by reducing inference costs on GSM8K and MATH500, while simultaneously enhancing accuracy on the challenging AIME2024 benchmark. Furthermore, PLP transfers effectively to diverse domains, including MMLU and GPQA. These results suggest that modeling reasoning length distributions and adapting penalties accordingly can mitigate the typical trade-off between efficiency and performance, enabling more reliable and cost-effective reasoning across tasks.

Large language model (LLM)-based text-to-speech (TTS) systems enable prompt-conditioned emotional control but struggle with fine-grained emotion intensity due to the semantic--acoustic gap between text and speech. To address this challenge, we formulate emotion intensity control in LLM-based TTS as a learning-to-rank problem and propose Emo-LiPO, a listwise preference optimization framework that aligns prompt-conditioned speech generation with relative emotion intensity expressed in text. Emo-LiPO explicitly models global intensity ordering within each emotion under fixed transcripts, enabling more faithful and continuous emotional expression. We further construct ESD-plus, a multi-speaker dataset with explicit emotion intensity variations, to support fine-grained emotion modeling and evaluation. Experiments on ESD-plus demonstrate that Emo-LiPO significantly improves emotion accuracy and intensity controllability over both supervised- and DPO-based LLM TTS baselines, with particularly pronounced gains at high intensity levels.

Zero-Shot Relation Triplet Extraction (ZSRTE) aims to extract relation triplets for unseen relation types without any annotated training data. Recent advancements in Large Language Models (LLMs) have significantly enhanced ZSRTE performance, enabling the direct generation of relational triplets from unstructured text. However, LLMs often introduce biases such as entity shift, relation confusion, and over-prediction, which limit the reliability of the extracted triplets. In this paper, we introduce DETRI, a novel debiasing framework for ZSRTE that addresses these biases by leveraging a discriminative structural expert model. DETRI treats LLMs as inspiration generators and refines their outputs via a three-stage debiasing pipeline, consisting of inspiration-based re-prediction, confidence-based filtering, and entity shift correction through span perturbation. The framework improves LLM-generated triplets without requiring fine-tuning the LLM, thus offering a transferable and scalable solution. Extensive experiments on two datasets show that DETRI outperforms state-of-the-art methods by reducing bias and improving extraction accuracy, achieving 2.89% F1 improvement over fine-tuned LLM methods while keeping the LLM parameters frozen. Our approach shows strong generalization across different LLMs, offering a solution to mitigate biases.