AAAI 2026 论文列表

Self-evaluation, a model's ability to assess the correctness of its own output, is crucial for Large Multimodal Models (LMMs) to achieve self-improvement in multi-turn conversations, yet largely absent in foundation models. Recent work has employed reinforcement learning (RL) to enhance self-evaluation; however, its fixed reward mechanism suffers from reward hacking when optimizing multiple training objectives, leading to model collapse. In this paper we propose AdaPO, an online reinforcement learning framework capable of adaptively adjusting training objective in real time according to the current training state for each task. Specifically, to mitigate reward hacking , AdaPO introduces an Adaptive Reward Model (ARM) and a Reward Aware Dynamic KL Regularization mechanism. ARM assesses the task's training state from the distribution of model generated multi-turn trajectories' performance. Reward Aware Dynamic KL replaces a fixed penalty with dynamic coefficients which is modulated by the reward gap between different multi-turn situations. Notably, our method automatically and smoothly adjusts its learning focus based on sub-tasks' training progress without manual intervention. Extensive experiments over 8 benchmarks and various models show that our method significantly enhances both direct reasoning and self-evaluation capability.

Recently, with the increasing capabilities of Large Language Models (LLMs), AI applications have gradually emerged to solve various problems in people's daily lives, so accurately measuring their performance and reliability is paramount. However, existing benchmarks predominantly rely on closed-ended, multiple-choice or short-answer question formats. While useful for assessment, these formats exhibit a significant gap compared to the diverse and open-ended nature of questions posed by real-world users. To bridge this gap, we produce OmniBench, a comprehensive open-domain benchmark. OmniBench is uniquely composed of authentic, user-generated questions harvested from real-world interactions on various websites and applications, covering 16 rigorously defined knowledge domains and 5 crucial user intents derived from a large-scale analysis of the mass corpus. Crucially, we propose three automated data construction pipelines that enable the continuous and periodic updating of the benchmark dataset. This approach not only ensures that the questions can keep up with current events, but also effectively mitigates the critical issue of data contamination prevalent in static benchmarks. Moreover, a multi-dimensional hybrid evaluation framework named OmniEval is proposed for evaluating the responses. This framework combines diverse metrics and evaluation methods to capture nuanced aspects of answer performance. Extensive validation demonstrates that this evaluation framework exhibits strong alignment with human judgments, ensuring the reliability of the benchmark results.

Large Language Models (LLMs) are widely adopted, but their high training cost leads many developers to fine-tune existing open-source models. While most adhere to open-source licenses, some falsely claim original training despite clear derivation from public models, raising pressing concerns about intellectual property protection and the need to verify model provenance. In this paper, we propose GhostSpec, a lightweight yet effective method for verifying LLM lineage without access to training data or modification of model behavior. Our approach constructs compact and robust fingerprints by applying singular value decomposition (SVD) to invariant products of internal attention weight matrices. Unlike watermarking or output-based methods, GhostSpec is fully data-free, non-invasive, and computationally efficient. Extensive experiments show it is robust to fine-tuning, pruning, expansion, and adversarial transformations, reliably tracing lineage with minimal overhead. By offering a practical solution for model verification, our method contributes to intellectual property protection and fosters a transparent, trustworthy LLM ecosystem.

Learning rate scheduling is crucial for training large language models, yet understanding the optimal annealing strategies across different model configurations remains challenging. In this work, we investigate the transferability of annealing dynamics in large language model training and refine a generalized predictive framework for optimizing annealing strategies under the Warmup-Steady-Decay (WSD) scheduler. Our improved framework incorporates training steps, maximum learning rate, and annealing behavior, enabling more efficient optimization of learning rate schedules. Our work provides a practical guidance for selecting optimal annealing strategies without exhaustive hyperparameter searches, demonstrating that smaller models can serve as reliable proxies for optimizing the training dynamics of larger models. We validate our findings on extensive experiments using both Dense and Mixture-of-Experts (MoE) models, demonstrating that optimal annealing ratios follow consistent patterns and can be transferred across different training configurations.

Our statistical analysis reveals a complementary phenomenon between large language model-based question answering (QA) and small model-based QA. To facilitate dual knowledge transfer between these two paradigms, this paper introduces a collaborative enhancement method of large and small models for question answering. The proposed method consists of two iterative steps: i) small4large step, in which the small model first predicts an answer for a given question along with its confidence, and these results are then leveraged as prompts to strengthen the large model's performance; ii) large4small step, where the large model enhances the small model through distillation, judgment and reflection. Through iteration of these two steps, the large and small models could enhance each other progressively. Experimental evaluations across eight datasets spanning five domains demonstrate that the proposed method effectively improves the question answering performance of both large and small models simultaneously.

Fine-tuning adapts pretrained models for specific tasks but poses the risk of catastrophic forgetting (CF), where critical knowledge from pretraining is overwritten. To address the issue of CF in a general-purpose framework, we propose Low-damage Knowledge Implanting (LoKI), a parameter-efficient fine-tuning (PEFT) technique that utilizes recent mechanistic understanding of how knowledge is stored in transformer architectures. We compare LoKI against state-of-the-art PEFT methods in two real-world fine-tuning scenarios. The results show that LoKI demonstrates significantly better preservation of general capabilities. At the same time, its task-specific performance is comparable to or even surpasses that of full parameter fine-tuning and these PEFT methods across various model architectures. Our work bridges the mechanistic insights of LLMs' knowledge storage with practical fine-tuning objectives, enabling an effective balance between task-specific adaptation and the retention of general-purpose capabilities.

The ability of Large Language Models (LLMs) to use ex ternal tools unlocks powerful real-world interactions, mak ing rigorous evaluation essential. However, current bench marks primarily report final accuracy, revealing what mod els can do but obscuring the cognitive bottlenecks that define their true capability boundaries. To move from simple per formance scoring to a diagnostic tool, we introduce a frame workgroundedinCognitive LoadTheory.Ourframeworkde constructs task complexity into two quantifiable components: Intrinsic Load, the inherent structural complexity of the solu tion path, formalized with a novel Tool Interaction Graph; and Extraneous Load, the difficulty arising from ambiguous task presentation. To enable controlled experiments, we construct ToolLoad-Bench, the first benchmark with parametrically ad justable cognitive load. Our evaluation reveals distinct per formance cliffs as cognitive load increases, allowing us to precisely map each model’s capability boundary. We validate that our framework’s predictions are highly calibrated with empirical results, establishing a principled methodology for understanding an agent’s limits and a practical foundation for building more efficient systems.

The static nature of knowledge within Large Language Models (LLMs) makes it difficult for them to adapt to evolving information, rendering knowledge editing a critical task. However, existing methods struggle with challenges of scalability and retrieval efficiency, particularly when handling complex, multi-hop questions that require multi-step reasoning. To address these challenges, this paper introduces ALEX (A Light Editing-knowledge Extractor), a lightweight knowledge editing framework. The core innovation of ALEX is its hierarchical memory architecture, which organizes knowledge updates (edits) into semantic clusters. This design fundamentally reduces retrieval complexity from a linear O(N) to a highly scalable O(K+N/C). Furthermore, the framework integrates an Inferential Query Synthesis (IQS) module to bridge the semantic gap between queries and facts , and a Dynamic Evidence Adjudication (DEA) engine that executes an efficient two-stage retrieval process. Experiments on the MQUAKE benchmark demonstrate that ALEX significantly improves both the accuracy of multi-hop answers (MultiHop-ACC) and the reliability of reasoning paths (HopWise-ACC). It also reduces the required search space by over 80% , presenting a promising path toward building scalable, efficient, and accurate knowledge editing systems.

Long-form books are among the most information-rich and structurally complex forms of written content, often exceeding 100,000 words. While recent methods have enabled basic long-text generation, they remain limited in two key aspects: the inability to generate ultra-long content at book scale, and the lack of mechanisms for integrating rich factual information. To address these limitations, we propose DeepWriter, a multi-agent collaborative framework that follows a structured planning-then-generation paradigm. It first constructs a detailed book outline with narrative arcs and chapter semantics, then incrementally generates content conditioned on retrieved knowledge and contextual signals. DeepWriter supports controllable generation of full-length books exceeding 100,000 words, enriched with citations, trivia and images. To support evaluation beyond surface-level fluency, we introduce DeepWriter-Bench, a bilingual benchmark of 18 annotated books designed to assess book-scale coherence, richness, and factual grounding. Additionally, we propose BookScore, a unified 100-point metric for quantifying book maturity. Experimental results show that DeepWriter achieves a state-of-the-art BookScore of 80.92, consistently outperforming strong baselines.

Large Language Models (LLMs) have demonstrated remarkable performance in code generation, offering new possibilities for translating natural language into executable programs. To further enhance LLMs’ code generation capabilities, Retrieval-Augmented Generation (RAG) has emerged as a promising strategy by retrieving code examples aligned with the generation intent to guide the process. However, existing RAG-based methods often suffer from unnecessary augmentation, preference misalignment, and surface-level mimicry, which undermine the effectiveness of retrieved examples in guiding LLMs toward accurate code generation. To address these challenges, we propose SRACG, a Selective Retrieval-Augmented Code Generation framework. SRACG begins with a necessity-aware selection mechanism to identify generation intents that genuinely require retrieval support, thereby avoiding degradation from indiscriminate augmentation. For intents identified as needing enhancement, it first employs a multi-objective retrieval strategy to select examples that are semantically aligned with the intent. These candidates are then further filtered by assessing their consistency with the LLM’s inherent generation preferences, ensuring alignment in both style and structure. Finally, it extracts execution plans from the filtered examples to uncover their underlying logic, guiding the LLM to better comprehend the examples instead of merely mimicking surface-level content. Experimental results on widely used benchmarks show that SRACG significantly improves the success rate of LLM-generated code and outperforms existing approaches.

Despite recent advances in the reasoning capabilities of Large Language Models (LLMs), improving the reasoning ability of Small Language Models (SLMs, e.g., up to 1.5B parameters) remains challenging. A key obstacle lies in the complexity and variability of natural language: essentially equivalent problems often appear in diverse surface forms, often obscured by redundant or distracting details. This imposes a dual burden on SLMs: they must first extract the core problem from complex linguistic input, and then perform reasoning based on that understanding. The resulting vast and noisy problem space hinders optimization, particularly for models with limited capacity. To address this, we propose a new framework that decouples understanding from reasoning by mapping natural language problems into a canonical problem space-a semantically simplified yet expressive domain. This enables SLMs to focus on reasoning over standardized inputs, free from linguistic variability. Within this framework, we introduce DURIT (Decoupled Understanding from Reasoning via Iterative Training), a three-step algorithm that iteratively: (1) mapping natural language problems via reinforcement learning, (2) aligns reasoning trajectories through self-distillation, and (3) trains reasoning policies in the problem space. The mapper and reasoner are co-trained in an alternating loop throughout this process. Experiments show that DURIT substantially improves SLMs' performance on both in-domain and out-of-domain mathematical and logical reasoning tasks. Beyond improving reasoning capabilities, DURIT also improves the robustness of reasoning, validating decoupling understanding from reasoning as an effective strategy for strengthening SLMs.

Large Language Models (LLMs) often exhibit sycophantic behavior, agreeing with user-stated opinions even when those contradict factual knowledge. While prior work has documented this tendency, the internal mechanisms that enable such behavior remain poorly understood. In this paper, we provide a mechanistic account of how sycophancy arises within LLMs. We first systematically study how user opinions induce sycophancy across different model families. We find that simple opinion statements reliably induce sycophancy, whereas user expertise framing has a negligible impact. Through logit-lens analysis and causal activation patching, we identify a two-stage emergence of sycophancy: (1) a late-layer output preference shift and (2) deeper representational divergence. We also verify that user authority fails to influence behavior because models do not encode it internally. In addition, we examine how grammatical perspective affects sycophantic behavior, finding that first-person prompts (“I believe...”) consistently induce higher sycophancy rates than third-person framings (“They believe...”) by creating stronger representational perturbations in deeper layers. These findings highlight that sycophancy is not a surface-level artifact but emerges from a structural override of learned knowledge in deeper layers, with implications for alignment and truthful AI systems.

Narrative comprehension on long stories and novels has been a challenging domain attributed to their intricate plotlines and entangled, often evolving relations among characters and entities. Given the LLM's diminished reasoning over extended context and its high computational cost, retrieval-based approaches remain a pivotal role in practice. However, traditional RAG methods could fall short due to their stateless, single-step retrieval process, which often overlooks the dynamic nature of capturing interconnected relations within long-range context. In this work, we propose ComoRAG, holding the principle that narrative reasoning is not a one-shot process, but a dynamic, evolving interplay between new evidence acquisition and past knowledge consolidation, analogous to human cognition on reasoning with memory-related signals in the brain. Specifically, when encountering a reasoning impasse, ComoRAG undergoes iterative reasoning cycles while interacting with a dynamic memory workspace. In each cycle, it generates probing queries to devise new exploratory paths, then integrates the retrieved evidence of new aspects into a global memory pool, thereby supporting the emergence of a coherent context for the query resolution. Across four challenging long-context narrative benchmarks (200K+ tokens), ComoRAG outperforms strong RAG baselines with consistent relative gains up to 11% compared to the strongest baseline. Further analysis reveals that ComoRAG is particularly advantageous for complex queries requiring global comprehension, offering a principled, cognitively motivated paradigm for retrieval-based stateful reasoning.

Cued Speech (CS) enhances lipreading via hand coding, offering visual phonemic cues that support precise speech perception for the hearing-impaired. The task of CS Video-to-Speech generation (CSV2S) aims to convert CS videos into intelligible speech signals. Most existing research focuses on CS Recognition (CSR), which transcribes video content into text. Consequently, a common solution for CSV2S is to integrate CSR with a text-to-speech (TTS) system. However, this pipeline relies on text as an intermediate medium, which may lead to error propagation and temporal misalignment between speech and CS video dynamics. In contrast, directly generating audio speech from CS video (direct CSV2S) often suffer from the inherent multimodal complexity and the limited availability of CS data. To address these challenges, we propose UniCUE, the first unified framework for CSV2S that directly generates speech from CS videos without relying on intermediate text. The core innovation of UniCUE lies in integrating a understanding task (CSR) that provides fine-grained CS visual-semantic cues to to guide the speech generation. Specifically, UniCUE incorporates a pose-aware visual processor, a semantic alignment pool that enables precise visual–semantic mapping, and a VisioPhonetic adapter to bridge the understanding and generation tasks within a unified architecture. To support this framework, we construct UniCUE-HI, a large-scale Mandarin CS dataset containing 11,282 videos from 14 cuers, including both hearing-impaired and normal-hearing individuals. Extensive experiments conducted on this dataset demonstrate that UniCUE achieves state-of-the-art (SOTA) performance across multiple evaluation metrics.

In the literature, existing human-centric emotional motion generation methods primarily focus on boosting performance within a single scale-fixed dataset, largely neglecting the flexible and scale-increasing motion scenarios (e.g., sports, dance), whereas effectively learning these newly emerging scenarios can significantly enhance the model’s real-world generalization ability. Inspired by this, this paper proposes a new LLM-Centric Lifelong Empathic Motion Generation (L2-EMG) task, which aims to equip LLMs with the capability to continually acquire emotional motion generation knowledge across different unseen scenarios, potentially contributing to building a closed-loop and self-evolving embodied agent equipped with both empathy and intelligence. Further, this paper poses two key challenges in the L2-EMG task, i.e., the emotion decoupling challenge and the scenario adapting challenge. To this end, this paper proposes an Emotion-Transferable and Scenario-Adapted Mixture of Experts (ES-MoE) approach which designs a causal-guided emotion decoupling block and a scenario-adapted expert constructing block to address the two challenges, respectively. Especially, this paper constructs multiple L2-EMG datasets to validate the effectiveness of the ES-MoE approach. Extensive evaluations show that ES-MoE outperforms advanced baselines.

The proliferation of large language models (LLMs) has significantly advanced intelligent systems. Unfortunately, LLMs often face knowledge conflicts between internal memory and retrieved external information, arising from misinformation, biases, or outdated knowledge. These conflicts undermine response reliability and introduce uncertainty in decision-making. In this work, we analyze how LLMs navigate knowledge conflicts from an information-theoretic perspective and reveal that when conflicting and supplementary information exhibit significant differences, LLMs confidently resolve their preferences and alleviate the uncertainty during their response generation. When this difference is ambiguous, LLMs experience considerable uncertainty about their generation. Based on this insight, we propose Swin-VIB, a novel framework that integrates a pipeline of variational information bottleneck models to adapt the retrieved information difference, facilitating robust response generation of LLMs even in conflicting contexts. Extensive experiments confirm our theoretical analysis and demonstrate the performance of Swin-VIB. Notably, Swin-VIB outperforms all competitive baselines in terms of the accuracy of the multiple-choice task, while improving the EM values in the open-ended QA task by at least 11.14%.

Existing benchmarks in e-commerce primarily focus on basic user intents, such as finding or purchasing products. However, real-world users often pursue more complex goals, such as applying vouchers, managing budgets, and finding multi-products seller. To bridge this gap, we propose ShoppingBench, a novel end-to-end shopping benchmark designed to encompass increasingly challenging levels of grounded intent. Specifically, we propose a scalable framework to simulate user instructions based on various intents derived from sampled real-world products. To facilitate consistent and reliable evaluations, we provide a large-scale shopping sandbox that serves as an interactive simulated environment, incorporating over 2.5 million real-world products. Experimental results demonstrate that even state-of-the-art language agents (such as GPT-4.1) achieve absolute success rates under 50% on our benchmark tasks, highlighting the significant challenges posed by our ShoppingBench. In addition, we propose a trajectory distillation strategy and leverage supervised fine-tuning, along with reinforcement learning on synthetic trajectories, to distill the capabilities of a large language agent into a smaller one. As a result, our trained agent achieves competitive performance compared to GPT-4.1.

Text-to-Audio (TTA) generation has made rapid progress, but current evaluation methods remain narrow, focusing mainly on perceptual quality while overlooking robustness, generalization, and ethical concerns. We present TTA-Bench, a comprehensive benchmark for evaluating TTA models across functional performance, reliability, and social responsibility. It covers seven dimensions including accuracy, robustness, fairness, and toxicity, and includes 2,999 diverse prompts generated through automated and manual methods. We introduce a unified evaluation protocol that combines objective metrics with over 118,000 human annotations from both experts and general users. Ten state-of-the-art models are benchmarked under this framework, offering detailed insights into their strengths and limitations. TTA-Bench establishes a new standard for holistic evaluation of TTA systems.

In goal-oriented dialogue tasks, the main challenge is to steer the interaction towards a given goal within a limited number of turns. Existing approaches either rely on elaborate prompt engineering, whose effectiveness is heavily dependent on human experience, or integrate policy networks and pre-trained policy models, which are usually difficult to adapt to new dialogue scenarios and costly to train. Therefore, in this paper, we present Nested Rollout Policy Adaptation for Goal-oriented Dialogue (NRPA-GD), a novel dialogue policy planning method that completely avoids specific model training by utilizing a Large Language Model (LLM) to simulate behaviors of user and system at the same time. Specifically, NRPA-GD constructs a complete evaluation mechanism for dialogue trajectories and employs an optimization framework of nested Monte Carlo simulation and policy self-adaptation to dynamically adjust policies during the dialogue process. The experimental results on four typical goal-oriented dialogue datasets show that NRPA-GD outperforms both existing prompt engineering and specifically pre-trained model-based methods. Impressively, NRPA-GD surpasses ChatGPT and pre-trained policy models with only a 0.6-billion-parameter LLM. The proposed approach further demonstrates the advantages and novelty of employing planning methods on LLMs to solve practical planning tasks.