AAAI 2026 Papers

The widespread adoption of reinforcement learning-based alignment highlights the growing importance of reward models. Various benchmarks have been built to evaluate reward models in various domains and scenarios. However, a significant gap remains in assessing reward models for long-form generation, despite its critical role in real-world applications. To bridge this, we introduce Long-form RewardBench, the first reward modeling testbed specifically designed for long-form generation. Our benchmark encompasses five key subtasks: QA, RAG, Chat, Writing, and Reasoning. We collected instruction and preference data through a meticulously designed multi-stage data collection process, and conducted extensive experiments on 20+ mainstream reward models, including both classifiers and generative models. Our findings reveal that current models still lack long-form reward modeling capabilities. Furthermore, we designed a novel Long-form Needle-in-a-Haystack Test, which revealed a correlation between reward modeling performance and the error's position within a response, as well as the overall response length, with distinct characteristics observed between classification and generative models. Finally, we demonstrate that classifier exhibit better generalizability compared to generative models trained on the same data. As the first benchmark for long-form reward modeling, this work aims to offer a robust platform for visualizing progress in this crucial area.

Generative LLMs typically improve Named Entity Recognition (NER) performance through instruction tuning. They excel at generating entities by semantic pattern matching but lack an explicit, verifiable reasoning mechanism. This "cognitive shortcutting" leads to suboptimal performance and weak generalization, especially in zero-shot and low-resource scenarios where reasoning from limited contextual cues is crucial. To address this issue, a reasoning framework is proposed for NER, which shifts the extraction paradigm from implicit pattern matching to explicit reasoning. This framework consists of three stages: Chain of Thought (CoT) generation, CoT tuning, and reasoning enhancement. First, a dataset annotated with NER-oriented CoTs is generated, which contain task-relevant reasoning chains. Then, they are used to tune the NER model to generate coherent rationales before deriving the final answer. Finally, a reasoning enhancement stage is implemented to optimize the reasoning process using a comprehensive reward signal. This stage ensures explicit and verifiable extractions. Experiments show that ReasoningNER demonstrates impressive cognitive ability in the NER task, achieving competitive performance. In zero-shot settings, it achieves SoTA performance, outperforming GPT-4 by 12.3 percentage points on the F1 score. Analytical results demonstrate its great potential to advance research in reasoning-oriented information extraction.

Event co-occurrences have been proven effective for event argument extraction (EAE) in previous studies; however, few have considered intra- and inter-event role correlations. Since role varies among different event types, event structure heterogeneity and overlap pose significant challenges to EAE. To address this issue, we propose a Role Correlation Structure-Enhanced model for Multi-Event Argument Extraction (RoSE), capable of capturing both heterogeneity and overlap of event structures through modeling role correlations. The proposed RoSE model employs a joint context-prompts input, role-centric graph-guided encoder (RoGE), and role-specific information fusion (RoIF). The RoGE is designed to enhance the intra- and inter-event role correlation between prompts and their corresponding event contexts. The RoIF module utilizes intra-event role information to improve multi-event arguments extraction. Extensive experiments on four widely-used benchmarks (RAMS, WikiEvents, MLEE, and ACE05) demonstrate that our proposed approach achieves state-of-the-art performance, validating the effectiveness of incorporating both intra- and inter-event role correlations.

While scaling the length of responses at test-time has been shown to markedly improve the reasoning abilities and performance of large language models (LLMs), it often results in verbose outputs and increases inference cost. Prior approaches for efficient test-time scaling, typically using universal budget constraints or query-level length optimization, do not leverage historical information from previous encounters with the same problem during training. We hypothesize that this limits their ability to progressively make solutions more concise over time. To address this, we present History-Aware Policy Optimization (HAPO), which keeps track of a history state (e.g., the minimum length over previously generated correct responses) for each problem. HAPO employs a novel length reward function based on this history state to incentivize the discovery of correct solutions that are more concise than those previously found. Crucially, this reward structure avoids overly penalizing shorter incorrect responses with the goal of facilitating exploration towards more efficient solutions. By combining this length reward with a correctness reward, HAPO jointly optimizes for correctness and efficiency. We use HAPO to train DeepSeek-R1-Distill-Qwen-1.5B, DeepScaleR-1.5B-Preview, and Qwen-2.5-1.5B-Instruct, and evaluate HAPO on several math benchmarks that span various difficulty levels. Experiment results demonstrate that HAPO effectively induces LLMs’ concise reasoning abilities, producing length reductions of 33-59% with accuracy drops of only 2-5%.

Agents based on Large Language Models (LLMs) have demonstrated strong capabilities across a wide range of tasks. However, deploying LLM-based agents in high-stakes domains comes with significant safety and ethical risks. Unethical behavior by these agents can directly result in serious real-world consequences, including physical harm and financial loss. To efficiently steer the ethical behavior of agents, we frame agent behavior steering as a model editing task, which we term Behavior Editing. Model editing is an emerging area of research that enables precise and efficient modifications to LLMs while preserving their overall capabilities. To systematically study and evaluate this approach, we introduce BehaviorBench, a multi-tier benchmark grounded in psychological moral theories. This benchmark supports both the evaluation and editing of agent behaviors across a variety of scenarios, with each tier introducing more complex and ambiguous scenarios. We first demonstrate that Behavior Editing can dynamically steer agents toward the target behavior within specific scenarios. Moreover, Behavior Editing enables not only scenario-specific local adjustments but also more extensive shifts in an agent’s global moral alignment. We demonstrate that Behavior Editing can be used to promote ethical and benevolent behavior or, conversely, to induce harmful or malicious behavior. Through extensive evaluations of agents built on frontier LLMs, BehaviorBench validates the effectiveness of behavior editing across a wide range of models and scenarios. Our findings offer key insights into a new paradigm for steering agent behavior, highlighting both the promise and perils of Behavior Editing.

Large Language Models (LLMs) have shown improved generation performance through retrieval-augmented generation (RAG) following the retriever-reader paradigm, which supplements model inputs with externally retrieved knowledge. However, prior work often evaluates RAG holistically, assessing the retriever and reader jointly, making it difficult to isolate the true contribution of retrieval, particularly given the prompt sensitivity of LLMs used as readers. We move beyond perplexity and introduce Spectrum Projection Score (SPS), a lightweight and supervision-free metric that allows the reader to gauge the semantic alignment of a retrieved summary with its hidden representation by comparing the area formed by generated tokens from the summary, and the principal directions of subspace in the reader and to measure the relevance. Building on SPS we present xCompress, an inference‑time controller framework that dynamically samples, ranks, and compresses retrieval summary candidates. Extensive experiments on five QA benchmarks with four open-sourced LLMs show that SPS not only enhances performance across a range of tasks but also provides a principled perspective on the interaction between retrieval and generation.

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) with external knowledge retrieval, improving factual accuracy and knowledge coverage. However, existing RAG approaches face a fundamental trade-off when handling complex reasoning: while traditional iterative retrieval methods offer flexibility, their local perspective limits their ability to establish global knowledge connections. In contrast, structure-augmented RAG methods capture global relationships but incur significant construction costs. To fill in this gap, we propose MGranRAG, an innovative framework designed to integrate precise local retrieval with structured global reasoning. Our approach circumvents expensive semantic extraction by employing a lightweight contextual hierarchical graph, effectively combining the local adaptability of iterative retrieval with the global consistency of structured knowledge. The framework adopts a novel iterative optimization scheme: at the local level, the LLM identifies multi-granular contextual evidence, such as key sentences and phrases, within retrieved passages to refine retrieval. At the global level, these multi-granularity evidence nodes are then mapped and propagated within the structured hierarchical graph, enabling the diffusion of rich contextual information at different levels to introduce global semantic constraints and reorder retrieval results. This coordination between local and global iterative processes dynamically balances retrieval accuracy and contextual coherence. Experimental results on challenging multi-hop and open-domain question answering datasets show that our proposal achieves new state-of-the-art performance in both retrieval and answer accuracy.

Large language models (LLMs) augmented with retrieval have shown impressive performance in open-domain question answering, yet struggle significantly with temporal knowledge graph question answering (TKGQA). The core issue lies in structural misalignment: treating structured, temporally sensitive graph queries as plain text often causes LLMs to retrieve or reason with semantically similar but structurally incorrect facts, resulting in critical inaccuracies. To address this, we introduce SAR (Structure-Aligned Reasoning), a novel TKGQA framework that integrates LLM reasoning tightly with the explicit subject–predicate–object–time schema inherent in knowledge graphs. SAR employs an LLM agent to first decompose natural language questions into structured queries, clearly delineating entities, relationships, and temporal constraints. It then conducts schema-consistent, time-aware retrieval from the knowledge graph to acquire candidate quadruples, which guide a subsequent iterative ReAct-style reasoning process by the LLM. A final verification stage ensures that proposed answers strictly adhere to temporal conditions, reinforcing accuracy and temporal coherence. Experiments on two benchmark datasets, MultiTQ and CronQuestions, demonstrate SAR’s effectiveness, achieving the best results. Specifically, with GPT-4.1, SAR achieves 78.2% Hits@1 on MultiTQ, significantly outperforming existing methods, and similarly establishes a new performance record on CronQuestions. Our results underscore the critical importance of structural alignment in temporal reasoning tasks, particularly in handling complex queries involving multiple temporal constraints and multi-hop reasoning.

The emergence of large language models (LLMs) marks a transformative era in artificial intelligence~(AI). However, systematically evaluating the capability of LLMs is challenging due to the necessity of a large number of labeled test data. To tackle this problem, in the conventional AI field, AutoEval has been proposed to estimate the capability of AI models without data labeling effort. Unfortunately, even though multiple AutoEval methods have been proposed, most are constructed for classification tasks and evaluated only on image datasets. As a result, their effectiveness for LLMs is unclear, as LLMs often target generation tasks. In this work, we introduce the first AutoEval benchmark specifically designed to estimate the capability of LLMs using unlabeled test data, AEBench. Besides existing AutoEval methods, AEBench also supports our designed method, which utilizes the correlation between data uncertainty and model ability for the capability estimation. In total, AEBench covers 12 AutoEval methods and 120 method combinations. Based on AEBench, we conducted a comprehensive study to explore the usefulness of AutoEval on LLMs. Experimental results on 10 datasets demonstrated that our designed uncertainty features-based methods perform the best in achieving the lowest estimation errors.

Recent studies have explored the capabilities of large language models (LLMs) in solving knowledge-intensive mathematical reasoning problems. However, existing benchmarks predominantly involve static theorems that LLMs have encountered during pretraining, failing to assess dynamic knowledge integration. In this work, we introduce TaxReasoning, a novel benchmark designed to evaluate LLMs’ abilities in real-world tax calculation scenarios. These tasks require not only mathematical reasoning and numerical computation, but also the extraction and application of complex, frequently updated tax regulations. Through extensive experiments with state-of-the-art LLMs using diverse prompting strategies and knowledge augmentation techniques, we uncover substantial limitations in their ability to handle dynamic, knowledge-intensive questions—primarily due to missing domain-specific knowledge and ineffective retrieval. Even the best-performing models fall significantly short of human-level performance. Our analysis points to key avenues for improvement, including enhancing LLMs' reasoning capabilities, developing more effective knowledge summarization techniques, and improving retrieval strategies. TaxReasoning offers a critical testbed for advancing LLMs in dynamic knowledge-intensive domains.

As backdoor attacks become more stealthy and robust, they reveal critical weaknesses in current defense strategies: detection methods often rely on coarse-grained feature statistics, and purification methods typically require full retraining or additional clean models. To address these challenges, we propose DUP (Detection-guided Unlearning for Purification), a unified framework that integrates backdoor detection with unlearning-based purification. The detector captures feature-level anomalies by jointly leveraging class-agnostic distances and inter-layer transitions. These deviations are integrated through a weighted scheme to identify poisoned inputs, enabling more fine-grained analysis. Based on the detection results, we purify the model through a parameter-efficient unlearning mechanism that avoids full retraining and does not require any external clean model. Specifically, we innovatively repurpose knowledge distillation to guide the student model toward increasing its output divergence from the teacher on detected poisoned samples, effectively forcing it to unlearn the backdoor behavior. Extensive experiments across diverse attack methods and language model architectures demonstrate that DUP achieves superior defense performance in detection accuracy and purification efficacy.

Mental health assessment is crucial for early intervention and effective treatment, yet traditional clinician-based approaches are limited by the shortage of qualified professionals. Recent advances in artificial intelligence have sparked growing interest in automated psychological assessment, yet most existing approaches are constrained by their reliance on static text analysis, limiting their ability to capture deeper and more informative insights that emerge through dynamic interaction and iterative questioning. Therefore, in this paper, we propose a multi-agent framework for mental health evaluation that simulates clinical doctor-patient dialogues, with specialized agents assigned to questioning, adequacy evaluation, scoring, and updating. In detail, we introduce an adaptive questioning mechanism in which an evaluation agent assesses the adequacy of user responses to determine the necessity of generating targeted follow-up queries to address ambiguity and missing information. Additionally, we employ a tree-structured memory in which the root node encodes the user's basic information, while child nodes (e.g., topic and statement) organize key information according to distinct symptom categories and interaction turns. This memory is dynamically updated throughout the interaction to reduce redundant questioning and enhance the information extraction and contextual tracking capabilities. Experimental results on the DAIC-WOZ dataset illustrate the effectiveness of our proposed method, which achieves better performance than existing approaches. Our code is released at \url{https://github.com/MindIntLab-HFUT/AgentMental}.

Significant progress has been made in spoken question answering (SQA) in recent years. However, many existing methods, including large audio language models, struggle with processing long audio. Follow the success of retrieval augmented generation, a speech-related retriever shows promising in help preprocessing long-form speech. But the performance of existing speech-related retrievers is lacking. To address this challenge, we propose CLSR, an end-to-end contrastive language-speech retriever that efficiently extracts question-relevant segments from long audio recordings for downstream SQA task. Unlike conventional speech-text contrastive models, CLSR incorporates an intermediate step that converts acoustic features into text-like representations prior to alignment, thereby more effectively bridging the gap between modalities. Experimental results across four cross-modal retrieval datasets demonstrate that CLSR surpasses both end-to-end speech related retrievers and pipeline approaches combining speech recognition with text retrieval, providing a robust foundation for advancing practical long-form SQA applications.

Retrieval-Augmented Generation (RAG) enhances the response quality and domain-specific performance of large language models (LLMs) by incorporating external knowledge to combat hallucinations. In recent research, graph structures have been integrated into RAG to enhance the capture of semantic relations between entities. However, it primarily focuses on low-order pairwise entity relations, limiting the high-order associations among multiple entities. Hypergraph-enhanced approaches address this limitation by modeling multi-entity interactions via hyperedges, but they are typically constrained to inter-chunk entity-level representations, overlooking the global thematic organization and alignment across chunks. Drawing inspiration from the top-down cognitive process of human reasoning, we propose a theme-aligned dual-hypergraph RAG framework (Cog-RAG) that uses a theme hypergraph to capture inter-chunk thematic structure and an entity hypergraph to model high-order semantic relations. Furthermore, we design a cognitive-inspired two-stage retrieval strategy that first activates query-relevant thematic content from the theme hypergraph, and then guides fine-grained recall and diffusion in the entity hypergraph, achieving semantic alignment and consistent generation from global themes to local details. Our extensive experiments demonstrate that Cog-RAG significantly outperforms existing state-of-the-art baseline approaches.

Recent studies have raised significant concerns regarding the reliability of current mathematical benchmarks, highlighting key limitations such as simplistic design and potential data contamination that undermine evaluation accuracy. Consequently, developing a reliable benchmark that effectively evaluates large language models' (LLMs) genuine capabilities in mathematical reasoning remains a critical challenge. To address these concerns, we propose RV-Bench, a novel evaluation methodology for Benchmarking LLMs with Random Variables in mathematical reasoning. Specifically, we develop question-generating functions to produce random variable questions (RVQs), whose background content mirrors the original benchmark problems, but with randomized variable combinations, rendering them "unseen" to LLMs. Models must completely understand the inherent question pattern to correctly answer RVQs with diverse variable combinations. Thus, an LLMs' genuine reasoning capability is reflected through its accuracy and robustness on RV-Bench. We conducted extensive experiments on over 30 representative LLMs across more than 1,000 RVQs. Our findings reveal that LLMs exhibit a proficiency imbalance between encountered and "unseen" data distributions. Furthermore, RV-Bench reveals that proficiency generalization across similar mathematical reasoning tasks is limited, but we verified that it can still be effectively elicited through test-time scaling.

With the growing number of submitted scientific papers, there is an increasing demand for systems that can assist reviewers in evaluating research claims. Experimental results are a core component of scientific work, often presented in varying formats such as tables or charts. Understanding how robust current multimodal large language models (multimodal LLMs) are at verifying scientific claims across different evidence formats remains an important and underexplored challenge. In this paper, we design and conduct a series of experiments to assess the ability of multimodal LLMs to verify scientific claims using both tables and charts as evidence. To enable this evaluation, we adapt two existing datasets of scientific papers by incorporating annotations and structures necessary for a multimodal claim verification task. Using this adapted dataset, we evaluate 12 multimodal LLMs and find that current models perform better with table-based evidence while struggling with chart-based evidence. We further conduct human evaluations and observe that humans maintain strong performance across both formats, unlike the models. Our analysis also reveals that smaller multimodal LLMs (under 8B) show weak correlation in performance between table-based and chart-based tasks, indicating limited cross-modal generalization. These findings highlight a critical gap in current models' multimodal reasoning capabilities. We suggest that future multimodal LLMs should place greater emphasis on improving chart understanding to better support scientific claim verification.

Simultaneous Speech Translation (SimulST) systems stream in audio while simultaneously emitting translated text or speech. Such systems face the significant challenge of balancing translation quality and latency. We introduce a strategy to optimize this tradeoff: wait for more input only if you gain information by doing so. Based on this strategy, we present Regularized Entropy INformation Adaptation (REINA), a novel loss to train an adaptive policy using an existing non-streaming translation model. We derive REINA from information theory principles and show that REINA helps push the reported Pareto frontier of the latency/quality tradeoff over prior works. Utilizing REINA, we train a SimulST model on French, Spanish and German, both from and into English. Training on only open source or synthetically generated data, we achieve state-of-the-art (SOTA) streaming results for models of comparable size. We also introduce a metric for streaming efficiency, quantitatively showing REINA improves the latency/quality trade-off by as much as 21 percent compared to prior approaches, normalized against non-streaming baseline BLEU scores.

Large language models have demonstrated remarkable capabilities across many tasks, yet face significant challenges when dealing with recursive reasoning problems—those requiring the resolution of nested hierarchical structures. While prior research has extensively studied length generalization (a model’s ability to handle longer sequences than seen during training), we investigate a distinct and underexplored limitation: depth generalization. Here, depth refers to the number of nested levels in a hierarchical problem, such as the layers of parentheses in a mathematical expression or the nesting of logical clauses in a Boolean formula. Our work reveals that standard transformer architectures struggle with problems involving deeper recursion than encountered during training, even when they perform well on longer but non-nested sequences. This limitation stems from their inability to maintain stack-like behavior—the capacity to track and resolve multiple levels of nested dependencies. Through systematic analysis, we demonstrate how this architectural constraint leads to rapid performance decay as the depth of recursion increases. To address this challenge, we develop a novel looped locate-and-replace pipeline that decomposes recursive problems into manageable subcomponents. The approach employs two specialized models: a locator that identifies solvable subexpressions and a replacer that evaluates these components while preserving the overall structure. We evaluate this method in three carefully designed domains—Boolean algebra, recursive arithmetic, and propositional logic—each with a controllable depth of recursion. Our results show that the proposed method effectively alleviates performance decay when tested on out-of-distribution recursion depth.

Many existing financial math reasoning benchmarks suffer from data contamination and high manual construction costs. To address this, we propose a novel formula-driven approach to dynamically construct math reasoning benchmarks in finance. Our two-stage approach: (1) generates single-formula questions by LLMs using a "Mask-for-Solve" paradigm for ground truth answers, and (2) synthesizes multi-formula questions through hierarchical tree-based DAGs. Our approach ensures novelty (via LLMs' creativity) and controllability of difficulty (via DAG structure). Based on a self-constructed financial formula bank, we utilize the proposed method to build FinMathBench, the first formula-driven and fully LLM-generated benchmark aimed at assessing LLMs' math reasoning abilities in finance, containing 946 questions across 4 complexity levels. Evaluation results on 40 LLMs demonstrate significant accuracy drops in multi-formula questions, e.g., 72.9% (1-Formula) to 14.0% (4-Formula) for GPT-4o under Chain-of-Thought prompting. Three critical flaws of LLMs are also observed: poor direct calculation performance, bias toward frequently solved variables in formulas, and erroneous "correction" of valid but extreme financial values. These findings highlight gaps in current LLMs' domain-specific reasoning and underscore FinMathBench's value for advancing robust financial LLMs.