AAAI 2026 Papers

Existing tool-augmented large language models (LLMs) encounter significant challenges when processing complex queries. Current frameworks such as ReAct are prone to local optimization traps due to their reliance on incremental decision-making processes. To address these limitations, we propose a novel Planner-centric Plan-Execute paradigm that fundamentally resolves local optimization bottlenecks through architectural innovation. Central to our approach is a novel Planner model that performs global Directed Acyclic Graph (DAG) planning for complex queries, enabling optimized execution beyond conventional tool coordination. We also introduce ComplexTool-Plan, a large-scale benchmark dataset featuring complex queries that demand sophisticated multi-tool composition and coordination capabilities. Additionally, we develop a two-stage training methodology that integrates Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO), systematically enhancing the Planner's tool selection accuracy and global planning awareness through structured DAG-based planning. When integrated with a capable executor, our framework achieves state-of-the-art performance on the StableToolBench benchmark for complex user queries, demonstrating superior end-to-end execution capabilities and robust handling of intricate multi-tool workflows.

Retrieval-augmented generation (RAG) has proven to be effective in mitigating hallucinations in large language models, yet its effectiveness remains limited in complex, multi-step reasoning scenarios. Recent efforts have incorporated search-based interactions into RAG, enabling iterative reasoning with real-time retrieval. Most approaches rely on outcome-based supervision, offering no explicit guidance for intermediate steps. This often leads to reward hacking and degraded response quality. We propose Bi-RAR, a novel retrieval-augmented reasoning framework that evaluates each intermediate step jointly in both forward and backward directions. To assess the information completeness of each step, we introduce a bidirectional information distance grounded in Kolmogorov complexity, approximated via language model generation probabilities. This quantification measures both how far the current reasoning is from the answer and how well it addresses the question. To optimize reasoning under these bidirectional signals, we adopt a multi-objective reinforcement learning framework with a cascading reward structure that emphasizes early trajectory alignment. Empirical results on seven question answering benchmarks demonstrate that Bi-RAR surpasses previous methods and enables efficient interaction and reasoning with the search engine during training and inference.

The capacity for social reasoning, particularly Theory of Mind (ToM), is a foundational prerequisite for aligning Large Language Models (LLMs) with human values. However, current evaluations are predominantly confined to simplistic, short-text scenarios, obscuring their true capabilities and potential failure modes in complex, long-range social dynamics. To address this deficit, we introduce MovieGraph-ToM, a large-scale benchmark for evaluating long-range ToM and social cognition within extended, multimodal narratives. We employ a "scaffold-and-probe" methodology: we construct a ground-truth Social-Causal Graph offline, which maps the narrative's latent mental states and causal chains. During evaluation, the model is denied access to this graph and must reason directly from raw multimodal inputs. This decoupling forces genuine inference over superficial pattern matching. Reasoning is probed via a hierarchical questioning framework designed to differentiate spontaneous understanding from logical robustness. Our empirical results reveal systematic vulnerabilities in even state-of-the-art models. We identify a critical "multiple-choice pitfall," where accuracy plummets against well-crafted distractors, and a stark "generative-discriminative divide," where models fail to construct coherent explanations for answers they correctly identify. These findings highlight a latent risk, as models that feign comprehension could lead to unpredictable and misaligned behaviors. MovieGraph-ToM thus offers a rigorous platform for assessing and advancing the robust social intelligence required for safely aligned AI systems.

Large reasoning models (LRMs) have shown significant progress in test-time scaling through chain-of-thought prompting. Current approaches like search-o1 integrate retrieval augmented generation (RAG) into multi-step reasoning processes but rely on a single, linear reasoning path while incorporating unstructured textual information in a flat, context-agnostic manner. As a result, these approaches can lead to error accumulation throughout the reasoning chain, which significantly limits its effectiveness in medical question-answering (QA) tasks where both accuracy and traceability are critical requirements. To address these challenges, we propose MIRAGE (Multi-path Inference with Retrieval-Augmented Graph Exploration), a novel test-time scalable reasoning framework that performs dynamic multi-path inference over structured medical knowledge graphs. Specifically, MIRAGE 1) decomposes complex queries into entity-grounded sub-questions, 2) executes parallel inference paths, 3) retrieves evidence adaptively via neighbor expansion and multi-hop traversal, and 4) integrates answers using cross-path verification to resolve contradictions. Experiments on three medical QA benchmarks (GenMedGPT-5k, CMCQA, and ExplainCPE) show that MIRAGE consistently outperforms GPT-4o, Tree-of-Thought variants, and other retrieval-augmented baselines in both automatic and human evaluations. Additionally, MIRAGE improves interpretability by generating explicit reasoning chains that trace each factual claim to concrete paths within the knowledge graph, making it especially suitable for complex medical reasoning scenarios.

For industrial-scale text-to-SQL, supplying the entire database schema to Large Language Models (LLMs) is impractical due to context window limits and irrelevant noise. Schema linking, which filters the schema to a relevant subset, is therefore critical. However, existing methods incur prohibitive costs, struggle to trade off recall and noise, and scale poorly to large databases. We present AutoLink, an autonomous agent framework that reformulates schema linking as an iterative, agent-driven process. Guided by an LLM, AutoLink dynamically explores and expands the linked schema subset, progressively identifying necessary schema components without inputting the full database schema. Our experiments demonstrate AutoLink's superior performance, achieving state-of-the-art strict schema linking recall of 97.4% on Bird-Dev and 91.2% on Spider 2.0-Lite, with competitive execution accuracy, i.e., 68.7% EX on Bird-Dev (better than CHESS) and 34.9% EX on Spider 2.0-Lite (ranking 2nd on the official leaderboard). Crucially, AutoLink exhibits exceptional scalability, maintaining high recall, efficient token consumption, and robust execution accuracy on large schemas (e.g., over 3,000 columns) where existing methods severely degrade—making it a highly scalable, high-recall schema-linking solution for industrial text-to-SQL systems.

Understanding and reasoning over complex spreadsheets remain fundamental challenges for large language models (LLMs), which often struggle with intricate structures and rely solely on neural computation. In this work, we propose SheetBrain, a neuro-symbolic dual-workflow agent framework for precise and interpretable reasoning over tabular data. SheetBrain consists of an understanding module that produces a comprehensive overview of the spreadsheet, including structural summaries and query-specific analyses to guide execution; an execution module that integrates a Python sandbox with preloaded table-processing libraries and an Excel helper toolkit for effective data manipulation; and a validation module that verifies the correctness of reasoning and answers, triggering re-execution if necessary. We evaluate SheetBrain on multiple public QA and manipulation benchmarks, and introduce SheetBench, a new benchmark targeting large, multi-table, and structurally complex spreadsheets. Experimental results show that SheetBrain significantly improves reasoning performance on both existing benchmarks and the more challenging scenarios presented in SheetBench.

We propose Knowledge Boundary Discovery (KBD), a reinforcement learning based framework to explore the knowledge boundaries of the Large Language Models (LLMs). We define the knowledge boundary by automatically generating two types of questions: (i) those the LLM can confidently answer within-knowledge boundary and (ii) those it cannot beyond-knowledge boundary. Iteratively exploring and exploiting the LLM's responses to find its knowledge boundaries is challenging because of the hallucination phenomenon. To find the knowledge boundaries of an LLM, the agent interacts with the LLM under the modeling of exploring a partially observable environment. The agent generates a progressive question as the action, adopts an entropy reduction as the reward, receives the LLM's response as the observation and updates its belief states. We demonstrate that the KBD detects knowledge boundaries of LLMs by automatically finding a set of non-trivial answerable and unanswerable questions. We validate the KBD by comparing its generated knowledge boundaries with manually crafted LLM benchmark datasets. Experiments show that our KBD-generated question set is comparable to the human-generated datasets. Our approach paves a new way to evaluate LLMs.

Chain-of-Thought (CoT) reasoning is a critical capability for large language models (LLMs), enabling them to tackle complex multi-step tasks. While base LLMs, pre-trained on general text corpora, often struggle with reasoning due to a lack of specialized training, recent studies reveal their latent reasoning potential tied to hidden states. However, existing hidden state manipulation methods, such as linear activation steering, suffer from limitations due to their rigid and unconstrained nature, often leading to distribution shifts and degraded text quality. In this work, we propose a novel approach for eliciting CoT reasoning from base LLMs through hidden state manipulation grounded in probabilistic conditional generation. By reformulating the challenge as an optimization problem with a balanced likelihood and prior regularization framework, our method guides hidden states toward reasoning-oriented trajectories while preserving linguistic coherence. Extensive evaluations across mathematical, commonsense, and logical reasoning benchmarks demonstrate that our approach consistently outperforms existing steering methods, offering a theoretically principled and effective solution for enhancing reasoning capabilities in base LLMs.

Benchmarks serve as standardized test systems to distinguish capabilities among large language models (LLMs). Discriminative items enable high-ability LLMs to favor correct answers, while causing low-ability models to assign lower plausibility to these answers and tend toward incorrect answers. Current methods for assessing benchmark quality primarily focus on coverage of difficulty levels and task diversity, yet lack direct quantification of discrimination—the core metric. Furthermore, large-scale benchmarks incur high evaluation costs. Although heuristic methods can reduce item counts to some extent, they cannot guarantee preservation of the benchmark’s original discriminative properties. To address these limitations, we propose MetaEval, a meta-evaluation framework designed to precisely quantify per-item discrimination and enable efficient assessment. Central to MetaEval is our novel Signal Detection and Item Response (SD-IR) model, which simulates LLMs’ detection of correct answers (signals) by representing each model’s perception through two latent ability states: “known” and “unknown”. For any item, discrimination is quantified as the difference in signal plausibility between these states. Leveraging these discrimination metrics, MetaEval introduces two strategies to replicate full-benchmark results using minimal subsets for efficient evaluation: (1) Distilling metaBench: a compact subset that retains discriminative power by removing redundant items; (2) Predicting performance on full-benchmark based on metaBench’s discrimination. Experiments across five benchmarks confirm that high-discrimination items capture greater performance variation among LLMs, align more closely with full-benchmark rankings, and exhibit superior predictive ability. Notably, in the best case, MetaEval achieves accurate full-benchmark estimation using only 2.5% of items, substantially reducing evaluation costs while preserving reliability.

Uncertainty quantification (UQ) in foundation models is crucial for identifying and mitigating hallucinations in automatically generated text. However, heuristic UQ approaches lack statistical guarantees for key metrics such as the false discovery rate (FDR) in selective prediction tasks. Previous research adopts the split conformal prediction (SCP) framework to ensure desired coverage of admissible answers by constructing data-driven prediction sets, yet these sets typically contain incorrect candidates, undermining their practical effectiveness. To address this, we introduce COIN, an uncertainty-guarding selection framework that calibrates statistically valid uncertainty thresholds to filter a single generated answer per question under user-specified FDR constraints. COIN estimates the empirical error rate on the calibration set and applies confidence interval methods such as Clopper–Pearson to establish a high-probability upper bound on the true error rate (i.e., FDR). This enables the selection of the largest threshold that ensures FDR control on test data while significantly increasing sample retention. We demonstrate COIN's robustness in risk control, strong test-time power in retaining admissible answers, and predictive efficiency under limited calibration data across both general and multimodal text generation tasks. Furthermore, we show that employing alternative UQ and upper bound construction strategies can further boost COIN's power performance, which underscores its extensibility and adaptability to diverse application scenarios.

Retrieval-Augmented Generation (RAG) has been demonstrated to effectively mitigate the knowledge recency issue in Large Language Models (LLMs) while significantly reducing hallucinations. However, existing RAG methods exhibit insufficient capability in modeling reasoning paths for complex multi-hop reasoning tasks. While Reinforcement Learning (RL) has demonstrated success in enhancing model reasoning ability, Token-level RL frameworks exhibit inherent limitations in maintaining coherent reasoning trajectories. This approach remains susceptible to the compounding accumulation of contextual errors during the retrieval process, ultimately resulting in erroneous output generation. To address this challenge, we propose Chain Progressive Search (CP-Search), a novel two-stage training framework designed to enhance the model's retrieval capability in complex scenarios. This framework models the entire retrieval process as a Retrieval-level Markov Decision Process, systematically optimizing the model's retrieval behavior at each step of the chained retrieval. Specifically, CP-Search first constructs a retrieval-cognitive behavioral dataset and employs Supervised Fine-Tuning (SFT) to endow the model with cognitive behaviors for searching. More importantly, by introducing a dense progressive procedural reward in reinforcement learning training, CP-Search significantly improves the model's reasoning consistency and feedback correction ability in chained retrieval. Experiments conducted on multiple multi-hop datasets demonstrate that CP-Search significantly outperforms existing RAG methods in complex multi-hop reasoning tasks.

Large language model (LLM) generated texts now rival human quality, creating four text categories: purely machine-generated, machine-rewritten, machine-polished, and human-written content. Traditional detection methods face significant challenges in human-machine hybrid scenarios where LLMs perform rewriting or polishing, as existing approaches focus on single-level features and fail to capture subtle, multi-layered machine traces. To address this, we propose the Multi-level Style Preference Optimization (MSPO) framework, capturing machine style features at multiple granularities: sequence-level (overall consistency), phrase-level (distinctive n-gram patterns), and lexical-level (word selection distributions). We further incorporate four text complexity indicators (Type-Token Ratio, Average Sentence Length, Average Word Length, and Punctuation Ratio) to dynamically adjust optimization parameters based on human-machine text complexity differences, enhancing adaptability across diverse text types. Additionally, we construct a comprehensive detection dataset spanning three representative domains (scientific writing, news articles, and creative writing) across four text types (human-written, purely machine-generated, machine-rewritten, and machine-polished), generated using state-of-the-art LLMs for robust evaluation. Experimental results demonstrate that MSPO significantly outperforms existing methods across all text types. On the challenging rewritten texts, MSPO achieves up to 82.14% AUROC, representing an improvement of 11.15 percentage points over the strongest baseline ImBD, while maintaining robust cross-domain generalizability across scientific, news, and creative writing domains.

With the rapid advancement of large language models (LLMs), retrieval-augmented generation (RAG) has emerged as a critical approach to supplement the inherent knowledge limitations of LLMs. However, due to the typically large volume of retrieved information, RAG tends to operate with long context lengths. From the perspective of entropy engineering, we identify unconstrained entropy growth and attention dilution due to long retrieval context as significant factors affecting RAG performance. In this paper, we propose the balanced entropy-engineered RAG (BEE-RAG) framework, which improves the adaptability of RAG systems to varying context lengths through the principle of entropy invariance. By leveraging balanced context entropy to reformulate attention dynamics, BEE-RAG separates attention sensitivity from context length, ensuring a stable entropy level. Building upon this, we introduce a zero-shot inference strategy for multi-importance estimation and a parameter-efficient adaptive fine-tuning mechanism to obtain the optimal balancing factor for different settings. Extensive experiments across multiple RAG tasks demonstrate the effectiveness of BEE-RAG.

Extending pre-trained text Large Language Models (LLMs)’s speech understanding or generation abilities by introducing various effective speech tokens has attracted great attention in the speech research community. However, building a unified speech understanding and generation model still faces the following challenges: (1) Due to the huge modality gap between speech and text tokens, extending text LLMs to unified speech LLMs relies on large-scale paired data for fine-tuning, and (2) Generation and understanding tasks prefer information at different levels, e.g., generation benefits from detailed acoustic features, while understanding favors high-level semantics. This divergence leads to difficult performance optimization in one unified model. To solve these challenges, in this paper, we present two key insights in speech tokenization and speech language modeling. Specifically, we first propose an Understanding-driven Speech Tokenizer (USTokenizer), which extracts high-level semantic information essential for accomplishing understanding tasks using text LLMs. In this way, USToken enjoys better modality commonality with text, which reduces the difficulty of modality alignment in adapting text LLMs to speech LLMs. Secondly, we present DualSpeechLM, a dual-token modeling framework that concurrently models USToken as input and acoustic token as output within a unified, end-to-end framework, seamlessly integrating speech understanding and generation capabilities. Furthermore, we propose a novel semantic supervision loss and a Chain-of-Condition (CoC) strategy to stabilize model training and enhance speech generation performance. Experimental results demonstrate that our proposed approach effectively fosters a complementary relationship between understanding and generation tasks, highlighting the promising strategy of mutually enhancing both tasks in one unified model.

Automatic medical report generation can greatly reduce the workload of doctors, but it is often unreliable for real-world deployment. Current methods can write formally fluent sentences but may be factually flawed, introducing serious medical errors known as clinical hallucinations, which make them untrustworthy for diagnosis. To bridge this gap, we introduce HiMed-RL, a Hierarchical Medical Reward Learning Framework designed to explicitly prioritize clinical quality. HiMed-RL moves beyond simple text matching by deconstructing reward learning into three synergistic levels: it first ensures linguistic fluency at the token-level, then enforces factual grounding at the concept-level by aligning key medical terms with expert knowledge, and finally assesses high-level diagnostic consistency at the semantic-level using a specialized LLM verifier. This hierarchical reward is implemented via a Human-inspired Dynamic Reward Adjustment, a strategy which first teaches the model to learn basic facts before progressing to more complex diagnostic reasoning. Experimentally, HiMed-3B achieves state-of-the-art performance on both in-domain and out-of-domain benchmarks, particularly on the latter, with an improvement of 10.8% over the second-best baseline. Our work provides a robust paradigm for generating reports that not only improve fluency but clinical fine-grained quality.

Inference-time scaling has emerged as a powerful technique for enhancing the reasoning performance of Large Language Models (LLMs). However, existing approaches often rely on heuristic strategies for parallel sampling, lacking a principled foundation. To address this gap, we propose a probabilistic framework that formalizes the optimality of inference-time scaling under the assumption that parallel samples are independently and identically distributed (i.i.d.), and where the Best-of-N selection strategy follows a probability distribution that can be estimated. Within this framework, we derive a theoretical lower bound on the required number of samples to achieve a target performance level, providing the first principled guidance for compute-efficient scaling. Leveraging this insight, we develop OptScale, a practical algorithm that dynamically determines the optimal number of sampled responses. OptScale employs a language model-based predictor to estimate probabilistic prior parameters, enabling the decision of the minimal number of samples needed that satisfy predefined performance thresholds and confidence levels. Extensive experiments on representative reasoning benchmarks (including MATH-500, GSM8K, AIME, and AMC) demonstrate that OptScale significantly reduces sampling overhead while remaining better or on par with state-of-the-art reasoning performance. Our work offers both a theoretical foundation and a practical solution for principled inference-time scaling, addressing a critical gap in the efficient deployment of LLMs for complex reasoning.

Despite its success in enriching LLMs with external knowledge, RAG remains plagued by faithfulness hallucinations, where generated text contradicts the retrieved source information. Previous research on faithfulness hallucination in LLMs is frequently hindered by prohibitive manual annotation costs and a dependency on static datasets, which caps their performance and adaptability. Furthermore, these models lack a clear training mechanism to explicitly promote contextual focus. In this work, we propose a novel iterative self-evolution framework to enhance model faithfulness. This framework autonomously generates high-quality data and leverages it for the continuous self-optimization of the model, leading to significant improvements in faithfulness. Our experimental analysis reveals that improving model faithfulness encourages a closer alignment of the attention distribution with the given context. Based on this finding, we design an attention-based loss function to further promote this process. Experimental results show that our model achieves state-of-the-art faithfulness on a range of context-based question-answering datasets, marking a significant advancement over previous approaches.

Reasoning is a key component of language understanding in Large Language Models. While Chain-of-Thought prompting enhances performance via explicit intermediate steps, it suffers from sufficient token overhead and a fixed reasoning trajectory, preventing step-wise refinement. Recent advances in latent reasoning address these limitations by refining internal reasoning processes directly in the model’s latent space, without producing explicit outputs. However, a key challenge remains: how to effectively update reasoning embeddings during post-training to guide the model toward more accurate solutions. To overcome this challenge, we propose a lightweight post-training framework that refines latent reasoning trajectories using two novel strategies: (1) Contrastive reasoning feedback, which compares reasoning embeddings against strong and weak baselines to infer effective update directions via embedding enhancement; (2) Residual embedding refinement, which stabilizes updates by progressively integrating current and historical gradients, enabling fast yet controlled convergence. Extensive experiments and case studies are conducted on five reasoning benchmarks to demonstrate the effectiveness of the proposed framework. Notably, a +5% accuracy gain on MathQA without additional training.

In multi-hop question answering (MHQA) tasks, Chain of Thought (CoT) improves the quality of generation by guiding large language models (LLMs) through multi-step reasoning, and Knowledge Graphs (KGs) reduce hallucinations via semantic matching. Outcome Reward Models (ORMs) provide feedback after generating the final answers but fail to evaluate the process for multi-step reasoning. Traditional Process Reward Models (PRMs) evaluate the reasoning process but require costly human annotations or rollout generation. While implicit PRM is trained only with outcome signals and derives step rewards through reward parameterization without explicit annotations, it is more suitable for multi-step reasoning in MHQA tasks. However, existing implicit PRM has only been explored for plain text scenarios. When adapting to MHQA tasks, it cannot handle the graph structure constraints in KGs and capture the potential inconsistency between CoT and KG paths. To address these limitations, we propose the DPRM (Dual Implicit Process Reward Model). It trains two implicit PRMs for CoT and KG reasoning in MHQA tasks. Both PRMs, namely KG-PRM and CoT-PRM, derive step-level rewards from outcome signals via reward parameterization without additional explicit annotations. Among them, KG-PRM uses preference pairs to learn structural constraints from KGs. DPRM further introduces a consistency constraint between CoT and KG reasoning steps, making the two PRMs mutually verify and collaboratively optimize the reasoning paths. We also provide a theoretical demonstration of the derivation of process rewards. Experimental results show that our method outperforms 13 baselines on multiple datasets with up to 16.6% improvement on Hit@1.