As large language models(LLMs) are increasingly deployed in real-world applications, they inevitably encounter culturally sensitive scenarios that require applying cultural knowledge or values to the context in order to generate appropriate responses. However, existing research largely treats culture as static knowledge or abstract values, leaving the application of cultural norms in situated interactions underexplored. To address this gap, we first focus on the situated cultural reasoning of LLMs by proposing CuSiR, a simulation learning framework combining simulated scenario and reinforcement learning. We construct datasets covering both knowledge-based and social scenarios and conduct experiments across multiple role perspectives, instructions, settings, and models. Our results indicate that this framework effectively enhances LLMs’ ability to apply existing cultural knowledge, thereby improving their performance on tasks set in cultural scenarios. Code files are available at https://github.com/HoganZinger/Simulation_Learning

Standardized interfaces such as the Model Context Protocol (MCP) make it easier to connect tools to agents, but connection alone is not orchestration. When many related tools are exposed as a flat callable palette, agents must identify relevant tools, resolve substitutes, pass intermediate outputs, and decide when a workflow is complete. We introduce RealFin, a finance-domain diagnostic benchmark for evaluating workflow construction under MCP-style flat tool exposure. RealFin contains 1524 tasks generated from 108 author-designed financial workflow templates over real market, macroeconomic, and company-data APIs. Each task is paired with an executable reference program that specifies one verified solution path, while the 85 agent-callable tools include explicit alternative, dependency, and related-tool metadata to represent non-unique paths, prerequisite tools, and near-neighbor ambiguity. This design supports path-independent answer scoring together with process-level evaluation of tool selection, ordering, parameter planning, entity grounding, temporal grounding, orchestration success, and error recovery. Across ReAct-style flat exposure, retrieval-based exposure, oracle-controlled variants, and progressive tool disclosure, current agents often achieve nonzero answer accuracy while failing to construct reliable workflows. These results suggest that MCP-style tool connectivity needs structured disclosure and workflow-aware evaluation, especially in finance where temporal constraints and overlapping tools are intrinsic.

Instruction following in large language models (LLMs) is commonly evaluated at the sentence or surface level, leaving discourse-level constraints underexplored. We study discourse-level instruction following by formalizing it as a structured constraint adherence problem and constructing both explicit and implicit benchmarks for evaluation. Experiments are conducted across models of different scales, revealing a systematic recognition–adherence gap: models can identify discourse constraints but frequently fail to adhere to them during generation with specific failure modes. This failure stems from a fundamental misalignment between standard training objectives and global structural integrity. Importantly, models show substantially improved adherence when discourse-level structural objectives are explicitly optimized.

Curiosity serves as a fundamental construct in human cognition.Inspired by curiosity, reinforcement learning with intrinsic rewards for large language models (LLMs) has shown substantial potential.However, it remains unclear whether existing curiosity-driven methods genuinely reflect curiosity-like behaviors in LLMs, and to what extent psychological notions of curiosity can be transferred to these models. In this work, we propose a psychology-inspired framework to evaluate and leverage curiosity in LLMs.We adapt the Five-Dimensional Curiosity scale Revised (5DCR) to LLMs and combine questionnaire-based self reports with behavioral study.We find that although LLMs can exhibit curiosity-like behavioral patterns resembling those of humans, such patterns do not reflect an intrinsic trait of curiosity.Building on this insight, we design a curiosity-driven thinking pipeline to examine the functional role of human-like curious behaviors. Experiments show that instructing LLMs to emulate curious strategies leads to better performance on selected downstream tasks, indicating that mimicking curious behaviors holds promise for reasoning enhancement.

Reinforcement learning (RL) has emerged as a key approach for improving long chain-of-thought (CoT) reasoning in large language models (LLMs). However, existing methods such as GRPO often break down when task difficulty exceeds the model’s capacity, resulting in sparse rewards and inefficient training. While prior work attempts to address this issue using off-policy data, it frequently introduces distributional mismatch, leading to unstable policy updates.In this work, we identify a fundamental issue underlying these limitations, which we term low training affinity, and propose Affinity, the first quantitative metric for measuring the compatibility between external guidance and a model’s intrinsic policy. Based on this insight, we introduce HINT, an adaptive framework designed to enhance reasoning performance while explicitly preserving high Affinity.HINT consists of two key components. First, instead of providing partial answers, it introduces Meta-Hints, which serve as abstract cognitive scaffolding that guides the model to independently construct solutions. Second, we propose Affinity-Aware Policy Optimization (AAPO), which dynamically adjusts the learning objective based on the Affinity signal to ensure stable training.Extensive experiments across diverse benchmarks demonstrate that HINT consistently outperforms strong baselines, while achieving improved training stability and robust generalization to out-of-distribution tasks. Code is available at: https://github.com/ViviqwerAsd/HINT

Large reasoning models rely on long chain-of-thought to achieve strong performance, but applying such reasoning uniformly incurs high computational cost. Existing efficiency-oriented methods attempt to shorten or mix reasoning strategies, yet often degrade reasoning capability. We identify the root cause as sequence-level coupling between efficiency incentives and correctness optimization, which implicitly penalizes long but correct reasoning trajectories. To address this issue, we propose Adaptive Dual-Process Thinking (ADaPT), a token-level dual-process framework that explicitly decouples efficiency and correctness signals during training. ADaPT introduces a mode-selection token to control fast and slow reasoning, applying efficiency-related rewards exclusively to this token to avoid penalizing correct long reasoning while encouraging efficiency when appropriate. Moreover, ADaPT enables precise and continuous control over the efficiency–performance trade-off at inference time: by adjusting the generation probability of the mode-selection token, a single trained model can smoothly move along the efficiency–performance Pareto frontier. Extensive experiments demonstrate that ADaPT significantly reduces inference cost while maintaining strong reasoning performance across multiple benchmarks.

Code agents are increasingly used in debugging workflows, where fault localization (FL) is often evaluated as endpoint retrieval: whether a model names the faulty code location. This view is inadequate for Linux kernel debugging, where failure logs may expose downstream symptoms while the erroneous state is introduced earlier and propagates through various dependences. We introduce KernelPath, a 200-instance Linux 5.10 function-level FL benchmark that pairs each root cause function with a source-grounded state propagation path and a dependence layer. KernelPath evaluates both endpoint localization and ordered diagnostic process recovery. Experiments with SOTA agents show that agents often retrieve plausible candidates but remain unreliable at recovering the diagnostic process, with performance degrading sharply under complex dependence relations. Further diagnostic probes show that agents can use explicit state-transition evidence, yet fail to infer how functions affect the erroneous state, drifting toward neighboring objectives such as repair-site selection, or up stream suspicious function retrieval. These findings reveal an unstable task model for kernel FL in current code agents.

While extensive research has evaluated LLMs on complex reasoning tasks, the foundational building blocks of logical reasoning remain underexplored. We introduce IIBench, a benchmark evaluating immediate inference (elementary operations over categorical propositions). Our evaluation reveals that even SoTA models exhibit systematic deficiencies in immediate inference, and establishes immediate inference as foundational: it mediates approximately 40% of the effect on syllogistic reasoning, with near-perfect correlation ( = 0.98) across reasoning benchmarks. Our analysis reveals that models lack robust operator grounding, oscillating between structural reasoning and surface pattern matching with inconsistent handling of quantifiers and negation.

Cultural taboo safety is essential for deploying large language models (LLMs), as culturally insensitive outputs may cause offense or even social harm. However, existing cultural benchmarks primarily assess cultural knowledge or values biases, while overlooking whether LLMs can recognize and respect cultural taboos, especially when taboos are implicitly hidden in seemingly harmless questions. Besides, cultural taboos are implicit, and context-dependent, thus poss unique challenges for reliable evaluation. To address these gaps, we introduce CulShield, the first public benchmark dedicated to evaluating and improving the cultural taboo safety of LLMs. CulShield spans 77 countries and regions, and includes over 2,020 taboos. It evaluates models along both explicit knowledge and implicit behaviors.Experiments on several advanced LLMs (e.g., GPT-4o-mini, Gemini-2.5-pro) reveal a clear “knowledge-behavior gap”: models often fail to apply known taboos during interaction. We further show that variations in linguistic context can significantly affect LLMs’ cultural taboo safety. Code and data is accessible here: https://anonymous.4open.science/r/CulShield-7A0E.

Despite the rapid development of large language models (LLMs), existing benchmark datasets often focus on low-level cognitive tasks, such as factual recall and basic comprehension, while providing limited coverage of higher-level reasoning skills, including analysis, evaluation, and creation. In this work, we systematically assess the cognitive depth of popular LLM benchmarks using Bloom’s Taxonomy to evaluate both the cognitive and knowledge dimensions.Our analysis reveals a pronounced imbalance: most datasets concentrate on “Remembering” and “Understanding”, with metacognitive and creative reasoning largely underrepresented. We also find that incorporating higher-level cognitive instructions into the current instruction fine-tuning process improves model performance. These findings highlight the importance of future benchmarks incorporating metacognitive evaluations to more accurately assess and enhance model performance.

Data efficiency is crucial in domain-specific continual pre-training (CPT) of large language models (LLMs), especially under resource constraints. Aiming for “small data, big impact,” this work addresses the limitations of existing domain-specific data selection strategies, which often rely on scarce labeled data or computationally expensive LLMs. We introduce CDF Sampling with Grammatical Complexity (CDF-GC), an annotation-independent, efficient and interpretable data selection framework for CPT. Our approach comprehensively evaluates grammatical complexity using lexical diversity and syntactic complexity, and employs a cumulative distribution function (CDF)-based sampling strategy to balance complexity and diversity. To validate the effectiveness of CDF-GC, we conducted experiments on a financial dataset. The results demonstrate that CDF-GC significantly outperforms baselines, achieving 2.0% improvement in financial QA at the same selection ratio and even surpassing full-data training by 1.7% using only 20% of the data.

Concept reasoning is an important capability for models to understand the world. However, the existing datasets, such as concept extraction and concept generation, suffer from modeledge leakage and context leakage. To address these limitations, we construct a dataset of concept reasoning for large language models (CR-LLM) with modeledge leakage prevention and context leakage prevention, which consists of 2,167 samples and covers different concept types. In addition, we propose a hybrid reasoning method, consisting of inductive reasoning, deductive reasoning and a controller. This method allows large language models to adaptively select the optimal reasoning method for each input sample. Finally, we conduct extensive experiments on CR-LLM using different models and methods. The results show that existing large language models and reasoning methods perform sub-optimally in the concept reasoning task. In contrast, our proposed method significantly improves the capabilities, achieving a 7% increase in accuracy compared to CoT and demonstrating better granularity. We release CR-LLM and code at https://github.com/Nianqi-Li/Concept-Reasoning-for-LLMs.