Step-by-step reasoning has become a standard approach for large language models (LLMs) to tackle complex tasks. While this paradigm has proven effective, it raises a fundamental question: How can we verify that an LLM’s reasoning is accurately grounded in knowledge? To address this question, we introduce a novel evaluation suite that systematically assesses the knowledge grounding of intermediate reasoning. Our framework comprises three key components. (1) Principal Knowledge Collection, a large-scale repository of atomic knowledge essential for reasoning. Based on the collection, we propose (2) knowledge-grounded evaluation metrics designed to measure how well models recall and apply prerequisite knowledge in reasoning. These metrics are computed by our (3) evaluator LLM, a lightweight model optimized for cost-effective and reliable metric computation. Our evaluation suite demonstrates remarkable effectiveness in identifying missing or misapplied knowledge elements, providing crucial insights for uncovering fundamental reasoning deficiencies in LLMs. Beyond evaluation, we demonstrate how these metrics can be integrated into preference optimization, showcasing further applications of knowledge-grounded evaluation. Our evaluation suite is publicly available.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
TurnBench-MS: A Benchmark for Evaluating Multi-Turn, Multi-Step Reasoning in Large Language Models
PDF ↗Despite impressive advances in large language models (LLMs), existing benchmarks often focus on single-turn or single-step tasks, failing to capture the kind of iterative reasoning required in real-world settings. To address this limitation, we introduce **TurnBench**, a novel benchmark that evaluates multi-turn, multi-step reasoning through an interactive code-breaking task inspired by the “Turing Machine Board Game.” In each episode, a model must uncover hidden logical or arithmetic rules by making sequential guesses, receiving structured feedback, and integrating clues across multiple rounds. This dynamic setup requires models to reason over time, adapt based on past information, and maintain consistency across steps—capabilities underexplored in current benchmarks. TurnBench includes two modes: *Classic*, which tests standard reasoning, and *Nightmare*, which introduces increased complexity and requires robust inferential chains. To support fine-grained analysis, we provide ground-truth annotations for intermediate reasoning steps. Our evaluation of state-of-the-art LLMs reveals significant gaps: the best model achieves 84% accuracy in Classic mode, but performance drops to 18% in Nightmare mode. In contrast, human participants achieve 100% in both, underscoring the challenge TurnBench poses to current models. By incorporating feedback loops and hiding task rules, TurnBench reduces contamination risks and provides a rigorous testbed for diagnosing and advancing multi-step, multi-turn reasoning in LLMs.
Vision-Language Models (VLMs) have demonstrated remarkable capabilities in interpreting visual layouts and text. However, a significant challenge remains in their ability to interpret robustly and reason over multi-tabular data presented as images, a common occurrence in real-world scenarios like web pages and digital documents. Existing benchmarks typically address single tables or non-visual data (text/structured). This leaves a critical gap: they don’t assess the ability to parse diverse table images, correlate information across them, and perform multi-hop reasoning on the combined visual data. To bridge this evaluation gap, we introduce MTabVQA, a novel benchmark specifically designed for multi-tabular visual question answering. MTabVQA comprises 3,745 complex question-answer pairs that necessitate multi-hop reasoning across several visually rendered table images. We provide extensive benchmark results for state-of-the-art VLMs on MTabVQA, revealing significant performance limitations. We further investigate post-training techniques to enhance these reasoning abilities and release MTabVQA-Instruct, a large-scale instruction-tuning dataset. Our experiments show that fine-tuning VLMs with MTabVQA-Instruct substantially improves their performance on visual multi-tabular reasoning. Code and dataset are available online: .
Recent advancements in Large Language Models (LLMs) have transformed natural language understanding and generation, leading to extensive benchmarking across diverse tasks. However, cryptanalysis—a critical area for data security and its connection to LLMs’ generalization abilities remains underexplored in LLM evaluations. To address this gap, we evaluate the cryptanalytic potential of state‐of‐the‐art LLMs on ciphertexts produced by a range of cryptographic algorithms. We introduce a benchmark dataset of diverse plaintexts—spanning multiple domains, lengths, writing styles, and topics—paired with their encrypted versions. Using zero‐shot and few‐shot settings along with chain‐of‐thought prompting, we assess LLMs’ decryption success rate and discuss their comprehension abilities. Our findings reveal key insights into LLMs’ strengths and limitations in side‐channel scenarios and raise concerns about their susceptibility to under-generalization related attacks. This research highlights the dual‐use nature of LLMs in security contexts and contributes to the ongoing discussion on AI safety and security.
Large Vision-Language Models (LVLMs) demonstrate strong visual question answering (VQA) capabilities but are shown to hallucinate. A reliable model should perceive its knowledge boundaries—knowing what it knows and what it does not. This paper investigates LVLMs’ perception of their knowledge boundaries by evaluating three types of confidence signals: probabilistic confidence, answer consistency-based confidence, and verbalized confidence. Experiments on three LVLMs across three VQA datasets show that, although LVLMs possess a reasonable perception level, there is substantial room for improvement. Among the three confidence, probabilistic and consistency-based signals are more reliable indicators, while verbalized confidence often leads to overconfidence. To enhance LVLMs’ perception, we adapt several established confidence calibration methods from Large Language Models (LLMs) and propose three effective methods. Additionally, we compare LVLMs with their LLM counterparts, finding that jointly processing visual and textual inputs decreases question-answering performance but reduces confidence, resulting in improved perception level compared to LLMs.
Persona-prompting is a growing strategy to steer LLMs toward simulating particular perspectives or linguistic styles through the lens of a specified identity. While this method is often used to personalize outputs, its impact on how LLMs represent social groups remains underexplored. In this paper, we investigate whether persona-prompting leads to different levels of linguistic abstraction—an established marker of stereotyping—when generating short texts linking socio-demographic categories with stereotypical or non-stereotypical attributes. Drawing on the Linguistic Expectancy Bias framework, we analyze outputs from six open-weight LLMs under three prompting conditions, comparing 11 persona-driven responses to those of a generic AI assistant. To support this analysis, we introduce Self-Stereo, a new dataset of self-reported stereotypes from Reddit. We measure abstraction through three metrics: concreteness, specificity, and negation. Our results highlight the limits of persona-prompting in modulating abstraction in language, confirming criticisms about the ecology of personas as representative of socio-demographic groups and raising concerns about the risk of propagating stereotypes even when seemingly evoking the voice of a marginalized groups.
MultiLingPoT: Boosting Mathematical Reasoning in LLMs through Multilingual Program Integration
PDF ↗Program-of-Thought, which aims to use program instead of natural language in reasoning, is an important way for LLMs to solve mathematical problems. Since different programming languages excel in different areas, it is natural to use the most suitable language for solving specific problems. However, current research only focuses on single language PoT, ignoring the differences between programming languages. Therefore, this paper proposes a multilingual programme reasoning method, MultiLingPoT, and deeply explores the impact of multilingual integration in the training and inference. This method allows the model to answer questions using multiple languages by fine-tuning on multilingual data and improving individual language’s reasoning accuracy by 2.5%. Additionally, prior and posterior selection methods are used to help the model select the most suitable language during inference, and achieves 8% performance gains. Finally, our code metric analysis shows that language differences manifest in encapsulation levels and implementation granularity, while strategic deviation from language conventions can enhances code performance.
Large language models (LLMs) have shown remarkable capabilities across various tasks, that are learned from massive amounts of text-based data. Although LLMs can control output sequence length, particularly in instruction-based settings, the internal mechanisms behind this control have been unexplored yet. In this study, we provide empirical evidence on how output sequence length information is encoded within the internal representations in LLMs. In particular, our findings show that multi-head attention mechanisms are critical in determining output sequence length, which can be adjusted in a disentangled manner. By scaling specific hidden units within the model, we can control the output sequence length without losing the informativeness of the generated text, thereby indicating that length information is partially disentangled from semantic information. Moreover, some hidden units become increasingly active as prompts become more length-specific, thus reflecting the model’s internal awareness of this attribute. Our findings suggest that LLMs have learned robust and adaptable internal mechanisms for controlling output length without any external control.
Voice Conversion research in recent times has increasingly focused on improving the zero-shot capabilities of existing methods. Despite remarkable advancements, current architectures still tend to struggle in zero-shot cross-lingual settings. They are also often unable to generalize for speakers of unseen languages and accents. In this paper, we adopt a simple yet effective approach that combines discrete speech representations from self-supervised models with a non-autoregressive Diffusion-Transformer based conditional flow matching speech decoder. We show that this architecture allows us to train a voice-conversion model in a purely textless, self-supervised fashion. Our technique works without requiring multiple encoders to disentangle speech features. Our model also manages to excel in zero-shot cross-lingual settings even for unseen languages. We provide our code, model checkpoint and demo samples here: https://github.com/ez-vc/ez-vc
Token Preference Optimization with Self-Calibrated Visual-Anchored Rewards for Hallucination Mitigation
PDF ↗Direct Preference Optimization (DPO) has been demonstrated to be highly effective in mitigating hallucinations in Large Vision Language Models (LVLMs) by aligning their outputs more closely with human preferences. Despite the recent progress, existing methods suffer from two drawbacks: 1) Lack of scalable token-level rewards; and 2) Neglect of visual-anchored tokens. To this end, we propose a novel Token Preference Optimization model with self-calibrated rewards (dubbed as TPO), which adaptively attends to visual correlated tokens without fine-grained annotations. Specifically, we introduce a token-level visual-anchored reward as the difference of the logistic distributions of generated tokens conditioned on the raw image and the corrupted one. In addition, to highlight the informative visual-anchored tokens, a visual-aware training objective is proposed to enhance more accurate token-level optimization. Extensive experimental results have manifested the state-of-the-art performance of the proposed TPO. For example, by building on top of LLaVA and Qwen, our TPO boosts the performance absolute improvement for hallucination benchmarks.
Inclusive Leadership in the Age of AI: A Dataset and Comparative Study of LLMs vs. Real-Life Leaders in Workplace Action Planning
PDF ↗Generative Large Language Models have emerged as useful tools, reshaping professional workflows. However, their efficacy in inherently complex and human-centric tasks such as leadership and strategic planning remains underexplored. In this interdisciplinary study, we present a novel dataset and compare LLMs and human leaders in the context of workplace action planning, specifically focusing on translating the abstract idea of inclusion into actionable SMART goals. We developed the Leader Success Bot, a script-based chatbot co-designed with domain experts, to guide more than 250 real-life leaders in generating inclusive workplace action plans. We systematically prompted seven state-of-the-art chat-based LLMs to perform the same task using the socio-demographic data of real-life leaders and instructions co-developed with domain experts. Our publicly released dataset enables direct comparison between human and LLM-generated workplace action plans, offering insights into their respective strengths, biases, and limitations. Our findings highlight critical gaps and opportunities for LLMs in leadership applications, fostering interdisciplinary collaboration and NLP applications.
Large Language Models (LLMs) are pivotal in enabling intelligent experiences across various applications, from summarization to advanced content organization and retrieval functionalities. However, deploying LLMs for diverse tasks is fundamentally constrained by memory and compute limitations, making it impractical to fine-tune separate models for each task. Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) offer a scalable solution for multi-task LLM deployment. Despite its potential, LoRA faces challenges in selecting optimal ranks and layers for each task-model pair, often resulting in inefficiencies and unnecessary parameters. We introduce Norm Adaptive Localized (NormAL) LoRA, a novel variant that employs rank-norm regularization to dynamically determine the optimal rank for each weight matrix, ensuring adaptation is concentrated where it is most impactful. Our approach reduces adapter parameters by 37% while preserving full fine-tuning performance, making NormAL LoRA a transformative tool for enabling efficient, scalable, and space-constrained AI deployments across diverse industries and applications.
LLM Agents at the Roundtable: A Multi-Perspective and Dialectical Reasoning Framework for Essay Scoring
PDF ↗The emergence of large language models (LLMs) has brought a new paradigm to automated essay scoring (AES), a long-standing and practical application of natural language processing in education. However, achieving human-level multi-perspective understanding and judgment remains a challenge. In this work, we propose Roundtable Essay Scoring (RES), a multi-agent evaluation framework designed to perform precise and human-aligned scoring under a zero-shot setting. RES constructs evaluator agents based on LLMs, each tailored to a specific prompt and topic context. Each agent independently generates a trait-based rubric and conducts a multi-perspective evaluation. Then, by simulating a roundtable-style discussion, RES consolidates individual evaluations through a dialectical reasoning process to produce a final holistic score that more closely aligns with human evaluation. By enabling collaboration and consensus among agents with diverse evaluation perspectives, RES outperforms prior zero-shot AES approaches. Experiments on the ASAP dataset using ChatGPT and Claude show that RES achieves up to a 34.86% improvement in average QWK over straightforward prompting (Vanilla) methods.
We introduce Spire, a speech-augmented language model (LM) capable of both translating and transcribing speech input from English into 10 other languages as well as translating text input in both language directions. Spire integrates the speech modality into an existing multilingual LM via speech discretization and continued pre-training using only 42.5 K hours of speech. In particular, we adopt the pretraining framework of multilingual LMs and treat discretized speech input as an additional translation language. This approach not only equips the model with speech capabilities, but also preserves its strong text-based performance. We achieve this using significantly less data than existing speech LMs, demonstrating that discretized speech input integration as an additional language is feasible during LM adaptation. We make our code and models available to the community.
Recent advances in large language model (LLM) fine‐tuning have shown that training data augmented with high-quality reasoning traces can remarkably improve downstream performance. However, existing approaches usually rely on expensive manual annotations or auxiliary models, and fail to address the unique constraints of smaller “weak” LLMs. To bridge these gaps, we introduce Weak2Wise, a fully automated, lightweight framework for synthesizing high‐quality, weak-LLM-friendly reasoning traces. Starting from a QA dataset, Weak2Wise filters out the samples that can already be correctly answered by the weak LLM, gathers diverse candidate reasoning traces from multiple strong LLMs, and leverages our Step‐Mask scoring to rank and truncate the most guidance‐effective traces. These reasoning traces are then used for fine‐tuning, yielding substantial improvements in the weak LLM’s reasoning abilities. The name Weak2Wise has two meanings: using a “weak” LLM to select the “wisest” reasoning traces generated by stronger LLMs, and fine‐tuning the same weak LLM on these reasoning traces to become “wiser”. We further use Weak2Wise to build GR-1K, a 1,000‐sample math and science QA‐reasoning dataset optimized for weak LLMs, and fine‐tune Qwen2.5‐7B on it to create GR‐7B, which achieves superior performance on AIME2024, MATH‐500, and GPQA Diamond benchmarks. Our codes are publicly released to facilitate further research.
Reasoning ability of large language models (LLMs) is a crucial ability,especially in complex decision-making tasks. One significant task to show LLMs’reasoning capability is code time complexity prediction, which involves variousintricate factors such as the input range of variables and conditional loops.Current benchmarks fall short of providing a rigorous assessment due to limiteddata, language constraints, and insufficient labeling. They do not consider timecomplexity based on input representation and merely evaluate whether predictionsfall into the same class, lacking a measure of how close incorrect predictionsare to the correct ones.To address these dependencies, we introduce CodeComplex, the first robust andextensive dataset designed to evaluate LLMs’ reasoning abilities in predictingcode time complexity. CodeComplex comprises 4,900 Java codes and an equivalentnumber of Python codes, overcoming language and labeling constraints, carefullyannotated with complexity labels based on input characteristics by a panel ofalgorithmic experts. Additionally, we propose specialized evaluation metrics forthe reasoning of complexity prediction tasks, offering a more precise andreliable assessment of LLMs’ reasoning capabilities. We release our dataset andbaseline models publicly to encourage the relevant (NLP, SE, and PL) communitiesto utilize and participate in this research. Our code and data are available athttps://github.com/sybaik1/CodeComplex.
We study how large language models (LLMs) reason about memorized knowledge through simple binary relations such as equality (=), inequality (<), and inclusion (\subset). Unlike in-context reasoning, the axioms (e.g., a < b, b < c) are only seen during training and not provided in the task prompt (e.g., evaluating a < c). The tasks require one or more reasoning steps, and data aggregation from one or more sources, showing performance change with task complexity. We introduce a lightweight technique, out-of-context representation learning, which trains only new token embeddings on axioms and evaluates them on unseen tasks. Across reflexivity, symmetry, and transitivity tests, LLMs mostly perform statistically significant better than chance, making the correct answer extractable when testing multiple phrasing variations, but still fall short of consistent reasoning on every single query. Analysis shows that the learned embeddings are organized in structured ways, suggesting real relational understanding. Surprisingly, it also indicates that the core reasoning happens during the training, not inference.
Stop Playing the Guessing Game! Evaluating Conversational Recommender Systems via Target-free User Simulation
PDF ↗Recent developments in Conversational Recommender Systems (CRSs) have focused on simulating real-world interactions between users and CRSs to create more realistic evaluation environments. Despite considerable advancements, reliably assessing the capability of CRSs in eliciting user preferences remains a significant challenge. We observe that user-CRS interactions in existing evaluation protocols resemble a guessing game, as they construct target-biased simulators pre-encoded with target item knowledge, thereby allowing the CRS to shortcut the elicitation process. Moreover, we reveal that current evaluation metrics, which predominantly emphasize single-turn recall of target items, suffer from target ambiguity in multi-turn settings and overlook the intermediate process of preference elicitation. To address these issues, we introduce PEPPER, a novel CRS evaluation protocol with target-free user simulators that enable users to gradually discover their preferences through enriched interactions, along with detailed measures for comprehensively assessing the preference elicitation capabilities of CRSs. Through extensive experiments, we validate PEPPER as a reliable simulation environment and offer a thorough analysis of how effectively current CRSs perform in preference elicitation and recommendation.
ALRPHFS: Adversarially Learned Risk Patterns with Hierarchical Fast & Slow Reasoning for Robust Agent Defense
PDF ↗LLM Agents are becoming central to intelligent systems. However, their deployment raises serious safety concerns. Existing defenses largely rely on “Safety Checks”, which struggle to capture the complex semantic risks posed by harmful user inputs or unsafe agent behaviors—creating a significant semantic gap between safety checks and real-world risks. To bridge this gap, we propose a novel defense framework, ALRPHFS (Adversarially Learned Risk Patterns with Hierarchical Fast & Slow Reasoning). ALRPHFS consists of two core components: (1) an offline adversarial self-learning loop to iteratively refine a generalizable and balanced library of risk patterns, substantially enhancing robustness without retraining the base LLM, and (2) an online hierarchical fast & slow reasoning engine that balances detection effectiveness with computational efficiency. Experimental results demonstrate that our approach achieves superior overall performance compared to existing baselines, achieving a best‐in‐class average accuracy of 80% and exhibiting strong generalizability across agents and tasks.