Step-by-step reasoning is crucial for solving complex visual tasks, yet existing approaches lack a comprehensive framework for evaluating this capability and do not emphasize step-wise problem-solving. To this end, we propose a comprehensive framework for advancing multi-step visual reasoning in large multimodal models (LMMs) through three key contributions. First, we introduce a Visual Reasoning Chain Benchmark, the most comprehensive benchmark for multi-step visual reasoning, covering eight diverse categories and over 4k reasoning steps. This enables rigorous evaluation of LMMs’ ability to reason accurately and interpretably across multiple steps. Second, we propose a fine-grained reasoning metric that evaluates correctness and logical coherence at each step, providing deeper insights beyond traditional accuracy metrics. Third, we introduce LlamaV-o1, a state-of-the-art multimodal reasoning model trained using a multi-step curriculum learning approach. LlamaV-o1 is optimized for structured, step-by-step reasoning and significantly outperforms existing open-source models. It surpasses Llava-CoT with a 3.8% absolute gain across six benchmarks, achieving an average score of 67.3 while being 5x faster during inference scaling. Our benchmark, model, and code is available at https://github.com/mbzuai-oryx/LlamaV-o1.
论文检索
输入标题、作者或关键词，从 3,086 篇学术成果中精准定位
Missing the Margins: A Systematic Literature Review on the Demographic Representativeness of LLMs
PDF ↗Many applications of Large Language Models (LLMs) require them to either simulate people or offer personalized functionality, making the demographic representativeness of LLMs crucial for equitable utility. At the same time, we know little about the extent to which these models actually reflect the demographic attributes and behaviors of certain groups or populations, with conflicting findings in empirical research. To shed light on this debate, we review 211 papers on the demographic representativeness of LLMs. We find that while 29% of the studies report positive conclusions on the representativeness of LLMs, 30% of these do not evaluate LLMs across multiple demographic categories or within demographic subcategories. Another 35% and 47% of the papers concluding positively fail to specify these subcategories altogether for gender and race, respectively. Of the articles that do report subcategories, fewer than half include marginalized groups in their study. Finally, more than a third of the papers do not define the target population to whom their findings apply; of those that do define it either implicitly or explicitly, a large majority study only the U.S. Taken together, our findings suggest an inflated perception of LLM representativeness in the broader community. We recommend more precise evaluation methods and comprehensive documentation of demographic attributes to ensure the responsible use of LLMs for social applications.
Information Extraction (IE) systems are traditionally domain-specific, requiring costlyadaptation that involves expert schema design,data annotation, and model training. WhileLarge Language Models have shown promisein zero-shot IE, performance degrades significantly in unseen domains where label definitions differ. This paper introduces GUIDEX,a novel method that automatically definesdomain-specific schemas, infers guidelines,and generates synthetically labeled instances,allowing for better out-of-domain generalization. Fine-tuning Llama 3.1 with GUIDEXsets a new state-of-the-art across seven zeroshot Named Entity Recognition benchmarks.Models trained with GUIDEX gain up to 7 F1points over previous methods without humanlabeled data, and nearly 2 F1 points higherwhen combined with it. Models trained onGUIDEX demonstrate enhanced comprehension of complex, domain-specific annotationschemas. Code, models, and synthetic datasetsare available at neilus03.github.io/guidex.com
The dissemination of Large Language Models (LLMs), trained at scale, and endowed with powerful text-generating abilities, has made it easier for all to produce harmful, toxic, faked or forged content. In response, various proposals have been made to automatically discriminate artificially generated from human-written texts, typically framing the problem as a binary classification problem. Early approaches evaluate an input document with a well-chosen detector LLM, assuming that low-perplexity scores reliably signal machine-made content. More recent systems instead consider two LLMs and compare their probability distributions over the document to further discriminate when perplexity alone cannot. However, using a fixed pair of models can induce brittleness in performance. We extend these approaches to the ensembling of several LLMs and derive a new, theoretically grounded approach to combine their respective strengths. Our experiments, using a variety of generator LLMs, suggest that this approach effectively harnesses each model’s capabilities, leading to strong detection performance on a variety of domains.
We introduce the concept of context-driven over-refusal, an abstention arising when model’s safety guardrails are triggered by the grounding knowledge provided alongside the user’s request. Distinct from question-driven over-refusal, this occurs in both retrieval-augmented generation (RAG) and natural language processing (NLP) task completion (e.g. summarization, translation) where external content can unexpectedly trigger refusals. In this work, we present a novel two-stage evaluation framework named COVER, designed to quantify and analyze this behavior. Through a comprehensive empirical study on two public corpora, we show that over-refusal rates strongly depend on the task, system prompts, model family, and the number of retrieved documents. We observe that tasks such as translation and summarization yield disproportionately high over-refusal rates, while question-answering remains relatively robust, especially in newer models. Moreover, increasing the number of contextual documents tends to reduce refusals, yet broadens the pool of prompts at risk of encountering at least one “unsafe” text. Interestingly, strict system prompts do not necessarily lead to higher over-refusal rates, suggesting that in the absence of explicit directives, some models may default to a more cautious behavior. These findings highlight the need for fine-grained alignment and benchmarking strategies sensitive to both user intent and contextual nuances, offering a roadmap for future research in model training and evaluation.
Large language models (LLMs) are arguably the most predictive models of human cognition available. Despite their impressive human-alignment, LLMs are often labeled as "*just* next-token predictors” that purportedly fall short of genuine cognition. We argue that these deflationary claims need further justification. Drawing on prominent cognitive and artificial intelligence research, we critically evaluate two forms of “Justaism” that dismiss LLM cognition by labeling LLMs as “just” simplistic entities without specifying or substantiating the critical capacities these models supposedly lack. Our analysis highlights the need for a more measured discussion of LLM cognition, to better inform future research and the development of artificial intelligence.
Parameter-efficient fine-tuning (PEFT) methods typically assume that Large Language Models (LLMs) are trained on data from a single device or client. However, real-world scenarios often require fine-tuning these models on private data distributed across multiple devices. Federated Learning (FL) offers an appealing solution by preserving user privacy, as sensitive data remains on local devices during training. Nonetheless, integrating PEFT methods into FL introduces two main challenges: communication overhead and data heterogeneity. In this paper, we introduce FedTT and FedTT+, methods for adapting LLMs by integrating tensorized adapters into client-side models’ encoder/decoder blocks. FedTT is versatile and can be applied to both cross-silo FL and large-scale cross-device FL. FedTT+, an extension of FedTT tailored for cross-silo FL, enhances robustness against data heterogeneity by adaptively freezing portions of tensor factors, further reducing the number of trainable parameters. Experiments on BERT and LLaMA models demonstrate that our proposed methods successfully address data heterogeneity challenges and perform on par or even better than existing federated PEFT approaches while achieving up to 10\times reduction in communication cost.
Seeing What Tastes Good: Revisiting Multimodal Distributional Semantics in the Billion Parameter Era
PDF ↗Human learning and conceptual representation is grounded in sensorimotor experience, in contrast to state-of-the-art foundation models. In this paper, we investigate how well such large-scale models, trained on vast quantities of data, represent the semantic feature norms of concrete object concepts, e.g. a ROSE is red, smells sweet, and is a flower. More specifically, we use probing tasks to test which properties of objects these models are aware of. We evaluate image encoders trained on image data alone, as well as multimodally-trained image encoders and language-only models, on predicting an extended denser version of the classic McRae norms and the newer Binder dataset of attribute ratings. We find that multimodal image encoders slightly outperform language-only approaches, and that image-only encoders perform comparably to the language models, even on non-visual attributes that are classified as “encyclopedic” or “function”. These results offer new insights into what can be learned from pure unimodal learning, and the complementarity of the modalities.
Large language models (LLMs) excel in natural language generation but also exhibit biases, particularly in gender, race, and religion, which can be amplified with widespread use. However, research on biases in specific domains, such as finance, remains limited. To address this gap, we conducted a comprehensive evaluation of 23 leading LLMs and found varying degrees of financial bias, including more pronounced biases in financial-specific LLMs (FinLLMs). In response, we propose the Financial Bias Indicators (FBI) framework, which includes components like the Bias Unveiler, Bias Detective, Bias Tracker, and Bias Antidote, designed to identify, detect, analyze, and mitigate financial biases. Our analysis explores the root causes of these biases and introduces a debiasing method based on financial causal knowledge, alongside three other debiasing techniques. For the most biased model, we successfully reduced bias by 68% according to key metrics. This study advances our understanding of LLM biases in finance and highlights the need for greater scrutiny in their application within this critical domain.
Multimodal multi-label emotion recognition (MMER) aims to identify the concurrent presence of multiple emotions in multimodal data. Existing studies primarily focus on improving fusion strategies and modeling modality-to-label dependencies. However, they often overlook the impact of aleatoric uncertainty, which is the inherent noise in the multimodal data and hinders the effectiveness of modality fusion by introducing ambiguity into feature representations.To address this issue and effectively model aleatoric uncertainty, this paper proposes Latent emotional Distribution Decomposition with Uncertainty perception (LDDU) framework from a novel perspective of latent emotional space probabilistic modeling. Specifically, we introduce a contrastive disentangled distribution mechanism within the emotion space to model the multimodal data, allowing for the extraction of semantic features and uncertainty. Furthermore, we design an uncertainty-aware fusion multimodal method that accounts for the dispersed distribution of uncertainty and integrates distribution information. Experimental results show that LDDU achieves state-of-the-art performance on the CMU-MOSEI and M^3ED datasets, highlighting the importance of uncertainty modeling in MMER. Code is available at https://github.com/201983290498/lddu_mmer.git.
Reinforcement Learning from Human Feedback (RLHF) has become an essential technique for enhancing pre-trained large language models (LLMs) to generate responses that align with human preferences and societal values. Although RLHF has shown promise, the training of reward models (RMs) still faces the challenge of reward hacking, motivating recent works to prevent RMs from finding shortcuts that bypass the intended optimization objectives by identifying simplistic patterns such as response length. Besides the issue of length bias, our work firstly reveals that prompt-template bias learned by RMs can also cause reward hacking when dealing with some marginal samples, resulting in LLMs preferring to generate responses in a specific format after RLHF fine-tuning, regardless of the format requested in the prompt. To this end, we propose a low-cost but effective method, namely Prompt Bias Calibration (PBC), to estimate the prompt-template bias term during reward modeling, which can be utilized to calibrate reward scores in the following RL fine-tuning process. Then, we show that our PBC method can be flexibly combined with existing algorithms of removing length bias, leading to a further improvement in the aspect of enhancing the quality of generated responses.
Are Multimodal Large Language Models Pragmatically Competent Listeners in Simple Reference Resolution Tasks?
PDF ↗We investigate the linguistic abilities of multimodal large language models in reference resolution tasks featuring simple yet abstract visual stimuli, such as color patches and color grids. Although the task may not seem challenging for today’s language models, being straightforward for human dyads, we consider it to be a highly relevant probe of the pragmatic capabilities of MLLMs. Our results and analyses indeed suggest that basic pragmatic capabilities, such as context-dependent interpretation of color descriptions, still constitute major challenges for state-of-the-art MLLMs.
KaFT: Knowledge-aware Fine-tuning for Boosting LLMs’ Domain-specific Question-Answering Performance
PDF ↗Supervised fine-tuning (SFT) is a common approach to improve the domain-specific question-answering (QA) performance of large language models (LLMs). However, recent literature reveals that due to the conflicts between LLMs’ internal knowledge and the context knowledge of training data, vanilla SFT using the full QA training set is usually suboptimal. In this paper, we first design a query diversification strategy for robust conflict detection and then conduct a series of experiments to analyze the impact of knowledge conflict. We find that 1) training samples with varied conflicts contribute differently, where SFT on the data with large conflicts leads to catastrophic performance drops; 2) compared to directly filtering out the conflict data, appropriately applying the conflict data would be more beneficial. Motivated by this, we propose a simple-yet-effective Knowledge-aware Fine-tuning (namely KaFT) approach to effectively boost LLMs’ performance. The core of KaFT is to adapt the training weight by assigning different rewards for different training samples according to conflict level. Extensive experiments show that KaFT brings consistent and significant improvements (up to +5.73% average scores) across four LLMs. More analyses prove that KaFT effectively improves the model generalization and alleviates the hallucination.
Large language models (LLMs) have achieved remarkable performance on knowledge graph question answering (KGQA) tasks by planning and interacting with knowledge graphs. However, existing methods often confuse tool utilization with knowledge reasoning, harming readability of model outputs and giving rise to hallucinatory tool invocations, which hinder the advancement of KGQA. To address this issue, we propose Memory-augmented Query Reconstruction for LLM-based Knowledge Graph Reasoning (MemQ) to decouple LLM from tool invocation tasks using LLM-built query memory. By establishing a memory module with explicit descriptions of query statements, the proposed MemQ facilitates the KGQA process with natural language reasoning and memory-augmented query reconstruction. Meanwhile, we design an effective and readable reasoning to enhance the LLM’s reasoning capability in KGQA. Experimental results that MemQ achieves state-of-the-art performance on widely used benchmarks WebQSP and CWQ.
RedundancyLens: Revealing and Exploiting Visual Token Processing Redundancy for Efficient Decoder-Only MLLMs
PDF ↗Current Multimodal Large Language Model (MLLM) architectures face a critical tradeoff between performance and efficiency: decoder-only architectures achieve higher performance but lower efficiency, while cross-attention-based architectures offer greater efficiency but lower performance. The key distinction lies in how visual tokens are processed. Decoder-only architectures apply self-attention and FFN operations on visual tokens, while cross-attention architectures skip these computations. To investigate whether redundancy exists in this computationally expensive process, we propose a training-free framework for analyzing trained MLLMs. It consists of Probe-Activated Dynamic FFN and Hollow Attention, which enable adjustable reductions in computations for visual tokens, as well as a Layer Ranking Algorithm that prioritizes layers for these reductions. Extensive experiments demonstrate substantial, structured, and clustered redundancy unique to decoder-only MLLMs, offering valuable insights for future MLLM architecture design. Furthermore, by leveraging our reduction framework as a training-free inference acceleration approach, we achieve performance comparable to or better than state-of-the-art methods while remaining compatible with them. Code is available at https://github.com/L-Hugh/RedundancyLens.
Instruction tuning is vital for enhancing the performance of large language models (LLMs), but existing text-to-text methods, referred to as TextTuning, struggle with issues such as generalization, robustness, and controllability due to their lack of explicit task structures. We introduce JsonTuning, a structure-to-structure approach that uses JSON structures to represent tasks. This method improves generalization by clarifying task elements and their relations, boosts robustness by minimizing ambiguity, and enhances controllability by allowing precise control over outputs. We conduct an extensive comparative analysis between JsonTuning and TextTuning using various language models and benchmarks. Our findings reveal that JsonTuning consistently surpasses TextTuning in terms of performance, robustness, and controllability across different scenarios. By overcoming the limitations of TextTuning, JsonTuning demonstrates significant potential for developing more effective and reliable LLMs capable of handling diverse scenarios.
This paper studies the problem of unsupervised time series representation learning, which aims to map unlabeled time series data into a low-dimensional latent space for various downstream tasks. Previous works usually combine a range of augmentation strategies with contrastive learning to generate discriminative representations. However, these augmentation strategies could alter the original semantics of time series data, which could degrade the performance of representation learning. To solve this problem, this paper incorporates the large language model (LLM) agent to guide unsupervised time series representation learning and proposes a novel framework named Multi-Agent Collaboration for Time-series Representation Learning (MERIT). The core of our MERIT is to utilize three LLM agents to collaboratively generate positive views for time series data. In particular, we first design a retrieval agent to automatically identify the relevant time series data from a coarse candidate set. Then, these selected sequences are further utilized to enhance an augmentation agent which automatically selects reliable augmentation strategies from an augmentation strategy library. We also design a review agent to evaluate the quality of generated views and stop the generation process. These three agents are designed to work in a loop for effective time series representation learning. Extensive experiments on multiple time series datasets demonstrate the effectiveness of our MERIT in comparison with state-of-the-art baselines.
Analogical reasoning is a unique ability of humans to address unfamiliar challenges by transferring strategies from relevant past experiences. One key finding in psychology is that compared with irrelevant past experiences, recalling relevant ones can help humans better handle new tasks. Coincidentally, the NLP community has also recently found that self-generating relevant examples in the context can help large language models (LLMs) better solve a given problem than hand-crafted prompts. However, it is yet not clear whether relevance is the key factor eliciting such capability, i.e., can LLMs benefit more from self-generated relevant examples than irrelevant ones? In this work, we systematically explore whether LLMs can truly perform analogical reasoning on a diverse set of reasoning tasks. With extensive experiments and analysis, we show that self-generated random examples can surprisingly achieve comparable or even better performance on certain tasks, e.g., 4% performance boost on GSM8K with random biological examples. We find that the accuracy of self-generated examples is the key factor and subsequently design two novel methods with improved performance and significantly reduced inference costs. Overall, we aim to advance a deeper understanding of LLM analogical reasoning and hope this work stimulates further research in the design of self-generated contexts.
Research on reference and naming suggests that humans can come up with very different ways of conceptualizing and referring to the same object, e.g. the same abstract tangram shape can be a “crab”, “sink” or “space ship”. Another common assumption in cognitive science is that scene context fundamentally shapes our visual perception of objects and conceptual expectations. This paper contributes SceneGram, a dataset of human references to tangram shapes placed in different scene contexts, allowing for systematic analyses of the effect of scene context on conceptualization. Based on this data, we analyze references to tangram shapes generated by multimodal LLMs, showing that these models do not account for the richness and variability of conceptualizations found in human references.
Enhancing the fine-grained instance spatiotemporal motion perception capabilities of Video Large Language Models is crucial for improving their temporal and general video understanding. However, current models struggle to perceive detailed and complex instance motions. To address these challenges, we have made improvements from both data and model perspectives. In terms of data, we have meticulously curated iMOVE-IT, the first large-scale instance-motion-aware video instruction-tuning dataset. This dataset is enriched with comprehensive instance motion annotations and spatiotemporal mutual-supervision tasks, providing extensive training for the model’s instance-motion-awareness. Building on this foundation, we introduce iMOVE, an instance-motion-aware video foundation model that utilizes Event-aware Spatiotemporal Efficient Modeling to retain informative instance spatiotemporal motion details while maintaining computational efficiency. It also incorporates Relative Spatiotemporal Position Tokens to ensure awareness of instance spatiotemporal positions. Evaluations indicate that iMOVE excels not only in video temporal understanding and general video understanding but also demonstrates significant advantages in long-term video understanding. We will release the data, code, and model weights after acceptance.