Fengyuan Liu

Tianping Zhang, Zheyu Zhang, Zhiyuan Fan et al.

The goal of automated feature generation is to liberate machine learning experts from the laborious task of manual feature generation, which is crucial for improving the learning performance of tabular data. The major challenge in automated feature generation is to efficiently and accurately identify effective features from a vast pool of candidate features. In this paper, we present OpenFE, an automated feature generation tool that provides competitive results against machine learning experts. OpenFE achieves high efficiency and accuracy with two components: 1) a novel feature boosting method for accurately evaluating the incremental performance of candidate features and 2) a two-stage pruning algorithm that performs feature pruning in a coarse-to-fine manner. Extensive experiments on ten benchmark datasets show that OpenFE outperforms existing baseline methods by a large margin. We further evaluate OpenFE in two Kaggle competitions with thousands of data science teams participating. In the two competitions, features generated by OpenFE with a simple baseline model can beat 99.3% and 99.6% data science teams respectively. In addition to the empirical results, we provide a theoretical perspective to show that feature generation can be beneficial in a simple yet representative setting. The code is available at https://github.com/ZhangTP1996/OpenFE.

Fengyuan Liu, Jay Gala, Nilaksh et al. · mila

Evaluating the real-world capabilities of AI systems requires grounding benchmark performance in human-interpretable measures of task difficulty. Existing approaches that rely on direct human task completion time annotations are costly, noisy, and difficult to scale across benchmarks. In this work, we propose BRIDGE, a unified psychometric framework that learns the latent difficulty scale from model responses and anchors it to human task completion time. Using a two-parameter logistic Item Response Theory model, we jointly estimate latent task difficulty and model capability from model performance data across multiple benchmarks. We demonstrate that latent task difficulty varies linearly with the logarithm of human completion time, allowing human task completion time to be inferred for new benchmarks from model performance alone. Leveraging this alignment, we forecast frontier model capabilities in terms of human task length and independently reproduce METR's exponential scaling results, with the 50% solvable task horizon doubling approximately every 6 months.

Gül Sena Altıntaş, Malikeh Ehghaghi, Brian Lester et al. · mila

Tokenizers provide the fundamental basis through which text is represented and processed by language models (LMs). Despite the importance of tokenization, its role in LM performance and behavior is poorly understood due to the challenge of measuring the impact of tokenization in isolation. To address this need, we present TokSuite, a collection of models and a benchmark that supports research into tokenization's influence on LMs. Specifically, we train fourteen models that use different tokenizers but are otherwise identical using the same architecture, dataset, training budget, and initialization. Additionally, we curate and release a new benchmark that specifically measures model performance subject to real-world perturbations that are likely to influence tokenization. Together, TokSuite allows robust decoupling of the influence of a model's tokenizer, supporting a series of novel findings that elucidate the respective benefits and shortcomings of a wide range of popular tokenizers.

Bo Wang, Haoyang Huang, Zhiying Lu et al.

This paper introduces StoryAnchors, a unified framework for generating high-quality, multi-scene story frames with strong temporal consistency. The framework employs a bidirectional story generator that integrates both past and future contexts to ensure temporal consistency, character continuity, and smooth scene transitions throughout the narrative. Specific conditions are introduced to distinguish story frame generation from standard video synthesis, facilitating greater scene diversity and enhancing narrative richness. To further improve generation quality, StoryAnchors integrates Multi-Event Story Frame Labeling and Progressive Story Frame Training, enabling the model to capture both overarching narrative flow and event-level dynamics. This approach supports the creation of editable and expandable story frames, allowing for manual modifications and the generation of longer, more complex sequences. Extensive experiments show that StoryAnchors outperforms existing open-source models in key areas such as consistency, narrative coherence, and scene diversity. Its performance in narrative consistency and story richness is also on par with GPT-4o. Ultimately, StoryAnchors pushes the boundaries of story-driven frame generation, offering a scalable, flexible, and highly editable foundation for future research.

Pin-Jie Lin, Rishab Balasubramanian, Fengyuan Liu et al. · mila

Modern LLMs struggle with efficient updates, as each new pretrained model version requires repeating expensive alignment processes. This challenge also applies to domain- or languagespecific models, where fine-tuning on specialized data must be redone for every new base model release. In this paper, we explore the transfer of fine-tuning updates between model versions. Specifically, we derive the diff vector (representing the weight changes from finetuning) from one source model version and apply it to the base model of a different target version. Through empirical evaluations on various open-weight model versions, we show that transferring diff vectors can significantly improve the performance of the target base model. For example, transferring the fine-tuning updates from Llama 3.0 8B improves Llama 3.1 8B by 46.9% on IFEval and 15.7% on LiveCodeBench without additional training, even surpassing Llama 3.1 8B Instruct. Furthermore, we demonstrate performance gains on multilingual tasks, with 4.7% and 15.5% improvements on Global MMLU for Malagasy and Turkish, respectively. We observe that these merged models provide stronger initializations for further fine-tuning. Lastly, our controlled experiments suggest that fine-tuning transfer is most effective when source and target models lie in a linearly connected region of parameter space, and we provide a theoretical analysis of our method. Taken together, fine-tuning transfer offers a cost-efficient and practical strategy for continuous LLM development. Our code is available at github.com/pjlintw/finetuning-transfer.

Juanxi Tian, Fengyuan Liu, Jiaming Han et al.

Aligning multimodal generative models with human preferences demands reward signals that respect the compositional, multi-dimensional structure of human judgment. Prevailing RLHF approaches reduce this structure to scalar or pairwise labels, collapsing nuanced preferences into opaque parametric proxies and exposing vulnerabilities to reward hacking. While recent Rubrics-as-Reward (RaR) methods attempt to recover this structure through explicit criteria, generating rubrics that are simultaneously reliable, scalable, and data-efficient remains an open problem. We introduce Auto-Rubric as Reward (ARR), a framework that reframes reward modeling from implicit weight optimization to explicit, criteria-based decomposition. Before any pairwise comparison, ARR externalizes a VLM's internalized preference knowledge as prompt-specific rubrics, translating holistic intent into independently verifiable quality dimensions. This conversion of implicit preference structure into inspectable, interpretable constraints substantially suppresses evaluation biases including positional bias, enabling both zero-shot deployment and few-shot conditioning on minimal supervision. To extend these gains into generative training, we propose Rubric Policy Optimization (RPO), which distills ARR's structured multi-dimensional evaluation into a robust binary reward, replacing opaque scalar regression with rubric-conditioned preference decisions that stabilize policy gradients. On text-to-image generation and image editing benchmarks, ARR-RPO outperforms pairwise reward models and VLM judges, demonstrating that explicitly externalizing implicit preference knowledge into structured rubrics achieves more reliable, data-efficient multimodal alignment, revealing that the bottleneck is the absence of a factorized interface, not a deficit of knowledge.

Fengyuan Liu, Nouar AlDahoul, Gregory Eady et al.

Previous studies proposed that the reasoning capabilities of large language models (LLMs) can be improved through self-reflection, i.e., letting LLMs reflect on their own output to identify and correct mistakes in the initial responses. However, earlier experiments offer mixed results when it comes to the benefits of self-reflection. Furthermore, prior studies on self-reflection are predominantly concerned with the reasoning capabilities of models, ignoring the potential for self-reflection in safety, bias, and ideological leaning. Here, by conducting a series of experiments testing LLM's self-reflection capability in various tasks using a variety of prompts and different LLMs, we make several contributions to the literature. First, we reconcile conflicting findings regarding the benefit of self-reflection, by demonstrating that the outcome of self-reflection is sensitive to prompt wording -- both the original prompt that are used to elicit an initial answer and the subsequent prompt used to self-reflect. Specifically, although self-reflection may improve the reasoning capability of LLMs when the initial response is simple, the technique cannot improve upon the state-of-the-art chain-of-thought (CoT) prompting. Second, we show that self-reflection can lead to safer (75.8\% reduction in toxic responses while preserving 97.8\% non-toxic ones), less biased (77\% reduction in gender biased responses, while preserving 94.3\% unbiased ones), and more ideologically neutral responses (100\% reduction in partisan leaning response, while preserving 87.7\% non-partisan ones). The paper concludes by discussing the implications of our findings on the deployment of large language models. We release our experiments at https://github.com/Michael98Liu/self-reflection.

Haochen Luo, Jindong Gu, Fengyuan Liu et al.

Different from traditional task-specific vision models, recent large VLMs can readily adapt to different vision tasks by simply using different textual instructions, i.e., prompts. However, a well-known concern about traditional task-specific vision models is that they can be misled by imperceptible adversarial perturbations. Furthermore, the concern is exacerbated by the phenomenon that the same adversarial perturbations can fool different task-specific models. Given that VLMs rely on prompts to adapt to different tasks, an intriguing question emerges: Can a single adversarial image mislead all predictions of VLMs when a thousand different prompts are given? This question essentially introduces a novel perspective on adversarial transferability: cross-prompt adversarial transferability. In this work, we propose the Cross-Prompt Attack (CroPA). This proposed method updates the visual adversarial perturbation with learnable prompts, which are designed to counteract the misleading effects of the adversarial image. By doing this, CroPA significantly improves the transferability of adversarial examples across prompts. Extensive experiments are conducted to verify the strong cross-prompt adversarial transferability of CroPA with prevalent VLMs including Flamingo, BLIP-2, and InstructBLIP in various different tasks. Our source code is available at \url{https://github.com/Haochen-Luo/CroPA}.

Zhongjian Hu, Peng Yang, Bing Li et al.

Recently, Large Language Models (LLMs) have been used for knowledge-based Visual Question Answering (VQA). Despite the encouraging results of previous studies, prior methods prompt LLMs to predict answers directly, neglecting intermediate thought processes. We argue that prior methods do not sufficiently activate the capacities of LLMs. We propose a framework called PLRH that Prompts LLMs with Rationale Heuristics for knowledge-based VQA. The PLRH prompts LLMs with Chain of Thought (CoT) to generate rationale heuristics, i.e., intermediate thought processes, and then leverages the rationale heuristics to inspire LLMs to predict answers. Experiments show that our approach outperforms the existing baselines by more than 2.2 and 2.1 on OK-VQA and A-OKVQA, respectively.

Sichun Luo, Yi Huang, Haochen Luo et al.

LLM-guided evolutionary search has emerged as a promising paradigm for automated algorithm discovery, yet most systems track search progress primarily through executable programs and scalar fitness. Even when natural-language reflection is used, it is often used locally in mutation prompts or stored without an explicit population-level organization of strategic directions. As a result, evolutionary search can struggle to distinguish syntactically different implementations of the same idea, preserve lower-fitness but strategically promising directions, or detect when an entire family of strategies has saturated. We introduce \model, a modular strategy-space layer that elevates natural-language strategy descriptions from transient prompt context to first-class population-level evolutionary state in LLM-driven program search. \model augments each candidate program with an explicit natural language strategy description and uses this representation in three ways: Strategy Articulation turns mutation into a diagnose-direct-implement process; Stratified Experience Retrieval organizes the archive into strategy clusters and selects inspirations by behavioral complementarity; and Strategic Landscape Navigation periodically summarizes effective, saturated, and underexplored strategy families to guide future mutations. Across mathematical algorithm discovery, systems optimization, and agent-scaffold benchmarks, \model improves the underlying evolutionary backbones in most settings, with particularly large gains (21% relative improvement) on open-ended system optimization tasks. These results suggest that persistent strategy representations provide a practical mechanism for improving the robustness and efficiency of LLM-guided evolutionary search, suggesting a path toward compound AI systems that accumulate algorithmic knowledge over time.

Sichun Luo, Yi Huang, Mukai Li et al.

Large language models (LLMs) are increasingly deployed as conversational assistants in open-domain, multi-turn settings, where users often provide incomplete or ambiguous information. However, existing LLM-focused clarification benchmarks primarily assume single-turn interactions or cooperative users, limiting their ability to evaluate clarification behavior in realistic settings. We introduce \textbf{ClarifyMT-Bench}, a benchmark for multi-turn clarification grounded in a five-dimensional ambiguity taxonomy and a set of six behaviorally diverse simulated user personas. Through a hybrid LLM-human pipeline, we construct 6,120 multi-turn dialogues capturing diverse ambiguity sources and interaction patterns. Evaluating ten representative LLMs uncovers a consistent under-clarification bias: LLMs tend to answer prematurely, and performance degrades as dialogue depth increases. To mitigate this, we propose \textbf{ClarifyAgent}, an agentic approach that decomposes clarification into perception, forecasting, tracking, and planning, substantially improving robustness across ambiguity conditions. ClarifyMT-Bench establishes a reproducible foundation for studying when LLMs should ask, when they should answer, and how to navigate ambiguity in real-world human-LLM interactions.

Fengyuan Liu, Haochen Luo, Yiming Li et al. · deepmind, oxford

Recent progress in visual generative models enables the generation of high-quality images. To prevent the misuse of generated images, it is important to identify the origin model that generates them. In this work, we study the origin attribution of generated images in a practical setting where only a few images generated by a source model are available and the source model cannot be accessed. The goal is to check if a given image is generated by the source model. We first formulate this problem as a few-shot one-class classification task. To solve the task, we propose OCC-CLIP, a CLIP-based framework for few-shot one-class classification, enabling the identification of an image's source model, even among multiple candidates. Extensive experiments corresponding to various generative models verify the effectiveness of our OCC-CLIP framework. Furthermore, an experiment based on the recently released DALL-E 3 API verifies the real-world applicability of our solution.

Fengyuan Liu, Nikhil Kandpal, Colin Raffel · mila

The influence of contextual input on the behavior of large language models (LLMs) has prompted the development of context attribution methods that aim to quantify each context span's effect on an LLM's generations. The leave-one-out (LOO) error, which measures the change in the likelihood of the LLM's response when a given span of the context is removed, provides a principled way to perform context attribution, but can be prohibitively expensive to compute for large models. In this work, we introduce AttriBoT, a series of novel techniques for efficiently computing an approximation of the LOO error for context attribution. Specifically, AttriBoT uses cached activations to avoid redundant operations, performs hierarchical attribution to reduce computation, and emulates the behavior of large target models with smaller proxy models. Taken together, AttriBoT can provide a >300x speedup while remaining more faithful to a target model's LOO error than prior context attribution methods. This stark increase in performance makes computing context attributions for a given response 30x faster than generating the response itself, empowering real-world applications that require computing attributions at scale. We release a user-friendly and efficient implementation of AttriBoT to enable efficient LLM interpretability as well as encourage future development of efficient context attribution methods.

Fengyuan Liu, Rui Zhao, Shuo Chen et al.

Individual Large Language Models (LLMs) have demonstrated significant capabilities across various domains, such as healthcare and law. Recent studies also show that coordinated multi-agent systems exhibit enhanced decision-making and reasoning abilities through collaboration. However, due to the vulnerabilities of individual LLMs and the difficulty of accessing all agents in a multi-agent system, a key question arises: If attackers only know one agent, could they still generate adversarial samples capable of misleading the collective decision? To explore this question, we formulate it as a game with incomplete information, where attackers know only one target agent and lack knowledge of the other agents in the system. With this formulation, we propose M-Spoiler, a framework that simulates agent interactions within a multi-agent system to generate adversarial samples. These samples are then used to manipulate the target agent in the target system, misleading the system's collaborative decision-making process. More specifically, M-Spoiler introduces a stubborn agent that actively aids in optimizing adversarial samples by simulating potential stubborn responses from agents in the target system. This enhances the effectiveness of the generated adversarial samples in misleading the system. Through extensive experiments across various tasks, our findings confirm the risks posed by the knowledge of an individual agent in multi-agent systems and demonstrate the effectiveness of our framework. We also explore several defense mechanisms, showing that our proposed attack framework remains more potent than baselines, underscoring the need for further research into defensive strategies.

Fengyuan Liu, Huang Yi, Sichun Luo et al.

Discovering effective predictive signals, or ``alphas,'' from financial data with high dimensionality and extremely low signal-to-noise ratio remains a difficult open problem. Despite progress in deep learning, genetic programming, and, more recently, large language model (LLM)--based factor generation, existing approaches still explore only a narrow region of the vast alpha search space. Neural models tend to produce opaque and fragile patterns, while symbolic or formula-based methods often yield redundant or economically ungrounded expressions that generalize poorly. Although different in form, these paradigms share a key limitation: none can conduct broad, structured, and human-like exploration that balances logical consistency with creative leaps. To address this gap, we introduce the Cognitive Alpha Mining Framework (CogAlpha), which combines code-level alpha representation with LLM-driven reasoning and evolutionary search. Treating LLMs as adaptive cognitive agents, our framework iteratively refines, mutates, and recombines alpha candidates through multi-stage prompts and financial feedback. This synergistic design enables deeper thinking, richer structural diversity, and economically interpretable alpha discovery, while greatly expanding the effective search space. Experiments on A-share equities demonstrate that CogAlpha consistently discovers alphas with superior predictive accuracy, robustness, and generalization over existing methods. Our results highlight the promise of aligning evolutionary optimization with LLM-based reasoning for automated and explainable alpha discovery. All source code will be released.

Zhangyu Ge, Xu He, Lingfei Mo et al.

The brain's Path Integration (PI) mechanism offers substantial guidance and inspiration for Brain-Inspired Navigation (BIN). However, the PI capability constructed by the Continuous Attractor Neural Networks (CANNs) in most existing BIN studies exhibits significant computational redundancy, and its operational efficiency needs to be improved; otherwise, it will not be conducive to the practicality of BIN technology. To address this, this paper proposes an efficient PI approach using representation learning models to replicate CANN neurodynamic patterns. This method successfully replicates the neurodynamic patterns of CANN-modeled Head Direction Cells (HDCs) and Grid Cells (GCs) using lightweight Artificial Neural Networks (ANNs). These ANN-reconstructed HDC and GC models are then integrated to achieve brain-inspired PI for Dead Reckoning (DR). Benchmark tests in various environments, compared with the well-known NeuroSLAM system, demonstrate that this work not only accurately replicates the neurodynamic patterns of navigation cells but also matches NeuroSLAM in positioning accuracy. Moreover, efficiency improvements of approximately 17.5% on the general-purpose device and 40~50% on the edge device were observed, compared with NeuroSLAM. This work offers a novel implementation strategy to enhance the practicality of BIN technology and holds potential for further extension.

Filippo Rinaldi, Aniello Panariello, Giacomo Salici et al. · mila

When a new release of a foundation model is published, practitioners typically need to repeat full fine-tuning, even if the same task has already been solved in the previous version. A promising alternative is to reuse the parameter changes (i.e., task vectors) that capture how a model adapts to a specific task. However, they often fail to transfer across different pre-trained models due to their misaligned parameter space. In this work, we show that the key to successful transfer lies in the sign structure of the gradients of the new model. Based on this insight, we propose GradFix, a novel method that approximates the ideal gradient sign structure and leverages it to transfer knowledge using only a handful of labeled samples. Notably, this requires no additional fine-tuning: the adaptation is achieved by computing a few gradients at the target model and masking the source task vector accordingly. This yields an update that is locally aligned with the target loss landscape, effectively rebasing the task vector onto the new pre-training. We provide a theoretical guarantee that our method ensures first-order descent. Empirically, we demonstrate significant performance gains on vision and language benchmarks, consistently outperforming naive task vector addition and few-shot fine-tuning.

Xinyi Shang, Peng Sun, Fengyuan Liu et al.

This paper pioneers a novel data-centric paradigm to maximize the utility of unlabeled data, tackling a critical question: How can we enhance the efficiency and sustainability of deep learning training by optimizing the data itself? We begin by identifying three key limitations in existing model-centric approaches, all rooted in a shared bottleneck: knowledge extracted from data is locked to model parameters, hindering its reusability and scalability. To this end, we propose CoOpt, a highly efficient, parallelized framework for collaborative unlabeled data optimization, thereby effectively encoding knowledge into the data itself. By distributing unlabeled data and leveraging publicly available task-agnostic models, CoOpt facilitates scalable, reusable, and sustainable training pipelines. Extensive experiments across diverse datasets and architectures demonstrate its efficacy and efficiency, achieving 13.6% and 6.8% improvements on Tiny-ImageNet and ImageNet-1K, respectively, with training speedups of $1.94 \times $ and $1.2 \times$.

Yu Zhang, Fengyuan Liu, Juan Lyu et al.

In the realm of intelligent maritime navigation, object detection from a shipborne perspective is paramount. Despite the criticality, the paucity of maritime-specific data impedes the deployment of sophisticated visual perception techniques, akin to those utilized in autonomous vehicular systems, within the maritime context. To bridge this gap, we introduce Navigation12, a novel dataset annotated for 12 object categories under diverse maritime environments and weather conditions. Based upon this dataset, we propose HMPNet, a lightweight architecture tailored for shipborne object detection. HMPNet incorporates a hierarchical dynamic modulation backbone to bolster feature aggregation and expression, complemented by a matrix cascading poly-scale neck and a polymerization weight sharing detector, facilitating efficient multi-scale feature aggregation. Empirical evaluations indicate that HMPNet surpasses current state-of-the-art methods in terms of both accuracy and computational efficiency, realizing a 3.3% improvement in mean Average Precision over YOLOv11n, the prevailing model, and reducing parameters by 23%.

Hazem Ibrahim, Fengyuan Liu, Rohail Asim et al.

The emergence of large language models has led to the development of powerful tools such as ChatGPT that can produce text indistinguishable from human-generated work. With the increasing accessibility of such technology, students across the globe may utilize it to help with their school work -- a possibility that has sparked discussions on the integrity of student evaluations in the age of artificial intelligence (AI). To date, it is unclear how such tools perform compared to students on university-level courses. Further, students' perspectives regarding the use of such tools, and educators' perspectives on treating their use as plagiarism, remain unknown. Here, we compare the performance of ChatGPT against students on 32 university-level courses. We also assess the degree to which its use can be detected by two classifiers designed specifically for this purpose. Additionally, we conduct a survey across five countries, as well as a more in-depth survey at the authors' institution, to discern students' and educators' perceptions of ChatGPT's use. We find that ChatGPT's performance is comparable, if not superior, to that of students in many courses. Moreover, current AI-text classifiers cannot reliably detect ChatGPT's use in school work, due to their propensity to classify human-written answers as AI-generated, as well as the ease with which AI-generated text can be edited to evade detection. Finally, we find an emerging consensus among students to use the tool, and among educators to treat this as plagiarism. Our findings offer insights that could guide policy discussions addressing the integration of AI into educational frameworks.