论文检索

Computer Vision · Image Generation

Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, Yu Zhou

Centred on content modification and style preservation, Scene Text Editing (STE) remains a challenging task despite considerable progress in text-to-image synthesis and text-driven image manipulation recently. GAN-based STE methods generally encounter a common issue of model generalization, while Diffusion-based STE methods suffer from undesired style deviations. To address these problems, we propose TextCtrl, a diffusion-based method that edits text with prior guidance control. Our method consists of two key components: (i) By constructing fine-grained text style disentanglement and robust text glyph structure representation, TextCtrl explicitly incorporates Style-Structure guidance into model design and network training, significantly improving text style consistency and rendering accuracy. (ii) To further leverage the style prior, a Glyph-adaptive Mutual Self-attention mechanism is proposed which deconstructs the implicit fine-grained features of the source image to enhance style consistency and vision quality during inference. Furthermore, to fill the vacancy of the real-world STE evaluation benchmark, we create the first real-world image-pair dataset termed ScenePair for fair comparisons. Experiments demonstrate the effectiveness of TextCtrl compared with previous methods concerning both style fidelity and text accuracy. Project page: https://github.com/weichaozeng/TextCtrl.

Computer Vision · Segmentation

Kangning Liu, Brian Price, Jason Kuen, Yifei Fan, Zijun Wei, Luis Figueroa, Krzysztof Geras, Carlos Fernandez-Granda

The Segment Anything Model (SAM) is a large-scale foundation model that has revolutionized segmentation methodology. Despite its impressive generalization ability, the segmentation accuracy of SAM on images with intricate structures is often unsatisfactory. Recent works have proposed lightweight fine-tuning using high-quality annotated data to improve accuracy on such images. However, here we provide extensive empirical evidence that this strategy leads to forgetting how to "segment anything": these models lose the original generalization abilities of SAM, in the sense that they perform worse for segmentation tasks not represented in the annotated fine-tuning set. To improve performance without forgetting, we introduce a novel framework that combines high-quality annotated data with a large unlabeled dataset. The framework relies on two methodological innovations. First, we quantify the uncertainty in the SAM pseudo labels associated with the unlabeled data and leverage it to perform uncertainty-aware fine-tuning. Second, we encode the type of segmentation task associated with each training example using a $\textit{task prompt}$ to reduce ambiguity. We evaluated the proposed Segmentation with Uncertainty Model (SUM) on a diverse test set consisting of 14 public benchmarks, where it achieves state-of-the-art results. Notably, our method consistently surpasses SAM by 3-6 points in mean IoU and 4-7 in mean boundary IoU across point-prompt interactive segmentation rounds. Code is available at https://github.com/Kangningthu/SUM

Computer Vision · Text Understanding

Yaoyuan Liang, Zhuojun Cai, Jian Xu, Guanbo Huang, Yiran Wang, Xiao Liang, Jiahao Liu, Ziran Li, Jingang Wang, Shao-Lun Huang

The Multi-modal Large Language Model (MLLM) based Referring Expression Generation (REG) task has gained increasing popularity, which aims to generate an unambiguous text description that applies to exactly one object or region in the image by leveraging foundation models. We empirically found that there exists a potential trade-off between the detailedness and the correctness of the descriptions for the referring objects. On the one hand, generating sentences with more details is usually required in order to provide more precise object descriptions. On the other hand, complicated sentences could easily increase the probability of hallucinations. To address this issue, we propose a training-free framework, named ``unleash-then-eliminate'', which first elicits the latent information in the intermediate layers, and then adopts a cycle-consistency-based decoding method to alleviate the production of hallucinations. Furthermore, to reduce the computational load of cycle-consistency-based decoding, we devise a Probing-based Importance Estimation method to statistically estimate the importance weights of intermediate layers within a subset. These importance weights are then incorporated into the decoding process over the entire dataset, intervening in the next token prediction from intermediate layers.Extensive experiments conducted on the RefCOCOg and PHD benchmarks show that our proposed framework could outperform existing methods on both semantic and hallucination-related metrics. Code will be made available in https://github.com/Glupayy/unleash-eliminate.

Optimization · Discrete and Combinatorial Optimization

Xiang Meng, Kayhan Behdin, Haoyue Wang, Rahul Mazumder

The impressive performance of Large Language Models (LLMs) across various natural language processing tasks comes at the cost of vast computational resources and storage requirements. One-shot pruning techniques offer a way to alleviate these burdens by removing redundant weights without the need for retraining. Yet, the massive scale of LLMs often forces current pruning approaches to rely on heuristics instead of optimization-based techniques, potentially resulting in suboptimal compression. In this paper, we introduce ALPS, an optimization-based framework that tackles the pruning problem using the operator splitting technique and a preconditioned conjugate gradient-based post-processing step. Our approach incorporates novel techniques to accelerate and theoretically guarantee convergence while leveraging vectorization and GPU parallelism for efficiency. ALPS substantially outperforms state-of-the-art methods in terms of the pruning objective and perplexity reduction, particularly for highly sparse models. On the LLaMA3-8B model with 70\% sparsity, ALPS achieves a 29\% reduction in test perplexity on the WikiText dataset and a 8\% improvement in zero-shot benchmark performance compared to existing methods. Our code is available at https://github.com/mazumder-lab/ALPS.

Deep Learning · Representation Learning

Bozhen Hu, Cheng Tan, Jun Xia, Yue Liu, Lirong Wu, Jiangbin Zheng, Yongjie Xu, Yufei Huang, Stan Z. Li

Learning effective representations is imperative for comprehending proteins and deciphering their biological functions. Recent strides in language models and graph neural networks have empowered protein models to harness primary or tertiary structure information for representation learning. Nevertheless, the absence of practical methodologies to appropriately model intricate inter-dependencies between protein sequences and structures has resulted in embeddings that exhibit low performance on tasks such as protein function prediction. In this study, we introduce CoupleNet, a novel framework designed to interlink protein sequences and structures to derive informative protein representations. CoupleNet integrates multiple levels and scales of features in proteins, encompassing residue identities and positions for sequences, as well as geometric representations for tertiary structures from both local and global perspectives. A two-type dynamic graph is constructed to capture adjacent and distant sequential features and structural geometries, achieving completeness at the amino acid and backbone levels. Additionally, convolutions are executed on nodes and edges simultaneously to generate comprehensive protein embeddings. Experimental results on benchmark datasets showcase that CoupleNet outperforms state-of-the-art methods, exhibiting particularly superior performance in low-sequence similarities scenarios, adeptly identifying infrequently encountered functions and effectively capturing remote homology relationships in proteins.

Probabilistic Methods · Bayesian Models and Methods

Julius Vetter, Guy Moss, Cornelius Schröder, Richard Gao, Jakob H Macke

Scientific modeling applications often require estimating a distribution of parameters consistent with a dataset of observations - an inference task also known as source distribution estimation. This problem can be ill-posed, however, since many different source distributions might produce the same distribution of data-consistent simulations. To make a principled choice among many equally valid sources, we propose an approach which targets the maximum entropy distribution, i.e., prioritizes retaining as much uncertainty as possible. Our method is purely sample-based - leveraging the Sliced-Wasserstein distance to measure the discrepancy between the dataset and simulations - and thus suitable for simulators with intractable likelihoods. We benchmark our method on several tasks, and show that it can recover source distributions with substantially higher entropy than recent source estimation methods, without sacrificing the fidelity of the simulations. Finally, to demonstrate the utility of our approach, we infer source distributions for parameters of the Hodgkin-Huxley model from experimental datasets with hundreds of single-neuron measurements. In summary, we propose a principled method for inferring source distributions of scientific simulator parameters while retaining as much uncertainty as possible.

Data-centric AI · Data augmentation

Tianyue Ou, Frank F. Xu, Aman Madaan, Jiarui Liu, Robert Lo, Abishek Sridhar, Sudipta Sengupta, Dan Roth, Graham Neubig, Shuyan Zhou

LLMs can now act as autonomous agents that interact with digital environments and complete specific objectives (e.g., arranging an online meeting). However, accuracy is still far from satisfactory, partly due to a lack of large-scale, direct demonstrations for digital tasks. Obtaining supervised data from humans is costly, and automatic data collection through exploration or reinforcement learning relies on complex environmental and content setup, resulting in datasets that lack comprehensive coverage of various scenarios. On the other hand, there is abundant knowledge that may indirectly assist task completion, such as online tutorials that were created for human consumption. In this work, we present Synatra, an approach that effectively transforms this indirect knowledge into direct supervision at scale. We define different types of indirect knowledge, and carefully study the available sources to obtain it, methods to encode the structure of direct demonstrations, and finally methods to transform indirect knowledge into direct demonstrations. We use 100k such synthetically-created demonstrations to finetune a 7B CodeLlama, and demonstrate that the resulting agent surpasses all comparably sized models on three web-based task benchmarks Mind2Web, MiniWoB++ and WebArena, as well as surpassing GPT-3.5 on WebArena and Mind2Web. In addition, while synthetic demonstrations prove to be only 3% the cost of human demonstrations (at $0.031 each), we show that the synthetic demonstrations can be more effective than an identical number of human demonstrations collected from limited domains.

Language · Knowledge

Pengcheng Jiang, Lang Cao, Cao (Danica) Xiao, Parminder Bhatia, Jimeng Sun, Jiawei Han

Knowledge Graph Embedding (KGE) techniques are crucial in learning compact representations of entities and relations within a knowledge graph, facilitating efficient reasoning and knowledge discovery. While existing methods typically focus either on training KGE models solely based on graph structure or fine-tuning pre-trained language models with classification data in KG, KG-FIT leverages LLM-guided refinement to construct a semantically coherent hierarchical structure of entity clusters. By incorporating this hierarchical knowledge along with textual information during the fine-tuning process, KG-FIT effectively captures both global semantics from the LLM and local semantics from the KG. Extensive experiments on the benchmark datasets FB15K-237, YAGO3-10, and PrimeKG demonstrate the superiority of KG-FIT over state-of-the-art pre-trained language model-based methods, achieving improvements of 14.4\%, 13.5\%, and 11.9\% in the Hits@10 metric for the link prediction task, respectively. Furthermore, KG-FIT yields substantial performance gains of 12.6\%, 6.7\%, and 17.7\% compared to the structure-based base models upon which it is built. These results highlight the effectiveness of KG-FIT in incorporating open-world knowledge from LLMs to significantly enhance the expressiveness and informativeness of KG embeddings.

Reinforcement Learning · Multi-agent

Yancheng Liang, Daphne Chen, Abhishek Gupta, Simon Du, Natasha Jaques

Training agents that can coordinate zero-shot with humans is a key mission in multi-agent reinforcement learning (MARL). Current algorithms focus on training simulated human partner policies which are then used to train a Cooperator agent. The simulated human is produced either through behavior cloning over a dataset of human cooperation behavior, or by using MARL to create a population of simulated agents. However, these approaches often struggle to produce a Cooperator that can coordinate well with real humans, since the simulated humans fail to cover the diverse strategies and styles employed by people in the real world. We show \emph{learning a generative model of human partners} can effectively address this issue. Our model learns a latent variable representation of the human that can be regarded as encoding the human's unique strategy, intention, experience, or style. This generative model can be flexibly trained from any (human or neural policy) agent interaction data. By sampling from the latent space, we can use the generative model to produce different partners to train Cooperator agents. We evaluate our method---Generative Agent Modeling for Multi-agent Adaptation (GAMMA)---on Overcooked, a challenging cooperative cooking game that has become a standard benchmark for zero-shot coordination. We conduct an evaluation with real human teammates, and the results show that GAMMA consistently improves performance, whether the generative model is trained on simulated populations or human datasets. Further, we propose a method for posterior sampling from the generative model that is biased towards the human data, enabling us to efficiently improve performance with only a small amount of expensive human interaction data.

Multimodal Models

An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, Sifei Liu

Vision Language Models (VLMs) have demonstrated remarkable performance in 2D vision and language tasks. However, their ability to reason about spatial arrangements remains limited. In this work, we introduce Spatial Region GPT (SpatialRGPT) to enhance VLMs’ spatial perception and reasoning capabilities. SpatialRGPT advances VLMs’ spatial understanding through two key innovations: (i) a data curation pipeline that enables effective learning of regional representation from 3D scene graphs, and (ii) a flexible ``plugin'' module for integrating depth information into the visual encoder of existing VLMs. During inference, when provided with user-specified region proposals, SpatialRGPT can accurately perceive their relative directions and distances. Additionally, we propose SpatialRGBT-Bench, a benchmark with ground-truth 3D annotations encompassing indoor, outdoor, and simulated environments, for evaluating 3D spatial cognition in Vision-Language Models (VLMs). Our results demonstrate that SpatialRGPT significantly enhances performance in spatial reasoning tasks, both with and without local region prompts. The model also exhibits strong generalization capabilities, effectively reasoning about complex spatial relations and functioning as a region-aware dense reward annotator for robotic tasks. Code, dataset, and benchmark are released at https://www.anjiecheng.me/SpatialRGPT.

Multimodal Models

Jianfeng Dong, Xiaoman Peng, Daizong Liu, Xiaoye Qu, Xun Yang, Cuizhu Bao, Meng Wang

As a widely explored multi-modal task, Temporal Sentence Grounding in videos (TSG) endeavors to retrieve a specific video segment matched with a given query text from a video. The traditional paradigm for TSG generally assumes that relevant segments always exist within a given video. However, this assumption is restrictive and unrealistic in real-world applications where the existence of a query-related segment is uncertain, easily resulting in erroneous grounding. Motivated by the research gap and practical application, this paper introduces a new task, named Temporal Sentence Grounding with Relevance Feedback (TSG-RF) in videos, which accommodates the possibility that a video may or may not include a segment related to the query. This task entails localizing precise video segments that semantically align with the query text when such content is present, while delivering definitive feedback on the non-existence of related segments when absent. Moreover, we propose a novel Relation-aware Temporal Sentence Grounding (RaTSG) network for addressing this challenging task. This network first reformulates the TSG-RF task as a foreground-background detection problem by investigating whether the query-related semantics exist in both frame and video levels. Then, a multi-granularity relevance discriminator is exploited to produce precise video-query relevance feedback and a relation-aware segment grounding module is employed to selectively conduct the grounding process, dynamically adapting to the presence or absence of query-related segments in videos. To validate our RaTSG network, we reconstruct two popular TSG datasets, establishing a rigorous benchmark for TSG-RF. Experimental results demonstrate the effectiveness of our proposed RaTSG for the TSG-RF task. Our source code is available at https://github.com/HuiGuanLab/RaTSG.

Trustworthy Machine Learning

Hao Dong, Yue Zhao, Eleni Chatzi, Olga Fink

Detecting out-of-distribution (OOD) samples is important for deploying machine learning models in safety-critical applications such as autonomous driving and robot-assisted surgery. Existing research has mainly focused on unimodal scenarios on image data. However, real-world applications are inherently multimodal, which makes it essential to leverage information from multiple modalities to enhance the efficacy of OOD detection. To establish a foundation for more realistic Multimodal OOD Detection, we introduce the first-of-its-kind benchmark, MultiOOD, characterized by diverse dataset sizes and varying modality combinations. We first evaluate existing unimodal OOD detection algorithms on MultiOOD, observing that the mere inclusion of additional modalities yields substantial improvements. This underscores the importance of utilizing multiple modalities for OOD detection. Based on the observation of Modality Prediction Discrepancy between in-distribution (ID) and OOD data, and its strong correlation with OOD performance, we propose the Agree-to-Disagree (A2D) algorithm to encourage such discrepancy during training. Moreover, we introduce a novel outlier synthesis method, NP-Mix, which explores broader feature spaces by leveraging the information from nearest neighbor classes and complements A2D to strengthen OOD detection performance. Extensive experiments on MultiOOD demonstrate that training with A2D and NP-Mix improves existing OOD detection algorithms by a large margin. To support accessibility and reproducibility, our source code and MultiOOD benchmark are available at https://github.com/donghao51/MultiOOD.

Robotics

Siddhant Haldar, Zhuoran Peng, Lerrel Pinto

Training generalist agents capable of solving diverse tasks is challenging, often requiring large datasets of expert demonstrations. This is particularly problematic in robotics, where each data point requires physical execution of actions in the real world. Thus, there is a pressing need for architectures that can effectively leverage the available training data. In this work, we present BAKU, a simple transformer architecture that enables efficient learning of multi-task robot policies. BAKU builds upon recent advancements in offline imitation learning and meticulously combines observation trunks, action chunking, multi-sensory observations, and action heads to substantially improve upon prior work. Our experiments on 129 simulated tasks across LIBERO, Meta-World suite, and the Deepmind Control suite exhibit an overall 18% absolute improvement over RT-1 and MT-ACT, with a 36% improvement on the harder LIBERO benchmark. On 30 real-world manipulation tasks, given an average of just 17 demonstrations per task, BAKU achieves a 91% success rate. Videos of the robot are best viewed at baku-robot.github.io.

Deep Learning · Representation Learning

Maitreya Patel, Naga Sai Abhiram Kusumba, Sheng Cheng, Changhoon Kim, Tejas Gokhale, Chitta Baral, 'YZ' Yezhou Yang

Contrastive Language-Image Pretraining (CLIP) models maximize the mutual information between text and visual modalities to learn representations. This makes the nature of the training data a significant factor in the efficacy of CLIP for downstream tasks. However, the lack of compositional diversity in contemporary image-text datasets limits the compositional reasoning ability of CLIP. We show that generating ``hard'' negative captions via in-context learning and synthesizing corresponding negative images with text-to-image generators offers a solution. We introduce a novel contrastive pre-training strategy that leverages these hard negative captions and images in an alternating fashion to train CLIP. We demonstrate that our method, named TripletCLIP, when applied to existing datasets such as CC3M and CC12M, enhances the compositional capabilities of CLIP, resulting in an absolute improvement of over 9% on the SugarCrepe benchmark on an equal computational budget, as well as improvements in zero-shot image classification and image retrieval. Our code, models, and data are available at: tripletclip.github.io.

Trustworthy Machine Learning

Aref Rekavandi, Farhad Farokhi, Olga Ohrimenko, Benjamin Rubinstein

Certified adversarial robustness of large-scale deep networks has progressed substantially after the introduction of randomized smoothing. Deep net classifiers are now provably robust in their predictions against a large class of threat models, including $\ell_1$, $\ell_2$, and $\ell_\infty$ norm-bounded attacks. Certified robustness analysis by randomized smoothing has not been performed for deep regression networks where the output variable is continuous and unbounded. In this paper, we extend the existing results for randomized smoothing into regression models using powerful tools from robust statistics, in particular, $\alpha$-trimming filter as the smoothing function. Adjusting the hyperparameter $\alpha$ achieves a smooth trade-off between desired certified robustness and utility. For the first time, we propose a benchmark for certified robust regression in visual positioning systems using the Cambridge Landmarks dataset where robustness analysis is essential for autonomous navigation of AI agents and self-driving cars. Code is publicly available at \url{https://github.com/arekavandi/Certified_adv_RRegression/}.

Datasets and Benchmarks

Niels Mündler, Mark Müller, Jingxuan He, Martin Vechev

Rigorous software testing is crucial for developing and maintaining high-quality code, making automated test generation a promising avenue for both improving software quality and boosting the effectiveness of code generation methods. However, while code generation with Large Language Models (LLMs) is an extraordinarily active research area, test generation remains relatively unexplored. We address this gap and investigate the capability of LLM-based Code Agents to formalize user issues into test cases. To this end, we propose a novel benchmark based on popular GitHub repositories, containing real-world issues, ground-truth bug-fixes, and golden tests. We find that LLMs generally perform surprisingly well at generating relevant test cases, with Code Agents designed for code repair exceeding the performance of systems designed specifically for test generation. Further, as test generation is a similar but more structured task than code generation, it allows for a more fine-grained analysis using issue reproduction rate and coverage changes, providing a dual metric for analyzing systems designed for code repair. Finally, we find that generated tests are an effective filter for proposed code fixes, doubling the precision of SWE-Agent. We release all data and code at https://github.com/logic-star-ai/SWT-Bench.

Generative Models · Misc

Abhinav Dutta, Sanjeev Krishnan, Nipun Kwatra, Ramachandran Ramjee

When Large Language Models (LLMs) are compressed using techniques such as quantization, the predominant way to demonstrate the validity of such techniques is by measuring the model's accuracy on various benchmarks. If the accuracies of the baseline model and the compressed model are close, it is assumed that there was negligible degradation in quality. However, even when the accuracy of baseline and compressed model are similar, we observe the phenomenon of flips, wherein answers change from correct to incorrect and vice versa in proportion. We conduct a detailed study of metrics across multiple compression techniques, models and datasets, demonstrating that the behavior of compressed models as visible to end-users is often significantly different from the baseline model, even when accuracy is similar. We further evaluate compressed models qualitatively and quantitatively using MT-Bench and show that compressed models exhibiting high flips are worse than baseline models in this free-form generative task. Thus, we argue that accuracy and perplexity are necessary but not sufficient for evaluating compressed models, since these metrics hide large underlying changes that have not been observed by previous work. Hence, compression techniques should also be evaluated using distance metrics. We propose two such distance metrics, KL-Divergence and flips, and show that they are well correlated.

Datasets and Benchmarks

Gabriel Rioux, Apoorva Nitsure, Mattia Rigotti, Kristjan Greenewald, Youssef Mroueh

Stochastic dominance is an important concept in probability theory, econometrics and social choice theory for robustly modeling agents' preferences between random outcomes. While many works have been dedicated to the univariate case,little has been done in the multivariate scenario, wherein an agent has to decide between different multivariate outcomes. By exploiting a characterization of multivariate first stochastic dominance in terms of couplings, we introduce a statistic that assesses multivariate almost stochastic dominance under the framework of Optimal Transport with a smooth cost. Further, we introduce an entropic regularization of this statistic, and establish a central limit theorem (CLT) and consistency of the bootstrap procedure for the empirical statistic. Armed with this CLT, we propose a hypothesis testing framework as well as an efficient implementation using the Sinkhorn algorithm. We showcase our method in comparing and benchmarking Large Language Models that are evaluated on multiple metrics. Our multivariate stochastic dominance test allows us to capture the dependencies between the metrics in order to make an informed and statistically significant decision on the relative performance of the models.

Generative Models · New Approaches

Vaskar Nath, Dylan Slack, Jeff Da, Yuntao Ma, Hugh Zhang, Spencer Whitehead, Sean Hendryx

Techniques that learn improved representations via offline data or self-supervised objectives have shown impressive results in traditional reinforcement learning.Nevertheless, it is unclear how improved representation learning can benefit reinforcement learning from human feedback on language models.In this work, we propose training reward models (RMs) in a contrastive, $\textit{goal-conditioned}$ fashion by increasing the representation similarity of future states along sampled preferred trajectories and decreasing the similarity along randomly sampled dispreferred trajectories.This objective significantly improves reward model performance by up to 0.09 AUROC across challenging benchmarks, such as MATH and GSM8k. These findings extend to general alignment as well -- on the Helpful-Harmless dataset, we observe 2.3\% increase in accuracy.Beyond improving reward model performance, we show this way of training RM representations enables improved steerability because it allows us to evaluate the likelihood of an action achieving a particular goal-state (e.g. whether a solution is correct or helpful).Leveraging this insight, we find that we can filter up to 55\% of generated tokens during majority voting by discarding trajectories likely to end up in an "incorrect" state, which leads to significant cost savings.We additionally find that these representations can perform fine-grained control by conditioning on desired future goal-states.For example, we show that steering a Llama 3 model towards helpful generations with our approach improves helpfulness by $9.6$\% over a supervised-fine-tuning trained baseline.Similarly, steering the model towards complex generations improves complexity by $21.6$\% over the baseline.Overall, we find that training RMs in this contrastive, goal-conditioned fashion significantly improves performance and enables model steerability.

Xiong-Hui Chen, Ziyan Wang, Yali Du, Shengyi Jiang, Meng Fang, Yang Yu, Jun Wang

When humans need to learn a new skill, we can acquire knowledge through written books, including textbooks, tutorials, etc. However, current research for decision-making, like reinforcement learning (RL), has primarily required numerous real interactions with the target environment to learn a skill, while failing to utilize the existing knowledge already summarized in the text. The success of Large Language Models (LLMs) sheds light on utilizing such knowledge behind the books. In this paper, we discuss a new policy learning problem called Policy Learning from tutorial Books (PLfB) upon the shoulders of LLMs’ systems, which aims to leverage rich resources such as tutorial books to derive a policy network. Inspired by how humans learn from books, we solve the problem via a three-stage framework: Understanding, Rehearsing, and Introspecting (URI). In particular, it first rehearses decision-making trajectories based on the derived knowledge after understanding the books, then introspects in the imaginary dataset to distill a policy network. We build two benchmarks for PLfB~based on Tic-Tac-Toe and Football games. In experiment, URI's policy achieves at least 44% net win rate against GPT-based agents without any real data; In Football game, which is a complex scenario, URI's policy beat the built-in AIs with a 37% while using GPT-based agent can only achieve a 6\% winning rate. The project page: https://plfb-football.github.io.