Datasets:

nhop
/

ReviewBench

License:

Dataset Preview

Split (1)

train

The full dataset viewer is not available (click to read why). Only showing a preview of the rows.

Could not read the parquet files: 404 Client Error: Not Found for url: https://hf-hub-lfs-us-east-1.s3.us-east-1.amazonaws.com/repos/81/7d/817d1fcd79b853072d54434b70f695a2ca66a7ce5cf7a4276c00b45c33f8639c/967e156938f8383610d64a9db049bbbace09cce6e48c90439f3e0f9ee511d2fd?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=AKIA2JU7TKAQLC2QXPN7%2F20250508%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20250508T020915Z&X-Amz-Expires=3600&X-Amz-Signature=83d6f689132e44146a669935da4d7ca17acc39de1a2a9ab757a0186d4e0af0e6&X-Amz-SignedHeaders=host&response-content-disposition=inline%3B%20filename%2A%3DUTF-8%27%27train-06002-of-07630.parquet%3B%20filename%3D%22train-06002-of-07630.parquet%22%3B&x-id=GetObject <?xml version="1.0" encoding="UTF-8"?> <Error><Code>NoSuchKey</Code><Message>The specified key does not exist.</Message><Key>repos/81/7d/817d1fcd79b853072d54434b70f695a2ca66a7ce5cf7a4276c00b45c33f8639c/967e156938f8383610d64a9db049bbbace09cce6e48c90439f3e0f9ee511d2fd</Key><RequestId>9S7PBQ8DVR9S4JMF</RequestId><HostId>9KLLEdHMohI9MrQjU3UYYq96J0LCvcz6ennpbYxe5oqt6E0M4T1ksDlvQR6wsS7NhCFve+AxN479hR0neH7Wc3YEZYzdvkkp</HostId></Error>

Error code:   FileSystemError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

title string	abstract string	field_of_study sequence	publication_date string	license string	reference_paperhash sequence	paperhash string	full_text string	figures images list	figures_metadata list	decision bool	detailed_decision int64	reviews list	comments list	score sequence	confidence sequence	correctness sequence	clarity sequence	impact sequence	ethics sequence
Customized Procedure Planning in Instructional Videos	Generating customized procedures for task planning in instructional videos poses a unique challenge for vision-language models. In this paper, we introduce Customized Procedure Planning in Instructional Videos, a novel task that focuses on generating a sequence of detailed action steps for task completion based on user requirements and the task's initial visual state. Existing methods often neglect customization and user directions, limiting their real-world applicability. The absence of instructional video datasets with step-level state and video-specific action plan annotations has hindered progress in this domain. To address these challenges, we introduce the Customized Procedure Planner (CPP) framework, a causal, open-vocabulary model that leverages a LlaVA-based approach to predict procedural plans based on a task's initial visual state and user directions. To overcome the data limitation, we employ a weakly-supervised approach, using the strong vision-language model GEMINI and the large language model (LLM) GPT-4 to create detailed video-specific action plans from the benchmark instructional video datasets (COIN, CrossTask), producing pseudo-labels for training. Discussing the limitations of the existing procedure planning evaluation metrics in an open-vocabulary setting, we propose novel automatic LLM-based metrics with few-shot in-context learning to evaluate the customization and planning capabilities of our model, setting a strong baseline. Additionally, we implement an LLM-based objective function to enhance model training for improved customization. Extensive experiments, including human evaluations, demonstrate the effectiveness of our approach, establishing a strong baseline for future research in customized procedure planning.	[ "Customized Procedure Planning", "multi-modal models", "vision-language models", "applications to robotics, autonomy, planning" ]	2024-10-04	CC BY 4.0	[ "\|pour_in_milk", "\|whisk_milk_and_eggs_together", "\|dip_homemade_ciabatta_bread_in_the_egg_mixture", "\|melt_butter_in_the_pan", "\|place_the_soaked_ciabatta_bread_in_the_pan", "\|flip_the_bread_to_cook_evenly", "\|glaze_the_top_with_cream_cheese._task:_make_french_toast_task:_make_bread_and_butter_pickles_...	zare\|customized_procedure_planning_in_instructional_videos\|ICLR_cc_2025_Conference	introduction Procedure planning in instructional videos (PPIV) involves generating a sequence of action steps, to transform an initial visual observation of a task into its completion (Chang et al., 2020; Bi et al., 2021a; Sun et al., 2022; Zhao et al., 2022; Wang et al., 2023a; b; Li et al., 2023; Niu et al., 2024; Zare et al., 2024; Nagasinghe et al., 2024) . Autonomous agents capable of performing this task can assist humans in efficiently completing complex, goal-oriented tasks and procedures in daily life. While humans intuitively understand the steps and reasoning needed to accomplish such tasks, machines face considerable challenges in replicating this ability. To overcome this gap, an autonomous agent requires a deep understanding of instructional procedures, their unique characteristics, related objects, the various states involved, and the transformations brought about by actions. This understanding is essential for generating a plausible, executable plan that leads to successful task completion.Despite considerable progress in recent studies, various obstacles still restrict its practical applications in the real world. Recent works on procedure planning in instructional videos have largely overlooked the importance of customization and user-specific directions. Most existing approaches rely on initial and final visual observations of a task, resulting in a non-causal formulation (Chang et al., 2020; Bi et al., 2021a; Sun et al., 2022; Zhao et al., 2022; Wang et al., 2023a; b; Li et al., 2023; Niu et al., 2024; Zare et al., 2024) , which limits their applicability in real-life scenarios. This reliance on visual information alone introduces a semantic gap, particularly in representing intermediate action steps that may depend on user-specific conditions but are not captured by the visual inputs. Consequently, the generated action plans often lack informativeness, producing generic se- quences blind to user-specific needs. This issue is illustrated in Fig. 1a , where the initial and final visual states do not distinguish between a generic and a more detailed plan, resulting in ambiguity and reinforcing the semantic gap.Some models, such as (Wang et al., 2023a) , which incorporate textual inputs, have made progress in bridging this semantic gap between visual observations and intermediate steps. However, they still fall short by conditioning planning solely on task-related textual information inferred from the observed states, without generating action steps tailored to user-specific directions or conditions necessary to complete a task from its current state.A model that fully addresses this limitation must go beyond simple visual inputs. It should be capable of processing both the current visual state of the task and user-specific requirements provided in textual form. This would allow the model to generate a more tailored plan, transforming the task toward completion in a way that aligns with both the visual state and the user's directions. Fig. 1b highlights the contrast between the two approaches: a model that incorporates user-specified needs, such as keyword conditions, can produce a more customized, detailed, and informative instructional plan with customized steps. This stands in contrast to a model that relies solely on task objectives, showcasing the practicality and relevance of customized procedure planning.Addressing this need cannot be adequately captured within the conventional closed-vocabulary setting under which this problem has been studied (Chang et al., 2020; Bi et al., 2021a; Sun et al., 2022; Zhao et al., 2022; Wang et al., 2023a; b; Li et al., 2023; Niu et al., 2024; Zare et al., 2024) , as it restricts plan prediction to predefined action labels. While recent works, such as Wu et al. (2024) , have made progress in expanding the problem of PPIV to an open-vocabulary setting, the challenge of generating detailed, user-specific action plans remains unresolved.A key challenge in extending PPIV to address user-specific needs has been the lack of suitable datasets for training. To train such a model, a large dataset of instructional videos is required, along with their corresponding detailed instructional plans, annotated with time-stamped procedural states. These detailed plans must be tailored to the specific characteristics of each video, which distinguish an instructional video from more generic ones, addressing unique user requirements. However, obtaining such annotations is both expensive and time-consuming. Existing benchmark datasets for this task, such as CrossTask and Coin (Zhukov et al., 2019; Tang et al., 2019) , provide step-level annotations of procedural states and generic plans, but they lack the detailed instructional plans and video-specific characteristics that make each instructional plans informative and unique in terms of user demands.Under review as a conference paper at ICLR 2025We tackle these challenges, by introducing the setting of Customized Procedure Planning in Instructional Videos (CPPIV) and proposing the Customized Procedure Planner (CPP) framework as a solution for this problem. We implement CPP as a LlaVa-based (Liu et al., 2023; 2024a; b) model, fine-tuned to generate detailed, open-vocabulary instructional plans for task completion, starting from an initial visual state and customized based on user-specified keywords.To overcome dataset limitations in training CPP, we adopt a weakly supervised approach. First, we leverage the powerful vision-language model, GEMINI (Team et al., 2023) , to extract video-specific, task-related keywords and generate descriptions that explain how these keywords are relevant to the video's action plan. This is applied to the CrossTask and COIN datasets. Using this customized information-key elements that differentiate the instructional content of each video from generic task plans-we conditionally generate a customized, video-specific instructional plan. To achieve this, we employ the strong LLM, GPT-4o (OpenAI, 2023) , to adapt the generic instructional ground truth plan for each video based on the extracted keywords. These customized plans serve as pseudolabels for training the CPP model. Additionally, during training, GPT-4o is integrated into the objective function to further enhance the model's ability to produce customized instructional plans.Extending Procedure Planning in Instructional Videos to an open-vocabulary setting presents challenges for traditional evaluation metrics, which rely on pre-defined, closed-vocabulary action step labels and fail to generalize effectively. To overcome this, we draw on recent works (Liang et al., 2023; Zhu et al., 2023; Wang et al., 2024; Huang et al., 2024) , and introduce a novel LLM-based approach-referred to as automatic metrics-to assess the quality of both planning and customization in detailed, varied, open-vocabulary plans. We evaluate our model on two widely used instructional video datasets, CrossTask and COIN. Additionally, we validate our model's performance by testing it on human-annotated customized plans from both datasets. Our model outperforms the state-ofthe-art (SoA) and establishes a strong baseline for the setting of customized procedure planning.Our main contributions are:-We emphasize the need for a more practical formulation of procedure planning in instructional videos that considers user directions and specific requirements and introduce the novel setting of customized procedure planning in instructional videos, aimed at generating instructional plans that cater to user task-specific needs rather than relying solely on generic task completion.-We propose the Customized Procedure Planner framework, which generates open-vocabulary instructional plans tailored to user-specified condition keywords, facilitating the transformation of initial visual states into task completion.-We propose a weakly supervised training approach that addresses the lack of customization annotations for CPPIV model training, allowing customized planning to be learned from unannotated videos.-We extend conventional procedure planning metrics to encompass open-vocabulary, varied, and detailed instructional plans, enabling a comprehensive assessment of planning and customization performance for predicted plans. related works 2.1 procedure planning Procedure planning from instructional videos involves generating effective task completion plans.Earlier works employed a two-branch architecture, sequentially predicting actions and states with recursive models (Jain & Medsker, 1999; Vaswani et al., 2017) to capture state transitions. More recent methods, such as Zhao et al. (2022) ; Wang et al. (2023b) , generate plans using a single-branch architecture that directly decodes actions, minimizing prediction error propagation. However, these approaches rely solely on visual observations of the initial and final states, resulting in a non-causal formulation that lacks adaptability to user-specific tasks and needs. Our work introduces CPP, a novel one-branch prediction framework that generates a detailed sequence of actions based on both the initial visual state and user-defined conditions, addressing this limitation in the existing literature.Under review as a conference paper at ICLR 2025The problem of customized procedure planning can be framed as Conditional Vision-Language Models for Sequence Generation. In this approach, the model generates an output sequence by conditioning on both visual input (i.e., the current visual state) and textual input (i.e., task and user requirements). To address the CPPIV challenge, we utilize the Conditional Vision-Language Models framework, leveraging models such as LLaVA (Liu et al., 2024a; 2023; 2024b) and GPT-4o (OpenAI, 2023) . The pipeline extracts task-specific keywords from the PPIV datasets using a vision-language model (VLM), which are then combined with a human-annotated generic plan to create pseudo labels for training customized instructional video datasets with the aid of a large language model (LLM). * Refer to prompts 1, 2, & 3 for the complete text.In this section, we introduce our proposed framework, the Customized Procedure Planner, designed for customized procedure planning in instructional videos. We also explore the weakly supervised learning approach employed to train the CPP in the absence of datasets containing customization annotations.Under review as a conference paper at ICLR 2025 setting: customized procedure planning We define the novel setting of customized procedure planning in instructional videos as follows: Given an initial visual observation o s , a task objective Task, and a sequence of userspecified customization keywords Keywords = {k 1 , k 2 , . . . , k K }, the model generates a plan p = {a 1 , a 2 , . . . , a T }, where T represents the plan's length (i.e., the action horizon) and a i (for 1 ≤ i ≤ T ) is the detailed customized text for the i-th action step. This plan should effectively transform o s into the task objective while satisfying the specified customization conditions outlined by Keywords (see the bottom scenario in Fig. 1b ).To implement the Customized Procedure Planner (CPP), we employ a vision-language model built on LLaVa. We experiment with LLaVa-1.5 (Liu et al., 2024a) and LLaVa-NeXT (Liu et al., 2024b) as the backbone of our framework, fine-tuning these models with pseudo-customized labels, as described in section 3.3. The operation of CPP is illustrated in Fig. 2a . The model takes as input o s , a prompt containing the task objective Task, and user-defined conditions Keywords, and generates a sequence of customized action steps p. The zero-shot input prompt structure is shown in Prompt 1.Prompt 1: 'Objective: Compose a detailed sequence of action steps, in order, to complete the task "{Task}" depicted in the image, starting from its current state. Conditions: {Keywords}. Instructions: Ensure that the steps align with the specified conditions and lead to successful task completion.'Customizing Instructional Datasets. Customized Procedure Planning suffers from a lack of sufficient datasets for training. We overcome this limitation by leveraging recent advancements in vision-language models (VLMs) and the capabilities of large language models (LLMs). As shown in Fig. 2b , our novel pipeline collects customizations from the PPIV datasets to build customized instructional video datasets, to use as pseudo labels for training. First, we employ the off-the-shelf vision-language model, GEMINI-1.5-Flash, to extract customization terms for each video sample in the datasets. These keywords are designed to be task-specific and tailored to the video's unique characteristics, as outlined in Prompt 2, which we implement along with a one-shot example response.Prompt 2:'You will be provided with an instructional video that demonstrates a task through a series of ordered action steps (i.e., an instructional plan). Your response should identify up to 3 keywords for the video that are directly related to both the task and the action steps. These keywords should emphasize what distinguishes the video's instructional plan from a generic plan on the same task. For each term, provide a brief explanation of its relevance to the video, the task, and the action steps in one sentence.'Next, we process the extracted keywords and their descriptions of how they relate to the video's instructional plan, alongside the corresponding human-annotated generic plan for the video. Using the GPT-4o LLM, guided by Prompt 3 and a one-shot example response, we generate a customized plan for each video (i.e., pseudo labels for training).'Compose a customized plan for an instructional video, based on the task and the video characteristics. The video includes a sequence of action steps, action-plan, in order. Format the response in one line. Your response should map each action step from the action-plan to a corresponding tailored customized step, maintaining the sequence order, in the format "'action step': tailored step", separated by commas. If you need to include an additional step, use the term "added step".' Weak supervision. With the generated pseudo-labels, we train the Customized Procedure Planner (CPP) using a cross-entropy loss function (Liu et al., 2024a; b) . To improve the model's customization, we further incorporate the large language model (LLM) GPT-4o during training. GPT-4o is tasked with selecting of the best related plan to the Keywords, between two plans, a and bone being the model's prediction and the other the pseudo-label, with the positions of a and b randomized. GPT-4o returns the error rate of its prediction, which is used to modify the overall batch loss.The error rate is computed over the entire batch. For each sample in the batch, if GPT-4o's selection matches the pseudo-label, the sample accuracy is 1; otherwise, the accuracy is 0. The batch accuracy is the average accuracy across all samples in the batch, and the batch error rate is the complement of this accuracy, given by:EQUATIONwhere: N is the batch size, ŷi is the LLM judge's selected plan for the i-th sample, y i is the corresponding pseudo-label for the i-th sample, 1(ŷ i = y i ) is an indicator function that equals 1 if ŷi = y i , and 0 otherwise.The batch error rate is calculated as:EQUATIONThis error rate is scaled by a set positive factor λ and added to the cross-entropy loss to adjust the training process, as described by the following equation:L batch = L CE + λ • Error batch (3)Where L CE is the cross-entropy loss between the predictions and the pseudo-labels, and λ is a learnable scaling factor that controls the impact of the error rate on the batch loss. experiments We conduct experiments on two benchmark datasets, using novel evaluation metrics to validate the effectiveness of our proposed model, and further support our results through human evaluation.We evaluate our methodology using two instructional video datasets: CrossTask (Zhukov et al., 2019) and COIN (Tang et al., 2019) . The CrossTask dataset includes videos across 18 topics, such as "Make French Toast," with an average of 7.6 actions per video. These topics are split into 18 primary and 65 related events. In our study, we focus on the primary subset, which provides precise timestamps for each action, enabling a clear sequence of instructional steps and encompassing 2,750 videos. The COIN dataset contains 11,827 videos covering 778 distinct actions, with an average of 3.6 actions per video. Following recent works (Wang et al., 2023b; Bi et al., 2021b; Chang et al., 2020; Zhao et al., 2022) , we create training and testing splits with a 70/30 ratio. To further enrich the datasets, we apply a moving window approach to organize videos into plans with varying action horizons. Starting from the i-th action, the window extends until the plan is complete (i.e., T = \|p\| -i).Next, we apply the pseudo-label generation pipeline, as detailed in section 3.3 and Fig. 2b , to obtain customized plans for each dataset. This process leads to a more diverse set of action plans across the datasets. Fig. 3 illustrates the expansion of vocabulary in the action plans through word clouds, comparing the generic plans with the added vocabulary for four sample tasks. This emphasizes the open-vocabulary setting and the degree of customization achieved. (Bird et al., 2009) in the visualization.The performance of PPIV models is typically assessed using three standard metrics (Chang et al., 2020; Zhao et al., 2022; Sun et al., 2022; Bi et al., 2021a; Wang et al., 2023b ): 1) Mean Intersection over Union (mIoU) evaluates the overlap between predicted and ground truth action sequences, defined as \|at∩ât\| \|at∪ât\| . This metric indicates whether the model identifies the correct steps but does not account for action order or repetitions. 2) Mean Accuracy (mAcc) measures the alignment of actions at each step, taking into account the order and repetitions of actions. And 3) Success Rate (SR), the strictest metric, which considers a plan successful only if it precisely matches the ground truth.However, all these metrics rely on action labels in both predicted and ground truth sequences, restricting the PPIV setting to a closed-vocabulary framework. This limitation impedes the evaluation of more practical open-vocabulary and varied plan sequences. In this study, we introduce four novel evaluation metrics that retain the essence of the conventional metrics while accommodating this new setting.Automatic Metrics. This study has to quantify the performance of proposed plans in two different dimensions: Planning quality and customization quality. As mentioned, the nature of an openvocabulary framework necessitates a novel approach to standard planning metrics found in previous literature. To this end, we combine Few-Shot-In-Context Learning and LLMs to create automatic metrics that are able to robustly score plans based on the two dimensions.With regards to the quality of planning, we use Few-Shot-In-Context Learning combined with GPT-4o to create two types of sequence mappings from the predicted sequence to the closed-vocabulary generic ground-truth sequence. The first mapping is order mapping. Order mapping is a sequential process that iterates over the ground truth sequence. For each step s n = a Generic n,GT in the sequence, it tries to map it to a corresponding step p m = a Customized m,P red in the predicted sequence. If unable to find a valid corresponding step p m , it denotes the step s n as missing. The mapping proceeds with the next ground truth step s n+1 , which is only able to map to predicted sequence steps p m+1 , ..., p M . This approach preserves the order of the sequence and can be used to calculate mean accuracy (mAcc) and success rate (SR) by aligning open-vocabulary customized plans with their closed-vocabulary counterparts and labels. The second mapping is overlap mapping. This procedure is identical to order mapping except that if s n maps to p m , a follow up step s n+1 can be mapped to any step p 1 , ..., p M as long as s n and s n+1 are distinct. For identical steps, s n+1 cannot map to p m . This type of mapping preserves an understanding of which steps in the ground-truth sequence are present in the predicted sequence, regardless of order. Thus, mIoU can be calculated from this mapping.Under review as a conference paper at ICLR 2025 For each mapping type and dataset, there are between 15 and 20 human-created training examples that are provided to ChatGPT-4o as few-shot examples. We refer to these metrics as automatic SR, mAcc and mIoU (a-SR, a-mAcc, a-mIoU).To assess the quality of customization, we use Few-Shot-In-Context Learning with GPT-4o to generate a "relevance score" that evaluates how well the plan incorporates input keywords. A rubric, scored from 1 to 5, measures this customization, rewarding plans that meaningfully integrate the keywords and penalizing those that lack customization, regardless of overall planning success.1: The plan is not relevant to any of the keywords. 2: The plan is somewhat relevant to a few keywords, but lacks depth. 3: The plan demonstrates a good balance of relevance, either highly relevant to one keyword or moderately relevant to all. 4: The plan is relevant to most keywords, demonstrating a strong application. 5: The plan is highly relevant to all keywords, thoroughly integrating them with clear and meaningful content.Using this rubric, we create 20 examples each for CrossTask and COIN, providing them to the LLM in a few-shot learning setup.Aligned BERT Score (aBERT-Score). To effectively measure the similarity between the predicted sequences and the generic plan, we further introduce a novel metric called aligned BERT Score. This metric is based on BERT similarity score (Zhang et al., 2020) and is calculated by applying an optimal alignment algorithm to both sequences, utilizing a similarity matrix M [i][j]. This matrix captures the cosine similarity between the embeddings of each action pair from the ground-truth sequence a Generic GT and the customized predicted sequence a Customized P red , as defined by the following equation:EQUATIONIn this equation, a Generic i,GT denotes the i th reference action, while a Customized j,P red represents the j th hypothesis action. We then derive the similarity score associated with the trajectory corresponding to the optimal alignment path between the two sequences, which serves as a measure of their similarities. For further details on the workings of this metric, please refer to appendix B.We implement the customized dataset using GEMINI-1.5-Flash as the VLM and GPT-4o mini as the LLM, which also serves as the judge for assessing customization loss during training (eq. ( 2)). To expand the training dataset, we generate pseudo-label customized plans for each sample by leveraging all combinations of sample Keywords, K k , where 0 < k ≤ K, effectively increasing the dataset size.We perform LoRA fine-tuning on the LlaVa-improved (13B parameter) (Liu et al., 2024a) model and full fine-tuning on Llava-Next (Liu et al., 2024b) , training each for three and four epochs to optimize performance on the validation set, respectively. Model evaluation is conducted on a heldout unseen test set. We use an initial learning rate of 2 × 10 -4 with a cosine learning rate scheduler and a training batch size of 16. The training process utilizes four NVIDIA A100 GPUs with 40GB of memory for LlaVa-improved and eight GPUs for Llava-Next. During training, we set λ in eq. ( 3) to 5 × 10 -3 .We assess CPP's performance in comparison to existing large models capable of customized procedure planning for instructional videos. Specifically, we use GPT-4o, a widely recognized and powerful vision-language model, as a baseline under zero-shot and few-shot regimes. Similar to CPP, GPT-4o is prompted with an initial visual observation and corresponding instructions for further details). We compare its performance to CPP models utilizing LlaVA-improved (i.e., LlaVA-1.5) and LlaVA-Next (i.e., LlaVA-1.6) backbones. To distinguish models trained with the customization loss introduced in eq. ( 2), we label them as "with CL" and "w/o CL" (CL referring to customization loss). The results presented in table 1 and table 2 highlight CPP's superiority across automatic metrics, including SR, Acc, mIoU, and aBERT-Score, for datasets CrossTask and Coin, demonstrating its advantages in planning, customization, and overall similarity to ground-truth plans.Notably, CPP with the LlaVA-1.6 backbone outperforms GPT-4o's few-shot performance by 13.08%, 18.32%, and 29.76% in a-SR, a-mAcc, and a-mIoU, respectively, on the CrossTask dataset, and by 14.96%, 20.4%, and 28.87% on the COIN dataset.In terms of customization, CPP performs competitively against GPT-4o, exceeding the a-Relevance score on CrossTask. GPT-4o's high score in this metric, however, results from over-customization of action steps based on the input keywords. GPT-4o leverages its vast prior knowledge to overcustomize plans, adapting them beyond the natural levels found in instructional videos in an attempt to fully satisfy the input prompt.Fig. 4 presents two sample predictions based on the given conditions and visual state for CPP (with the LlaVA-1.6 backbone and CL). As shown, the model accurately understands the initial task state and generates a plan that successfully meets the keyword conditions through to completion.The integration of the customization loss into the overall objective function of the model (eq. ( 3)) significantly enhances CPP's performance, as illustrated in tables 3 and 4. In the CrossTask dataset, the a-Relevance score increases by 14 points, while the COIN dataset sees a rise of 25 points. Furthermore, this loss functions as a regularization mechanism, contributing to an overall improvement in planning scores, with a 1.26% increase in success rate for the COIN dataset. The tables also include the p-values for improvements in the a-Relevance score, highlighting the significance of these enhancements for each backbone. In this study, we tackled the novel challenge of customized procedure planning in instructional videos by developing the Customized Procedure Planner (CPP) framework. Unlike previous approaches in CPPIV, which were limited to using only initial and final visual observations for procedure induction, CPP generates plans in a causal setting based on initial observations, along with user and task-specific requirements. CPP surpasses the state-of-the-art in existing models. A key innovation is the use of weak supervision by customizing existing PPIV datasets. This is achieved by extracting video-specific customization information from video samples and utilizing advanced LLMs. Our model also incorporates a novel LLM-based objective function during training to further enhance customization. We evaluate CPP using new metrics designed specifically for this setting, demonstrating its superiority in CPPIV. Looking ahead, we see potential for applying CPP to more diverse scenarios and generating customized plans for unseen tasks. Additionally, developing a high-quality customized dataset will pave the way for more advanced applications in this field.		[ { "caption": "Figure 3: Expansion of vocabulary in action plans as the result of customization pipeline. The word clouds compare generic plans (top) with the added vocabulary (bottom) for four sample tasks, showcasing the open-vocabulary setting and customization on the CrossTask dataset. Stop-words are exclude...	false	null	[ { "limitations": null, "main_review": null, "paper_summary": "paper_summary: The paper addresses the issue of generating customized procedures for task planning in instructional videos. Existing methods face challenges like overlooking customization and lacking proper datasets. The contributions are sig...	[ { "comment": "Thanks for your response. My concerns are partially addressed. I decide to keep my score, as the paper still some distance from the acceptance threshold and needs further refinement.", "title": "Response to the authors" }, { "comment": "Thank you for your response. My concerns are part...	[ 0.6000000000000001, 0.6000000000000001, 0.4, 0.4 ]	[ 0.75, 0.75, 1, 0.75 ]	[ 0.6666666666666666, 0.6666666666666666, 0.3333333333333333, 0.6666666666666666 ]	[ 0.3333333333333333, 0.3333333333333333, 0.6666666666666666, 0.6666666666666666 ]	[ 0.3333333333333333, 0.3333333333333333, 0.3333333333333333, 0.3333333333333333 ]	[ null, null, null, null ]
A Black Swan Hypothesis: The Role of Human Irrationality in AI Safety	"Black swan events are statistically rare occurrences that carry extremely high risks. A typical vie(...TRUNCATED)	["AI Safety","Risk","Reinforcement Learning","alignment, fairness, safety, privacy, and societal con(...TRUNCATED)	2024-10-04	CC BY 4.0	["agarwal\|model-based_reinforcement_learning_with_a_generative_model_is_minimax_optimal","agrawal\|re(...TRUNCATED)	lee\|a_black_swan_hypothesis_the_role_of_human_irrationality_in_ai_safety\|ICLR_cc_2025_Conference	"introduction\nTo successfully deploy machine learning (ML) systems in open-ended environments, thes(...TRUNCATED)	[{"src":"https://huggingface.co/static-proxy/datasets-server.huggingface.co/assets/nhop/ReviewBench/--/{dataset_git_revision}/--(...TRUNCATED)	[{"caption":"Figure 1: Value distortion function u and probability distortion function w. The gray l(...TRUNCATED)	true	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: This paper challenges the co(...TRUNCATED)	[{"comment":"Dear Reviewer WW6X, \n\nThank you for your thoughtful feedback, especially on how the r(...TRUNCATED)	[ 0.6000000000000001, 0.6000000000000001, 0.6000000000000001, 0.4 ]	[ 0.5, 0.25, 0.25, 0.5 ]	[ 0.6666666666666666, 0.6666666666666666, 0.6666666666666666, 0.6666666666666666 ]	[ 1, 0.3333333333333333, 0.6666666666666666, 0.3333333333333333 ]	[ 0.6666666666666666, 0.3333333333333333, 0.3333333333333333, 0.3333333333333333 ]	[ null, null, null, null ]
Network-based Active Inference and its Application in Robotics	"This paper introduces Network-based Active Inference (NetAIF), a novel robotic framework that enabl(...TRUNCATED)	["Active Inference (AIF)","Free Energy Principle (FEP)","Robotics","Trajectory generation","Random d(...TRUNCATED)	2024-10-04	CC BY 4.0	["anonymous\|network-based_active_inference_for_adaptive_and_cost-efficient_real-world_applications:_(...TRUNCATED)	yoon\|networkbased_active_inference_and_its_application_in_robotics\|ICLR_cc_2025_Conference	"introduction 1.overcoming automation challenges with advanced learning methods\nThe World Energy Em(...TRUNCATED)	[{"src":"https://huggingface.co/static-proxy/datasets-server.huggingface.co/assets/nhop/ReviewBench/--/{dataset_git_revision}/--(...TRUNCATED)	[ { "caption": "Table 2: Summary of time taken to generate values by the network", "figType": "Table" } ]	false	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: The work at first sight is v(...TRUNCATED)	[{"comment":"Dear Authors,\nThanks so much for your effort to address answer to my questions. While (...TRUNCATED)	[ 0.2, 0.6000000000000001, 0.2, 0 ]	[ 1, 0.75, 0.5, 0.75 ]	[ 0.3333333333333333, 0.6666666666666666, 0.3333333333333333, 0 ]	[ 0.6666666666666666, 0.6666666666666666, 0, 0 ]	[ 0.6666666666666666, 1, 0.3333333333333333, 0 ]	[ null, null, null, null ]
VR-Sampling: Accelerating Flow Generative Model Training with Variance Reduction Sampling	"Recent advancements in text-to-image and text-to-video models, such as Stable Diffusion 3 (SD3), Fl(...TRUNCATED)	[ "Flow Generative Models", "Training Acceleration", "Diffusion Models", "generative models" ]	2024-10-04	CC BY 4.0	["abramson\|accurate_structure_prediction_of_biomolecular_interactions_with_alphafold_3","arjevani\|lo(...TRUNCATED)	"pan\|vrsampling_accelerating_flow_generative_model_training_with_variance_reduction_sampling\|ICLR_cc(...TRUNCATED)	"introduction\nDiffusion models (Song et al., 2021b; Ho et al., 2020; Dhariwal & Nichol, 2021; Song (...TRUNCATED)		[]	false	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: The authors theoretically id(...TRUNCATED)	[{"comment":"Dear Authors,\n\nThank you for your clarification. My concerns are mostly addressed. Th(...TRUNCATED)	[ 0.2, 0.8, 0.4, 0.8 ]	[ 0.75, 0.5, 0.5, 0.5 ]	[ 0, 0.6666666666666666, 0.6666666666666666, 0.6666666666666666 ]	[ 0.3333333333333333, 0.3333333333333333, 0.6666666666666666, 0.6666666666666666 ]	[ 0.3333333333333333, 0.6666666666666666, 0.3333333333333333, 0.6666666666666666 ]	[ null, null, null, null ]
What Matters in Hierarchical Search for Combinatorial Reasoning Problems?	"Combinatorial reasoning problems, particularly the notorious NP-hard tasks, remain a significant ch(...TRUNCATED)	["deep learning","search","subgoals","hierarchical reinforcement learning","imitation learning","rei(...TRUNCATED)	2024-10-04	CC BY 4.0	["achiam\|constrained_policy_optimization","andrychowicz\|what_matters_in_on-policy_reinforcement_lear(...TRUNCATED)	"zawalski\|what_matters_in_hierarchical_search_for_combinatorial_reasoning_problems\|ICLR_cc_2025_Conf(...TRUNCATED)	"introduction\nFigure 1 : Performance comparison of hierarchical methods (AdaSubS, kSubS) and low-le(...TRUNCATED)	[{"src":"https://huggingface.co/static-proxy/datasets-server.huggingface.co/assets/nhop/ReviewBench/--/{dataset_git_revision}/--(...TRUNCATED)	[{"caption":"Figure 1: Performance comparison of hierarchical methods (AdaSubS, kSubS) and low-level(...TRUNCATED)	false	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: This paper reports on an emp(...TRUNCATED)	[{"comment":"Thank you for your constructive feedback that helped us strengthen our work. Let us sum(...TRUNCATED)	[ 0.4, 0.6000000000000001, 0.4, 0.6000000000000001 ]	[ 0.75, 1, 0.5, 0.75 ]	[ 0.6666666666666666, 1, 0.3333333333333333, 0.6666666666666666 ]	[ 0.3333333333333333, 1, 0.3333333333333333, 0.3333333333333333 ]	[ 0.3333333333333333, 0.3333333333333333, 0.3333333333333333, 0.3333333333333333 ]	[ null, null, null, null ]
Self-Play Preference Optimization for Language Model Alignment	"Standard reinforcement learning from human feedback (RLHF) approaches relying on parametric models (...TRUNCATED)	["self play","preference optimization","large language model","RLHF","alignment, fairness, safety, p(...TRUNCATED)	2024-10-04	CC BY 4.0	["ahmadian\|back_to_basics:_revisiting_reinforce-style_optimization_for_learning_from_human_feedback_(...TRUNCATED)	wu\|selfplay_preference_optimization_for_language_model_alignment\|ICLR_cc_2025_Conference	"introduction\nLarge Language Models (LLMs) (e.g., Ouyang et al., 2022; OpenAI et al., 2023) , have (...TRUNCATED)	[{"src":"https://huggingface.co/static-proxy/datasets-server.huggingface.co/assets/nhop/ReviewBench/--/{dataset_git_revision}/--(...TRUNCATED)	[{"caption":"Table 6: Another generation example of our fine-tuned model by SPPO at different iterat(...TRUNCATED)	true	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: The paper proposes a novel s(...TRUNCATED)	[{"comment":"Thanks for providing the detailed explanation. I increase my contribution score to 3. T(...TRUNCATED)	[ 0.6000000000000001, 0.6000000000000001, 0.6000000000000001, 0.6000000000000001 ]	[ 0.75, 0.75, 0.5, 0.75 ]	[ 0.6666666666666666, 1, 0.6666666666666666, 0.6666666666666666 ]	[ 0.6666666666666666, 0.6666666666666666, 0.6666666666666666, 0.6666666666666666 ]	[ 0.6666666666666666, 0.6666666666666666, 0.6666666666666666, 0.6666666666666666 ]	[ null, null, null, null ]
ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy	"Large-scale cell microscopy screens are used in drug discovery and molecular biology research to st(...TRUNCATED)	["MAE","microscopy","transformers","SSL","linear probing","biology","high-content screening","founda(...TRUNCATED)	2024-10-04	CC BY 4.0	["spearman\|rxrx3),_b__=_12_mae-l/8_(rpi-93m),_b_=_24_mae-l/8_(rpi-93m),_b__=_15_mae-l/8_(pp-16m),_(...TRUNCATED)	"kenyondean\|vitally_consistent_scaling_biological_representation_learning_for_cell_microscopy\|ICLR_c(...TRUNCATED)	"introduction\nLarge-scale cell microscopy assays are used to discover previously unknown biological(...TRUNCATED)	[{"src":"https://huggingface.co/static-proxy/datasets-server.huggingface.co/assets/nhop/ReviewBench/--/{dataset_git_revision}/--(...TRUNCATED)	[{"caption":"Table 4: Overview of vision transformer (ViT) encoders used and evaluated in this work.(...TRUNCATED)	false	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: The authors presented a new (...TRUNCATED)	[{"comment":"I've changed my score to 5, however, I still think that the current version is not yet (...TRUNCATED)	[ 0.2, 0.4, 0.8, 0.4 ]	[ 0.75, 1, 0.5, 0.75 ]	[ 0.6666666666666666, 0.3333333333333333, 0.6666666666666666, 0 ]	[ 0.6666666666666666, 0.6666666666666666, 1, 1 ]	[ 0.3333333333333333, 0.3333333333333333, 0.6666666666666666, 0 ]	[ null, null, null, null ]
Synthetic continued pretraining	"Pretraining on large-scale, unstructured internet text enables language models to acquire a signifi(...TRUNCATED)	["large language model","synthetic data","continued pretraining","foundation or frontier models, inc(...TRUNCATED)	2024-10-04	CC BY 4.0	["abdin\|phi-2:_the_surprising_power_of_small_language_models","akyürek\|deductive_closure_training_o(...TRUNCATED)	yang\|synthetic_continued_pretraining\|ICLR_cc_2025_Conference	"introduction\nLanguage models (LMs) have demonstrated a remarkable ability to acquire knowledge fro(...TRUNCATED)	[{"src":"https://huggingface.co/static-proxy/datasets-server.huggingface.co/assets/nhop/ReviewBench/--/{dataset_git_revision}/--(...TRUNCATED)	[{"caption":"Figure 2: Accuracy on the QuALITY question set Qtest (y-axis) as a function of the synt(...TRUNCATED)	true	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: This paper addresses the pro(...TRUNCATED)	[{"comment":"Thanks! Will keep my positive score.","title":null},{"comment":"Thank you for your clar(...TRUNCATED)	[ 0.8, 0.8, 0.8, 0.8 ]	[ 0.75, 0.5, 0.75, 0.5 ]	[ 1, 1, 0.3333333333333333, 1 ]	[ 1, 0.6666666666666666, 0.6666666666666666, 0.6666666666666666 ]	[ 0.6666666666666666, 0.6666666666666666, 0.3333333333333333, 0.6666666666666666 ]	[ null, null, null, null ]
On the Optimization Landscape of Low Rank Adaptation Methods for Large Language Models	"Training Large Language Models (LLMs) poses significant memory challenges, making low-rank adaptati(...TRUNCATED)	[ "large language model", "LoRA", "optimization", "foundation or frontier models, including LLMs" ]	2024-10-04	CC BY 4.0	["aghajanyan\|intrinsic_dimensionality_explains_the_effectiveness_of_language_model_fine-tuning","all(...TRUNCATED)	"liu\|on_the_optimization_landscape_of_low_rank_adaptation_methods_for_large_language_models\|ICLR_cc_(...TRUNCATED)	"introduction\nLarge Language Models (LLMs) have demonstrated impressive performance across various (...TRUNCATED)	[{"src":"https://huggingface.co/static-proxy/datasets-server.huggingface.co/assets/nhop/ReviewBench/--/{dataset_git_revision}/--(...TRUNCATED)	[{"caption":"Table 11: The mean and standard deviation of GaRare on various sizes of LLaMA models on(...TRUNCATED)	true	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: This paper builds on previou(...TRUNCATED)	[{"comment":"Dear Reviewer eja2,\n\nAs the discussion period is nearing its close, we would like to (...TRUNCATED)	[ 0.6000000000000001, 0.8, 0.8, 0.4, 0.6000000000000001, 0.4 ]	[ 0.75, 0.5, 0.25, 0.5, 0.75, 0.75 ]	[ 0.6666666666666666, 1, 0.6666666666666666, 0.6666666666666666, 0.6666666666666666, 0.6666666666666666 ]	[0.6666666666666666,0.6666666666666666,0.6666666666666666,0.3333333333333333,0.6666666666666666,0.66(...TRUNCATED)	[0.6666666666666666,0.6666666666666666,0.6666666666666666,0.3333333333333333,0.3333333333333333,0.66(...TRUNCATED)	[ null, null, null, null, null, null ]
OSCAR: Operating System Control via State-Aware Reasoning and Re-Planning	"Large language models (LLMs) and large multimodal models (LMMs) have shown great potential in autom(...TRUNCATED)	["Large Language Model","Autonomous Agent","Graphical User Interface","applications to robotics, aut(...TRUNCATED)	2024-10-04	CC BY 4.0	["achiam\|shyamal_anadkat,_et_al._gpt-4_technical_report","chen\|a_dataset_for_gui-oriented_multimodal(...TRUNCATED)	"wang\|oscar_operating_system_control_via_stateaware_reasoning_and_replanning\|ICLR_cc_2025_Conference(...TRUNCATED)	"introduction\nLarge Language Models (LLMs) (Ouyang et al., 2022; Achiam et al., 2023; Dubey et al.,(...TRUNCATED)	[{"src":"https://huggingface.co/static-proxy/datasets-server.huggingface.co/assets/nhop/ReviewBench/--/{dataset_git_revision}/--(...TRUNCATED)	[{"caption":"Figure 4: Illustration of task-driven re-planning and code-centric control in OSCAR. Ba(...TRUNCATED)	true	null	[{"limitations":null,"main_review":null,"paper_summary":"paper_summary: - This work presents an LLM+(...TRUNCATED)	[{"comment":"Thanks for addressing my concerns. I've updated my rating.","title":null},{"comment":"T(...TRUNCATED)	[ 0.8, 0.6000000000000001, 0.8, 0.6000000000000001 ]	[ 0.5, 0.75, 0.5, 0.75 ]	[ 0.6666666666666666, 1, 0.6666666666666666, 0.3333333333333333 ]	[ 1, 0.6666666666666666, 0.6666666666666666, 0.3333333333333333 ]	[ 0.6666666666666666, 0.6666666666666666, 0.6666666666666666, 0.3333333333333333 ]	[ null, null, null, null ]

End of preview.

README.md exists but content is empty.

Downloads last month: 56