\\\\sigma_{\\\\mathrm{min}}^{2}$. Note that when $\\\\sigma_{\\\\mathrm{max}}^{2}=\\\\sigma_{\\\\mathrm{min}}^{2}$, we recover the same probability path as in Sec. 2.2. However, we require $\\\\sigma_{\\\\mathrm{max}}^{2}>\\\\sigma_{\\\\mathrm{min}}^{2}$ to obtain a score function in our vector field.\\n\\nUsing Theorem 3 from Lipman et al. (), we construct the corresponding vector field:\\n\\n$$ $\\\\displaystyle u_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{z})=\\\\mathbf{x}_{1}-\\\\mathbf{x}_{0 }+\\\\frac{\\\\sigma_{\\\\mathrm{min}}^{2}-\\\\sigma_{\\\\mathrm{max}}^{2}}{(1-t)\\\\sigma_{ \\\\mathrm{max}}^{2}+t\\\\sigma_{\\\\mathrm{min}}^{2}}\\\\left(\\\\mathbf{x}_{t}-(1-t)\\\\mathbf {x}_{0}-t\\\\mathbf{x}_{1}\\\\right).$ (18) $$\\n\\nWe observe that the vector field can be expressed using the score function of the conditional probability paths in Eq. 17:\\n\\n$$ $\\\\displaystyle u_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{z})=\\\\mathbf{x}_{1}-\\\\mathbf{x}_{0 }-(\\\\sigma_{\\\\mathrm{min}}^{2}-\\\\sigma_{\\\\mathrm{max}}^{2})\\\\nabla_{\\\\mathbf{x}_{t}} \\\\log p_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{z})$ (19) $$\\n\\nHere, we used the fact that the score function of the Gaussian in Eq. 17 is:\\n\\n$$ $\\\\displaystyle\\\\nabla_{\\\\mathbf{x}_{t}}\\\\log p_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{z})=- \\\\frac{\\\\mathbf{x}_{t}-(1-t)\\\\mathbf{x}_{0}-t\\\\mathbf{x}_{1}}{(1-t)\\\\sigma_{\\\\mathrm {max}}^{2}+t\\\\sigma_{\\\\mathrm{min}}^{2}}.$ (20) $$\\n\\n##### Marginal Vector Field.\\n\\nBy integrating over $\\\\mathbf{z}$ we can rewrite the marginal vector field:\\n\\n$$ $\\\\displaystyle u_{t}(\\\\mathbf{x}_{t})=$ $\\\\displaystyle\\\\int\\\\frac{u_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{z})p_{t}(\\\\mathbf{x}_{t} \\\\mid\\\\mathbf{z})q(\\\\mathbf{z})}{p_{t}(\\\\mathbf{x}_{t})}\\\\mathrm{d}\\\\mathbf{z}$ (21) $\\\\displaystyle=$ $\\\\displaystyle\\\\int\\\\frac{(\\\\mathbf{x}_{1}-\\\\mathbf{x}_{0}-(\\\\sigma_{\\\\mathrm{min}}^{ 2}-\\\\sigma_{\\\\mathrm{max}}^{2})\\\\nabla_{\\\\mathbf{x}_{t}}\\\\log p_{t}(\\\\mathbf{x}_{t} \\\\mid\\\\mathbf{z}))p_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{z})q(\\\\mathbf{z})}{p_{t}( \\\\mathbf{x}_{t})}\\\\mathrm{d}\\\\mathbf{z}$ (22) $\\\\displaystyle=$ $\\\\displaystyle\\\\int\\\\frac{(\\\\mathbf{x}_{1}-\\\\mathbf{x}_{0})p_{t}(\\\\mathbf{x}_{t}\\\\mid \\\\mathbf{z})q(\\\\mathbf{z})}{p_{t}(\\\\mathbf{x}_{t})}\\\\mathrm{d}\\\\mathbf{z}-(\\\\sigma_{ \\\\mathrm{min}}^{2}-\\\\sigma_{\\\\mathrm{max}}^{2})\\\\int\\\\frac{\\\\nabla_{\\\\mathbf{x}_{t}} \\\\log p_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{z})p_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{z})q( \\\\mathbf{z})}{p_{t}(\\\\mathbf{x}_{t})}\\\\mathrm{d}\\\\mathbf{z}$ (23) $\\\\displaystyle=$ $\\\\displaystyle v_{t}(\\\\mathbf{x}_{t})-(\\\\sigma_{\\\\mathrm{min}}^{2}-\\\\sigma_{\\\\mathrm {max}}^{2})\\\\nabla_{\\\\mathbf{x}_{t}}\\\\log p_{t}(\\\\mathbf{x}_{t}),$ (24) $$\\n\\nwhere $v_{t}(\\\\mathbf{x}_{t})$ represents the expected drift term $\\\\mathbb{E}_{\\\\mathbf{z}\\\\mid\\\\mathbf{x}_{t}}\\\\left[\\\\mathbf{x}_{1}-\\\\mathbf{x}_{0}\\\\right]$.\\n\\n##### Conditional Score Function.\\n\\nTo incorporate the observed past $\\\\mathbf{y}^{p}$, we substitute the score function with its conditional counterpart:\\n\\n$$ $\\\\displaystyle\\\\nabla_{\\\\mathbf{x}_{t}}\\\\log p_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{y}^{p })=\\\\nabla_{\\\\mathbf{x}_{t}}\\\\log p_{t}(\\\\mathbf{x}_{t})+\\\\nabla_{\\\\mathbf{x}_{t}} \\\\log p_{t}(\\\\mathbf{y}^{p}\\\\mid\\\\mathbf{x}_{t}),$ (25) $$\\n\\nleading to the following modified vector field:\\n\\n$$ $\\\\displaystyle\\\\tilde{u}_{t}(\\\\mathbf{x}_{t})=\\\\underbrace{v_{t}(\\\\mathbf{x}_{t})-( \\\\sigma_{\\\\mathrm{min}}^{2}-\\\\sigma_{\\\\mathrm{max}}^{2})\\\\nabla_{\\\\mathbf{x}_{t}} \\\\log p_{t}(\\\\mathbf{x}_{t})}_{=u_{t}(x_{t})}-\\\\underbrace{\\\\lambda(\\\\sigma_{ \\\\mathrm{min}}^{2}-\\\\sigma_{\\\\mathrm{max}}^{2})}_{:=s}\\\\nabla_{\\\\mathbf{x}_{t}}\\\\log p _{t}(\\\\mathbf{y}^{p}\\\\mid\\\\mathbf{x}_{t}),$ (26) $$\\n\\nwhere $\\\\lambda$ and $s$ are scaling parameters introduced to control the strength of the conditioning. We can summarize this as:\\n\\n$$ $\\\\displaystyle\\\\tilde{u}_{t}(\\\\mathbf{x}_{t})=u_{t}(\\\\mathbf{x}_{t})-s\\\\nabla_{ \\\\mathbf{x}_{t}}\\\\log p_{t}(\\\\mathbf{y}^{p}\\\\mid\\\\mathbf{x}_{t}).$ (27) $$\\n\\nThus, to sample from the conditional generative model $p_{t}(\\\\mathbf{x}_{t}\\\\mid\\\\mathbf{y}^{p})$, we use the modified vector field $\\\\tilde{u}_{t}(\\\\mathbf{x}_{t})$ from Eq. 27 with our trained vector field $u_{\\\\bm{\\\\theta}}(t,\\\\mathbf{x}_{t})$.\",\n \"metadata\": {\n \"authors\": [\n \"Marcel Kollovieh, Marten Lienen, David Lüdke, Leo Schwinn, Stephan Günnemann\",\n \"Department of Informatics & Munich Data Science Institute\",\n \"Technical University of Munich, Germany\",\n \"{\",\n \"}\",\n \"@tum.de\"\n ],\n \"title\": \"Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting\",\n \"total_chars\": 64999,\n \"total_lines\": 578,\n \"updated_at\": \"2026-01-25T15:06:48.228340\"\n }\n}"},"reviews":{"kind":"list like","value":[{"id":"tBf42IhNBQ","initial_rating":6,"confidence":2,"soundness":3,"contribution":3,"presentation":3,"summary":"The existing diffusion models have problems in the time series generation since the data and prior distributions differ. The authors handle this problem by utilizing conditional flow matching framework. They propose TSFlow, which sets the prior distribution as Gaussian process to make the prior distribution close to the data distribution. Also, they propose conditional prior sampling which makes an unconditionally trained model possible for probabilistic forecasting.","strengths":"• By utilizing Gaussian process to the conditional flow matching, the model reflects the temporal dependencies of the given time series data better.\n\n• The model enables both unconditional and conditional generations.\n\n• By conditional prior sampling, the unconditionally trained model could follow the given guidance.","weaknesses":"Please refer to the Questions section.","questions":"•\tThe problem only considers the univariate case. Can the model extend to the multivariate time series problem?\n\n•\tI want some more explanation about the effectiveness of informed prior distributions. Why does closedness of the prior and data distribution imply easy learning. Do you have any experiments about train efficiency or path efficiency? \n\n•\tCan the given prior (Gaussian process) extend to the arbitrary prior ? for example, refer to [1].\n\n•\tDo you have any theoretical evidence about how the selection of kernel functions effect to the model performance? (ex. OU kernel is better when the data follows OU process)\n\n[1] Leveraging Priors via Diffusion Bridge for Time Series Generation, Arxiv24"},{"id":"r8kTEkO0yD","initial_rating":6,"confidence":4,"soundness":4,"contribution":2,"presentation":4,"summary":"This paper studies the use of flow matching for time series forecasting. The authors first propose the use of flow matching techniques for unconditional generation using Gaussian process priors (section 3.1.1), followed by a technique for conditioning an unconditional model by sampling a relevant prior $x_0$ (section 3.1.2) or guidance (section 3.1.3), and finally the authors discuss a technique which uses a data-dependent Gaussian process prior for conditional sampling (section 3.2). The proposed methodology is empirically validated on several univariate time series datasets.","strengths":"- The investigation of techniques for conditional sampling is pretty thorough, and I think the proposed methods are quite interesting\n- The proposed method obtains fairly strong empirical results, and the empirical evaluation is convincing \n- Throughout the paper is very clear and well-written","weaknesses":"- There is some highly relevant related work that the authors do not discuss. [Functional Flow Matching, AISTATS 2024](https://arxiv.org/abs/2305.17209) proposes the use of GP priors in conjunction with flow matching and studies techniques for forecasting with these models. Similarly, [Conditional Flow Matching for Time Series Modeling, SPIGM@ICML 2024](https://openreview.net/forum?id=Hqn4Aj7xrQ) uses GPs with flow matching for time series. The authors should cite these works and discuss the differences with their proposed method.\n- There are some (relatively minor) clarity issues throughout\n - The use of equation 9 was a bit unclear to me. Why is this specific form of $q_1$ chosen? Some justification for this modeling choice would be good.\n - In Section 3.1.2, I am guessing that once we sample $x_0 \\sim q_0(x_0 \\mid y^p)$, then we use $x_0$ as an initial condition for the flow model to generate new samples $y \\mid x_0$. Is this the case? If so, it would help to state this explicitly somewhere in the paper.\n - In Line 303, the authors write \"approximating q_0(x_0 \\mid y^p)$ with $q_0(x_0 \\mid y^p)$. I think I understand what is meant here, but this seems to be a typo.","questions":"- The exact problem statement was a bit unclear to me. In lines 154-156 the authors describe a time series as a vector in $\\mathbb{R}^L$. Does this mean that the authors only work on time series having a fixed length $L$, or does the method allow for variable-length time series? Similarly, are the authors assuming that the time series all share a fixed discretization (i.e., there are some fixed times $t_1, \\dots, t_L$ corresponding to $y_1, \\dots, y_L$), or can the discretization vary across time series? Is this discretization assumed to be uniform, or can it be irregular as well?\n- It seems to me that the setup is not limited just to forecasting, but could be applied to general conditional generation tasks, e.g., imputation. Have the authors tried anything beyond forecasting?"},{"id":"Ht8UdSWjui","initial_rating":5,"confidence":2,"soundness":3,"contribution":2,"presentation":2,"summary":"This paper introduces TSFlow, a model for probabilistic time series forecasting that enhances generative modeling by incorporating Gaussian Process (GP) priors within the Conditional Flow Matching (CFM) framework. The use of more informative GP priors helps align the prior distribution with the temporal structure of the data, potentially improving both performance and leading to runtime efficiency by simplifying the probability paths. TSFlow is flexible and can be used for both, conditional and unconditional, time-series generation.\n\nThe model demonstrates strong performance across several benchmark datasets, but there are aspects of the paper that could benefit from further clarification and improvement. Depending on the author's response, I am willing to increase my score from \"weak reject\" to \"weak accept\".","strengths":"**Incorporating Gaussian Process Priors:** The main contribution—replacing the typical isotropic Gaussian prior $q(x_0)$ with a data-dependent conditional prior $q(x_0∣y^p)$ is well-motivated. GP priors are naturally suited for time series due to their ability to model temporal dependencies, and this idea is a clear innovation over existing flow matching methods.\n\n**Empirical Performance:** The empirical results show that TSFlow performs well compared to state-of-the-art models across various benchmark tasks.\n\n**Flexibility in Conditional and Unconditional Modeling:** The approach supports both unconditional and conditional generation. While unconditional generation has fewer use cases compared to conditional generation, it is a feature often overlooked in time-series analysis. The ability to use the same model for both tasks—by applying conditioning only during inference—adds versatility to the paper.","weaknesses":"**Majors**\n\n**Difficulty in Parsing for Non-Experts:** The paper assumes substantial familiarity with flow matching and related generative methods, making it challenging for readers without a deep background in these specific techniques. It took for me considerable time to fully grasp the concepts, which suggests that the paper might also be difficult to read for a broader audience.\n\n**Lack of Runtime Analysis:** The authors propose replacing the isotropic Gaussian prior with a more complex GP prior. Howeer, they do not address the increased computational cost that comes with using GP priors. Although the GP prior is more suited to time-series tasks, this advantage must be weighed against the computational overhead. The paper should clearly state the theoretical runtime complexity of using GP priors and provide empirical runtime comparisons in the experiments. The reduction in NFEs is a positive, but its trade-off with the computational cost of the GP prior must be considered.\n\n**Baseline Comparison:** A simple baseline for the forecasting tasks could involve using Eq. 6 but with an isotropic Gaussian prior. Please include this method as a comparison partner in Table 3. If it is not a valid approach, it should be explained why such a baseline is excluded.\n\n**Inconsistent Findings on Kernel Choice:** The periodic kernel minimizes the Wasserstein distance in Figure 2, suggesting it aligns well with the data distribution. However, in Table 1, the periodic kernel does not significantly outperform other kernels, and it is not even considered in Table 4.2. This inconsistency is counterintuitive and warrants further discussion. Moreover, the necessity for different hyperparameter choices across tasks (e.g., generative modeling vs. forecasting) weakens the \"one model for all\" argument, suggesting more task-specific tuning may be required.\n\n**Minors:**\n\n**Missing Experiments on Guided Generation (3.1.3):** I could not find experiments on Guided Generation (3.1.3) in the paper. Where can I find them?\n\n**LPS Score Clarification:** The Linear Predictive Score (LPS) is not sufficiently explained. Please provide more details on how the LPS is calculated, specifically on what the linear model is regressed against and what the output represents.\n\n**GP Hyperparameter Optimization:** You state in your paper that you did not fit the GP hyperparameters. This is somewhat unexpected, I would have had a strong guess that learning the hyperparameters of the GP brings a large benefit to the model. Can you perform a small experiment to evaluate the difference in performance with/without GP hyperparameter optimization?","questions":"See above."},{"id":"oywKq5KOZr","initial_rating":5,"confidence":3,"soundness":3,"contribution":3,"presentation":2,"summary":"This paper proposes a model that uses flow-matching algorithms to generate and forecast univariate time-series. When trained unconditionally, it replaces the isotropic Gaussian prior distribution with an informed one and uses conditional sampling and inferencing. The model itself can also be made conditional. Experiments show that the model surpasses existing ones on benchmark tasks.","strengths":"1. The model is flexible with many possible options; the informed prior sampling is intuitive.\n\n2. The model shows promising results on univariate benchmark tasks.","weaknesses":"1. The model right now is only discussed and validated on univariate time-series.\n\n2. While there are many proposed options in the model, there lacks an ablation or a comprehensive comparison of different choices. Neither is there any theoretical insights in the proposed model.\n\n3. The presentation of the model needs a bit more clarification. The problem formulation can potentially be expanded a bit more (see questions below). The discussion of different training/inference choices is a bit dense. Maybe some concise pseudocode or flowchart in the main text could be helpful.","questions":"1. If my understanding is correct, the overall idea of the model is to push an initial GP to the target time-series. Therefore, the sequence length $L$ in the problem formulation section is equivalent to the vector field dimension $d$ in the background section. Can the author(s) kindly confirm if this is true? It would be nice to clarify the setting a little bit. I think the main source of confusion is that there are two time indices in this paper: the time in flow-matching and the time in the time-series, and these are orthogonal to each other. I think this should be made clear somewhere in the paper.\n\n2. Following the first question, it would be nice to indicate in Figure 1 the time-series index and the flow-matching index. Moreover, different notations should be used. (For instance, on line 202-203, $t$ should not be used for the kernel because it has already been used in Eq. (1).)\n\n3. How would the model respond to a growing $L$? That is, there are two subquestions in this query:\n\ta. What is the time complexity of the model and how does it compare to other models that you benchmarked against?\n\tb. When $L$ is large, Eq. (6) seems to suffer from the curse of dimensionality and the integral would be impossible to discretize. How does the model see this issue, or is it not relevant?\n\n4. The section \"Effect on the Optimal Transport Problem\" only considers the distance between sequences but not anything about the training. Can you show some experiments where models that use the periodic kernel are indeed easier to train on tasks that involve periodicity?\n\n5. On line 302-303, you wrote \"we additionally condition the prior distribution on the observed past data by approximating $q_0(\\mathbf{x}_0 | \\mathbf{y}^p)$ with $q_0(\\mathbf{x}_0 | \\mathbf{y}^p)$.\" I assume there is a typo. What would be the intended sentence?"}],"string":"[\n {\n \"id\": \"tBf42IhNBQ\",\n \"initial_rating\": 6,\n \"confidence\": 2,\n \"soundness\": 3,\n \"contribution\": 3,\n \"presentation\": 3,\n \"summary\": \"The existing diffusion models have problems in the time series generation since the data and prior distributions differ. The authors handle this problem by utilizing conditional flow matching framework. They propose TSFlow, which sets the prior distribution as Gaussian process to make the prior distribution close to the data distribution. Also, they propose conditional prior sampling which makes an unconditionally trained model possible for probabilistic forecasting.\",\n \"strengths\": \"• By utilizing Gaussian process to the conditional flow matching, the model reflects the temporal dependencies of the given time series data better.\\n\\n• The model enables both unconditional and conditional generations.\\n\\n• By conditional prior sampling, the unconditionally trained model could follow the given guidance.\",\n \"weaknesses\": \"Please refer to the Questions section.\",\n \"questions\": \"•\\tThe problem only considers the univariate case. Can the model extend to the multivariate time series problem?\\n\\n•\\tI want some more explanation about the effectiveness of informed prior distributions. Why does closedness of the prior and data distribution imply easy learning. Do you have any experiments about train efficiency or path efficiency? \\n\\n•\\tCan the given prior (Gaussian process) extend to the arbitrary prior ? for example, refer to [1].\\n\\n•\\tDo you have any theoretical evidence about how the selection of kernel functions effect to the model performance? (ex. OU kernel is better when the data follows OU process)\\n\\n[1] Leveraging Priors via Diffusion Bridge for Time Series Generation, Arxiv24\"\n },\n {\n \"id\": \"r8kTEkO0yD\",\n \"initial_rating\": 6,\n \"confidence\": 4,\n \"soundness\": 4,\n \"contribution\": 2,\n \"presentation\": 4,\n \"summary\": \"This paper studies the use of flow matching for time series forecasting. The authors first propose the use of flow matching techniques for unconditional generation using Gaussian process priors (section 3.1.1), followed by a technique for conditioning an unconditional model by sampling a relevant prior $x_0$ (section 3.1.2) or guidance (section 3.1.3), and finally the authors discuss a technique which uses a data-dependent Gaussian process prior for conditional sampling (section 3.2). The proposed methodology is empirically validated on several univariate time series datasets.\",\n \"strengths\": \"- The investigation of techniques for conditional sampling is pretty thorough, and I think the proposed methods are quite interesting\\n- The proposed method obtains fairly strong empirical results, and the empirical evaluation is convincing \\n- Throughout the paper is very clear and well-written\",\n \"weaknesses\": \"- There is some highly relevant related work that the authors do not discuss. [Functional Flow Matching, AISTATS 2024](https://arxiv.org/abs/2305.17209) proposes the use of GP priors in conjunction with flow matching and studies techniques for forecasting with these models. Similarly, [Conditional Flow Matching for Time Series Modeling, SPIGM@ICML 2024](https://openreview.net/forum?id=Hqn4Aj7xrQ) uses GPs with flow matching for time series. The authors should cite these works and discuss the differences with their proposed method.\\n- There are some (relatively minor) clarity issues throughout\\n - The use of equation 9 was a bit unclear to me. Why is this specific form of $q_1$ chosen? Some justification for this modeling choice would be good.\\n - In Section 3.1.2, I am guessing that once we sample $x_0 \\\\sim q_0(x_0 \\\\mid y^p)$, then we use $x_0$ as an initial condition for the flow model to generate new samples $y \\\\mid x_0$. Is this the case? If so, it would help to state this explicitly somewhere in the paper.\\n - In Line 303, the authors write \\\"approximating q_0(x_0 \\\\mid y^p)$ with $q_0(x_0 \\\\mid y^p)$. I think I understand what is meant here, but this seems to be a typo.\",\n \"questions\": \"- The exact problem statement was a bit unclear to me. In lines 154-156 the authors describe a time series as a vector in $\\\\mathbb{R}^L$. Does this mean that the authors only work on time series having a fixed length $L$, or does the method allow for variable-length time series? Similarly, are the authors assuming that the time series all share a fixed discretization (i.e., there are some fixed times $t_1, \\\\dots, t_L$ corresponding to $y_1, \\\\dots, y_L$), or can the discretization vary across time series? Is this discretization assumed to be uniform, or can it be irregular as well?\\n- It seems to me that the setup is not limited just to forecasting, but could be applied to general conditional generation tasks, e.g., imputation. Have the authors tried anything beyond forecasting?\"\n },\n {\n \"id\": \"Ht8UdSWjui\",\n \"initial_rating\": 5,\n \"confidence\": 2,\n \"soundness\": 3,\n \"contribution\": 2,\n \"presentation\": 2,\n \"summary\": \"This paper introduces TSFlow, a model for probabilistic time series forecasting that enhances generative modeling by incorporating Gaussian Process (GP) priors within the Conditional Flow Matching (CFM) framework. The use of more informative GP priors helps align the prior distribution with the temporal structure of the data, potentially improving both performance and leading to runtime efficiency by simplifying the probability paths. TSFlow is flexible and can be used for both, conditional and unconditional, time-series generation.\\n\\nThe model demonstrates strong performance across several benchmark datasets, but there are aspects of the paper that could benefit from further clarification and improvement. Depending on the author's response, I am willing to increase my score from \\\"weak reject\\\" to \\\"weak accept\\\".\",\n \"strengths\": \"**Incorporating Gaussian Process Priors:** The main contribution—replacing the typical isotropic Gaussian prior $q(x_0)$ with a data-dependent conditional prior $q(x_0∣y^p)$ is well-motivated. GP priors are naturally suited for time series due to their ability to model temporal dependencies, and this idea is a clear innovation over existing flow matching methods.\\n\\n**Empirical Performance:** The empirical results show that TSFlow performs well compared to state-of-the-art models across various benchmark tasks.\\n\\n**Flexibility in Conditional and Unconditional Modeling:** The approach supports both unconditional and conditional generation. While unconditional generation has fewer use cases compared to conditional generation, it is a feature often overlooked in time-series analysis. The ability to use the same model for both tasks—by applying conditioning only during inference—adds versatility to the paper.\",\n \"weaknesses\": \"**Majors**\\n\\n**Difficulty in Parsing for Non-Experts:** The paper assumes substantial familiarity with flow matching and related generative methods, making it challenging for readers without a deep background in these specific techniques. It took for me considerable time to fully grasp the concepts, which suggests that the paper might also be difficult to read for a broader audience.\\n\\n**Lack of Runtime Analysis:** The authors propose replacing the isotropic Gaussian prior with a more complex GP prior. Howeer, they do not address the increased computational cost that comes with using GP priors. Although the GP prior is more suited to time-series tasks, this advantage must be weighed against the computational overhead. The paper should clearly state the theoretical runtime complexity of using GP priors and provide empirical runtime comparisons in the experiments. The reduction in NFEs is a positive, but its trade-off with the computational cost of the GP prior must be considered.\\n\\n**Baseline Comparison:** A simple baseline for the forecasting tasks could involve using Eq. 6 but with an isotropic Gaussian prior. Please include this method as a comparison partner in Table 3. If it is not a valid approach, it should be explained why such a baseline is excluded.\\n\\n**Inconsistent Findings on Kernel Choice:** The periodic kernel minimizes the Wasserstein distance in Figure 2, suggesting it aligns well with the data distribution. However, in Table 1, the periodic kernel does not significantly outperform other kernels, and it is not even considered in Table 4.2. This inconsistency is counterintuitive and warrants further discussion. Moreover, the necessity for different hyperparameter choices across tasks (e.g., generative modeling vs. forecasting) weakens the \\\"one model for all\\\" argument, suggesting more task-specific tuning may be required.\\n\\n**Minors:**\\n\\n**Missing Experiments on Guided Generation (3.1.3):** I could not find experiments on Guided Generation (3.1.3) in the paper. Where can I find them?\\n\\n**LPS Score Clarification:** The Linear Predictive Score (LPS) is not sufficiently explained. Please provide more details on how the LPS is calculated, specifically on what the linear model is regressed against and what the output represents.\\n\\n**GP Hyperparameter Optimization:** You state in your paper that you did not fit the GP hyperparameters. This is somewhat unexpected, I would have had a strong guess that learning the hyperparameters of the GP brings a large benefit to the model. Can you perform a small experiment to evaluate the difference in performance with/without GP hyperparameter optimization?\",\n \"questions\": \"See above.\"\n },\n {\n \"id\": \"oywKq5KOZr\",\n \"initial_rating\": 5,\n \"confidence\": 3,\n \"soundness\": 3,\n \"contribution\": 3,\n \"presentation\": 2,\n \"summary\": \"This paper proposes a model that uses flow-matching algorithms to generate and forecast univariate time-series. When trained unconditionally, it replaces the isotropic Gaussian prior distribution with an informed one and uses conditional sampling and inferencing. The model itself can also be made conditional. Experiments show that the model surpasses existing ones on benchmark tasks.\",\n \"strengths\": \"1. The model is flexible with many possible options; the informed prior sampling is intuitive.\\n\\n2. The model shows promising results on univariate benchmark tasks.\",\n \"weaknesses\": \"1. The model right now is only discussed and validated on univariate time-series.\\n\\n2. While there are many proposed options in the model, there lacks an ablation or a comprehensive comparison of different choices. Neither is there any theoretical insights in the proposed model.\\n\\n3. The presentation of the model needs a bit more clarification. The problem formulation can potentially be expanded a bit more (see questions below). The discussion of different training/inference choices is a bit dense. Maybe some concise pseudocode or flowchart in the main text could be helpful.\",\n \"questions\": \"1. If my understanding is correct, the overall idea of the model is to push an initial GP to the target time-series. Therefore, the sequence length $L$ in the problem formulation section is equivalent to the vector field dimension $d$ in the background section. Can the author(s) kindly confirm if this is true? It would be nice to clarify the setting a little bit. I think the main source of confusion is that there are two time indices in this paper: the time in flow-matching and the time in the time-series, and these are orthogonal to each other. I think this should be made clear somewhere in the paper.\\n\\n2. Following the first question, it would be nice to indicate in Figure 1 the time-series index and the flow-matching index. Moreover, different notations should be used. (For instance, on line 202-203, $t$ should not be used for the kernel because it has already been used in Eq. (1).)\\n\\n3. How would the model respond to a growing $L$? That is, there are two subquestions in this query:\\n\\ta. What is the time complexity of the model and how does it compare to other models that you benchmarked against?\\n\\tb. When $L$ is large, Eq. (6) seems to suffer from the curse of dimensionality and the integral would be impossible to discretize. How does the model see this issue, or is it not relevant?\\n\\n4. The section \\\"Effect on the Optimal Transport Problem\\\" only considers the distance between sequences but not anything about the training. Can you show some experiments where models that use the periodic kernel are indeed easier to train on tasks that involve periodicity?\\n\\n5. On line 302-303, you wrote \\\"we additionally condition the prior distribution on the observed past data by approximating $q_0(\\\\mathbf{x}_0 | \\\\mathbf{y}^p)$ with $q_0(\\\\mathbf{x}_0 | \\\\mathbf{y}^p)$.\\\" I assume there is a typo. What would be the intended sentence?\"\n }\n]","isConversational":false},"scores":{"kind":"string","value":"{\n \"rating\": \"5;5;6;6\",\n \"rating_avg\": 5.5,\n \"confidence\": \"3;2;4;2\",\n \"confidence_avg\": 2.75,\n \"soundness\": \"3;3;4;3\",\n \"soundness_avg\": 3.25,\n \"contribution\": \"3;2;2;3\",\n \"contribution_avg\": 2.5,\n \"presentation\": \"2;2;4;3\",\n \"presentation_avg\": 2.75\n}"},"metadata":{"kind":"string","value":"{\n \"primary_area\": \"\",\n \"track\": \"main\",\n \"venue\": \"ICLR 2025 Poster\",\n \"venueid\": \"ICLR.cc/2025/Conference\",\n \"processed_at\": \"2026-01-14T22:16:03.997821\"\n}"},"meta_review":{"kind":"string","value":"{\n \"id\": \"g83aQncGrU\",\n \"metareview\": \"This paper introduces a new generative framework for time series forecasting, based on conditional flow matching and leveraging conditional Gaussian Processes as informed priors. By aligning the prior distribution more closely with the data distribution, this approach simplifies the probability paths. In turn, the experimental results demonstrate that this approach helps to improve the generative performance for forecasting.\\n\\nOverall, the reviewers are highly positive about this work. They emphasize its originality, strong empirical results, and the intuitive nature of the informed prior sampling strategy. The paper is also well written and clearly presented. One limitation is that the approach has primarily been validated on univariate time series problems.\\n\\nGiven the importance of machine learning for time series, this paper makes a valuable contribution. The positive and supportive reviews underline its merit, and I feel that this paper will motivate future work. I therefore recommend accepting this paper.\",\n \"additional_comments\": \"The authors provided a detailed rebuttal that addressed the reviewers’ concerns, including an extended ablation study, a revised manuscript clarifying several key points, and an expanded related work section.\"\n}"},"decision":{"kind":"string","value":"{\n \"decision\": \"Accept (Poster)\"\n}"}}},{"rowIdx":3698,"cells":{"paper_id":{"kind":"string","value":"uy31tqVuNo"},"arxiv_id":{"kind":"string","value":"2410.18975v2"},"title":{"kind":"string","value":"Unbounded: A Generative Infinite Game of Character Life Simulation"},"markdown":{"kind":"string","value":"{\n \"content\": \"## Abstract\\n\\nAbstract We introduce the concept of a generative infinite game , a video game that transcends the traditional boundaries of finite, hard-coded systems by using generative models. Inspired by James P. Carse’s distinction between finite and infinite games (Carse, 1986 ) , we leverage recent advances in generative AI to create Unbounded : a game of character life simulation that is fully encapsulated in generative models. Specifically, Unbounded draws inspiration from sandbox life simulations and allows you to interact with your autonomous virtual character in a virtual world by feeding, playing with and guiding it - with open-ended mechanics generated by an LLM, some of which can be emergent. In order to develop Unbounded , we propose technical innovations in both the LLM and visual generation domains. Specifically, we present: (1) a specialized, distilled large language model (LLM) that dynamically generates game mechanics, narratives, and character interactions in real-time, and (2) a new dynamic regional image prompt Adapter (IP-Adapter) for vision models that ensures consistent yet flexible visual generation of a character across multiple environments. We evaluate our system through both qualitative and quantitative analysis, showing significant improvements in character life simulation, user instruction following, narrative coherence, and visual consistency for both characters and the environments compared to traditional related approaches.\\n\\n## 1 Introduction\\n\\nIn his work “Finite and Infinite Games: A Vision of Life as Play and Possibility” (Carse, ), James P. Carse introduces a distinction between two types of games. Carse defines finite games as those “played for the purpose of winning,” with boundaries, fixed rules, and a definitive endpoint. In contrast, infinite games are “played for the purpose of continuing the play,” with no fixed boundaries and evolving rules. Traditional video games are, inherently, finite games due to the limitations of programming and computer graphics. For example, game mechanics have to be fully pre-defined in the programming language and graphics assets usually have to be pre-designed (modulo traditional procedural generation which still grapples with structural limitations). This allows for only a finite, and sometimes predefined, set of actions and paths that can be taken. They also feature predefined rules, boundaries, and win conditions, which run counter to the infinite game definition.\\n\\nRecent advances in generative models have been impressive. We hypothesize that these developments have finally opened up the possibility of creating the first generative infinite video game - an infinite game that is also fully subsumed in generative models with no logic or graphics being controlled by other, more traditional, processes. Two advancements have made this achievable: (1) large language models (LLMs) that can encode persistent video game mechanics (e.g. interactions with the game environment or characters, character state tracking, object permanence), generate interactive stories, and produce spontaneous, and sometimes emergent, behaviors; and (2) visual generative models capable of producing high-quality images that follow prompts. In this work, we present Unbounded, what we believe to be the first interactive generative infinite game, where all game behaviors and graphics are generated by AI models, transcending the constraints of hard-coded systems. Unbounded draws inspiration from sandbox life simulations and digital pet games such as Little Computer People, The Sims and Tamagotchi. It incorporates elements from tabletop roleplaying games like Dungeons & Dragons, which offer unconstrained storytelling experiences that have been unattainable in video games.\\n\\nSome useful prior work has proposed ideas that are related to our concept of generative infinite games. Gaudl et al. () proposes fluidic games that merge gameplay with game design, automated game design (Cook et al., ) explores automated systems that can help in making new games, and work on AI-based games (Treanor et al., ) proposes design techniques for games that put AI at the forefront of user experience. These concepts are different from our definition of generative infinite games since they emphasize structured exploration within predefined game parameters and tools, where the design space is extensive yet ultimately bounded. In contrast, infinite games, as we envision them, aim to continuously evolve beyond predefined structures, with generative models dynamically creating both content and mechanics in real-time, enabling an open-ended gameplay experience with no fixed endpoint or limitations on possible interactions. There is also important work on using machine learning models as components that generate parts of a game. Agarwal () generates novel interactions, Sun et al. () generates new mechanics and items as well as some graphical assets, and there is also work on orchestration of different models for game design (Liapis et al., ). These works do not generate all components of the game. In contrast, in our work, all of the game mechanics, characters, environments, narrative, and graphics are fully produced by generative models. This is more similar in spirit to recent work Genie (Bruce et al., ) and GameNGen (Valevski et al., ), although in contrast to these works that mainly generate platformers with diffuse mechanics or regenerate behaviors of one pre-existing game, our work proposes an open-ended narrative experience with stable game mechanics enabled by an LLM-based game engine.\\n\\nUnbounded offers a gameplay loop centered around character simulation and open-ended interaction (Figure 2). Players can insert their characters into the game, defining their appearance and personality. The game generates a world where these characters can explore environments, interact with objects, and engage in conversations. The game generates new scenarios, stories, and challenges based on the player’s actions and choices, creating a personalized and infinite gaming experience. Some generative game examples are shown in Figure 3.\\n\\nSpecifically, Unbounded has the following capabilities: (1) Character Personalization: players can insert their characters into the game, defining their appearance and personality. (2) Game Environment Generation: Unbounded generates a persistent world that the characters can explore and interact. (3) Open-Ended Interaction: Players can interact with the character using natural language instructions, and there are no pre-defined rules to constraint the interaction. (4) Real-Time Generation: we pay special attention to the speed of the game and achieve 5-10x speedups over a naive implementation, serving each new scene with a latency of about one second.\\n\\nOur approach introduces technical innovations in both LLM and vision generation domains. On the language side, we developed a LLM based game engine capable of maintaining consistent game mechanics, generating coherent narratives, and producing contextual character responses in real-time. Our distilled specialized model is fine-tuned on data automatically generated with two collaborative strong LLM agents, without the need for human annotation in the loop. Our distilled LLM model handles the dynamic generation of game rules and scenarios, adapting to player input and game state. In visual generation, we introduce a new regional IP-Adapter, which allows for the consistent generation of characters and environments while maintaining visual coherence across multiple images. Specifically, our regional IP-Adapter conditions the image generation on the game environment and character appearance encoded modulated by a dynamic mask obtained from attention outputs in cross-attention layers. This is in order to mitigate the interference between the environment and character, in order to have both reliably appear in the scene. This approach enables real-time image generation that reflects the game state and player actions.\\n\\nThe contributions of this work are conceptual and technical. We introduce the notion of a generative infinite game, demonstrating its feasibility and potential impact on the future of interactive entertainment. We present a new paradigm for game design where the game logic and content are encapsulated within generative models. Our main technical contributions include the specialized distilled LLM for game logic and narrative generation and the regional IP-Adapter for consistent visual generation. We demonstrate the effectiveness of our regional IP-Adapter through both quantitative and qualitative evaluations, surpassing state-of-the-art in both character and environment consistency. Furthermore, we show that our distilled LLM performs comparably to a very large LLM while having interactive speed. These advancements enable the creation of Unbounded and lay the groundwork for future research and development in the field of AI-driven interactive experiences.\\n\\nFigure: Figure 2: Example of Unbounded. Based on an initial user input, Unbounded sets up game simulation environments, and generates character actions in the environments. Users can interact with the character with natural language instructions, exploring the game with unlimited options.\\nRefer to caption: x2.png\\n\\n## 2 Related Work\\n\\n##### Controllable Text-to-Image Generation.\\n\\nControllable text-to-image generation becomes a key research direction in diffusion model applications, enabling diverse ways to guide the generation process. For instance, ControlNet (Zhang et al., ) introduces conditioning mechanisms that utilize control signals such as depth maps, poses, edges, and segmentation maps to guide image generation. Other works focus on layout control using bounding boxes to control object placement within the generated images (Li et al., ; Shin et al., ). Beyond these control signals, another major area of research involves personalization, where the goal is to generate consistent characters (Ruiz et al., ; Gal et al., ; Kumari et al., ; Avrahami et al., ) or consistent face identities (Li et al., ; Wang et al., ; Yan et al., ; Ruiz et al., ) across multiple generations. However, most existing approaches lack support for conditioning both characters and environments separately, often requiring predefined masks for generating characters (Chen et al., ; Lugmayr et al., ; Yang et al., ), with the environment remaining identical to the input image. This limitation makes it difficult to seamlessly integrate characters into different environments while ensuring both consistency and alignment with the input prompts.\\nIP-Adapter (Ye et al., ) tackles this task by conditioning the generation on the environment and character images. However, IP-Adapter tends to over reconstruct the conditions, which causes interference between them. In this paper, we build our approach on IP-Adapter and propose an improved regional IP-Adapter with block drop, separating character and environment generation to enhance consistency.\\n\\n##### Large Language Models in Image Generation.\\n\\nLarge language models have demonstrated strong in-context learning capability (Brown, ), which enables them to solve diverse customized tasks based on human instructions and in-context examples. In the field of image generation, large language models have been employed for various tasks, such as image layout generator based on prompts (Cheng et al., ; ), interactive multi-turn image generation (Zeqiang et al., ; Huang et al., ; Gong et al., ; Wang et al., ), and interleaved text and image generator (Team, ; Zhou et al., ). Unlike these applications, our work focuses on distilling a specialized LLM, serving as a game engine, responsible for generating game mechanics, narratives and character interactions.\\n\\n##### Game Generation.\\n\\nIn the field of Procedural Content Generation (PCG), diverse approaches have been explored to create dynamic game content (Summerville et al., ).\\nEarly works employed concept maps (Treanor et al., ), Markov Chains (Snodgrass & Ontañón, ), Bayes Nets (Guzdial & Riedl, ), and LSTMs (Summerville & Mateas, ) to simulate interactive game environments.\\nRecent research has advanced to using Generative Adversarial Networks (GANs) for generating game levels or dynamic environments (Volz et al., ; Kumaran et al., ; Schubert et al., ; Kim et al., ), exploiting the use of diffusion models for game generation (Zhou et al., ), utilizing LLMs to design and generate the game environments or mechanics (Sudhakaran et al., ; Todd et al., ; Nasir & Togelius, ; Hu et al., ; Zala et al., ; Anjum et al., ; Chung & Kreminski, ; Chung et al., ) and employing LLMs to generate items and narratives, as well as using diffusion models for scene generation in 1001 Nights (Sun et al., ).\\nOne major feature of this body of work is that AI is often employed to help in game design, or to generate one component of the game, for example environments or interactions. Our work aims to fully generate all behaviors of the game, including all the game’s graphics, characters, environments, and narrative, using generative models.\\n\\nBruce et al. () explore synthesizing a completely new interactive game with a video diffusion model conditioned on previous game frames and user actions. This work achieves impressive initial results for this direction, at present it has limitations on the scope of the games being generated, with mostly 2D platformers being represented. This differs from our work since these are not infinite games. We also believe that generating compelling infinite games by only modeling pixels is a hard problem and hence we opt to use language models to generate all the open-ended game mechanics instead. Valevski et al. () train a diffusion model to be a game engine running Doom, and achieve consistent generation of game mechanics by only modeling pixels. Nevertheless, this work currently fits the diffusion model to a single finite game that already exists.\\n\\nBesides generating the game content, there has been exploration in letting AI take different roles (e.g., competitor, designer, or teammate) in games (Zhu et al., ; Gallotta et al., ; Pell, ; Agarwal et al., ), and a system with continuous creativity for automated game design for games that are played on a grid of up to 8x8 squares is built in Cook ().\\nAlongside generating visual game environments, another line of research builds narrative text games with conditional language generation models to complete a knowledge graph (Ammanabrolu et al., ) or utilizes LLMs or other AI tools to generate the game (Li et al., ; Latitude Inc., ; Wang et al., ) or act in different roles in the game (Zhu et al., ; Kreminski et al., ; Cui et al., ; Zhou et al., ; Dambekodi et al., ). Our work diverges from these by (1) tackling both generation of all game mechanics, graphics, characters, environments, and narrative of the game using generative models (2) introducing personalization into generative games, where a user can insert a custom character and personalize the game world and initial story arc, and (3) achieving interactive real-time refresh speed through our innovations in the vision and language domains.\\n\\nFigure: Figure 3: Generative game examples of Unbounded. The user can insert a custom character into the game, engage with the character through natural language instructions, bring the character to different environments, and interact with it to maintain a healthy state under the games’ mechanics.\\nRefer to caption: x3.png\\n\\n## 3 Method\\n\\nWe introduce Unbounded, an interactive generative infinite game powered by text-to-image generation models and large language models. Unbounded offers: (1) Custom Character Personalization: users create unique characters with customizable appearances and personalities; (2) Dynamic World Creation: the system generates a persistent, interactive game world for exploration; (3) Open-Ended Interaction and Gameplay: players interact with their characters via natural language, with the game dynamically generating new scenarios and storylines based on player actions; and (4) Generation in Interactive Speed: the game runs with near real-time interactivity, achieving a refresh rate close to one second. We detail the methods enabling these capabilities in this section.\\n\\n### 3.1 Personalization of Latent Consistency Models for Character Consistency\\n\\nA key feature of Unbounded is its ability to serve a fully generative model-based game with real-time interaction. This is made possible through the use of latent consistency models (LCM) (Luo et al., ) which allow for high-resolution image generation with as few as two diffusion steps. By utilizing LCMs, Unbounded achieves real-time text-to-image (T2I) generation, critical for delivering an interactive gaming experience with a refresh rate close to one second.\\n\\nTo support the use of custom characters in the game, we incorporate DreamBooth (Ruiz et al., ) into the T2I model. Given a set of character images, we fine-tune the diffusion model using LoRA modules (Hu et al., ). During fine-tuning, we append a unique identifier, “[V]”, which has a weak prior in the model, to denote the subject. DreamBooth personalization is performed on the base diffusion model, and we merge the subject-specific LoRA with the LCM LoRA trained for few-step diffusion. This simple arithmetic LoRA merging works surprisingly well in maintaining both inference speed and subject preservation. Other alternatives for personalization exist, and many do not need setup time, yet we find that they often fail in strongly preserving the character’s features, which is a critical component in order to have a satisfactory experience in this type of game.\\n\\nFigure: Figure 4: (a) Our overall image generation method. We achieve real-time image generation with LCM LoRA, maintain character consistency with DreamBooth LoRAs, and introduce a regional IP-Adapter (shown in (c)) for improved environment and character consistency. (b) Our proposed dynamic mask genreation separating the environment and character conditioning, preventing interference between the two.\\nRefer to caption: x4.png\\n\\n### 3.2 Regional IP-Adapter with Block Drop for Environment Consistency\\n\\nAnother key feature of Unbounded is generating the character in pre-defined environments performing different actions based on user instructions. Thus, maintaining both character and environment consistency is essential. While character consistency is handled as discussed in Sec. 3.1, two additional challenges arise: ensuring the environment consistency across different generations and accurately placing the character within the environment, without losing alignment with text prompts. We find that existing method fails to consistently perform well for all requirements in interactive speed. As one of our main technical contributions we propose a novel regional IP-Adapter in order to consistently implant a character in pre-defined environments following text prompt.\\n\\n#### 3.2.1 Regional IP-Adapter\\n\\nWe propose an improved version of IP-Adapters (Ye et al., ) that enables dual-conditioning on both subjects and environments, allowing for the generation of a pre-defined character in a user-specified environment. Unlike the original IP-Adapters, which focus on single-image conditioning, our approach introduces dual-conditioning and dynamic regional injection mechanism to represent both concepts simultaneously in the generated images.\\n\\nLet us start with an example. As shown in Figure 4, given the text prompt “Beneath the desert sky, [V] witch coaxes the cacti to blossom with vibrant, glowing flowers” and the desert environment image, the model needs to know that the character should be generated beside cacti and within the desert environment. This requires the model to correctly (1) preserve the environment (2) preserve the character (3) follow the prompt. Utilizing IP-Adapter to encode the environment greatly harms both (2) and (3) (Figure 8). Our regional IP-Adapter solves this by implementing a novel attention separation mechanism for generating the two elements. Specifically, we introduce a dynamic mask-based approach that leverages cross-attentions between the character text embedding and the hidden states at each layer of the model. As shown in Figure 4, our approach applies the adapter to the regions corresponding to the environment and character separately, preventing the environment conditioning to interfere with the character’s appearance, and vice versa. At each cross-attention layer, we calculate the attention map between a pre-defined character text embedding $K_{c}$ and the output hidden states from the text cross-attention layer $O_{t}$:\\n\\n$$ $A_{c}=\\\\frac{W_{q}O_{t}*W_{k}K_{c}^{T}}{\\\\sqrt{d}}$ (1) $$\\n\\nwhere $W_{q}$ and $W_{k}$ are projection weights adopted from the text cross-attention layers. The dynamic mask we use for the regional IP-Adapter is defined as:\\n\\n$$ $M_{c}=\\\\left\\\\{\\\\begin{array}[]{ll}1&A_{c}\\\\leq threshold\\\\\\\\ 0&A_{c}>threshold\\\\\\\\ \\\\end{array}\\\\right.$ (2) $$\\n\\nWe rank the attention scores $A_{c}$ and set the $threshold$ at top $r\\\\%$, dynamically updating the mask at each attention layer throughout the diffusion process.\\nThe output of the cross-attention blocks is calculated as:\\n\\n$$ $O=O_{t}+\\\\alpha_{e}M_{c}*O_{e}+\\\\alpha_{c}(1-M_{c})*O_{c}$ (3) $$\\n\\nwhere $O_{e}$ and $O_{c}$ represent the outputs of the IP-Adapter image cross-attention layers of environment and character, and $\\\\alpha_{e}$ and $\\\\alpha_{c}$ are the IP-Adapter scales to adjust the strength of the environment and character conditioning respectively. Our regional IP-Adapter with dynamic masks allows the model to generate the character without the interference from the environment conditioning, greatly enhancing character consistency preservation (Figure 8). Besides, it preserves environment consistency by generating the background around the character conditioning on the environment image, while maintaining the implicit layout information in the hidden states from the text cross-attention layers $O_{t}$, ensuring the character is placed accurately following the text prompt.\\n\\nFigure: Figure 5: Attention map between character embedding and hidden states in cross-attention layers in different blocks. The character embedding we use is “A [V] witch”.\\nRefer to caption: x5.png\\n\\n#### 3.2.2 Block-Wise Drop on Environment Conditioning\\n\\nFor our regional IP-Adapter, we use a dynamic mask from the cross-attention between character text and hidden states. This mask’s quality is key for separating character and environment generation. Figure 5 shows attention maps between character embeddings and hidden states in cross-attention layers of down sample blocks. We observe that the attention doesn’t focus on the character but spreads across the full image for these blocks. We take this as a strong indication that the diffusion model doesn’t separate character and environment generation in these layers and instead focuses on overall image structure based on text prompts. This aligns with Wang et al. ()’s finding that down sample blocks capture spatial layouts more, while up sample blocks capture style. We drop the regional IP-Adapter in down sample blocks, using it only in mid and up sample blocks. This allows for better spatial layout generation between character and environment.\\nAdditionally, we find that adding dynamic regional IP-Adapters in up sample blocks more strongly aligns the generated background with the conditioning environment by extracting relevant semantic information while preserving character-specific details like appearance and poses.\\n\\n### 3.3 Language Model Game Engine with Open Ended Interactions and Integrated Game Mechanics\\n\\nUnbounded simulates character actions in pre-defined environments with images generated based on scene text descriptions while monitoring the character’s state and providing the user with natural language interaction with the character and game environment. For example, the user can take the character to different environments, interact with the character (e.g. “I pet the character on its head”), or essentially take any open ended action e.g. “I pet the character and then take them for a rocket ride at the space station.”. Since the game is ultimately built on a language model, these expansive capabilities present several challenges: (1) Environment Binding: the model needs to place the character in the correct environment based on the natural language instructions. (2) Coherent Story Generation: the model generates coherent narrative descriptions and character responses that align with user-specified character traits. (3) Game Mechanics Generation: the model needs to monitor the state of the character (i.e., hunger, energy, fun, hygiene), and update them based on user interactions and story events. (4) Prompt Rewriting: the model needs to rewrite the narratives for the diffusion model (i.e., append special token “[V]” for the character, align environment descriptions to the pre-generated environments for better environment consistency).\\n\\nSurprisingly, we find that a very large language model (e.g., GPT-4, GPT-4o) with detailed instructions and using in-context learning (Brown, ) can exhibit these capabilities. Nevertheless, using such large models as game engines is not directly feasible due to the large latency (e.g., 5 seconds for a 7B model to give one response). Given this, we propose to distill these capabilities from a very large model into a smaller model based on Gemma-2B (Team et al., ) for game logic and narrative generation that supports real-time interaction. In this section we propose two key technical contributions (1) a design for a character life simulation game using two very large language models that control for world modeling and user interaction respectively (2) a framework for distilling this knowledge into one smaller language model that is fast enough to achieve interactive speeds.\\n\\n#### 3.3.1 Character Life Simulation with Multi-LLM Collaboration\\n\\nWe build a character life simulation game with two LLMs agents. One agent serves as the world simulation model, responsible for setting up game environments, generating narratives and image descriptions, tracking character states and simulating the character’s behavior. The second agent functions as a user model, simulating the player’s interactions with the world simulation model. It has three types of interactions: continuing the story within the current environment, moving the character into different environments, or interacting with the character to maintain the healthy state of the character. In each interaction category, the user has the option to provide personality details of the character or guide the character actions that, in turn, guide the simulator’s narrative generation. This interaction between the world simulation LLM and the user LLM allows for a dynamic character life simulation game with virtually unlimited interaction possibilities and narrative paths.\\n\\nFigure: Figure 6: Overview of our user-simulation data collection process for LLM distillation. (a) We begin by collecting diverse topic and character data, filtered using ROUGE-L for diversity. (b) The World LLM and User LLM interact to generate user-simulation data through multi-round exchanges.\\nRefer to caption: x6.png\\n\\n#### 3.3.2 Framework for Small LLM Distillation\\n\\nWe propose a framework for distilling the capabilities of larger LLMs into the smaller, more efficient model using synthetic data generated by the multiple stronger LLMs. Our framework contains two steps: (1) automated data collection (Figure 6) and (2) small model distillation.\\n\\n##### Automated Data Collection\\n\\nOur goal is to build a general-purpose character life simulator capable of understanding a wide range of character traits and generating games across diverse topics. To achieve this, the first step is to gather a diverse dataset of topics and characters. We prompt a large LLM to generate pairs of topics and corresponding main characters. To ensure data diversity, we retain only the generated pairs whose ROUGE-L similarity to existing data is below 0.7, following (Wang et al., ), which demonstrated the importance of diverse data for enhancing an LLM’s ability to follow instructions. This process results in 5,000 unique topic-character pairs, which serve as the basis for user-simulator interaction data.\\n\\nIn the second step, we collect multi-round interaction data between the world simulation LLM and the user LLM. The process begins with the world simulation LLM setting up the game environment and initiating a character action based on a randomly sampled topic-character pair from the dataset. The user LLM is then prompted to provide interaction inputs, while the world simulation LLM generates updated character actions, states, and responses. This iterative process continues for five interaction rounds per session, resulting in a total of 5,000 user-simulator interaction examples. All the prompt templates are in Appendix.\\n\\n##### Distillation\\n\\nOnce the interaction data has been collected, we fine-tune the smaller Gemma-2B model using the 5,000 synthetic user-simulator interaction samples. During supervised fine-tuning, we mask out the loss on user input data, focusing the optimization on learning the world simulation model’s behavior based on multi-round interaction history and current user input. This approach enables Gemma-2B to replicate the capabilities of larger LLMs as a game engine while supporting real-time interaction. Our distilled Gemma-2B demonstrates performance comparable to GPT-4o, effectively following user input and supporting unbounded interactions.\\n\\n## 4 Experimental Setup\\n\\n### 4.1 Evaluation Benchmarks\\n\\n##### Evaluation of Image Generations\\n\\nTo evaluate our image generation approach, we collect an evaluation dataset consisting of 5,000 (character image, environment description, text prompt) triplets with GPT4o (OpenAI, ). It includes 5 characters (dog, cat, panda, witch, and wizard), 100 diverse environments, and 1,000 text prompts (10 per environment). We evaluate the image generation performance under three criteria: environment consistency, character consistency and semantic alignment with the text prompt. To measure similarity between images, we employ CLIP-I (Radford et al., ), DINO (Caron et al., ), and DreamSim (Fu et al., ). We denote the similarity between environment reference image and the generated images as CLIP-I^E, DINO^E, DreamSim^E, and the similarity between the character reference image and the generated images as CLIP-I^C, DINO^C, DreamSim^C. Additionally, we use CLIP-T (Radford et al., ) to evaluate the semantic alignment with the text prompt. Given that Unbounded is a character life simulation game, ensuring the presence of the character in the image is important. Therefore, we further utilize Grounding-DINO (Liu et al., ) to detect the presence of the character in the generated images. We set similarity scores to 0 and distance scores to 1 if there is no character in the generated image.\\n\\n##### Evaluation of LLM Generations\\n\\nWe collect an additional evaluation dataset with 100 user-simulator interaction samples using the pipeline in Sec.\\n3.3. Each user-simulator interaction sample contains five rounds of interaction between the user and the world model. We use GPT-4 (OpenAI, ) as a judge, scoring the response between two models (baseline model vs. our model) in overall score, and then in four aspects: accuracy of character state update, environment relevance, story coherence, and user input instruction following. Scores range from 0 to 10.\\n\\n### 4.2 Implementation Details\\n\\nOur image generator is built on SDXL (Podell et al., ).\\nWe train a DreamBooth LoRA of rank 16 with batch size 1 and a constant learning rate 1e-4 for 500 steps on a single A100, which takes approximately 30 mins. The special token we append before the character is “sks”. During inference, we merge the LCM-LoRA with DreamBooth LoRA with scale 1.0 for each. We use IP-Adapter-plus-sdxl-vit-h for encoding the environment, and IP-Adapter-plus-face-sdxl-vit-h for encoding the character. The dynamic mask ratio $r\\\\%$ in set to be 60%.\\n\\nOur LLM is built on Gemma-2B (Team et al., ). We distill the LLM using 5,000 user-simulator interaction samples collected from GPT-4o. We train the LLM for 6,500 steps, with batch size 8, distributed across 4 A100s, and learning rate 1e-4. The learning rate scheduler is set to be cosine annealing (Loshchilov & Hutter, ), and the warmup steps ratio is 0.03. In evaluation, we use an LLM to generate responses with sampling. The sampling hyperparameters are set to be default.\\n\\n## 5 Results and Analysis\\n\\n**Table 1: Comparison of Unbounded and other methods for maintaining environment consistency and character consistency. Unbounded achieves the best performance in maintaining consistency, while maintaining comparable semantic alignment with the text prompt. Best scores are in bold.**\\n| Methods | Environment Consistency | Character Consistency | Semantic Alignment | | | | |\\n| --- | --- | --- | --- | --- | --- | --- | --- |\\n| CLIP-I^E $\\\\uparrow$ | DINO^E $\\\\uparrow$ | DreamSim^E $\\\\downarrow$ | CLIP-I^C $\\\\uparrow$ | DINO^C $\\\\uparrow$ | DreamSim^C $\\\\downarrow$ | CLIP-T $\\\\uparrow$ | |\\n| IP-Adapter (Ye et al., ) | 0.470 | 0.381 | 0.595 | 0.366 | 0.139 | 0.832 | 0.168 |\\n| IP-Adapter-Instruct (Rowles et al., ) | 0.334 | 0.151 | 0.832 | 0.246 | 0.124 | 0.872 | 0.098 |\\n| StoryDiffusion (Zhou et al., ) | 0.528 | 0.257 | 0.733 | 0.629 | 0.464 | 0.545 | 0.242 |\\n| Ours | 0.563 | 0.322 | 0.675 | 0.676 | 0.470 | 0.488 | 0.242 |\\n\\nFigure: Figure 7: Comparison with other approaches for generating environment and character consistent images based on text prompts. We observe that our method strongly outperfoms related work.\\nRefer to caption: x7.png\\n\\n**Table 2: Ablations of the effectiveness of dynamic regional IP-Adapter and block drop in our consistent image generation approach.**\\n| No. | Block | Regional | Scale | Environment Consistency | Character Consistency | Alignment | | | | |\\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\\n| Drop | IP-Adapter | CLIP-I^E $\\\\uparrow$ | DINO^E $\\\\uparrow$ | DreamSim^E $\\\\downarrow$ | CLIP-I^C $\\\\uparrow$ | DINO^C $\\\\uparrow$ | DreamSim^C $\\\\downarrow$ | CLIP-T $\\\\uparrow$ | | |\\n| 1. | ✗ | ✗ | 1.0 | 0.123 | 0.111 | 0.885 | 0.073 | 0.024 | 0.973 | 0.034 |\\n| 2. | ✓ | ✗ | 1.0 | 0.414 | 0.331 | 0.647 | 0.337 | 0.147 | 0.832 | 0.149 |\\n| 3. | ✓ | ✓ | 1.0 | 0.563 | 0.322 | 0.675 | 0.676 | 0.470 | 0.488 | 0.242 |\\n| 4. | ✗ | ✗ | 0.5 | 0.470 | 0.381 | 0.595 | 0.366 | 0.139 | 0.832 | 0.168 |\\n| 5. | ✓ | ✗ | 0.5 | 0.577 | 0.332 | 0.640 | 0.627 | 0.374 | 0.575 | 0.252 |\\n| 6. | ✓ | ✓ | 0.5 | 0.549 | 0.263 | 0.726 | 0.705 | 0.514 | 0.450 | 0.246 |\\n\\n### 5.1 Comparison with Different Approaches for Maintaining Environment Consistency and Character Consistency\\n\\nQuantitative Results We compare our regional IP-Adapter with block drop with previous approaches in maintaining environment consistency and character consistency. For all the approaches, we merge the character LoRA and LCM LoRA with the model to support fast inference and improve character consistency and provide an apples-to-apples comparison.\\nAs shown in Table 1, our approach consistently outperforms previous approach in maintaining environment consistency and character consistency, while achieving comparable performance in maintaining semantic alignment. Specifically, our approach significantly overtakes StoryDiffusion (Zhou et al., ) by 0.047 in CLIP-I^C, and 0.057 in DreamSim^C for character consistency, and 0.035 in CLIP-I^E, 0.065 in DINO^E, and 0.058 in DreamSim^E for environment consistency, demonstrating the effectiveness of our approach. Besides, our approach also achieves comparable performance in maintaining semantic alignment, suggesting strong text following capabilities.\\n\\nQualitative Results\\nWe present a qualitative comparison with other approaches in Figure 7. Our regional IP-Adapter with block drop consistently generates images with high character consistency, whereas other methods may fail to include the character or generate characters with inconsistent appearances (see Example 1 & 2). Furthermore, we show that our approach balances environment consistency and character consistency well, while other approaches might generate environments that differ from the condition environment (e.g., StoryDiffusion in Example 1 & 3).\\n\\nFigure: Figure 8: Comparison between our regional IP-Adapter approach and baseline approaches.\\nRefer to caption: x8.png\\n\\n### 5.2 Effectiveness of Dynamic Regional IP-Adapter with Block Drop\\n\\nQuantitative Results We demonstrate that our regional IP-Adapter with block drop is essential for placing the character in the environments following the text prompt, while maintaining both environment and character consistency with ablation studies. As shown in Table 2, adding block drop improves both environment and character consistency compared with multi-IP-Adapter (No. 2. vs. No. 1.), with an increase of 0.291 in CLIP-I^E and 0.264 in CLIP-I^C, alongside better alignment between the text prompt and the generated image. Furthermore, our regional IP-Adapter enhances character consistency and text alignment while maintaining comparable performance in environment consistency (No. 3 vs. No. 2). We also explore the effect of the environment IP-Adapter scale, and our findings indicate that using a smaller scale (e.g., 0.5) generally improves character consistency, though it slightly compromises environment consistency (No. 6 vs. No. 3).\\n\\nQualitative Results As shown in Figure 8, conditioning on the environment using IP-Adapter achieves good environment reconstruction, but the character consistency is influenced by the environment style. Introducing block drop improves adherence to the text prompt, resulting in images with the correct spatial layout for both the character and the environment. However, the character’s appearance remains influenced by the surrounding environment. By incorporating our proposed regional injection mechanism with our proposed dynamic mask scheme, the generated images achieve strong character consistency while maintaining effective conditioning on the environment.\\n\\n**Table 3: Comparison of Unbounded and different LLMs on serving as game engines for open ended interactions and integrated game mechanics. We use GPT-4 to provide pairwise scores between our model and other LLMs.**\\n| Model | Overall | State | Environment | Story | Instruction | | | | | |\\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\\n| Update | Relevance | Coherence | Following | | | | | | | |\\n| | Base | Ours | Base | Ours | Base | Ours | Base | Ours | Base | Ours |\\n| Gemma-2B (Team et al., ) | 6.22 | 7.44 | 5.60 | 7.47 | 6.12 | 7.94 | 6.34 | 7.57 | 6.43 | 7.67 |\\n| Gemma-7B (Team et al., ) | 6.80 | 7.39 | 6.29 | 7.43 | 7.07 | 7.91 | 6.90 | 7.48 | 6.89 | 7.53 |\\n| Llama3.2-3B (Meta, ) | 7.21 | 7.50 | 6.86 | 7.38 | 7.63 | 7.93 | 7.36 | 7.56 | 7.31 | 7.67 |\\n| Ours-1k | 7.65 | 7.82 | 7.50 | 7.74 | 8.10 | 8.19 | 7.78 | 7.93 | 7.82 | 7.97 |\\n| GPT-4o (OpenAI, ) | 7.76 | 7.68 | 7.69 | 7.66 | 8.20 | 8.10 | 7.95 | 7.82 | 7.85 | 7.82 |\\n\\n### 5.3 Effectiveness of Distilling Specialized Large Language Model\\n\\nWe show that our diverse user-simulator interaction data effectively distills Gemma-2B into a capable game engine. As shown in Table 3, zero-shot inference with small LLMs (i.e., Gemma-2B, Llama3.2-3B), or a slightly larger LLM (i.e., Gemma-7B) results in lower performance compared to ours, highlighting the importance of distillation from a stronger LLM for game world and character action simulation. Furthermore, we show that our model achieves performance comparable to GPT-4o, validating the effectiveness of our approach. We also investigate the impact of distillation data size on performance by comparing a Gemma-2B model distilled with 1K data and 5K data. Results show that using a larger dataset consistently improves performance across all aspects, highlighting the potential for further enhancements with more data to fully match GPT-4o’s performance.\\n\\n## 6 Conclusion\\n\\nWe introduce Unbounded, an interactive generative infinite game based on generative models. Unbounded is built on two main components, a specialized, distilled LLM for real-time interaction, and a fast diffusion model with our proposed regional IP-Adapter for consistent generation across multiple scenes. We show that our proposed approach allows for an interactive game subsumed in generative models, with consistent characters, environments and story and an expansive gameplay characteristic of an infinite game.\\n\\n## Acknowledgments\\n\\nWe thank Shiran Zada, Peyman Millanfar, Shlomi Fruchter, Michael Goin and Matthew Guzdial for the thoughtful feedback and discussion.\\n\\n## Appendix\\n\\nIn this appendix, we present the following:\\n\\n- •\\nPrompts we use for user-simulation data collection in Sec. A.\\n- •\\nEvaluation prompts we use for querying GPT-4 as a judge for LLM evaluation in Sec. B.\\n- •\\nReproducibility statement in Sec. C.\\n\\n## Appendix A Prompt for synthetic user-simulator interaction data collection\\n\\nIn this section, we provide the prompt templates we use to collect the user-simulation data. Specifically, we query GPT-3.5 to generate diverse topics and character descriptions using the template shown in Figure 9. The prompt template for guiding the potential user interactions for user LLM is shown in Figure 10. To ensure user interactions align with the dialogue history, we include the interaction history as input to the user LLM. The world simulation LLM prompt template, shown in Figure 11, also takes in the dialogue history and generates the next character actions, states and narratives. We constraint the world simulation LLM to generate one storyline at a time, allowing users to choose how to continue the story.\\n\\n## Appendix B Evaluation Prompt\\n\\nWe include the prompt template we use to compare the outputs from two LLMs in Figure 12. The prompt is adapted from Vicuna (Chiang et al., ), and has been validated as an effective tool for comparing the performance of two LLMs on a given task.\\n\\n## Appendix C Reproducibility Statement\\n\\nFirst, we include the implementation details of Unbounded in Sec. 4.2, covering the training hyperparameters for Dreambooth fine-tuning, and LLM distillation, and the hyperparameters we use during inference for both LLM and image generation. Second, we provide detailed description of the user-simulation data we collect for training in Sec. 3.3, and further include the prompt template used to query GPT models in Appendix Sec. A. Lastly, for the LLM-based evaluation, the prompt template for querying GPT-4 to compare two LLM outputs is provided in Appendix Sec. B.\\n\\nFigure: Figure 9: Prompt template used to collect diverse topic and character data.\\nRefer to caption: x9.png\\n\\nFigure: Figure 10: Prompt template used to query user LLM.\\nRefer to caption: x10.png\\n\\nFigure: Figure 11: Prompt template used to query world simulation LLM.\\nRefer to caption: x11.png\\n\\nFigure: Figure 12: Prompt template used to query GPT4 to compare the performance between two LLM outputs. The prompt is adapted from Vicuna (Chiang et al., ).\\nRefer to caption: x12.png\",\n \"metadata\": {\n \"authors\": [\n \"Michael Rubinstein\",\n \"Mohit Bansal\",\n \"Nataniel Ruiz\"\n ],\n \"title\": \"Unbounded : A Generative Infinite Game of Character Life Simulation\",\n \"total_chars\": 43398,\n \"total_lines\": 239,\n \"updated_at\": \"2026-01-25T15:06:46.872794\"\n }\n}"},"reviews":{"kind":"list like","value":[{"id":"B5fG2n5Q8V","initial_rating":6,"confidence":4,"soundness":2,"contribution":2,"presentation":3,"summary":"The authors present \"UNBOUNDED\" a virtual pet game based on LLMs. The authors present two novel technical elements to support this game: a regional IP-Adapter for character-environment consistency and a domain-specific distillation approach for finetuning a smaller LLM to allow for real time gameplay. The authors then present experiments comparing their approach to baselines in terms of visual quality and consistency, and overall game quality as scored by GPT-4o.","strengths":"The main strengths of the paper are (1) the originality of entirely basing a virtual pet game on LLMs, (2) the significance of the two primary technical contributions to figure coherent image generation and real time LLM application uses, and (3) the quality of the image portion of the evaluation/experiments. The paper is also overall well-written, especially in terms of the technical aspects. These primarily justify the positive aspects of my above scores.","weaknesses":"The paper has two major weaknesses in its present draft, the way it positions itself in terms of prior work and the evaluation. \n\n### Prior Work\n\nThe authors do not appear to have engaged with the field of technical games research at all. This is a shame, as their game can be understood as an AI-based game [1] or specifically a NN game [2]. More broadly, the authors' work in terms of generation would fit into the Procedural Content Generation paradigm [3]. There is a 3+ decades long history of work in this space that is relevant to the authors' work [4], with significant recent work at the intersection of LLMs and game generation [5], and it is crucial that the authors update the introduction and related work to situate their work within this field.\n\n### Evaluation\n\nThe authors primarily evaluate their work in terms of image quality and I have no major concern with these results (though the DreamSim results seem to incorrectly attribute the best score to their system). However, I have major concerns with the \"Evaluation of LLM Generations\", which is functionally the closest thing presented to a holistic evaluation of UNBOUNDED. Firstly, the norm within the area would be to make use of a user study for evaluation purposes [6,7]. Particularly since the authors make such an effort to allow for real time control, this would seem to follow naturally. Secondly, the evaluation is somewhat concerning due to the training process. Specifically, since the distillation approach makes use of GPT-4 to simulate user interaction data and the authors sample for uniqueness/variety, its fairly likely that the newly sampled user interaction data approximations will be very similar to these. Further, the largest model the authors compare to is again a GPT-4 variant. Given that GPT-4 is being used for both generation of data, as the system, and as the evaluation, there's a clear risk for bias. This could be addressed by any number of ways, including automated metrics for measuring game quality developed in prior game generation-like work [8,9]. But at present, it is difficult as an outside reader to get a sense of the actual quality of UNBOUNDED as a game. Failing this, it might be beneficial to include longer examples of gameplay in the supplementary materials at least.\n\n### Figures\n\nWhile these are not of major concern, I wanted to note two aspects of figures that I felt could be improved. First, the authors make use of a large number of fonts in their figures, and their primary font is a bit difficult to read. I'd suggest updating the figures to use a single, more legible font. Second, Figure 4 is largely superfluous. If the authors wish to retain it, it might be a better fit for an appendix.\n\n1. Treanor, M., Zook, A., Eladhari, M. P., Togelius, J., Smith, G., Cook, M., ... & Smith, A. (2015). AI-based game design patterns.\n2. Zhu, J., Villareale, J., Javvaji, N., Risi, S., Löwe, M., Weigelt, R., & Harteveld, C. (2021, May). Player-AI interaction: What neural network games reveal about AI as play. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (pp. 1-17).\n3. Shaker, N., Togelius, J., & Nelson, M. J. (2016). Procedural content generation in games.\n4. Pell, B. (1992). METAGAME: A new challenge for games and learning.\n5. Gallotta, R., Todd, G., Zammit, M., Earle, S., Liapis, A., Togelius, J., & Yannakakis, G. N. (2024). Large language models and games: A survey and roadmap. arXiv preprint arXiv:2402.18659.\n6. Anjum, A., Li, Y., Law, N., Charity, M., & Togelius, J. (2024, May). The Ink Splotch Effect: A case study on ChatGPT as a co-creative game designer. In Proceedings of the 19th International Conference on the Foundations of Digital Games (pp. 1-15).\n7. Guzdial, M., & Riedl, M. O. (2021). Conceptual game expansion. IEEE Transactions on Games, 14(1), 93-106.\n8. Khalifa, A., Green, M. C., Perez-Liebana, D., & Togelius, J. (2017, August). General video game rule generation. In 2017 IEEE Conference on Computational Intelligence and Games (CIG) (pp. 170-177). IEEE.\n9. Guzdial, M., & Riedl, M. (2018, September). Automated game design via conceptual expansion. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (Vol. 14, No. 1, pp. 31-37).","questions":"1. How do the authors position their work in the field of technical game design? \n2. What do longer gameplay examples look like? \n3. How do the authors respond to my concerns around their \"Evaluation of LLM Generations\"?"},{"id":"wmOOPEEhwh","initial_rating":6,"confidence":4,"soundness":3,"contribution":3,"presentation":3,"summary":"This work combines various aspects of Generative AI to develop a LLM-based game engine. The design principles refer to open-world, continuous interaction games with limited simultaneous characters but significant opportunities for situation generation and interaction. Capitalizing upon recent research in the field, the system introduces a number of technical innovations both in terms of architecture (in the form of a dynamic regional prompt adapter) and in term of implementation (distillation of Small LLM). With an existing implementation, the paper includes elements of comparative evaluation, both quantitative and qualitative.","strengths":"This is an ambitious piece of work, tackling difficult technical issues in balancing generation and interaction, as well as various training and fine-tuning protocols. It appears to have innovated on a number of aspects, for instance the dynamic mask to balance character and environment generation. It gives some interesting, detailed, insights such as those of Figure 4. The improvement over IP-Adapter thus appears more than incremental. Another strong points is the Small LLM distillation framework.\nThe comparative LLM evaluation has staged an appropriate number of LLM, capturing size and diversity. \nFinally, the paper demonstrates a good command of recent relevant work and compares its approach to really recent alternative methods (2023, 2024), which is welcome.\nHaving considered the authors' rebuttal and additional data, I have raised my score for contribution to 3 ('good').","weaknesses":"From a fundamental perspective, there is a lack of awareness of some game design issues. The introduction is relatively naïve considering how these issues are addressed in game/play theory (Caillois, Huizinga). With the fine line between simulation and entertainment, and notwithstanding the reference to The Sims and Tamagotchi, contemporary audiences tend to have gameplay expectations that can only be met with sustained and meaningful (in the narrative sense) interactions. This feeds into the overall perspective of the evaluation methods which in the paper are primarily focused individual ‘event’ generation. While sustained UX evaluation may be at odds with the system’s current level of maturity, some of this could have been addressed through even slightly longer exchanges or chains of events construed as minimal narrative experience (which is claimed in the abstract but does not really materialize in the examples). The small examples given, including on Figure 1, do not convey a very strong sense of gameplay. \n\nThere is no evidence that the number of user-simulator examples is at all sufficient for the intended purpose of multi-topic data collection. Nor is there any discussion of some hierarchical organization of topics that would come close to scenarios, game genres or elementary forms of plot backbones.\n\nDespite the statement on replicability, the level of implementation details is unclear. Implementation details are rather cursory and uneven with some in-depth details but no systematic discussion, and minor duplications from other sections.\n\nEvaluation has both positive (see above) and negative aspects.\nUnless I am mistaken, Table I shows a couple of inconsistencies for results.\nOn Environment Consistency, how can 0.322 be the best score when it is in-between 0.381 and 0.257? How can 0.675 be the best score in-between 0.595 and 0.832?\nIt is difficult to find the Qualitative component of evaluation very convincing in the absence of a more systematic approach for generating test situations. From a similar perspective, the example Prompts do not contain advanced mechanisms for controlling response, or specific examples. One would thus wonder how LLM response can be kept consistent across runs, and what type of statistical sampling has been associated to LLM use?","questions":"To which extent are results from Figure 6 consistent with results from Table I?\nHow have you ensured the quality and fairness (for evaluation) of the text prompts sample?\nHow can the current approach maintain narrative consistency over a number of interactions?"}],"string":"[\n {\n \"id\": \"B5fG2n5Q8V\",\n \"initial_rating\": 6,\n \"confidence\": 4,\n \"soundness\": 2,\n \"contribution\": 2,\n \"presentation\": 3,\n \"summary\": \"The authors present \\\"UNBOUNDED\\\" a virtual pet game based on LLMs. The authors present two novel technical elements to support this game: a regional IP-Adapter for character-environment consistency and a domain-specific distillation approach for finetuning a smaller LLM to allow for real time gameplay. The authors then present experiments comparing their approach to baselines in terms of visual quality and consistency, and overall game quality as scored by GPT-4o.\",\n \"strengths\": \"The main strengths of the paper are (1) the originality of entirely basing a virtual pet game on LLMs, (2) the significance of the two primary technical contributions to figure coherent image generation and real time LLM application uses, and (3) the quality of the image portion of the evaluation/experiments. The paper is also overall well-written, especially in terms of the technical aspects. These primarily justify the positive aspects of my above scores.\",\n \"weaknesses\": \"The paper has two major weaknesses in its present draft, the way it positions itself in terms of prior work and the evaluation. \\n\\n### Prior Work\\n\\nThe authors do not appear to have engaged with the field of technical games research at all. This is a shame, as their game can be understood as an AI-based game [1] or specifically a NN game [2]. More broadly, the authors' work in terms of generation would fit into the Procedural Content Generation paradigm [3]. There is a 3+ decades long history of work in this space that is relevant to the authors' work [4], with significant recent work at the intersection of LLMs and game generation [5], and it is crucial that the authors update the introduction and related work to situate their work within this field.\\n\\n### Evaluation\\n\\nThe authors primarily evaluate their work in terms of image quality and I have no major concern with these results (though the DreamSim results seem to incorrectly attribute the best score to their system). However, I have major concerns with the \\\"Evaluation of LLM Generations\\\", which is functionally the closest thing presented to a holistic evaluation of UNBOUNDED. Firstly, the norm within the area would be to make use of a user study for evaluation purposes [6,7]. Particularly since the authors make such an effort to allow for real time control, this would seem to follow naturally. Secondly, the evaluation is somewhat concerning due to the training process. Specifically, since the distillation approach makes use of GPT-4 to simulate user interaction data and the authors sample for uniqueness/variety, its fairly likely that the newly sampled user interaction data approximations will be very similar to these. Further, the largest model the authors compare to is again a GPT-4 variant. Given that GPT-4 is being used for both generation of data, as the system, and as the evaluation, there's a clear risk for bias. This could be addressed by any number of ways, including automated metrics for measuring game quality developed in prior game generation-like work [8,9]. But at present, it is difficult as an outside reader to get a sense of the actual quality of UNBOUNDED as a game. Failing this, it might be beneficial to include longer examples of gameplay in the supplementary materials at least.\\n\\n### Figures\\n\\nWhile these are not of major concern, I wanted to note two aspects of figures that I felt could be improved. First, the authors make use of a large number of fonts in their figures, and their primary font is a bit difficult to read. I'd suggest updating the figures to use a single, more legible font. Second, Figure 4 is largely superfluous. If the authors wish to retain it, it might be a better fit for an appendix.\\n\\n1. Treanor, M., Zook, A., Eladhari, M. P., Togelius, J., Smith, G., Cook, M., ... & Smith, A. (2015). AI-based game design patterns.\\n2. Zhu, J., Villareale, J., Javvaji, N., Risi, S., Löwe, M., Weigelt, R., & Harteveld, C. (2021, May). Player-AI interaction: What neural network games reveal about AI as play. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (pp. 1-17).\\n3. Shaker, N., Togelius, J., & Nelson, M. J. (2016). Procedural content generation in games.\\n4. Pell, B. (1992). METAGAME: A new challenge for games and learning.\\n5. Gallotta, R., Todd, G., Zammit, M., Earle, S., Liapis, A., Togelius, J., & Yannakakis, G. N. (2024). Large language models and games: A survey and roadmap. arXiv preprint arXiv:2402.18659.\\n6. Anjum, A., Li, Y., Law, N., Charity, M., & Togelius, J. (2024, May). The Ink Splotch Effect: A case study on ChatGPT as a co-creative game designer. In Proceedings of the 19th International Conference on the Foundations of Digital Games (pp. 1-15).\\n7. Guzdial, M., & Riedl, M. O. (2021). Conceptual game expansion. IEEE Transactions on Games, 14(1), 93-106.\\n8. Khalifa, A., Green, M. C., Perez-Liebana, D., & Togelius, J. (2017, August). General video game rule generation. In 2017 IEEE Conference on Computational Intelligence and Games (CIG) (pp. 170-177). IEEE.\\n9. Guzdial, M., & Riedl, M. (2018, September). Automated game design via conceptual expansion. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (Vol. 14, No. 1, pp. 31-37).\",\n \"questions\": \"1. How do the authors position their work in the field of technical game design? \\n2. What do longer gameplay examples look like? \\n3. How do the authors respond to my concerns around their \\\"Evaluation of LLM Generations\\\"?\"\n },\n {\n \"id\": \"wmOOPEEhwh\",\n \"initial_rating\": 6,\n \"confidence\": 4,\n \"soundness\": 3,\n \"contribution\": 3,\n \"presentation\": 3,\n \"summary\": \"This work combines various aspects of Generative AI to develop a LLM-based game engine. The design principles refer to open-world, continuous interaction games with limited simultaneous characters but significant opportunities for situation generation and interaction. Capitalizing upon recent research in the field, the system introduces a number of technical innovations both in terms of architecture (in the form of a dynamic regional prompt adapter) and in term of implementation (distillation of Small LLM). With an existing implementation, the paper includes elements of comparative evaluation, both quantitative and qualitative.\",\n \"strengths\": \"This is an ambitious piece of work, tackling difficult technical issues in balancing generation and interaction, as well as various training and fine-tuning protocols. It appears to have innovated on a number of aspects, for instance the dynamic mask to balance character and environment generation. It gives some interesting, detailed, insights such as those of Figure 4. The improvement over IP-Adapter thus appears more than incremental. Another strong points is the Small LLM distillation framework.\\nThe comparative LLM evaluation has staged an appropriate number of LLM, capturing size and diversity. \\nFinally, the paper demonstrates a good command of recent relevant work and compares its approach to really recent alternative methods (2023, 2024), which is welcome.\\nHaving considered the authors' rebuttal and additional data, I have raised my score for contribution to 3 ('good').\",\n \"weaknesses\": \"From a fundamental perspective, there is a lack of awareness of some game design issues. The introduction is relatively naïve considering how these issues are addressed in game/play theory (Caillois, Huizinga). With the fine line between simulation and entertainment, and notwithstanding the reference to The Sims and Tamagotchi, contemporary audiences tend to have gameplay expectations that can only be met with sustained and meaningful (in the narrative sense) interactions. This feeds into the overall perspective of the evaluation methods which in the paper are primarily focused individual ‘event’ generation. While sustained UX evaluation may be at odds with the system’s current level of maturity, some of this could have been addressed through even slightly longer exchanges or chains of events construed as minimal narrative experience (which is claimed in the abstract but does not really materialize in the examples). The small examples given, including on Figure 1, do not convey a very strong sense of gameplay. \\n\\nThere is no evidence that the number of user-simulator examples is at all sufficient for the intended purpose of multi-topic data collection. Nor is there any discussion of some hierarchical organization of topics that would come close to scenarios, game genres or elementary forms of plot backbones.\\n\\nDespite the statement on replicability, the level of implementation details is unclear. Implementation details are rather cursory and uneven with some in-depth details but no systematic discussion, and minor duplications from other sections.\\n\\nEvaluation has both positive (see above) and negative aspects.\\nUnless I am mistaken, Table I shows a couple of inconsistencies for results.\\nOn Environment Consistency, how can 0.322 be the best score when it is in-between 0.381 and 0.257? How can 0.675 be the best score in-between 0.595 and 0.832?\\nIt is difficult to find the Qualitative component of evaluation very convincing in the absence of a more systematic approach for generating test situations. From a similar perspective, the example Prompts do not contain advanced mechanisms for controlling response, or specific examples. One would thus wonder how LLM response can be kept consistent across runs, and what type of statistical sampling has been associated to LLM use?\",\n \"questions\": \"To which extent are results from Figure 6 consistent with results from Table I?\\nHow have you ensured the quality and fairness (for evaluation) of the text prompts sample?\\nHow can the current approach maintain narrative consistency over a number of interactions?\"\n }\n]","isConversational":false},"scores":{"kind":"string","value":"{\n \"rating\": \"5;6\",\n \"rating_avg\": 5.5,\n \"confidence\": \"4;4\",\n \"confidence_avg\": 4,\n \"soundness\": \"2;3\",\n \"soundness_avg\": 2.5,\n \"contribution\": \"2;2\",\n \"contribution_avg\": 2,\n \"presentation\": \"2;3\",\n \"presentation_avg\": 2.5\n}"},"metadata":{"kind":"string","value":"{\n \"primary_area\": \"\",\n \"track\": \"main\",\n \"venue\": \"ICLR 2025 Poster\",\n \"venueid\": \"ICLR.cc/2025/Conference\",\n \"processed_at\": \"2026-01-14T22:16:03.998593\"\n}"},"meta_review":{"kind":"string","value":"{\n \"id\": \"VB2N4zOWA8\",\n \"metareview\": \"The paper presents an interesting experiment in creating a generative game where the core \\\"game engine\\\" is based on LLMs and diffusion models. While obviously there is a long way to go, there are interesting things to learn from the paper. The \\\"core technical\\\" contribution of the paper is perhaps the regional IP-adapter used to ensure that placement of characters in images work, but the bigger contribution is the system as a whole. It would have been nice with a proper user study to understand how people play with the system, but... next paper?\\nThe first version of this paper was seriously lacking in its related work section, essentially ignoring the most relevant research community, but this has been much improved in the updated version.\",\n \"additional_comments\": \"The authors engaged well with the reviewers and updated their paper appropriately.\"\n}"},"decision":{"kind":"string","value":"{\n \"decision\": \"Accept (Poster)\"\n}"}}},{"rowIdx":3699,"cells":{"paper_id":{"kind":"string","value":"uy4EavBEwl"},"arxiv_id":{"kind":"string","value":"2405.19667v1"},"title":{"kind":"string","value":"Reconciling Model Multiplicity for Downstream Decision Making"},"markdown":{"kind":"string","value":"{\n \"content\": \"## Abstract\\n\\nAbstract We consider the problem of model multiplicity in downstream decision-making, a setting where two predictive models of equivalent accuracy cannot agree on the best-response action for a downstream loss function. We show that even when the two predictive models approximately agree on their individual predictions almost everywhere, it is still possible for their induced best-response actions to differ on a substantial portion of the population. We address this issue by proposing a framework that calibrates the predictive models with regard to both the downstream decision-making problem and the individual probability prediction. Specifically, leveraging tools from multi-calibration, we provide an algorithm that, at each time-step, first reconciles the differences in individual probability prediction, then calibrates the updated models such that they are indistinguishable from the true probability distribution to the decision-maker. We extend our results to the setting where one does not have direct access to the true probability distribution and instead relies on a set of i.i.d data to be the empirical distribution. Finally, we provide a set of experiments to empirically evaluate our methods: compared to existing work, our proposed algorithm creates a pair of predictive models with both improved downstream decision-making losses and agrees on their best-response actions almost everywhere.\\n\\n## 1 Introduction\\n\\nIn many applications, individual probability prediction is at the heart of a decision-making process.\\nFor example, in the Job Training Partnership Act (JTPA) training program (Bloom et al., ), a decision-maker may want to predict whether an individual is employed or not before assigning them to training; or in medical trials, a doctor wants to predict the probability that the patient has contracted a disease before recommending them a treatment. Since the hospital does not know the true individual probability that a particular patient is ill, they can only evaluate the individual probability predictions through its average outcome over a sufficiently large sample set. For a predictive task, the standard convention is to choose the model that maximizes *accuracy*. However, previous work has shown that it is common to have multiple predictive models with similar accuracy but substantially different properties (Chen et al., ; Rodolfa et al., ; D’Amour et al., ). This phenomenon is called *predictive* (or model) multiplicity, a line of work studied by Breiman (); Marx et al. (); Black et al. ().\\n\\nIn a predictive multiplicity scenario, the decision-maker may have two or more predictive models that are nearly equivalent in terms of accuracy but disagree on their predictions on many individual samples. In our motivating example (Figure 1), the hospital has access to two models $f_{1}$ and $f_{2}$ predicting the probability of disease which are equally accurate on average over the entire population, but their predictions on a subpopulation may vastly differ. This disagreement in outcome prediction may have a disparate impact on the subpopulation if the hospital has to choose one predictor over the other to make important downstream decisions. For example, they might select a treatment based on the predicted probability that a patient has contracted a disease. Formally, given a predictive model $f$ and a decision-making loss function $\\\\ell(y,\\\\cdot)$, the decision-maker wants to choose a best-response action, i.e., the action $a$ that minimizes $\\\\mathbb{E}_{y\\\\sim f}[\\\\ell(^{\\\\prime},a)]$. When two models $f_{1}$ and $f_{2}$ have nearly equivalent accuracy but lead to different best-response actions, the decision-maker would not be able to identify which best-response action to take for individual patients. While predictive multiplicity offers great flexibility for the decision-maker in the model selection process, it also places an additional burden on the decision-maker to correctly navigate such freedom and justify how they use a predictive model to make downstream decisions.\\n\\nRoth et al. () attempts to address the model multiplicity issue by resolving prediction disagreement between models. Adapting techniques from the literature of multi-calibration (Hebert-Johnson et al., ), they provide a procedure called \\\"Reconcile\\\" that updates the predictive models to minimize their disagreements and improve the accuracy of each model. However, we show simple settings where the reconciled predictions from Roth et al. () can lead the downstream decision-makers to take actions with substantially higher losses.\\nWe visually demonstrate this scenario in Figure 1. For a more detailed discussion on the limitation of prior work, see Section 2.4 and Appendix A. This motivates the study of how to reconcile predictive multiplicity with an explicit focus on its impact on downstream decisions.\\n\\nFigure: Figure 1: An illustrative example of the drawback in a prior work’s attempt at addressing model multiplicity. Consider a stylized binary classification problem on a dataset with $8$ units (patients) and the hospital deciding between two actions (treatment vs. no treatment). Treatment is assigned if the predicted probability is above 1/2. Left: The true probability that each patient is labeled ‘ill’. Middle: The predicted probability that each patient is ill according to $f_{1}$ (white) and $f_{2}$ (blue). While these two predictors have almost the same accuracy, their individual probability predictions for patients $3$ and $6$ vastly differ. Right: After running the Reconcile procedure of Roth et al. (), the individual probability predictions agree everywhere. However, the best-response action of unit $3$ changed from correct (no treatment) to incorrect (treatment). If the hospital uses the updated $f_{1}$ to make their treatment recommendation, they would incur more loss than before had they not updated the predictor using Reconcile. This example is formalized in Theorem 2.7.\\nRefer to caption: extracted/5630318/plots/counterexample/counterexample2.png\\n\\nIn this work, our goal is to leverage tools from multi-calibration (Hebert-Johnson et al., ) to alleviate the model multiplicity issue in high-dimensional decision-making tasks for multiple decision-makers with multiple decision-making loss functions. Specifically, we show how the decision-maker can update a pair of predictors so they approximately agree on (1) individual predictions and (2) best-response actions for each individual in the downstream decision-making task. Our procedure ensures that the number of disagreements in best-response actions decreases over time, which enables the decision-makers to confidently use either of the updated predictors to justify their decisions.\\n\\n#### Overview of Paper.\\n\\nWe study the problem of reconciling model multiplicity for multiple downstream decision-making tasks, where the decision-makers have two predictive models with nearly equivalent accuracy but may lead to vastly different best-response actions for a significant number of individuals in the population. Our key contributions are summarized as follows.\\n\\n- •\\nIn Section 2, we formulate the problem of model multiplicity from the perspective of the decision-makers.\\nIn Section 2.4, we formalize our motivating example in Figure 1 and show that it is insufficient to only update two predictive models so that they have improved accuracy and nearly agree on their individual predictions almost everywhere.\\n- •\\nIn Section 3, we introduce an algorithm, ReDCal, that outputs predictive models that are (1) calibrated to a finite set of downstream decision-making tasks and (2) approximately agree on their predictions and best-response actions almost everywhere for each downstream task.\\n- •\\nIn Section 3.2, we extend our analysis to the setting where one does not have direct access to the true distribution and instead only has a validation dataset with samples drawn i.i.d from the underlying distribution. We show that the guarantees obtained using the empirical distribution can be translated to the unknown underlying distribution.\\n- •\\nFinally, in Section 4, we empirically evaluate the performance of the proposed algorithm on real-world datasets and show our improvement over the benchmark prior work in resolving disagreement in downstream decision-making tasks.\\n\\n### 1.1 Related Work\\n\\n#### Model multiplicity.\\n\\nWithin the literature on predictive multiplicity, our work builds off the line of work focusing on predicting individual probabilities (Marx et al., ; D’Amour et al., ; Black et al., ; Breiman, ), where solving an error minimization problem for some prediction tasks can lead to multiple solutions with roughly similar performance in terms of accuracy. Sandroni () showed that one cannot empirically distinguish the outcomes from a predictor encoding the true individual probabilities from one without in isolation. Al-Najjar and Weinstein (); Feinberg and Stewart () provided comparative tests to differentiate between the true probability predictor and one that is not. Particularly, Feinberg and Stewart () relied on *cross-calibration*, i.e., calibration conditional on the predictions of both models to empirically falsify one of them. For downstream decision-making, Garg et al. () worked on refining predictors and provided an algorithm that produces a predictor $f_{3}$ that is cross-calibrated with respect to both $f_{1}$ and $f_{2}$. An alternative framework studied by Globus-Harris et al. () seeks to update models that are sub-optimal for different subsets of the population, following the ’bug bounties’ approach used by the software and security communities.\\n\\nParticularly relevant to our work is Roth et al. (), which aims to reconcile different predictors with equivalent errors such that the updated predictors both have lower errors compared to the initial models and approximately agree on their prediction on almost all units. Despite the similarity in our motivation, our results go beyond the binary classification setting considered in Roth et al. (). In our model, we consider reconciling predictors for both regression and multi-class classification problems, and their impact on the downstream decision-making tasks. In Section 2.4, we provide a numerical example where simply reconciling the probability predictions according to Roth et al. () can lead to additional losses in downstream decision-making tasks.\\n\\n#### Calibration and Multi-calibration.\\n\\nOur work draws on techniques from the growing literature on multi-calibration (Hebert-Johnson et al., ; Kim et al., ; Dwork et al., ; Shabat et al., ; Jung et al., ; Dwork et al., ; Jung et al., ; Haghtalab et al., ; Deng et al., ; Noarov et al., ). Multi-calibration has been used as a notion of fairness as it guarantees calibration for any identifiable group.\\n\\nWithin the framework of multi-calibration, the work most related to ours is that of Zhao et al. (), who considered decision calibration with respect to all classes of loss functions. Similar to us, Zhao et al. () takes the perspective of a decision-maker who wants to ensure the predictive models are *indistinguishable* from the true probability when they are used to make downstream decisions. However, two decision-calibrated models can still disagree on their individual predictions and best-response actions for sufficiently many units. In our motivating example, a hospital with two decision-calibrated predictors may still want to make their predictive models approximately agree on their predictions for almost all individuals in the population and lead to the same downstream decision, i.e., whether to recommend treatment or not based on the predicted probability that a patient has contracted a disease. We formalize this example in Section 2.4 and provide empirical experiments to show our improvement over their result in Section 4.\\n\\nAn independent and concurrent work by Globus-Harris et al. () considers ensembling multiple predictive models for high-dimensional downstream decision-making tasks. While their work also leverages techniques from multi-calibration, their goal is to output an ensembled predictor whose self-estimated expected payoff is accurate and whose induced policy has a payoff at least as high as the maximum self-assessed payoff of individual models. In contrast, our interest is in reducing the downstream decision losses by resolving the differences between equivalent predictors and mitigating model multiplicity for decision-making.\\n\\n## 2 Problem Formulation\\n\\n#### Notation.\\n\\nThroughout this paper, we use subscripts $i$ to index different predictions, superscripts $t$ to index different time-steps, and $a$ to index actions. For $K\\\\in\\\\mathbb{N}$, we use the shorthand $[K]\\\\coloneqq\\\\{1,2,\\\\cdots,K\\\\}$. $\\\\Delta(\\\\mathcal{X})$ denote the set of possible distributions over $\\\\mathcal{X}$.\\n\\nWe consider the prediction problem with random variables $x$ and $y$, where $x\\\\in\\\\mathcal{X}$ represents the features and $y\\\\in\\\\mathcal{Y}$ represents the labels. We focus on the regression problem in which the label domain is real-valued and bounded: $\\\\mathcal{Y}\\\\subset[0,1]^{d}$. Our formulation also permits the multi-class classification problem by writing the label $y$’s as one-hot vectors, e.g., for $\\\\left\\\\lvert\\\\mathcal{Y}\\\\right\\\\rvert=3$, we can write $\\\\mathcal{Y}=\\\\{[1,0,0],[0,1,0],[0,0,1]\\\\}$.\\n\\nWe denote $\\\\mathcal{D}\\\\in\\\\Delta(\\\\mathcal{X}\\\\times\\\\mathcal{Y})$ as the true distribution over the pairs of features-label $(x,y)$. In practice, we will not have access to $\\\\mathcal{D}$, and instead only know a set of $n$ data points $D$ sampled i.i.d from $\\\\mathcal{D}$. In such case, we consider the dataset $D=\\\\{(x_{1},y_{1}),\\\\cdots,(x_{n},y_{n})\\\\}$ to be the *empirical distribution* over $D$, which is a discrete distribution that place uniform weight $\\\\nicefrac{{1}}{{n}}$ on each sample $(x,y)\\\\in D$.\\n\\nA predictor is a map $f:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$. Our goal is to find the Bayes optimal predictor $f^{*}:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$ such that for all $x\\\\in\\\\mathcal{X}$, $f^{*}(x)=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[y|x]$ is the *conditional label expectation given $x$*.\\n\\n### 2.1 Model Evaluation\\n\\nGiven a predictor $f\\\\in[0,1]^{d}$, we evaluate $f$ via its squared error, i.e., its expected deviation from the true label. We formalize this objective in the following definition.\\n\\n###### Definition 2.1 (Brier Score) .\\n\\nThe squared error (also known as Brier score) of a predictor $f$ evaluated on distribution $\\\\mathcal{D}$ is given as:\\n\\n$$ $B(f,\\\\mathcal{D})=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\left\\\\lVert f(x)-y\\\\right \\\\rVert_{2}^{2}]$ $$\\n\\nWhen we only have a dataset $D=\\\\{(x_{1},y_{1}),\\\\cdots,(x_{n},y_{n})\\\\}$, the empirical Brier score is given as:\\n\\n$$ $B(f,D)=\\\\frac{1}{n}\\\\sum_{i=1}^{n}\\\\left\\\\lVert f(x_{i})-y_{i}\\\\right\\\\rVert_{2}^{2}$ $$\\n\\nNote that we use the Brier score as our metric because it can be accurately estimated given access to only the samples from the distribution. Moreover, among all possible predictors, the Brier score is minimized by the Bayes optimal predictor $f^{*}$.\\n\\n###### Lemma 2.2 .\\n\\nFix any distribution $\\\\mathcal{D}$ and let $f^{*}(x)=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[y|x]$ represent the true conditional label encoded by $\\\\mathcal{D}$. Let $f:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$ be any other model. Then we have $B(f^{*},\\\\mathcal{D})\\\\leq B(f,\\\\mathcal{D})$.\\n\\nHence, given two predictors $f_{1}$ and $f_{2}$, if we can verify empirically from the observable data that $B(f_{1},\\\\mathcal{D})\\\\leq B(f_{2},\\\\mathcal{D})$, then we can empirically falsify that $f_{2}$ encodes the true conditional label.\\n\\n###### Definition 2.1 (Brier Score) .\\n\\n###### Lemma 2.2 .\\n\\n### 2.2 Downstream Decision-Making Tasks and Loss Functions\\n\\nBeyond our initial goal of finding a good estimate for the true conditional predictor $f^{*}$, we are also interested in using our predictors for downstream decision-making problems. Formally, we consider a loss minimization problem, where the decision-maker has a set of possible actions $\\\\mathcal{A}$ and a loss function $\\\\ell:\\\\mathcal{Y}\\\\times\\\\mathcal{A}\\\\rightarrow[0,d]$. Wlog, we only consider action set $\\\\mathcal{A}=[K]$, i.e., there are $K$ possible actions. In this paper, we assume that the loss function does not directly depend on the features $x$ and is linear in $\\\\mathcal{Y}$. That is, for each action $a\\\\in\\\\mathcal{A}$, there exists some $\\\\ell_{a}\\\\in[0,1]^{d}$ such that\\n\\n$$ $\\\\ell(y,a)=\\\\langle y,\\\\ell_{a}\\\\rangle$ $$\\n\\nWe write $\\\\mathcal{L}=\\\\{\\\\ell:\\\\mathcal{Y}\\\\times\\\\mathcal{A}\\\\rightarrow[0,d]\\\\}$ to denote a finite family of loss functions. In general, we consider the setting with multiple different decision-makers, each using a different linear loss function in $\\\\mathcal{L}$. For any loss function $\\\\ell:\\\\mathcal{Y}\\\\times\\\\mathcal{A}\\\\rightarrow\\\\mathbb{R}$, we can rescale each coordinate of $\\\\ell_{a}$ to be between $[0,1]$.\\n\\nGiven a predictor $f$ and a loss function $\\\\ell$, the decision-maker selects an action $a\\\\in\\\\mathcal{A}$ that minimizes the expected loss. We define the best-response policy taken by the decision-maker as follows.\\n\\n###### Definition 2.3 (Best-response policy) .\\n\\nGiven a loss function $\\\\ell$, a predictor $f$ and the action set $\\\\mathcal{A}$, the best-response policy for $\\\\ell$ is given as\\n\\n$$ $\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f(x))=\\\\mathrm{argmin}_{a\\\\in\\\\mathcal{A}}\\\\langle f(x), \\\\ell_{a}\\\\rangle.$ $$\\n\\n###### Definition 2.3 (Best-response policy) .\\n\\n### 2.3 Calibration\\n\\nIn our setting, we consider the decision-maker only having access to some pre-trained predictors $f$ given by a third-party. For instance, a data scientist trained a pair of models on an image dataset without exact knowledge of how the downstream decision-maker will use such predictors. We may imagine the decision-maker as a hospital considering whether to recommend treatment to certain patients based on the predicted probability that the patient has contracted a skin disease. Since the hospital’s treatment-recommendation algorithm is not known to the public (and the data scientist), we assume that the data scientist initially aim to minimize the squared error in their predictions.\\n\\nSince the hospital believes that the input predictors may not perform well according to their own loss function, they want the data scientist to convey trust through other performance guarantees of the predictors. One such guarantee is multi-calibration with respect to a finite set of loss functions $\\\\mathcal{L}\\\\ni\\\\ell$ and a set of events $\\\\mathcal{E}$ on the best-response policy, i.e., if the loss function $\\\\ell$ belongs to $\\\\mathcal{L}$, the decision-maker should be able to accurately compute the expected loss of choosing an action using the best-response policy $\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}$. Formally, we let $E_{a,\\\\ell}(f(x),x)$ denote the action selection events:\\n\\n###### Definition 2.4 (Best-response Events) .\\n\\nGiven a predictor $f$ and a loss function $\\\\ell$, for each action $a\\\\in[K]$, define the event\\n\\n$$ $E_{\\\\ell,a}(f(x),x)=\\\\mathbf{1}\\\\{x:\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f(x))=a\\\\}$ $$\\n\\nand let $\\\\mathcal{E}=\\\\{E_{\\\\ell,a}\\\\}_{a\\\\in[K],\\\\ell\\\\in\\\\mathcal{L}}$.\\n\\nGiven a set of events $\\\\mathcal{E}$, we can define an approximate notion of multi-calibration with respect to $\\\\mathcal{E}$.\\n\\n###### Definition 2.5 ( $\\\\beta$ -approximate decision calibration) .\\n\\nA predictor $f$ is $\\\\beta-$decision calibrated with respect to the set of best-response events $\\\\mathcal{E}$ if for all $E_{\\\\ell,a}\\\\in\\\\mathcal{E}$, we have:\\n\\n$$ $\\\\left\\\\lVert\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(y-f(x))\\\\cdot E_{\\\\ell,a}(f(x),x)] \\\\right\\\\rVert_{2}\\\\leq\\\\beta.$ $$\\n\\nThis definition follows from an equivalent definition of decision calibration in Zhao et al. (). The main difference is we define calibration with respect to a set of events on the best-response actions following the formulation of multi-calibration for online learning in Noarov et al. () and a generalization of multi-calibration in Deng et al. ().\\nThis definition implies that if a predictor $f$ is $\\\\beta-$decision calibrated with respect to the best-response events $\\\\mathcal{E}$, then the decision-maker can accurately estimate the expected loss from using $f$ to make decisions.\\n\\n###### Lemma 2.6 ( (Zhao et al., 2021 ) ) .\\n\\nFor all $a,a^{\\\\prime}\\\\in\\\\mathcal{A},\\\\ell\\\\in\\\\mathcal{L}$, if $f$ is $\\\\beta$-decision-calibrated with respect to the best-response events $\\\\mathcal{E}$ , then the loss estimation satisfies\\n\\n$$ $\\\\displaystyle\\\\left|\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,a^{\\\\prime})\\\\cdot E _{\\\\ell,a}(f(x),x)]-\\\\mathbb{E}_{x\\\\sim\\\\mathcal{D}_{\\\\mathcal{X}}}[\\\\langle f(x), \\\\ell_{a^{\\\\prime}}\\\\rangle\\\\cdot E_{\\\\ell,a}(f(x),x)]\\\\right|\\\\leq\\\\beta\\\\sqrt{d}$ $$\\n\\n###### Definition 2.4 (Best-response Events) .\\n\\n###### Definition 2.5 ( β 𝛽 \\\\beta italic_β -approximate decision calibration) .\\n\\n###### Lemma 2.6 ( (Zhao et al., 2021 ) ) .\\n\\n### 2.4 Limitations of Prior Works\\n\\nIn this section, we show that improving the accuracy until the two predictors agree on their predictions almost everywhere is not a sufficient solution to our problem. In our analysis below, we consider a stylized problem with $\\\\mathcal{Y}=\\\\{0,1\\\\}$ and $\\\\mathcal{A}=\\\\{0,1\\\\}$, i.e., binary class and binary action space. A predictor here is $f:\\\\mathcal{X}\\\\rightarrow[0,1]$, and the optimal predictor is $f^{*}(x)=\\\\Pr_{(x^{\\\\prime},y^{\\\\prime})\\\\sim\\\\mathcal{D}}[y^{\\\\prime}=1|x^{\\\\prime}=x]$. As shorthand, we denote $f_{1}(1)$ as the probability of unit $1$ being labeled $1$. The loss is defined as\\n\\n$$ $\\\\displaystyle\\\\ell(0,0)=\\\\langle[1,0],[0,1]\\\\rangle=0,\\\\quad\\\\ell(1,0)=\\\\langle[0,1] ,[0,1]\\\\rangle=1,$ $\\\\displaystyle\\\\ell(0,1)=\\\\langle[1,0],[1,0]\\\\rangle=1,\\\\quad\\\\ell(1,1)=\\\\langle[0,1] ,[1,0]\\\\rangle=0,$ (1) $$\\n\\nThat is, for any $x$, the best-response policy is to take action 0 0 if $f(x)\\\\leq\\\\nicefrac{{1}}{{2}}$ and action $1$ otherwise.\\n\\n#### Reconcile individual predictions.\\n\\nPrior work by Roth et al. () considers the model multiplicity problem for individual probability predictions. Their proposed algorithm, Reconcile (Algorithm 3), returns a pair of predictors that has a smaller Brier score than the input predictors and approximately agree on their predictions on almost all units. In the following theorem, we show that the best-response policy induced by the predictors updated by Reconcile might lead to a higher expected loss than the ones they started with.\\n\\n###### Theorem 2.7 .\\n\\nFor any $\\\\alpha\\\\in(0,\\\\nicefrac{{1}}{{3}})$, $\\\\eta\\\\in(0,1)$, there exists a pair of predictors $f_{1},f_{2}$ with equivalent accuracy such that after running Algorithm 3, the output models $f_{1}^{T},f_{2}^{T}$ agree on their individual predictions everywhere, but there exists a loss function $\\\\ell(y,a)$ such that $f_{1}^{T}$ and $f_{2}^{T}$ induce worse losses compared to the original models.\\n\\n###### Proof.\\n\\nConsider a setting with $\\\\mathcal{X}=[2]$ and $\\\\Pr[x=1]=\\\\Pr[x=2]=0.5$. For any $0<\\\\alpha<\\\\nicefrac{{1}}{{3}}$, let $\\\\phi\\\\geq\\\\alpha$, we consider the two predictors $f_{1},f_{2}$ defined as follows:\\n\\n$$ $\\\\displaystyle f_{1}(1)=\\\\frac{1}{2}-\\\\frac{\\\\phi}{2},\\\\quad f_{1}(2)=\\\\frac{1}{2}- \\\\frac{3\\\\phi}{2},\\\\quad f_{2}(1)=\\\\frac{1}{2}+\\\\frac{\\\\phi}{2},\\\\quad f_{2}(2)=\\\\frac {1}{2}-\\\\frac{\\\\phi}{2}$ (2) $$\\n\\nand the true probability of each unit being labeled $1$ are $f^{*}(1)=0$ and $f^{*}(2)=1$.\\n\\nThe Brier scores of $f_{1}$ and $f_{2}$ differ only by $\\\\phi^{2}$, but their individual predictions differ for both feature $x=1$ and $x=2$. We can run Algorithm 3 and patch $f_{1}$ to get the updated model $f_{1}^{T}$ with\\n\\n$$ $f_{1}^{T}(1)=\\\\frac{1}{2}+\\\\frac{\\\\phi}{2}=f_{2}(1),\\\\quad f_{1}^{T}(2)=\\\\frac{1}{2 }-\\\\frac{\\\\phi}{2}=f_{2}(2).$ $$\\n\\nConsider the loss function $\\\\ell$ defined as Equation 1.\\nThe change in expected loss after patching $f_{1}$ is\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{1}^{T}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{1}(x)))]=\\\\frac{1}{2}>0.$ $$\\n\\nTherefore, no matter how small we let $\\\\alpha$ and $\\\\eta$ be, the loss of predictor $f_{1}$ increases by a constant amount after running Algorithm 3. For the theoretical guarantees of Algorithm 3, see Appendix A.\\n∎\\n\\nMoreover, we provide a counterexample to show that it is insufficient to only ensure each individual predictor is approximately decision-calibrated using Algorithm 1.\\n\\n###### Theorem 2.7 .\\n\\n###### Proof.\\n\\n#### Decision-Calibrated predictions.\\n\\nAnother baseline algorithm we consider is to run Decision Calibration (Algorithm 1) separately for both $f_{1},f_{2}$. However, in the following theorem, we show that the updated predictors $f_{1}^{\\\\prime}$ and $f_{2}^{\\\\prime}$ can still disagree with each other on the best-response actions for substantially many units, indicating room for further improvement.\\n\\n###### Theorem 2.8 .\\n\\nFor any $\\\\eta\\\\in(0,\\\\nicefrac{{1}}{{4}})$ and $\\\\beta\\\\in(0,\\\\nicefrac{{1}}{{2}})$, there exists a pair of predictors $f_{1},f_{2}$ and a loss function $\\\\ell(y,a)$ such that after running Decision-Calibration (Zhao et al., ), the resulting models $f_{1}^{T},f_{2}^{T}$ are $\\\\beta$-decision-calibrated with respect to the loss function $\\\\ell$. There exists a set of units $x$ with probability mass $2\\\\eta$ where $f_{1}^{T}$ and $f_{2}^{T}$ disagree on the individual best-response actions.\\n\\n###### Proof.\\n\\nFor any $\\\\eta\\\\in(0,\\\\nicefrac{{1}}{{4}}),\\\\beta\\\\in(0,\\\\nicefrac{{1}}{{2}})$, let $\\\\mathcal{X}=[4]$, with $\\\\Pr[1]=\\\\Pr[4]=\\\\nicefrac{{1}}{{2}}-\\\\eta$ and $\\\\Pr[x=2]=\\\\Pr[x=3]=\\\\eta$. Consider the predictors $f_{1},f_{2}$ as follows:\\n\\n$$ $\\\\displaystyle f_{1}(1)=f_{1}(2)=f_{2}(1)=f_{2}(3)=\\\\frac{\\\\beta}{4}-2\\\\eta\\\\beta+2\\\\eta,$ (3) $\\\\displaystyle f_{1}(3)=f_{1}(4)=f_{2}(2)=f_{2}(4)=1-\\\\frac{\\\\beta}{2},$ (4) $\\\\displaystyle f^{*}(1)=\\\\frac{\\\\beta}{2},~{}~{}f^{*}(2)=f^{*}(3)=f^{*}(4)=1- \\\\frac{\\\\beta}{2}.$ (5) $$\\n\\nNotice that\\n\\n$$ $\\\\displaystyle\\\\frac{\\\\alpha}{4}-2\\\\eta\\\\alpha+2\\\\eta=\\\\frac{\\\\alpha}{4}+2\\\\eta(1- \\\\alpha)<\\\\frac{1}{4}+\\\\frac{1}{4}=\\\\frac{1}{2}.$ $$\\n\\nThe best-response policy for each predictor is\\n\\n$$ $\\\\displaystyle\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{1}(1))=\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{1} (2))=0,~{}~{}$ $\\\\displaystyle\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{1}(3))=\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{1} (4))=1,$ (6) $\\\\displaystyle\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{2}(1))=\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{2} (3))=0,~{}~{}$ $\\\\displaystyle\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{2}(2))=\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{1} (4))=1.$ (7) $$\\n\\nFor each best-response event, we have\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[f^{*}(x)E_{0}(f_{1}(x),x)]$ $\\\\displaystyle=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[f_{1}(x)E_{0}(f_{1}(x),x)],$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[f^{*}(x)E_{1}(f_{1}(x),x)]$ $\\\\displaystyle=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[f_{1}(x)E_{1}(f_{1}(x),x)],$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[f^{*}(x)E_{0}(f_{2}(x),x)]$ $\\\\displaystyle=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[f_{2}(x)E_{0}(f_{2}(x),x)],$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[f^{*}(x)E_{1}(f_{2}(x),x)]$ $\\\\displaystyle=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[f_{2}(x)E_{1}(f_{2}(x),x)].$ $$\\n\\nThat is, $f_{1},f_{2}$ are already decision-calibrated, so running Decision Calibration (Algorithm 1) will not further improve either of the two predictors. However, based on our definition of disagreement events (Definition 3.1), we still have\\n\\n$$ $\\\\displaystyle E_{0,1}=\\\\{2\\\\},~{}~{}E_{1,0}=\\\\{3\\\\},$ $$\\n\\neach with size\\n\\n$$ $\\\\displaystyle\\\\mu(E_{0,1})=\\\\mu(E_{1,0})=\\\\eta.$ $$\\n\\nWe observe that $f_{1}$ and $f_{2}$ still disagree on the best-response action for units $x=2$ and $x=3$.\\nWe can further reduce the differences in best-response actions using our algorithm Algorithm 2.\\n∎\\n\\n###### Theorem 2.8 .\\n\\n###### Proof.\\n\\n## 3 Reconcile for Decision Making\\n\\nSuppose we are given two predictors $f_{1},f_{2}:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$. We consider the model multiplicity problem with respect to the downstream decision-making problem – where $f_{1},f_{2}$ with nearly equivalent accuracy differ in their induced decision-making policies, but we cannot falsify either of the two from the data. Informally, our goal is to return a pair of models $f_{1}^{\\\\prime},f_{2}^{\\\\prime}$ such that: (1) for both $i\\\\in\\\\{1,2\\\\}$, $f_{i}^{\\\\prime}$ is more accurate than $f_{i}$ in terms of Brier score; (2) for both $i\\\\in\\\\{1,2\\\\}$, the best-response policy induced by $f_{i}^{\\\\prime}$ has no larger expected loss than that of $f_{i}$; (3) $f_{1}^{\\\\prime}$ and $f_{2}^{\\\\prime}$ approximately agree almost everywhere, indicating limited room for additional improvement.\\n\\nTo this end, we are interested in the region where the two predictors disagree substantially with respect to the downstream decision-making task. We define the disagreement region as follows:\\n\\n###### Definition 3.1 (Disagreement Event) .\\n\\nFor $f_{1},f_{2}$, margin $\\\\alpha>0$, and a loss function $\\\\ell$, the disagreement event is defined for a pair of best-response actions $a_{1},a_{2}\\\\in\\\\mathcal{A}$ where $a_{1}\\\\neq a_{2}$ as\\n\\n$$ $\\\\displaystyle E^{\\\\alpha}_{\\\\ell,a_{1},a_{2}}(f_{1}(x),f_{2}(x),x)=\\\\mathbb{I} \\\\Big{[}x\\\\in\\\\big{\\\\{}x:$ $\\\\displaystyle\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{1}(x))=a_{1},\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{2}(x))=a_{2},$ $\\\\displaystyle\\\\langle f_{1}(x),\\\\ell_{a_{2}}-\\\\ell_{a_{1}}\\\\rangle>\\\\alpha~{}\\\\text{ or}~{}\\\\langle f_{2}(x),\\\\ell_{a_{1}}-\\\\ell_{a_{2}}\\\\rangle>\\\\alpha\\\\big{\\\\}}\\\\Big{]},$ $$\\n\\nAs shorthand, we denote $E_{\\\\ell,a_{1},a_{2}}(x)=E^{\\\\alpha}_{\\\\ell,a_{1},a_{2}}(f_{1}(x),f_{2}(x),x)$ when the predictors $f_{1},f_{2}$ and the margin $\\\\alpha$ are clear from context. For a finite family of loss functions $\\\\mathcal{L}$, we can always iterate through $\\\\mathcal{L}$ to identify the tuple $(\\\\ell,a_{1},a_{2})$ that defines a disagreement region between $f_{1}$ and $f_{2}$.\\n\\nWe say the two models approximately agree with each other when the size of the disagreement event is small enough, i.e., its probability mass $\\\\mu(E_{\\\\ell,a_{1},a_{2}})$ on the underlying distribution $\\\\mathcal{D}$ is small.\\n\\n###### Definition 3.1 (Disagreement Event) .\\n\\n### 3.1 The Reconcile Procedure\\n\\nIn this section, we propose our main algorithm, ReDCal (Algorithm 2). Whenever the decision-maker observes a large disagreement event $E_{\\\\ell,a_{1},a_{2}}$, the best-response action and its corresponding expected loss given by at least one of the predictors must be incorrect.\\nFor example, at time step $t$ and a unit $x$, if the gap between the losses after taking $a_{1}$ and $a_{2}$ according to $f_{2}$ is substantially different from the loss gap observed on the data, then the decision-maker can induce that $f_{2}$ must have been wrong in its prediction for $x$.\\nThen, the decision-maker would want to ’patch’ predictor $f_{2}$ in this time-step.\\n\\nThe calibration procedure within each time-step is divided into two stages. In the first stage, we update model $f_{2}$ to $f_{2}^{\\\\prime}$ by minimizing the mean prediction error on the disagreement event, i.e., minimizing $\\\\mathbb{E}[\\\\|f_{2}^{\\\\prime}(x)-y|E_{\\\\ell,a_{1},a_{2}}(x)=1\\\\|]$. Following the intuition from multi-calibration, updating predictor $f_{2}$ in this manner would improve the Brier score and produce a more accurate predictor.\\nHowever, the updated model $f^{\\\\prime}_{2}$ is not guaranteed to induce the correct best-response action and could instead induce some other actions that might lead to a larger expected loss. To cope with this, in the second stage, we further update $f_{2}^{\\\\prime}$ to a model $f_{2}^{\\\\prime\\\\prime}$ that is approximately decision-calibrated within event $E_{\\\\ell,a_{1},a_{2}}$ using Algorithm 1. Since the loss estimation given by $f_{2}^{\\\\prime\\\\prime}$ is accurate for all best-response events within $E_{\\\\ell,a_{1},a_{2}}$ and we are taking actions to minimize estimated loss, we can now safely take the best-response action induced by $f_{2}^{\\\\prime\\\\prime}$. The formal description of the algorithm is given by Algorithm 2.\\n\\nFigure: Algorithm 1 Decision Calibration\\n\\nFigure: Algorithm 2 Reconcile Decision Calibration (ReDCal)\\n\\nWe provide the theoretical guarantees of our proposed algorithm below. At a high level, Algorithm 2 produces a pair of models with improved accuracy and approximately agrees on the best-response action almost everywhere. For the formal proofs of this section, see Appendix B.\\n\\n###### Theorem 3.2 .\\n\\nFor any pair of models $f_{1},f_{2}:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$, any distribution $\\\\mathcal{D}$, family of loss functions $\\\\mathcal{L}$, any loss margin $\\\\alpha>0$, disagreement region mass $\\\\eta>0$, and decision-calibration tolerance $\\\\beta>0$, Algorithm 2 updates $f_{1}$ and $f_{2}$ for $T_{1}$ and $T_{2}$ time-steps, respectively,\\nand outputs a pair of models $(f_{1}^{T},f_{2}^{T})$, such that:\\n\\n- 1.\\nAlgorithm 2 terminates within $T=T_{1}+T_{2}\\\\leq\\\\frac{4\\\\cdot d\\\\cdot\\\\left(B(f_{1},\\\\mathcal{D})+B(f_{2},\\n\\\\mathcal{D})\\\\right)}{\\\\alpha^{2}\\\\eta}$ time-steps.\\n- 2.\\nThe Brier scores of the final models are lower than that of the input models $(f_{1},f_{2})$:\\n$B(f_{1}^{T},\\\\mathcal{D})\\\\leq B(f_{1},\\\\mathcal{D})-T_{1}\\\\cdot\\\\frac{\\\\alpha^{2}\\n\\\\eta}{4d}\\\\quad\\\\text{and }\\\\quad B(f_{2}^{T},\\\\mathcal{D})\\\\leq B(f_{2},\\\\mathcal{D\\n})-T_{2}\\\\cdot\\\\frac{\\\\alpha^{2}\\\\eta}{4d}$\\n- 3.\\nAll the downstream decision-making losses of the final models do not increase by much compared to that of the input models $(f_{1},f_{2})$: for each $i\\\\in\\\\{1,2\\\\}$ and for all $\\\\ell\\\\in\\\\mathcal{L}$,\\n$\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{T}(x)\\n)]-\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}(x))\\n]\\\\leq T_{i}\\\\beta\\\\sqrt{d}K$\\n- 4.\\nThe final models approximately agree on their best-response actions almost everywhere. That is, the disagreement region $E_{\\\\ell,a_{1},a_{2}}$ calculated using $f_{1}^{T},f_{2}^{T}$ has small mass. For all $\\\\ell\\\\in\\\\mathcal{L}$,\\n$\\\\mu(E_{\\\\ell,a_{1},a_{2}})<\\\\eta\\\\quad\\\\text{ for all }a_{1},a_{2}\\\\in\\\\mathcal{A}\\n\\\\quad\\\\text{s.t }a_{1}\\\\neq a_{2}$\\n\\n###### Remark 3.3 .\\n\\nNote that in the third result of Theorem 3.2, the increase in downstream decision-making loss at each time-step only depends on the decision-calibrate tolerance $\\\\beta$, dimension $d$, and number of actions $K$. Since the total number of time-steps in Theorem 3.2 does not depend on $\\\\beta$, we can set $\\\\beta=\\\\frac{\\\\alpha}{T\\\\sqrt{d}K}$ to ensure the loss of taking the best-response action does not degrade by more than $\\\\alpha$. Moreover, in our empirical experiments (Section 4), we observe that the loss of taking the best-response action only increases minimally.\\n\\n###### Theorem 3.2 .\\n\\n###### Remark 3.3 .\\n\\n### 3.2 Finite Sample Analysis\\n\\nIn Section 3.1, we have presented an algorithm, ReDCal, to reconcile two predictors assuming the decision-makers have direct access to the probability distribution $\\\\mathcal{D}$. However, in practice, the decision-makers will only have access to a dataset $D=\\\\{(x_{1},y_{1}),\\\\cdots,(x_{n},y_{n})\\\\}$ containing $n$ i.i.d samples drawn from $\\\\mathcal{D}$. In this section, we will instead run Algorithm 2 on the empirical distribution over $D$ and show that its guarantees can translate to the underlying distribution $\\\\mathcal{D}$ with high probability. To prevent data leakage, it is important to assume that the dataset $D$ is drawn independently of the predictors $f_{1}$ and $f_{2}$, i.e., the dataset contains freshly drawn data that was not used to train either of the predictors that we want to reconcile. For the formal proofs, see Appendix C.\\n\\nAt a high level, since the samples in $D$ are independently and identically distributed, we can apply Chernoff-Hoeffding inequality to show that, with high probability, the in-sample quantities are approximately equal to out-sample quantities. We summarize the results in the theorem below.\\n\\n###### Theorem 3.4 .\\n\\nFix any distribution $\\\\mathcal{D}$ and dataset $D\\\\sim\\\\mathcal{D}$ containing $n$ samples drawn i.i.d from $\\\\mathcal{D}$. For any pair of predictors $f_{1},f_{2}:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$, family of loss functions $\\\\mathcal{L}$, loss margin $\\\\alpha>0$, disagreement region mass $\\\\eta>0$, and decision-calibration tolerance $\\\\beta>0$, Algorithm 2 run over the empirical distribution $D$ updates $f_{1}$ and $f_{2}$ for $T_{1}$ and $T_{2}$ time-steps, respectively, and outputs a pair of predictors $(f_{1}^{T},f_{2}^{T})$ such that, with probability at least $1-\\\\delta$ over the randomness of $D\\\\sim\\\\mathcal{D}^{n}$,\\n\\n- 1.\\nThe total number of time-steps for Algorithm 2 and Algorithm 1 is\\n$T=T_{1}+T_{2}\\\\leq\\\\frac{2d}{\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}$\\n- 2.\\nFor $i\\\\in\\\\{1,2\\\\}$, the Brier scores of the final models are lower than that of the input models:\\n$\\\\displaystyle B(f_{i}^{T_{i}},\\\\mathcal{D})\\\\leq B(f_{i},\\\\mathcal{D})-(T_{i}/2)\\n\\\\cdot\\\\min\\\\left\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/(4d)\\\\right\\\\}$\\n- 3.\\nFor $i\\\\in\\\\{1,2\\\\}$ and for all $\\\\ell\\\\in\\\\mathcal{L}$, the downstream decision-making losses of the final models do not increase by much compared to that of the input models:\\n$\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}\\n(f_{i}^{T}(x))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}(x))]\\\\leq 2T_{i}\\\\beta\\\\sqrt\\n{d}K$\\n- 4.\\nThe final models approximately agree on their best-response actions almost everywhere. That is, the disagreement region $E_{\\\\ell,a_{1},a_{2}}$ calculated using $f_{1}^{T},f_{2}^{T}$ has small mass: $\\\\forall\\\\ell\\\\in\\\\mathcal{L}$,\\n$\\\\mu(E_{\\\\ell,a_{1},a_{2}})\\\\leq 2\\\\eta\\\\quad\\\\text{ for all }a_{1},a_{2}\\\\in\\\\mathcal\\n{A}\\\\quad\\\\text{s.t }a_{1}\\\\neq a_{2}$\\n\\nif $n\\\\geq\\\\Omega\\\\left(d/(\\\\eta^{2}\\\\min\\\\{\\\\beta,\\\\eta\\\\alpha^{2}/d\\\\})\\\\cdot\\\\left(\\\\ln(K)+\\n\\\\ln(\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert)+d\\\\ln\\\\left(d/\\\\min\\\\{\\\\beta,\\\\eta\\\\alpha^{2}\\n/d\\\\}\\\\right)+\\\\ln(1/\\\\delta)\\\\right)\\\\right).$\\n\\n###### Theorem 3.4 .\\n\\n## 4 Experiments\\n\\nIn this section, we complement our theoretical results with a set of experiments on real-world datasets to show our improvement in decreasing decision-making loss compared to prior work in calibration.\\n\\n### 4.1 Imagenet Multi-class Classification\\n\\n#### Experiment Setup.\\n\\nWe use the ImageNet dataset (Deng et al., ) and two pre-trained models provided by pyTorch (inception-v3 (Szegedy et al., ) and resnet50 (He et al., )). Among the $50000$ validation samples, we use $40000$ samples for calibration and $10000$ samples for testing.\\n\\nWe investigate how the downstream decision loss changes with the four calibration algorithms: Reconcile (Algorithm 3), Decision-Calibration (Algorithm 1), ReDCal (Algorithm 2), and the combination of running ReDCal after Decision Calibration as post-process.\\nWe run each calibration algorithm $500$ times. For each run, we\\nfirst randomly draw $100$ classes from the $1000$ classes of ImageNet. Then, we randomly generate a loss function such that, for each $y\\\\in\\\\mathcal{Y}$, $a\\\\in\\\\mathcal{A}$, $\\\\ell(y,a)\\\\sim\\\\mathsf{Normal}(0,1)$. For each randomly generated loss function $\\\\ell$, we compare the expected losses derived from the best-response policies based on predictors $f_{1}$ and $f_{2}$ against those based on the optimal predictor $f^{*}$. Formally, the loss gap at timestep $t$ using predictor $f_{i}$ is defined as\\n\\n$$ $\\\\text{LossGap}(f_{i}^{t})=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^ {\\\\mathsf{BR}}(f_{i}^{t}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f^{*}(x)))].$ $$\\n\\nThe hyperparameters are chosen as follows: loss margin $\\\\alpha=0.001$, disagreement region mass $\\\\eta=0.01$, decision-calibration tolerance $\\\\beta=0.00001$, and the number of actions $K=10$.\\n\\n#### Results.\\n\\nWe compare the performance of ReDCal with the two baseline algorithms in terms of Brier scores and decision loss reduction. Furthermore, in Figure 4, we provide a comparison of the calibration algorithms’ performance when the number of dimensions increases.\\n\\n#### Brier score.\\n\\nIn Figure 2, we compare the Brier score of ReDCal (Algorithm 2) with the two baseline algorithms on both the calibration and the test datasets. Compared to Reconcile, our algorithm decreases the Brier score by a smaller amount on the test dataset. The combined algorithm of Decision-Calibration with ReDCal as post-process achieves the most substantial decrease in the Brier score.\\n\\nFigure: Figure 2: ReDCal decreases Brier score on Imagenet. Compared to Reconcile, our algorithm decreases the Brier score by a smaller amount on the test dataset. Decision-Calibration with ReDCal as post-process achieves the most substantial decrease in the Brier score.\\nRefer to caption: extracted/5630318/plots/ImageNet/withgap_brier_100classes_10actions_500.png\\n\\n#### Decision loss on calibration dataset.\\n\\nIn Figure 3, we compare the decision gap of our proposed algorithm with the two baseline algorithms on the training dataset.\\nCompared to Reconcile, ReDCal converges within a similar number of time-step and decreases the loss by a larger amount on the test dataset. Moreover, ReDCal further decrease the loss when used as a post-process after Decision-Calibration terminates.\\n\\nFigure: (a) ReDCal decreases the decision loss on the validation partition of the ImageNet dataset.\\nRefer to caption: extracted/5630318/plots/ImageNet/withgap_loss_100classes_10actions_500_train.png\\n\\n#### Decision loss comparison for high-dimensional classification problem.\\n\\nIn Figure 4, we compare the decision loss gap of our proposed algorithm with the two baseline algorithms on the testing dataset, using $d=10,100$, and $1000$ classes. We plot the average loss gap of the two predictors.\\nThe hyperparameters are: disagreement margin $\\\\alpha=0.1/d$, decision-calibration tolerance $\\\\beta=0.001/d$, disagreement region mass $\\\\eta=0.01$, number of actions $K=10$.\\n\\nFigure: Figure 4: ReDCal decreases decision loss on Imagenet. The takeaway results are similar to Figure 3(b). As the number of classes in the multi-class classification problem grows from $10$ to $1000$, ReDCal still outperforms Reconcile in decreasing decision loss on the test dataset. When we have $1000$ classes, ReDCal converges slower than Reconcile. Furthermore, ReDCal can further decrease the decision loss when it is used as a post-process after Decision Calibration terminates.\\nRefer to caption: extracted/5630318/plots/ImageNet/compare_loss_10actions_500_test.png\\n\\n### 4.2 HAM10000 Multi-class Classification\\n\\n#### Experiment Setup.\\n\\nWe use the HAM10000 dataset (Tschandl et al., ) (licensed CC BY-NC $4.0$) on pigmented skin lesions to predict the probability that a patient has contracted one of $7$ possible skin diseases: ’akiec’, ’bcc’, ’bkl’, ’df’, ’nv’, ’vasc’, and ’mel’. We split the dataset into train/validation/test sets, with $20\\\\%$ of the data are used for validation and $20\\\\%$ are used for testing. We use the train set to train two neural networks using pyTorch with resnet50 (He et al., ) and densenet121 (Huang et al., ) architectures and learn two models with around $88\\\\%$ top-$1$ accuracy. From each model, we output the individual probability prediction for each of the $7$ possible labels. We use the validation set to calibrate the predictors using our proposed algorithm and the two baseline algorithms, and the test set to measure the final performance.\\n\\nWe run each calibration algorithm $10$ times. At each run, we draw a fresh loss function created based on the loss function motivated by medical domain knowledge in Zhao et al. () and additional random noise drawn from $\\\\mathsf{Normal}(0,1)$. There are two possible actions for the decision-maker: treatment ($a=[1,0]$) or no treatment ($a=[0,1]$). Given a loss function $\\\\ell$ and a predictor $f$, the decision-maker will choose an action that minimizes their loss.\\n\\nFor each calibration algorithm, we calculate (1) the Brier score of the updated predictors and (2) the differences between the optimal loss had we known $y$ and the actual loss from taking the best-response actions induced by each predictor.\\n\\nThe hyperparameters for Algorithm 2 are chosen as follows: loss margin $\\\\alpha=0.1$, target disagreement region mass $\\\\eta=0.01$, and decision-calibration tolerance $\\\\beta=0.000001$.\\n\\n#### Results.\\n\\nSimilar to the experiment on the ImageNet dataset, we observe lower decision loss after running ReDCal compared to the baseline algorithm Reconcile. Moreover, we compare the performance of running Decision-Calibration on its own and using ReDCal as a post-process.\\n\\n#### Brier score.\\n\\nIn Figure 5, we compare the Brier score of ReDCal (Algorithm 2) with the two baseline algorithms on both the calibration and the test datasets. Compared to Reconcile, our proposed algorithm decreases the Brier score by a smaller amount.\\n\\nFigure: Figure 5: Brier score of the updated predictors using Algorithm 2 (orange) and two benchmark algorithms: Algorithm 3 (dashed-blue) and Algorithm 1 (dashed-green). Our algorithm reduces the Brier score by a smaller amount compared to Algorithm 2. Results are averaged over $10$ runs and the shaded region indicates $\\\\pm 1$ standard error.\\nRefer to caption: extracted/5630318/plots/HAM/HAM_brier.png\\n\\n#### Decision loss.\\n\\nIn Figure 6, we compare the decision loss gap of our proposed algorithm with the two baseline calibration algorithms. Compared to Reconcile, our algorithm decreases the decision loss by a larger amount on the test dataset. Furthermore, while Decision-Calibration already decreases the decision loss, our algorithm can further improve upon their result when it is used as a post-process after Decision-Calibration terminates.\\n\\nFigure: (a) ReDCal decreases the decision loss on the validation partition of HAM10000 dataset.\\nRefer to caption: extracted/5630318/plots/HAM/HAM_decision_loss_train.png\\n\\n## 5 Conclusion\\n\\nPredictive multiplicity is a phenomenon in machine learning where the decision-makers have two predictors with nearly equivalent accuracy but vastly different individual predictions. We propose an algorithm, ReDCal, that updates the pair of predictors using either true distribution or an i.i.d validation set until they approximately agree almost everywhere on (1) individual predictions, (2) best-response actions in the downstream decision-making task, and (3) following the best-response actions incur losses that are close to the optimal loss. This result helps alleviate the problem of predictive multiplicity in model selection. Finally, we provide experiments using real-world datasets to show that our proposed algorithm achieves lower decision loss compared to existing work. While we do not provide examples of domain-specific loss functions as part of our analysis and experiments, we hope that our findings can aid future studies on the impact of model multiplicity in decision-making.\\n\\n## Appendix A Limitation of Prior Work (Continue)\\n\\nWe provide the Algorithm 3 from [Roth et al., ] and their theoretical guarantees for completion. First, given two predictors $f_{1}$ and $f_{2}$, define the disagreement region as:\\n\\n$$ $\\\\displaystyle U_{\\\\epsilon}(f_{1},f_{2})\\\\coloneqq\\\\{x:\\\\left\\\\lvert f_{1}(x)-f_{2} (x)\\\\right\\\\rvert>\\\\epsilon\\\\}$ $$\\n\\nwhich can be further divided into two partitions:\\n\\n$$ $\\\\displaystyle U_{\\\\epsilon}^{>}(f_{1},f_{2})$ $\\\\displaystyle=\\\\{x\\\\in U_{\\\\epsilon}(f_{1},f_{2}):f_{1}(x)>f_{2}(x)\\\\}$ $\\\\displaystyle U_{\\\\epsilon}^{<}(f_{1},f_{2})$ $\\\\displaystyle=\\\\{x\\\\in U_{\\\\epsilon}(f_{1},f_{2}):f_{1}(x)<f_{2}(x)\\\\}$ $$\\n\\nFigure: Algorithm 3 Reconcile [Roth et al., ]\\n\\n###### Theorem A.1 (Reconcile [Roth et al., 2023 ] ) .\\n\\nFor any pair of models $f_{1},f_{2}:\\\\mathcal{X}\\\\rightarrow[0,1]$, any distribution $\\\\mathcal{D}$, and any $\\\\alpha,\\\\eta>0$, Algorithm 3 runs for $T=T_{1}+T_{2}$ many rounds and outputs a pair of models $(f_{1}^{T_{1}},f_{2}^{T_{2}})$ such that:\\n\\n- 1.\\n$T\\\\leq(B(f_{1},\\\\mathcal{D})+B(f_{2},\\\\mathcal{D}))\\\\cdot\\\\frac{16}{\\\\eta\\\\alpha^{2}}$\\n- 2.\\n$B(f_{1}^{T_{1}},\\\\mathcal{D})\\\\leq B(f_{1},\\\\mathcal{D})-T_{1}\\\\cdot\\\\frac{\\\\eta\\n\\\\alpha^{2}}{16}$ and $B(f_{2}^{T_{2}},\\\\mathcal{D})\\\\leq B(f_{2},\\\\mathcal{D})-T_{2}\\\\cdot\\\\frac{\\\\eta\\n\\\\alpha^{2}}{16}$\\n- 3.\\n$\\\\mu(U_{\\\\epsilon}(f_{1}^{T_{1}},f_{2}^{t_{2}}))\\\\leq\\\\eta$\\n\\n###### Theorem A.1 (Reconcile [Roth et al., 2023 ] ) .\\n\\n## Appendix B Proofs of Section 3.1 : Reconcile for Decision Making\\n\\nFirst, we show that if a disagreement event has a large probability mass, then at least one of $f_{1}$, $f_{2}$ has a large prediction error within the region:\\n\\n###### Lemma B.1 .\\n\\nFix any two predictors $f_{1},f_{2}:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$ and $\\\\alpha,\\\\eta>0$. If $\\\\mu(E_{\\\\ell,a_{1},a_{2}})>\\\\eta$ for some $a_{1},a_{2}\\\\in\\\\mathcal{A}$, then we have\\n\\n$$ $\\\\displaystyle\\\\|\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[f_{i}(x)-f^{*}(x)\\\\big{|}E_{ \\\\ell,a_{1},a_{2}}(x)\\\\right]\\\\|\\\\geq\\\\frac{\\\\alpha}{2\\\\sqrt{d}}$ (8) $$\\n\\nfor some $i\\\\in\\\\{1,2\\\\}$.\\n\\n###### Proof.\\n\\nBy definition of event $E_{\\\\ell,a_{1},a_{2}}$, we have\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[\\\\ell(f_{1}(x),a_{2})-\\\\ell(f_{1 }(x),a_{1})+\\\\ell(f_{2}(x),a_{1})-\\\\ell(f_{2}(x),a_{2})]|E_{\\\\ell,a_{1},a_{2}}(x) \\\\right]\\\\geq\\\\alpha$ $$\\n\\nAlso, we have\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[\\\\ell(f_{1}(x),a_{2})-\\\\ell(f_{1 }(x),a_{1})+\\\\ell(f_{2}(x),a_{1})-\\\\ell(f_{2}(x),a_{2})]|E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]$ $\\\\displaystyle=$ $\\\\displaystyle\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[\\\\langle f_{1}(x)-f_{2}(x),\\\\ell _{a_{2}}-\\\\ell_{a_{1}}\\\\rangle|E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]$ $\\\\displaystyle=$ $\\\\displaystyle\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[\\\\langle f_{1}(x)-f^{*}(x),\\\\ell _{a_{2}}-\\\\ell_{a_{1}}\\\\rangle|E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]$ $\\\\displaystyle\\\\quad+\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[\\\\langle f^{*}(x)-f_{2}(x ),\\\\ell_{a_{2}}-\\\\ell_{a_{1}}\\\\rangle|E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]$ $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\|\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[f_{1}(x)-f^{*}(x)\\\\big{|}E_{ \\\\ell,a_{1},a_{2}}(x)\\\\right]\\\\|_{2}\\\\sqrt{d}+\\\\|\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left [f_{2}(x)-f^{*}(x)\\\\big{|}E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]\\\\|_{2}\\\\sqrt{d}$ $\\\\displaystyle=$ $\\\\displaystyle\\\\sqrt{d}\\\\cdot\\\\left(\\\\|\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[f_{1}(x)- f^{*}(x)\\\\big{|}E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]\\\\|_{2}+\\\\|\\\\mathbb{E}_{x\\\\sim \\\\mathcal{X}}\\\\left[f_{2}(x)-f^{*}(x)\\\\big{|}E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]\\\\|_{2}\\\\right)$ $$\\n\\nwhere the last inequality comes from Cauchy-Schwartz and that $\\\\ell$ is bounded in $[0,1]$.\\n\\nCombining the above inequalities, we have\\n\\n$$ $\\\\displaystyle\\\\sqrt{d}\\\\cdot\\\\left(\\\\left\\\\lVert\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[ f_{1}(x)-f^{*}(x)\\\\big{|}E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]\\\\right\\\\rVert_{2}+\\\\left \\\\lVert\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[f_{2}(x)-f^{*}(x)\\\\big{|}E_{\\\\ell,a_{1} ,a_{2}}(x)\\\\right]\\\\right\\\\rVert_{2}\\\\right)\\\\geq\\\\alpha.$ $$\\n\\nTherefore, for some $i\\\\in\\\\{1,2\\\\}$, we have\\n\\n$$ $\\\\displaystyle\\\\left\\\\lVert\\\\mathbb{E}_{x\\\\sim\\\\mathcal{X}}\\\\left[f_{i}(x)-f^{*}(x) \\\\big{|}E_{\\\\ell,a_{1},a_{2}}(x)\\\\right]\\\\right\\\\rVert_{2}\\\\geq\\\\frac{\\\\alpha}{2\\\\sqrt{ d}}.$ $$\\n\\n∎\\n\\nThis lemma indicates that, if we have two predictors $f_{1},f_{2}$ that create a large disagreement event, we can falsify at least one of the models. We now show that these events also provide a directly actionable way to improve one of the models.\\n\\n###### Lemma B.2 .\\n\\nFor any predictor $f:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$, any event $E\\\\in\\\\mathcal{E}$, and distribution $\\\\mathcal{D}$. Let $\\\\phi=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[y-f(x)|E(x)=1]$. We patch $f$ as\\n\\n$$ $f^{\\\\prime}(x)=\\\\mathrm{proj}_{[0,1]^{d}}(f(x)+\\\\phi E(x)),~{}~{}\\\\text{where}~{}~ {}\\\\mathrm{proj}_{[0,1]^{d}}(y)=\\\\mathrm{argmin}_{y^{\\\\prime}\\\\in[0,1]^{d}}\\\\|y-y^{ \\\\prime}\\\\|_{2}.$ $$\\n\\nThen,\\n\\n$$ $B(f,\\\\mathcal{D})-B(f^{\\\\prime},\\\\mathcal{D})\\\\geq\\\\|\\\\phi\\\\|^{2}_{2}\\\\mu(E).$ $$\\n\\n###### Proof.\\n\\n$$ $\\\\displaystyle B(f,\\\\mathcal{D})-B(f^{\\\\prime},\\\\mathcal{D})=$ $\\\\displaystyle\\\\mathbb{E}\\\\left[\\\\|f(x)-y\\\\|_{2}^{2}-\\\\|f^{\\\\prime}(x)-y\\\\|_{2}^{2}\\\\right]$ $\\\\displaystyle\\\\geq$ $\\\\displaystyle\\\\mathbb{E}\\\\left[\\\\|f(x)-y\\\\|_{2}^{2}-\\\\|f(x)+\\\\phi E(x)-y\\\\|_{2}^{2}\\\\right]$ (since projection is non-expansive) $\\\\displaystyle=$ $\\\\displaystyle\\\\mathbb{E}\\\\left[2\\\\langle y-f(x),\\\\phi E(x)\\\\rangle-\\\\|\\\\phi E(x)\\\\|_{2 }^{2}\\\\right]$ $\\\\displaystyle\\\\geq$ $\\\\displaystyle\\\\|\\\\phi\\\\|_{2}^{2}\\\\cdot\\\\mu(E)$ $$\\n\\n∎\\n\\nTherefore, whenever we have two predictors that have a large disagreement event, we can always falsify at least one of the predictors and improve it through patching, causing the Brier score to decrease by a large amount. Similarly, for a fixed predictor, if one of its best-response events has a large calibration error, we can patch the predictor within the event to decrease the Brier score. As the Brier score is bounded in $[0,d]$, these two observations imply that the number of time-steps for both Algorithm 2 and its subroutine Algorithm 1 are bounded.\\n\\nOther than the Brier score, we also care about minimizing the loss of the downstream decision-making task. We now show that, after a further update through the subroutine Algorithm 1, the loss does not increase much at each time-step of Algorithm 2:\\n\\n###### Lemma B.3 .\\n\\nFor any predictors $f_{1},f_{2}$, loss function $\\\\ell\\\\in\\\\mathcal{L}$ and any distribution $\\\\mathcal{D}$, at any time-step $t$ of Algorithm 2, the predictors satisfies\\n\\n$$ $\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t+1}( x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t}(x)))]\\\\leq\\\\beta\\\\sqrt{d}K,$ $$\\n\\nfor all $i\\\\in\\\\{1,2\\\\}$.\\n\\n###### Proof.\\n\\nAt each round, we define the set $\\\\Delta_{a}^{t}\\\\subseteq E^{t}$ as\\n\\n$$ $\\\\displaystyle\\\\Delta_{a}^{t}=\\\\{x\\\\in E^{t}:\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f^{t+1}_{i}( x))=a\\\\}.$ $$\\n\\nThen, we have\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{t+1}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t}(x)))]$ $\\\\displaystyle=$ $\\\\displaystyle\\\\sum_{a\\\\in\\\\mathcal{A}}\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(\\\\ell(y,a )-\\\\ell(y,a_{i}^{t}))\\\\Delta_{a}^{t}(x)].$ $$\\n\\nFor each term in the summation, we can upper-bound it as\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(\\\\ell(y,a)-\\\\ell(y,a_{i}^{t})) \\\\Delta_{a}^{t}(x)]$ (9) $\\\\displaystyle=$ $\\\\displaystyle\\\\langle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[y\\\\Delta_{a}^{t}(x)],\\\\ell _{a}-\\\\ell_{a_{i}^{t}}\\\\rangle$ (Linearity of Expectation) $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\langle\\\\mathbb{E}_{x\\\\sim\\\\mathcal{D}_{\\\\mathcal{X}}}[f_{i}^{t+1}(x) \\\\Delta_{a}^{t}(x)],\\\\ell_{a}-\\\\ell_{a_{i}^{t}}\\\\rangle+\\\\beta\\\\sqrt{d}$ (Since $f_{i}^{t+1}$ is $\\\\beta$ -calibrated) $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\beta\\\\sqrt{d}.$ (Since $a$ is the new Best-response action) $$\\n\\nSumming these actions together, we have\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{t+1}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t}(x)))]\\\\leq\\\\beta \\\\sqrt{d}K.$ (10) $$\\n\\n∎\\n\\nInstead of setting a fixed $\\\\beta$, we can calculate a different $\\\\beta^{t}$ at each round, which allows a smaller increase in loss.\\n\\n###### Lemma B.4 .\\n\\nAt each round $t$, if $a_{i}^{t}$ is not the best action on $E^{t}$ in average, i.e.\\n\\n$$ $\\\\delta^{t}=\\\\max_{a\\\\in\\\\mathcal{A}}\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(\\\\ell(y,a_{ i}^{t})-\\\\ell(y,a))E^{t}(x)]>0,$ $$\\n\\nthen we can set $\\\\beta^{t}\\\\leq\\\\delta^{t}/\\\\sqrt{d}$, such that\\n\\n$$ $\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t+1}( x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t}(x)))]\\\\leq 0.$ $$\\n\\n###### Proof.\\n\\nWe can write the change in loss at each round as\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{t+1}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t}(x)))]$ $\\\\displaystyle=$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{t+1}(x)))-\\\\ell(y,a_{i}^{t}))E^{t}(x)]$ $\\\\displaystyle=$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{t+1}(x)))-\\\\ell(y,a^{\\\\prime}))E^{t}(x)]+\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D }}[\\\\ell(y,a^{\\\\prime})-\\\\ell(y,a_{i}^{t}))E^{t}(x)]$ $\\\\displaystyle=$ $\\\\displaystyle\\\\sum_{a\\\\in\\\\mathcal{A}}\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(\\\\ell(y,a )-\\\\ell(y,a^{\\\\prime}))\\\\Delta_{a}^{t}(x)]+\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell (y,a^{\\\\prime})-\\\\ell(y,a_{i}^{t}))E^{t}(x)],$ $$\\n\\nfor any $a^{\\\\prime}\\\\in\\\\mathcal{A}$.\\n\\nWe can use the same analysis as in Lemma B.3 to get\\n\\n$$ $\\\\sum_{a\\\\in\\\\mathcal{A}}\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(\\\\ell(y,a)-\\\\ell(y,a^{ \\\\prime}))\\\\Delta_{a}^{t}(x)]\\\\leq\\\\beta^{t}\\\\sqrt{d}.$ $$\\n\\nFor the second term, we would want the loss to be as small as possible, so we can choose $\\\\displaystyle a^{\\\\prime}=\\\\mathrm{argmin}_{a\\\\in\\\\mathcal{A}}\\\\mathbb{E}_{(x,y)\\n\\\\sim\\\\mathcal{D}}[\\\\ell(y,a)\\\\cdot E^{t}(x)]$ and let\\n\\n$$ $\\\\displaystyle\\\\delta^{t}=-\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(\\\\ell(y,a^{\\\\prime}) -\\\\ell(y,a_{i}^{t}))E^{t}(x)],$ (11) $$\\n\\nthen $\\\\delta^{t}$ is maximized and $\\\\delta^{t}\\\\geq 0$ by definition.\\n\\nThe total change in loss in this round can be written as\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{t+1}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t}(x)))]\\\\leq\\\\beta^{t} \\\\sqrt{d}-\\\\delta^{t}.$ $$\\n\\nIf $\\\\delta^{t}>0$, we can set $\\\\beta^{t}\\\\leq\\\\delta^{t}/\\\\sqrt{d}$ to ensure the loss does not increase at this round.\\n\\n∎\\n\\n###### Lemma B.1 .\\n\\n###### Proof.\\n\\n###### Lemma B.2 .\\n\\n###### Proof.\\n\\n###### Lemma B.3 .\\n\\n###### Proof.\\n\\n###### Lemma B.4 .\\n\\n###### Proof.\\n\\n### B.1 Proof of Theorem 3.2\\n\\n###### Proof.\\n\\nBy Lemma B.1 and B.2, for any $i\\\\in\\\\{1,2\\\\}$, at time-step $t$, we have the inequality\\n\\n$$ $\\\\displaystyle B(f_{i}^{t},\\\\mathcal{D})-B(f_{i}^{t+1},\\\\mathcal{D})\\\\geq\\\\frac{ \\\\alpha^{2}\\\\eta}{4d}.$ $$\\n\\nTaking the sum over all time-steps, we have for any $i\\\\in\\\\{1,2\\\\}$,\\n\\n$$ $\\\\displaystyle B(f_{i},\\\\mathcal{D})-B(f_{i}^{T},\\\\mathcal{D})\\\\geq T_{i}\\\\cdot \\\\frac{\\\\alpha^{2}\\\\eta}{4d}.$ $$\\n\\nSince the Brier score is always non-negative, we have\\n\\n$$ $\\\\displaystyle T_{i}\\\\leq\\\\frac{4d\\\\cdot B(f_{i},\\\\mathcal{D})}{\\\\alpha^{2}\\\\eta}.$ $$\\n\\nSecond, using Lemma B.3 and summing over all time-steps, we have\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{T}(x))]-\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{ BR}}(f_{i}(x))]$ (12) $\\\\displaystyle=$ $\\\\displaystyle\\\\sum_{t=1}^{T}\\\\mathbb{I}[i_{t}=i]\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D} }[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t+1}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{ BR}}(f_{i}^{t}(x)))]$ (13) $\\\\displaystyle\\\\leq$ $\\\\displaystyle T_{i}\\\\cdot\\\\beta\\\\sqrt{d}K.$ (14) $$\\n\\nFinally, the halting condition implies that $\\\\mu(E_{\\\\ell,a_{1},a_{2}})<\\\\eta$ for all $a_{1},a_{2}\\\\in\\\\mathcal{A}$.\\n∎\\n\\n###### Proof.\\n\\n## Appendix C Proofs of Section 3.2 : Finite Sample Analysis\\n\\nFirst, to make our argument that in-sample quantities translate to out-sample quantities, it is useful for the patching operations to use values that are rounded to a finite grid, rather than the precise value from the arbitrary sample. We define the finite grid as follows:\\n\\n###### Definition C.1 .\\n\\nFor any integer $m>0$, let $1/m$ denote the $m+1$ grid points,\\n\\n$$ $\\\\left[\\\\frac{1}{m}\\\\right]=\\\\left\\\\{0,\\\\frac{1}{m},\\\\frac{2}{m},\\\\ldots,\\\\frac{m-1}{m} ,1\\\\right\\\\}.$ $$\\n\\nFor any value $v\\\\in[0,1]^{d}$, let $Round(v;m)=\\\\mathrm{argmin}_{v^{\\\\prime}\\\\in[1/m]^{d}}\\\\left\\\\lVert v-v^{\\\\prime}\\n\\\\right\\\\rVert_{2}$ denote the closest grid point to $v$ in $[1/m]^{d}$.\\n\\nAt each time-step in Algorithm 1 and Algorithm 2, denote $\\\\widetilde{\\\\phi}^{t}=Round(\\\\phi;m)$, and we patch the predictors using $\\\\widetilde{\\\\phi}^{t}$ instead of $\\\\phi^{t}$, i.e., we update $f^{t}$ to $f^{t+1}$ as\\n\\n$$ $f^{t+1}(x)=\\\\mathrm{proj}_{[0,1]^{d}}(f^{t}(x)+\\\\widetilde{\\\\phi}^{t}E^{t}(f^{t}( x),x)).$ $$\\n\\nWith this new patching operation, we can perform a similar analysis in Section 3.1 to show that the Brier score decreases at each iteration, and therefore the algorithm terminates within a finite number of time-steps. We denote the maximum number of time-steps, counting both Algorithm 1 and Algorithm 2, as $T_{\\\\max}$.\\n\\nThen, we can count the total number of possible predictors outputted by\\nAlgorithm 2 by observing that, for a fixed pair of input predictors, each pair of output predictors can be encoded as a sequence of tuple, $\\\\{(i^{t},E^{t},\\\\Delta^{t})\\\\}_{t\\\\in[T]}$. Here, index $i^{t}\\\\in\\\\{1,2\\\\}$, event $E^{t}\\\\in\\\\{E_{\\\\ell,a_{1},a_{2}}:\\\\ell\\\\in\\\\mathcal{L},a_{1},a_{2}\\\\in\\\\mathcal{A}\\\\}\\n\\\\cup\\\\{E_{\\\\ell,a}\\\\cap E_{\\\\ell^{\\\\prime},a_{1},a_{2}}:\\\\ell,\\\\ell^{\\\\prime}\\\\in\\n\\\\mathcal{L},a_{1},a_{2},a\\\\in\\\\mathcal{A}\\\\}$, and $\\\\Delta_{t}\\\\in[1/m]^{d}$ are all chosen from a finite set, and the length of the sequence, $T$, is also bounded. Specifically, for a fixed input $f_{1},f_{2}$, we denote $S$ to be the set of all possible predictors outputted by Algorithm 2. Then, its size satisfies\\n\\n$$ $\\\\displaystyle|S|\\\\leq(4|\\\\mathcal{L}|^{2}K^{3}(m+1)^{d})^{T_{\\\\max}+1}$ $$\\n\\nWe show that the number of predictors outputted by Algorithm 2 is bounded:\\n\\n###### Lemma C.2 .\\n\\nFix any pair of predictors $f_{1},f_{2}:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$ and any $\\\\eta,\\\\alpha,\\\\beta>0$. Then the total number of possible predictors outputted by Algorithm 2 is at most $\\\\left\\\\lvert S\\\\right\\\\rvert$ such that, for any distribution $\\\\mathcal{D}$ on which Algorithm 2 is run, the output predictors $(f_{1}^{t},f_{2}^{t})\\\\in S$.\\n\\n###### Proof.\\n\\nFirst, notice that a sequence of quantities $\\\\{(i^{t},E^{t},\\\\Delta^{t})\\\\}_{t\\\\in[T]}$ defines the pair of predictors outputted by Algorithm 2.\\n\\nLet $S$ denote the pairs of functions induced by all such trajectories defined above. Here, $i^{t}\\\\in\\\\{1,2\\\\},E^{t}\\\\in\\\\{E_{\\\\ell,a_{1},a_{2}}:\\\\ell\\\\in\\\\mathcal{L},a_{1},a_{2}\\n\\\\in\\\\mathcal{A}\\\\}\\\\cup\\\\{E_{\\\\ell,a}\\\\cap E_{\\\\ell^{\\\\prime},a_{1},a_{2}}:\\\\ell,\\\\ell^{\\n\\\\prime}\\\\in\\\\mathcal{L},a_{1},a_{2},a\\\\in\\\\mathcal{A}\\\\}$, and $\\\\Delta_{t}\\\\in[1/m]^{d}$. Therefore, there are\\n\\n$$ $\\\\displaystyle|S|\\\\leq\\\\sum_{t=1}^{T}\\\\left(2(|\\\\mathcal{L}|K^{2}+|\\\\mathcal{L}|^{2} K^{3})(m+1)^{d}\\\\right)^{t}\\\\leq(4|\\\\mathcal{L}|^{2}K^{3}(m+1)^{d})^{T_{\\\\max}+1}$ $$\\n\\noutput predictors.\\n∎\\n\\n###### Definition C.1 .\\n\\n###### Lemma C.2 .\\n\\n###### Proof.\\n\\n### C.1 Finite Grid\\n\\nWith this new patching operation, we can show that the Brier score decreases on the empirical distribution $D$, corresponding to Lemma B.2:\\n\\n###### Lemma C.3 .\\n\\nFix any event $E$. Let $\\\\phi=\\\\mathbb{E}_{(x,y)\\\\sim D}[y-f(x)|E(x)=1]$. For any predictor $f$, we patch $f$ as $f^{\\\\prime}(x)=\\\\mathrm{proj}_{\\\\Delta}(f(x)+\\\\tilde{\\\\phi}E(x))$. Then,\\n\\n$$ $B(f,D)-B(f^{\\\\prime},D)\\\\geq\\\\|\\\\phi\\\\|_{2}^{2}\\\\mu(E)-\\\\frac{d}{4m^{2}}$ $$\\n\\n###### Proof.\\n\\nLet $\\\\tilde{f}^{\\\\prime}(x)=f(x)+\\\\phi E(x)$. Then, we have\\n\\n$$ $\\\\displaystyle B(f,D)-B(f^{\\\\prime},D)=$ $\\\\displaystyle B(f,D)-B(\\\\tilde{f}^{\\\\prime},D)+B(\\\\tilde{f}^{\\\\prime},D)-B(f^{ \\\\prime},D)$ $\\\\displaystyle\\\\geq$ $\\\\displaystyle\\\\|\\\\phi\\\\|_{2}^{2}\\\\mu(E)+\\\\mathbb{E}[\\\\left\\\\lVert f(x)+\\\\phi E(x)-y \\\\right\\\\rVert_{2}^{2}-\\\\left\\\\lVert f(x)+\\\\tilde{\\\\phi}E(x)-y\\\\right\\\\rVert_{2}^{2}]$ ( Lemma B.2 ) $\\\\displaystyle=$ $\\\\displaystyle\\\\|\\\\phi\\\\|_{2}^{2}\\\\mu(E)-\\\\mathbb{E}\\\\left[\\\\left\\\\lVert\\\\tilde{\\\\phi}- \\\\phi\\\\right\\\\rVert_{2}^{2}\\\\right]\\\\mu(E)$ $$\\n\\nBy definition $\\\\tilde{\\\\phi}$, we know that each index of $|\\\\tilde{\\\\phi}-\\\\phi|$ is in $[0,\\\\frac{1}{2m}]$. Therefore, we have\\n\\n$$ $\\\\displaystyle B(f,D)-B(f^{\\\\prime},D)\\\\geq$ $\\\\displaystyle\\\\|\\\\phi\\\\|_{2}^{2}\\\\cdot\\\\mu(E)-\\\\frac{d}{4m^{2}}.$ $$\\n\\n∎\\n\\nSince the Brier score is within the range $[0,d]$, and it decreases at each iteration, we can show that, if we set $m$ large enough, Algorithm 2 terminates within a finite number of iterations:\\n\\n###### Lemma C.4 .\\n\\nFor any predictor $f_{1},f_{2}$. Let $m\\\\geq\\\\left\\\\lceil\\\\sqrt{\\\\frac{d}{2\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}}\\\\right\\\\rceil$. The Brier score at each iteration of Algorithm 1 and Algorithm 2 satisfies\\n\\n$$ $B(f^{t},D)-B(f^{t+1},D)>\\\\frac{\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/(4d)\\\\}}{2}.$ $$\\n\\nCounting both Algorithm 2 and its subroutine Algorithm 1, the total number of iterations $T$ satisfies\\n\\n$$ $\\\\displaystyle T\\\\leq\\\\frac{2d}{\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}.$ $$\\n\\n###### Proof.\\n\\nBy Algorithm 1, we have by definition of $\\\\beta$-decision calibration that\\n\\n$$ $\\\\displaystyle\\\\mu(E^{t}\\\\cap E_{\\\\ell,a})\\\\cdot\\\\left\\\\lVert\\\\mathbb{E}_{(x,y)\\\\sim D} [(y-f(x)|E^{t}(x)\\\\cdot E_{\\\\ell,a}(x)=1]\\\\right\\\\rVert_{2}^{2}$ (15) $\\\\displaystyle\\\\geq$ $\\\\displaystyle\\\\mu(E^{t}\\\\cap E_{\\\\ell,a})^{2}\\\\cdot\\\\left\\\\lVert\\\\mathbb{E}_{(x,y) \\\\sim D}[(y-f(x)|E^{t}(x)\\\\cdot E_{\\\\ell,a}(x)=1]\\\\right\\\\rVert_{2}^{2}$ (16) $\\\\displaystyle=$ $\\\\displaystyle\\\\left\\\\lVert\\\\mathbb{E}_{(x,y)\\\\sim D}[(y-f(x)\\\\cdot E^{t}(x)\\\\cdot E_ {\\\\ell,a}(x)]\\\\right\\\\rVert_{2}^{2}>\\\\beta^{2}.$ (17) $$\\n\\nIn Algorithm 2, we have by Lemma B.1 that\\n\\n$$ $\\\\displaystyle\\\\mu(E^{t})\\\\left\\\\lVert\\\\mathbb{E}_{(x,y)\\\\sim D}[(y-f(x)|E^{t}(x)=1] \\\\right\\\\rVert_{2}^{2}>\\\\frac{\\\\eta\\\\alpha^{2}}{4d}.$ (18) $$\\n\\nTherefore, for any $\\\\phi$ and event $E$ that we patch in Algorithm 1 or Algorithm 2, they satisfy\\n\\n$$ $\\\\|\\\\phi\\\\|_{2}^{2}\\\\cdot\\\\mu(E)>\\\\min\\\\{\\\\beta^{2},\\\\frac{\\\\eta\\\\alpha^{2}}{4d}\\\\}.$ $$\\n\\nLetting $m\\\\geq\\\\left\\\\lceil\\\\sqrt{\\\\frac{d}{2\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}}\\\\right\\\\rceil$, we can ensure\\n\\n$$ $\\\\displaystyle B(f^{t},D)-B(f^{t+1},D)\\\\geq$ $\\\\displaystyle\\\\frac{\\\\|\\\\phi\\\\|_{2}^{2}\\\\cdot\\\\mu(E)}{2}>\\\\frac{\\\\min\\\\{\\\\beta^{2},\\\\eta \\\\alpha^{2}/(4d)\\\\}}{2}.$ $$\\n\\nSince the Brier score is in the range $[0,d]$, we can bound the total number of iterations of both Algorithm 1 and Algorithm 2 as\\n\\n$$ $\\\\displaystyle T\\\\leq\\\\frac{2d}{\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}.$ $$\\n\\n∎\\n\\n###### Lemma C.3 .\\n\\n###### Proof.\\n\\n###### Lemma C.4 .\\n\\n###### Proof.\\n\\n### C.2 Proof of Theorem 3.4\\n\\nFirst, we show that, for a fixed predictor $f$ and event $E_{\\\\ell,a}$, the in-sample prediction error is approximately accurate. The deviation bound of the Brier score and calibration error can then be directly implied.\\n\\n###### Lemma C.5 .\\n\\nFix any $f$, $E_{\\\\ell,a}$, with probability at least $1-\\\\delta^{\\\\prime}$, we have\\n\\n$$ $\\\\displaystyle\\\\left\\\\lVert\\\\mathbb{E}_{\\\\mathcal{D}}[(y-f(x))E_{\\\\ell,a}(f(x),x)]- \\\\frac{1}{n}\\\\sum_{i=1}^{n}[(y_{i}-f(x_{i}))E_{\\\\ell,a}(f(x_{i}),x_{i})]\\\\right \\\\rVert_{2}\\\\leq\\\\sqrt{\\\\frac{d\\\\ln(2d/\\\\delta^{\\\\prime})}{2n}}.$ $$\\n\\n###### Proof.\\n\\nFix an index $j\\\\in[d]$,\\nwe know $\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(y-f(x))_{j}\\\\cdot E_{\\\\ell,a}(f(x),x)]\\\\in[0,1]$ and\\n\\n$$ $\\\\mathbb{E}_{D}\\\\left[\\\\frac{1}{n}\\\\sum_{i=1}^{n}[(y_{i}-f(x_{i}))_{j}\\\\cdot E_{ \\\\ell,a}(f(x_{i}),x_{i})]\\\\right]=\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(y-f(x))_{j} \\\\cdot E_{\\\\ell,a}(f(x),x)].$ $$\\n\\nSince $(x_{i},y_{i})$ is drawn i.i.d. from $\\\\mathcal{D}$, we can use Hoeffding’s inequality to get, with probability $\\\\delta^{\\\\prime}/d$,\\n\\n$$ $\\\\displaystyle\\\\left|\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(y-f(x))_{j}\\\\cdot E_{\\\\ell ,a}(f(x),x)]-\\\\frac{1}{n}\\\\sum_{i=1}^{n}[(y_{i}-f(x_{i}))_{j}\\\\cdot E_{\\\\ell,a}(f( x_{i}),x_{i})]\\\\right|\\\\leq\\\\sqrt{\\\\frac{\\\\ln(2d/\\\\delta^{\\\\prime})}{2n}}$ $$\\n\\nUsing union bound, we have that with probability $1-\\\\delta^{\\\\prime}$, the above inequality holds for all $j\\\\in[d]$. Then, we have\\n\\n$$ $\\\\displaystyle\\\\left\\\\lVert\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(y-f(x))\\\\cdot E_{ \\\\ell,a}(f(x),x)]-\\\\frac{1}{n}\\\\sum_{i=1}^{n}[(y_{i}-f(x_{i}))\\\\cdot E_{\\\\ell,a}(f( x_{i}),x_{i})]\\\\right\\\\rVert_{2}$ $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\sqrt{\\\\sum_{j\\\\in d}\\\\left(\\\\sqrt{\\\\frac{\\\\ln(2d/\\\\delta^{\\\\prime})}{2n} }\\\\right)^{2}}=\\\\sqrt{\\\\frac{d\\\\ln(2d/\\\\delta^{\\\\prime})}{2n}}.$ $$\\n\\n∎\\n\\nThe deviation bound of the Brier score and calibration error can be directly implied by Lemma C.5. We summarize them in the lemmas below:\\n\\n###### Lemma C.6 .\\n\\nFor a fixed $f$, with probability at least $1-\\\\delta^{\\\\prime}$, $|B(f,\\\\mathcal{D})-B(f,D)|\\\\leq\\\\sqrt{\\\\frac{d\\\\ln(2d/\\\\delta^{\\\\prime})}{2n}}$.\\n\\n###### Proof.\\n\\nUsing triangle inequality, we have\\n\\n$$ $\\\\displaystyle|B(f,\\\\mathcal{D})-B(f,D)|=$ $\\\\displaystyle\\\\left|\\\\left\\\\lVert\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[y-f(x)]\\\\right \\\\rVert_{2}-\\\\left\\\\lVert\\\\frac{1}{n}\\\\sum_{i=1}^{n}[y_{i}-f(x_{i})]\\\\right\\\\rVert_{2 }\\\\right|$ $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\left\\\\lVert\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[y-f(x)]-\\\\frac{1}{n} \\\\sum_{i=1}^{n}[y_{i}-f(x_{i})]\\\\right\\\\rVert_{2}$ $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\sqrt{\\\\frac{d\\\\ln(2d/\\\\delta^{\\\\prime})}{2n}}.$ $$\\n\\n∎\\n\\n###### Lemma C.7 .\\n\\nFor a fixed $f$, any loss function $\\\\ell\\\\in\\\\mathcal{L}$ and $\\\\mathcal{E}=\\\\{E_{\\\\ell,a}\\\\cap E_{\\\\ell^{\\\\prime},a_{1},a_{2}}:\\\\ell,\\\\ell^{\\\\prime}\\n\\\\in\\\\mathcal{L},a,a_{1},a_{2}\\\\in\\\\mathcal{A}\\\\}$, with probability $1-\\\\delta^{\\\\prime}$, we have\\n\\n$$ $\\\\displaystyle\\\\left\\\\lVert\\\\mathbb{E}_{\\\\mathcal{D}}[(y-f(x))E(x)]-\\\\frac{1}{n}\\\\sum _{i=1}^{n}[(y_{i}-f(x_{i}))E(x)]\\\\right\\\\rVert_{2}\\\\leq\\\\sqrt{\\\\frac{3d\\\\ln(2dK\\\\left \\\\lvert\\\\mathcal{L}\\\\right\\\\rvert/\\\\delta^{\\\\prime})}{2n}}$ $$\\n\\nfor all $E\\\\in\\\\mathcal{E}$.\\n\\n###### Proof.\\n\\nThe claim follows by using a union bound over the events in $\\\\mathcal{E}^{t}$, using Lemma C.5, and that $|\\\\mathcal{E}^{t}|=K^{3}\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert^{2}$.\\n∎\\n\\nFor a fixed pair of predictors, we can also show that the empirical size of the disagreement events $E_{\\\\ell,a_{1},a_{2}}$ is approximately correct with high probability:\\n\\n###### Lemma C.8 .\\n\\nFix any pair of predictors $(f_{1},f_{2})\\\\in S$, with probability at least $1-\\\\delta^{\\\\prime}$ over $D$, we have\\n\\n$$ $\\\\displaystyle\\\\left|\\\\mu(E_{\\\\ell,a_{1},a_{2}})-\\\\frac{1}{n}\\\\sum_{i=1}^{n}\\\\mathbb{ I}[E_{\\\\ell,a_{1},a_{2}}(x_{i})=1]\\\\right|\\\\leq\\\\sqrt{\\\\frac{2\\\\ln(2K\\\\left\\\\lvert \\\\mathcal{L}\\\\right\\\\rvert/\\\\delta^{\\\\prime})}{2n}}.$ $$\\n\\nfor all $a_{1},a_{2}\\\\in\\\\mathcal{A}$ with $a_{1}\\\\neq a_{2}$ and for all $\\\\ell\\\\in\\\\mathcal{L}$.\\n\\n###### Proof.\\n\\nWe know $\\\\mathbb{I}[E_{\\\\ell,a_{1},a_{2}}(x_{i})=1]\\\\in[0,1]$ and\\n\\n$$ $\\\\mathbb{E}_{D}\\\\left[\\\\frac{1}{n}\\\\sum_{i=1}^{n}\\\\mathbb{I}[E_{\\\\ell,a_{1},a_{2}}(x _{i})\\\\right]=\\\\mu(E_{\\\\ell,a_{1},a_{2}}).$ $$\\n\\nSince $(x_{i},y_{i})$ is drawn i.i.d. from $\\\\mathcal{D}$, we can use Hoeffding’s inequality to get, with probability $1-\\\\delta^{\\\\prime}/(K^{2}|\\\\mathcal{L}|)$,\\n\\n$$ $\\\\displaystyle\\\\left|\\\\mu(E_{\\\\ell,a_{1},a_{2}})-\\\\frac{1}{n}\\\\sum_{i=1}^{n}\\\\mathbb{ I}[E_{\\\\ell,a_{1},a_{2}}(x_{i})]\\\\right|\\\\leq\\\\sqrt{\\\\frac{2\\\\ln(2K\\\\left\\\\lvert \\\\mathcal{L}\\\\right\\\\rvert/\\\\delta^{\\\\prime})}{2n}}.$ $$\\n\\nUsing union bound over all pairs of $a_{1},a_{2}\\\\in\\\\mathcal{A}$ and $\\\\ell\\\\in\\\\mathcal{L}$, we know the above inequality holds for all $a_{1},a_{2}$ and $\\\\ell\\\\in\\\\mathcal{L}$ with probability at least $1-\\\\delta^{\\\\prime}$.\\n∎\\n\\nWe summarize the above results in the theorem below. Theorem 3.4 follows by solving for $n$ in the 2-4th guarantees below.\\n\\n###### Theorem C.9 .\\n\\nFix any distribution $\\\\mathcal{D}$ and dataset $D\\\\sim\\\\mathcal{D}$ containing $n$ samples drawn i.i.d from $\\\\mathcal{D}$. For any pair of predictors $f_{1},f_{2}:\\\\mathcal{X}\\\\rightarrow[0,1]^{d}$, loss margin $\\\\alpha>0$, disagreement region mass $\\\\eta>0$, and decision-calibration tolerance $\\\\beta>0$, Algorithm 2 run over the empirical distribution $D$ updates predictors $f_{1}$ and $f_{2}$ for $T_{1}$ and $T_{2}$ time-steps, respectively, and outputs a pair of predictors $(f_{1}^{T},f_{2}^{T})$ such that, with probability at least $1-\\\\delta$ over the randomness of $D\\\\sim\\\\mathcal{D}^{n}$,\\n\\n- 1.\\nThe total number of time-steps for Algorithm 2 and Algorithm 1 is\\n$T=T_{1}+T_{2}\\\\leq\\\\frac{2d}{\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}$\\n- 2.\\nFor $i\\\\in\\\\{1,2\\\\}$, the Brier scores of the final models are lower than that of the input models:\\n$\\\\displaystyle B(f_{i}^{T_{i}},\\\\mathcal{D})\\\\leq B(f_{i},\\\\mathcal{D})-T_{i}\\\\cdot\\n\\\\min\\\\left\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/(4d)\\\\right\\\\}+\\\\sqrt{(d\\\\ln(6d|S|/\\\\delta)/(2n)}$\\n- 3.\\nFor $i\\\\in\\\\{1,2\\\\}$, the downstream decision-making losses of the final models do not increase by much compared to that of the input models:\\n$\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi^{\\\\mathsf{BR}}(f_{i}^\\n{T}(x))-\\\\ell(y,\\\\pi^{\\\\mathsf{BR}}(f_{i}(x))]\\\\leq\\\\left(\\\\beta+\\\\sqrt{(3d\\\\ln(6dK|S|\\n\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert/\\\\delta)/(2n)}\\\\right)\\\\sqrt{d}KT_{i}$\\n- 4.\\nThe final models approximately agree on their best-response actions almost everywhere. That is, the disagreement region $E_{a_{1},a_{2}}$ calculated using $f_{1}^{T},f_{2}^{T}$ has small mass.\\n$\\\\mu(E_{\\\\ell,a_{1},a_{2}})\\\\leq\\\\eta+\\\\sqrt{(2\\\\ln(6K|S|\\\\left\\\\lvert\\\\mathcal{L}\\n\\\\right\\\\rvert/\\\\delta)/(2n)}\\\\quad\\\\text{ for all }a_{1},a_{2}\\\\in\\\\mathcal{A}\\\\quad\\n\\\\text{s.t }a_{1}\\\\neq a_{2}$\\n\\nHere, $S$ is the set of all possible predictors outputted by Algorithm 2 satisfying\\n\\n$$ $\\\\displaystyle\\\\ln(|S|)\\\\leq\\\\left(\\\\frac{2d}{\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}+ 1\\\\right)\\\\ln\\\\left(4|\\\\mathcal{L}|^{2}K^{3}\\\\left(\\\\left\\\\lceil\\\\sqrt{\\\\frac{d}{2\\\\min \\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}}\\\\right\\\\rceil+1\\\\right)^{d}\\\\right)$ $$\\n\\n###### Proof.\\n\\nThe upper bound on $T$ holds true with probability $1$. For the remaining three guarantees, we show that each of them holds with probability at least $1-\\\\delta/3$ over the randomness of $D$.\\n\\nBrier Score. First, by Lemma C.6 and using union bound over all possible output predictors $(f_{1},f_{2})\\\\in S$, we have with probability at least $1-\\\\delta/3$ that\\n\\n$$ $\\\\displaystyle|B(f_{i},\\\\mathcal{D})-B(f_{i},D)|\\\\leq\\\\sqrt{\\\\frac{d\\\\ln(6d|S|/ \\\\delta)}{2n}}.$ $$\\n\\nBy Lemma C.4, and summing over all iterations, we have\\n\\n$$ $\\\\displaystyle B(f_{i}^{T_{i}},D)\\\\leq B(f_{i},D)-T_{i}\\\\cdot\\\\min\\\\left\\\\{\\\\frac{ \\\\beta^{2}}{2},\\\\frac{\\\\eta\\\\alpha^{2}}{8d}.\\\\right\\\\}$ $$\\n\\nTherefore,\\n\\n$$ $\\\\displaystyle B(f_{i}^{T_{i}},\\\\mathcal{D})\\\\leq$ $\\\\displaystyle B(f_{i}^{T_{i}},D)+\\\\sqrt{\\\\frac{d\\\\ln(6d|S|/\\\\delta)}{2n}}$ $\\\\displaystyle\\\\leq$ $\\\\displaystyle B(f_{i},\\\\mathcal{D})-T_{i}\\\\cdot\\\\min\\\\left\\\\{\\\\frac{\\\\beta^{2}}{2}, \\\\frac{\\\\eta\\\\alpha^{2}}{8d}\\\\right\\\\}+\\\\sqrt{\\\\frac{d\\\\ln(6d|S|/\\\\delta)}{2n}}$ $$\\n\\nfor $i\\\\in\\\\{1,2\\\\}$.\\n\\nExpected Loss.\\nUsing union bound over all predictors in $S$, by Lemma C.7, we have, with probability at least $1-\\\\delta/3$,\\n\\n$$ $\\\\displaystyle\\\\left\\\\lVert\\\\mathbb{E}_{\\\\mathcal{D}}[(y-f(x))E_{\\\\ell,a}(f(x),x)]- \\\\frac{1}{n}\\\\sum_{i=1}^{n}[(y_{i}-f(x_{i}))E_{\\\\ell,a}(f(x_{i}),x_{i})]\\\\right \\\\rVert_{2}\\\\leq\\\\sqrt{\\\\frac{3d\\\\ln(6dK|S|\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert/ \\\\delta)}{2n}}$ $$\\n\\nfor all predictors $f$, action $a\\\\in\\\\mathcal{A}$ and loss $\\\\ell\\\\in\\\\mathcal{L}$.\\nUsing similar method as in Lemma 2.6, we define the set $\\\\Delta_{a}^{t}\\\\subseteq E^{t}$ as\\n\\n$$ $\\\\displaystyle\\\\Delta_{a}^{t}=\\\\{x\\\\in E^{t}:\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f^{t+1}_{i}( x))=a\\\\}.$ $$\\n\\nThen, we have\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{t+1}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t}(x)))]$ $\\\\displaystyle=$ $\\\\displaystyle\\\\sum_{a\\\\in\\\\mathcal{A}}\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(\\\\ell(y,a )-\\\\ell(y,a_{i}^{t}))\\\\Delta_{a}^{t}(x)].$ $$\\n\\nFor each term in the summation,\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[(\\\\ell(y,a)-\\\\ell(y,a_{i}^{t})) \\\\Delta_{a}^{t}(x)]$ (19) $\\\\displaystyle=$ $\\\\displaystyle\\\\langle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[y\\\\Delta_{a}^{t}(x)],\\\\ell _{a}-\\\\ell_{a_{i}^{t}}\\\\rangle$ (Linearity of Expectation) $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\langle\\\\mathbb{E}_{x\\\\sim\\\\mathcal{D}_{\\\\mathcal{X}}}[f_{i}^{t+1}(x) \\\\Delta_{a}^{t}(x)],\\\\ell_{a}-\\\\ell_{a_{i}^{t}}\\\\rangle+\\\\left(\\\\beta+\\\\sqrt{\\\\frac{3d \\\\ln(6dK|S|\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert/\\\\delta)}{2n}}\\\\right)\\\\sqrt{d}$ ( Lemma C.7 and $\\\\beta$ -calibrated) $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\left(\\\\beta+\\\\sqrt{\\\\frac{3d\\\\ln(6dK|S|\\\\left\\\\lvert\\\\mathcal{L}\\\\right \\\\rvert/\\\\delta)}{2n}}\\\\right)\\\\sqrt{d}.$ (Since $a$ is the new Best-response action) $$\\n\\nSumming these actions together, we have\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}} (f_{i}^{t+1}(x)))-\\\\ell(y,\\\\pi_{\\\\ell}^{\\\\mathsf{BR}}(f_{i}^{t}(x)))]\\\\leq\\\\beta \\\\sqrt{d}K+\\\\sqrt{\\\\frac{3\\\\ln(6dK|S|\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert/\\\\delta)}{2 n}}dK.$ (20) $$\\n\\nSumming over all iterations, we conclude that, with probability at least $1-\\\\delta/3$,\\n\\n$$ $\\\\displaystyle\\\\mathbb{E}_{(x,y)\\\\sim\\\\mathcal{D}}[\\\\ell(y,\\\\pi^{\\\\mathsf{BR}}(f_{i}^ {T_{i}}(x)))-\\\\ell(y,\\\\pi^{\\\\mathsf{BR}}(f_{i}(x)))]\\\\leq T_{i}(\\\\beta\\\\sqrt{d}K+ \\\\sqrt{\\\\frac{3\\\\ln(6dK|S|\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert/\\\\delta)}{2n}}dK)$ $$\\n\\nfor all $i\\\\in\\\\{1,2\\\\}$.\\n\\nDisagreement Event.\\nBy Lemma C.8, with probability at least $1-\\\\delta/(3|S|)$, we have, for all $\\\\ell\\\\in\\\\mathcal{L}$, $a_{1},a_{2}\\\\in\\\\mathcal{A}$ with $a_{1}\\\\neq a_{2}$ and all $(f_{1},f_{2})\\\\in S$,\\n\\n$$ $\\\\displaystyle\\\\left|\\\\mu(E_{\\\\ell,a_{1},a_{2}}(f_{1}(x),f_{2}(x),x)-\\\\frac{1}{n} \\\\sum_{i=1}^{n}\\\\mathbb{I}[E_{\\\\ell,a_{1},a_{2}}(f_{1}(x_{i}),f_{2}(x_{i}),x_{i}) )]\\\\right|\\\\leq\\\\sqrt{\\\\frac{2\\\\ln(6K|S|\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert/\\\\delta)} {2n}}.$ $$\\n\\nFrom the while loop condition in Algorithm 2, we know that\\n\\n$$ $\\\\frac{1}{n}\\\\sum_{i=1}^{n}\\\\mathbb{I}[E_{\\\\ell,a_{1},a_{2}}(f_{1}^{T_{1}}(x_{i}), f_{2}^{T_{2}}(x_{i}),x_{i})]\\\\leq\\\\eta$ $$\\n\\nThen, using union bound over all $(f_{1},f_{2})\\\\in S$ and $\\\\ell\\\\in\\\\mathcal{L}$, with probability at least $1-\\\\delta/3$, we have the guarantee\\n\\n$$ $\\\\displaystyle\\\\mu(E_{\\\\ell,a_{1},a_{2}}(f_{1}^{T_{1}}(x),f_{2}^{T_{2}}(x),x))\\\\leq$ $\\\\displaystyle\\\\frac{1}{n}\\\\sum_{i=1}^{n}\\\\mathbb{I}[E_{\\\\ell,a_{1},a_{2}}(f_{1}^{T _{1}}(x_{i}),f_{2}^{T_{2}}(x_{i}),x_{i})]+\\\\sqrt{\\\\frac{2\\\\ln(6K|S|\\\\left\\\\lvert \\\\mathcal{L}\\\\right\\\\rvert/\\\\delta)}{2n}}$ (21) $\\\\displaystyle\\\\leq$ $\\\\displaystyle\\\\eta+\\\\sqrt{\\\\frac{2\\\\ln(6K|S|\\\\left\\\\lvert\\\\mathcal{L}\\\\right\\\\rvert/ \\\\delta)}{2n}}$ (22) $$\\n\\nfor all $a_{1},a_{2}\\\\in\\\\mathcal{A},a_{1}\\\\neq a_{2}$ and $\\\\ell\\\\in\\\\mathcal{L}$.\\n\\nFinally, using results from Lemma C.2, value of $T_{\\\\max}$, and $m=\\\\left\\\\lceil\\\\sqrt{\\\\frac{d}{2\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}}\\\\right\\\\rceil$, we conclude by showing\\n\\n$$ $\\\\displaystyle\\\\ln(|S|)\\\\leq$ $\\\\displaystyle\\\\ln\\\\left((4|\\\\mathcal{L}|^{2}K^{3}(m+1)^{d})^{T_{\\\\max}+1}\\\\right)$ $\\\\displaystyle=$ $\\\\displaystyle(T_{\\\\max}+1)\\\\ln(4|\\\\mathcal{L}|^{2}K^{3}(m+1)^{d})$ $\\\\displaystyle=$ $\\\\displaystyle\\\\left(\\\\frac{2d}{\\\\min\\\\{\\\\beta^{2},\\\\eta\\\\alpha^{2}/4d\\\\}}+1\\\\right)\\\\ln \\\\left(4|\\\\mathcal{L}|^{2}K^{3}\\\\left(\\\\left\\\\lceil\\\\sqrt{\\\\frac{d}{2\\\\min\\\\{\\\\beta^{2}, \\\\eta\\\\alpha^{2}/4d\\\\}}}\\\\right\\\\rceil+1\\\\right)^{d}\\\\right)$ $$\\n\\n∎\\n\\n###### Lemma C.5 .\\n\\n###### Proof.\\n\\n###### Lemma C.6 .\\n\\n###### Proof.\\n\\n###### Lemma C.7 .\\n\\n###### Proof.\\n\\n###### Lemma C.8 .\\n\\n###### Proof.\\n\\n###### Theorem C.9 .\\n\\n###### Proof.\",\n \"metadata\": {\n \"authors\": [\n \"Ally Yalei Du\",\n \"Carnegie Mellon University,\",\n \"{aydu, zstevenwu}@cmu.edu\",\n \"Daniel Ngo\",\n \"University of Minnesota,\",\n \"Zhiwei Steven Wu\"\n ],\n \"title\": \"Reconciling Model Multiplicity for Downstream Decision Making\",\n \"total_chars\": 76772,\n \"total_lines\": 902,\n \"updated_at\": \"2026-01-25T15:06:49.888290\"\n }\n}"},"reviews":{"kind":"list like","value":[{"id":"FSZcgEYDki","initial_rating":6,"confidence":3,"soundness":3,"contribution":2,"presentation":3,"summary":"This paper studied the problem of model multiplicity in downstream decision-making. In this setting, two predictive models of equivalent accuracy do not agree on their predictions for the downstream decision-making problem.\nThe paper proposed a new calibration framework which calibrates the predictive models with respect to both a finite set of downstream decision-making problems and the individual probability prediction. Further, the paper proposed an algorithm that first reconciles the differences in individual probability prediction then calibrates the updated models so that they are indistinguishable from the true probability distribution to the decision-makers together with its finite-sample analysis. Numerical experiments on two datasets demonstrate the effectiveness of the proposed method.","strengths":"- This paper studied an important and practical problem of model multiplicity. The paper overall is well-organized and presented with a good clarity. \n- It is in particular helpful to have the illustrative example in Figure 1, which directly shows that it is insufficient to only update two predictive models so that they have improved squared loss and nearly agree on their individual predictions almost everywhere.\n- Theoretical guarantee shows that the new algorithm ReDCal provides an improved accuracy and approximately agrees on the best-response action almost everywhere compared to the prior work.","weaknesses":"- The experimental results with the HAM10000 dataset show substantially larger error bars, and much less smooth convergence. It is helpful to provide more details on this differences between the two sets of results.\n- The experiments only compared to one other baseline proposed in (Roth et al 2023). How does the proposed algorithm compared to other related works in the model multiplicity?","questions":"- W.r.t. the finite sample analysis, is the proposed algorithm robust towards any noise in the samples?"},{"id":"yK6jfvgsxH","initial_rating":3,"confidence":4,"soundness":2,"contribution":2,"presentation":1,"summary":"This paper studies the Reconcile algorithm proposed in [1] which takes two prediction models and outputs two models that have better prediction accuracy which is quantified by the Brier score. They expand on the existing method by proposing a new reconcile algorithm that looks to improve prediction accuracy while also preserving the downstream decision-loss. They provide theoretical bounds that show their method trades off improvement to prediction accuracy at the cost of downstream decision-loss. They provide numerical results to show their algorithm helps reduce the loss gap after running their version of the Reconcile algorithm. \n\n\n\n-----\n[1] Aaron Roth, Alexander Tolbert, and Scott Weinstein. Reconciling individual probability fore- casts. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Trans- parency, FAccT ’23, pp. 101–110, New York, NY, USA, 2023. Association for Computing Ma- chinery. ISBN 9798400701924. doi: 10.1145/3593013.3593980. URL https://doi.org/ 10.1145/3593013.3593980.","strengths":"The paper highlights an important problem, that improvements to prediction models can hurt downstream decision-making since downstream decision-makers may have loss functions that do not necessarily align with prediction accuracy. The paper combines existing work in multi-calibration with work in model multiplicity to solve this problem. The algorithm proposed by the paper seems novel and provides what seems to be sensible theoretical guarantees that trade-off between improvements to prediction accuracy and preserving decision-loss. This seems to be a relatively significant improvement to the Reconcile algorithm.","weaknesses":"The paper has weaknesses in its presentation as well as results that seem somewhat suspicious/hard to interpret precisely.\n\nThe following items could be addressed and improved for presentation:\n- In the paper's introduction, the authors mention calibration several times, but for someone not immediately familiar with the literature it's hard to understand what it is formally. It becomes a little better defined at Lemma 2.6, but having extra background or explanation in the introduction would be helpful in understanding the high-level concepts used to construction the paper's proposed algorithm. \n- The paper could provide better background on the Reconcile algorithm which seems to be a key motivation for pursuing the work. There is no high level description of how the algorithm works, so Theorem 2.7 for example feels impossible to verify. Additionally, the Reconcile algorithm reproduced in the appendix seems to be missing key elements, like defining $h$ and $\\mu$. The former seems to be an important update step.\n- The paper does not define notation well. For example $\\mu$ and $A$ are not defined or are defined offhandedly, making it hard to decipher when it shows up later in the text.\n- In Theorem 2.7 the distribution $\\mathcal{D}$ is not defined which makes the last expectation in the proof hard to verify. I could not immediately come up with a distribution where the expectation is 1/2. \n- Some notation seems to be incorrect or have a typo. For example the Definition 2.4 of $E_{\\ell,a}$ is an indicator outside a set. I don't think this is standard notation and is not precise. My guess is it's a typo.\n- Is the notation $\\ell_a$ necessary? If so how do you obtain $\\ell_a$ given $\\ell$?\n- The descriptions of Decision-Calibration, ReDCal, and Decision-Calibration + ReDCal are confusing to me. What exactly is the algorithm for the first and last algorithms. When you count the number of time steps for ReDCal, which loop are you counting? Does it include the loops in decision-calibration? \n\nThe following items can be addressed to help improve the results of the paper:\n\n**Questions related to Theorem 3.4**\n- In Theorem 3.4, does item 1. have a typo? \n- Define $A$ the number of actions in the theorem statement for part 3. since it is off-handedly defined somewhere else in the paper. \n- How do I choose $\\beta$, $\\alpha$, and $\\eta$ if I wanted to guarantee a certain decision-making loss level while optimizing the Brier scores? Managing the trade-off between prediction accuracy and decision-loss would help improve the impact of this theoretical result. Right now it seems hard since the term $T_i \\beta$ seems hard to control.\n- The decision gap in the numerical results don't seem to correspond to the bound in 3. Shouldn't it be close to 0 if you choose $\\beta = 0.00001$?","questions":"1. One thing I didn't understand was if you are given two models at the end of ReDCal, how do you choose which model to use? Why is it important that the two models agree in the decisions they make if you already control the decision-loss gap?\n2. If you don't want to affect the decision-loss, it makes sense to me that you just wouldn't change the two predictors. Is that the case and is it reflected in your algorithm if you choose the correct parameters?\n3. Why does the algorithm converge so fast in the numerics? The upper bound on the time steps seems to suggest the number of time steps that should be run before convergence should be large given the choices of $\\beta$ and $\\alpha$. Does this imply the bound is is potentially very loose and is it possible for it to be made tighter?"},{"id":"TbmGeckPjy","initial_rating":6,"confidence":4,"soundness":3,"contribution":3,"presentation":3,"summary":"The paper presents a new algorithm for reconciling two or more models in a supervised setting; the goal is to address the model multiplicity problem (which refers to the phenomenon that there may be > 1 model with similar accuracy but substantial disagreements among predictions) in a decision-aware way. Specifically, the paper shows how to perform model reconciliation in a way that (1) makes the models agree in the prediction space, (2) does not hurt (and possibly even improves) the squared loss of either model, and (3) does not hurt (and possibly even improves) the decision loss of a given downstream decision-maker. (The decision loss is an arbitrary linear loss that the decision maker uses to map predictions to actions.) In addition, the paper performs preliminary experiments showing that the method outperforms (with respect to decision loss) a prior reconciliation algorithm on some vision datasets.","strengths":"This paper is overall a good and novel contribution to the predictive multiplicity literature. Specifically:\n\n+ The contribution of this paper appears new in the model multiplicity literature: it gives a rigorous method for reconciliation of any two models with provable guarantees in terms of the resulting model loss (for which only one algorithm exists in the literature), while at the same time ensuring in a rigorous way that downstream decision making is not affected negatively (which is new); and it moreover extends this method to more than 2 models in a natural way. \n\n+ In fact, from what I can tell, it has an even broader scope than typically considered in the model multiplicity context, as it does not require the to-be-reconciled models to start off having similar or equal accuracy. When the to-be-reconciled models do have similar accuracy to begin with, then the paper's contribution intuitively appears to offer the interesting and thought-provoking semantics of: \"given two different models with similar performance, you can bypass the multiplicity issue by finding an even higher-performing model (which ensembles/combines the two initial models), with similar or better downstream decision making guarantees\".","weaknesses":"While overall I believe this to be a good-quality paper, there is the following (relatively non-major) consideration that I would call a weakness:\n\n- The paper currently appears written with a primarily theoretical audience in mind, but I think it could still do a better/more thorough job coming up with/describing experiments. It currently gives two semi-synthetic ones. In the first one, linear decision losses are generated in a Gaussian manner --- so that the two vision models are essentially being calibrated to realizations of random noise (granted, the experiment does illustrate the point that the new method achieves better decision-making properties than the old one, but the setup still sounds strange). In the second one, the decision loss is a \"loss function motivated by medical domain knowledge in Zhao et al (2021) and additional random noise\" --- and in this case, the paper should at the very least clearly state what the loss function in Zhao et al (2021) is, and perhaps provide more principled perturbations of it than Gaussian noise ones.\n\nNeither of these experiments, however, connects in any way to the examples given earlier in the paper on how disregarding decision loss can lead to its deterioration during the reconciliation procedure. These examples are in some sense prototypical of how reconciliation could hurt downstream decision making; and given how easy it is to come up with a synthetic task/models + synthetic loss clearly showcasing them (i.e. making it to where some significant mass of items will flip across the decision boundary as a result of the reconciliation), I would like to urge the authors to do just that as it will make the paper more logically coherent.","questions":"The paper currently appears mostly in good shape; my main small improvement suggestions are about the experimental section (at least by lightly tweaking existing experiments/adding another simple but conceptual one). \n\nIn addition, even though the writing of the paper is mostly quite good, it could still be improved in a few places. For instance, the subsection about handling more than 2 models is somewhat confusingly written: e.g. consider the sentence about exponentially many output predictors in the number of models k, followed by the sentence that union and Chernoff-Hoeffding bounds \"suggest\" that the sample complexity scales linearly in k --- a reader who may not have checked the details in the appendix may easily confuse this for saying that the method may not necessarily be scalable in k, even though it actually is. Also, the event collection setup in that case, which involves reconciling a single base model with every other model, is not described that well in prose --- so consider e.g. including a quick diagram which could e.g. have the base model with \"reconciliation arrows\" pointing to the other models, or something similar."},{"id":"xM6sILiTmP","initial_rating":6,"confidence":3,"soundness":3,"contribution":3,"presentation":3,"summary":"The paper provides a framework using multi-calibration to reconcile model multiplicity in downstream decisions. It seeks to develop a framework to address the inherent discrepancies between predictive models, which often result in varying decision recommendations despite equivalent accuracy levels. The authors propose an algorithm that aligns predictive models with decision-making, improving agreement on best-response actions and losses. Empirical validation shows enhancements in downstream decision performance and addressing prediction disagreements, relevant to multi-calibration advancements. Proposed approach improves the consistency and utility of models in decision-based applications, including cases where only empirical data is available.","strengths":"The paper presents a novel framework to address the issue of model multiplicity in predictive modeling for decision-making, using multi-calibration techniques. It identifies specific hyperparameters, such as loss margin and decision-calibration tolerance as the driving parameters of key results. In practice, these parameters will need to be adjusted to mediate trade-offs between model fairness and computational efficiency. The authors provide both theoretical insights and empirical validations of their approach. The paper makes a reasonable contribution by integrating theoretical multi-calibration concepts into empirical settings and testing them across multiple datasets.","weaknesses":"Comments/Questions to Authors: \n•\tThe initial concept of multi-calibration was introduced by Hebert-Johnson et al. in 2017 which focused on ensuring fairness across overlapping subpopulations. Recent advancement includes extension of multi-calibration to game dynamics and multi-objective learning – for example, work by Nika Haghtalab and Eric Zhao, which utilize game dynamics to connect fairness and optimization. Another line of work by Roth et al deals with a wide array of issues from online multicalibrated learning to omnipredictors. While the article does a good job at comparing and showing differences with respect to work of Roth, the conceptual difference with respect to work of Haghtalab et al. would be much appreciated. \n\n•\tThe algorithm’s performance depends on several hyperparameters (such as loss margin and decision-calibration tolerance). It seems techniques to fine-tune parameters such as loss margin and decision-calibration tolerance are crucial for achieving the results and trade-offs that is important in practice (e.g., between prediction fairness and accuracy). These are critical for adapting multi-calibration models to robustly function across heterogeneous data distributions and complex decision environments, the paper should clearly discuss how the choice of these hyperparameters impact the performance of proposed algorithm. I am also curious about the third result of Theorem 3.2 (remark 3.3). How does the impact of this result show up in practice/experiments? I believe a major limitation of the paper is lack of explicit focus on loss margin and decision-calibration tolerance hyperparameters.\n\n•\tThe paper could better motivate the empirical tests conducted and discuss how they confirm the theoretical claims and show improvements in decision-making outcomes and model agreement. Avenues for future work on validating the proposed algorithm's effectiveness in real-world data scenarios should be discussed. \n\n•\tAdditional literature – please clarify the relationships / novelty of this paper with respect to work of these authors. I would be especially interested in knowing the relevance of this paper with issues of fairness-accuracy tradeoff. \n\n“Inference for an Algorithmic Fairness-Accuracy Frontier,” Yiqi Liu and Francesca Molinari (2024). This work provides a consistent estimator for a fairness-accuracy frontier. Method for testing fairness-related hypotheses in algorithmic decision-making seems relevant. \n\n“Fair Representation: Guaranteeing Approximate Multiple Group Fairness for Unknown Tasks,” Xudong Shen and Yongkang Wong and Mohan S. Kankanhalli, IEEE Transactions on Pattern Analysis and Machine Intelligence (2021). Explores approximate fairness using fair representation across multiple fairness notions. \n\n“Multigroup Robustness,” Lunjia Hu and Charlotte Peale and Judy Hanwen Shen (2024). This work establishes a connection between multigroup fairness and robustness and discusses robustness algorithms tailored for subpopulations under data corruption.\n\n“Inherent Trade-Offs in the Fair Determination of Risk Scores,” J. Kleinberg and S. Mullainathan and Manish Raghavan (2016). Shows inherent trade-offs in algorithmic fairness in risk scores and formalizes three fairness conditions and proves their incompatibility in most cases.\n\n“Multi-group Agnostic PAC Learnability,” G. Rothblum and G. Yona, International Conference on Machine Learning (2021). Provides a framework for multi-group agnostic PAC learning.\n\nAlso important is work by Michael P. Kim et al. on Multiaccuracy: Black-Box Post-Processing for Fairness in Classification (2018): It introduces a foundational framework for ensuring fairness in classification, with a focus on multi-group fairness using multiaccuracy post-processing techniques. This paper sets a critical baseline for future research on fairness in predictions.","questions":"Please see my questions above. Additionally: \n\n1) How does the framework perform across different domains or datasets that exhibit varying levels of complexity and noise?\n2) As mentioned, determining optimal ranges and conditions for key hyperparameters parameters across varying contexts can be challenging. \n3) Although the paper addresses computational efficiency, the scalability of the algorithm for very large datasets remains an open question. How will the computational complexity grow with increased data size?"}],"string":"[\n {\n \"id\": \"FSZcgEYDki\",\n \"initial_rating\": 6,\n \"confidence\": 3,\n \"soundness\": 3,\n \"contribution\": 2,\n \"presentation\": 3,\n \"summary\": \"This paper studied the problem of model multiplicity in downstream decision-making. In this setting, two predictive models of equivalent accuracy do not agree on their predictions for the downstream decision-making problem.\\nThe paper proposed a new calibration framework which calibrates the predictive models with respect to both a finite set of downstream decision-making problems and the individual probability prediction. Further, the paper proposed an algorithm that first reconciles the differences in individual probability prediction then calibrates the updated models so that they are indistinguishable from the true probability distribution to the decision-makers together with its finite-sample analysis. Numerical experiments on two datasets demonstrate the effectiveness of the proposed method.\",\n \"strengths\": \"- This paper studied an important and practical problem of model multiplicity. The paper overall is well-organized and presented with a good clarity. \\n- It is in particular helpful to have the illustrative example in Figure 1, which directly shows that it is insufficient to only update two predictive models so that they have improved squared loss and nearly agree on their individual predictions almost everywhere.\\n- Theoretical guarantee shows that the new algorithm ReDCal provides an improved accuracy and approximately agrees on the best-response action almost everywhere compared to the prior work.\",\n \"weaknesses\": \"- The experimental results with the HAM10000 dataset show substantially larger error bars, and much less smooth convergence. It is helpful to provide more details on this differences between the two sets of results.\\n- The experiments only compared to one other baseline proposed in (Roth et al 2023). How does the proposed algorithm compared to other related works in the model multiplicity?\",\n \"questions\": \"- W.r.t. the finite sample analysis, is the proposed algorithm robust towards any noise in the samples?\"\n },\n {\n \"id\": \"yK6jfvgsxH\",\n \"initial_rating\": 3,\n \"confidence\": 4,\n \"soundness\": 2,\n \"contribution\": 2,\n \"presentation\": 1,\n \"summary\": \"This paper studies the Reconcile algorithm proposed in [1] which takes two prediction models and outputs two models that have better prediction accuracy which is quantified by the Brier score. They expand on the existing method by proposing a new reconcile algorithm that looks to improve prediction accuracy while also preserving the downstream decision-loss. They provide theoretical bounds that show their method trades off improvement to prediction accuracy at the cost of downstream decision-loss. They provide numerical results to show their algorithm helps reduce the loss gap after running their version of the Reconcile algorithm. \\n\\n\\n\\n-----\\n[1] Aaron Roth, Alexander Tolbert, and Scott Weinstein. Reconciling individual probability fore- casts. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Trans- parency, FAccT ’23, pp. 101–110, New York, NY, USA, 2023. Association for Computing Ma- chinery. ISBN 9798400701924. doi: 10.1145/3593013.3593980. URL https://doi.org/ 10.1145/3593013.3593980.\",\n \"strengths\": \"The paper highlights an important problem, that improvements to prediction models can hurt downstream decision-making since downstream decision-makers may have loss functions that do not necessarily align with prediction accuracy. The paper combines existing work in multi-calibration with work in model multiplicity to solve this problem. The algorithm proposed by the paper seems novel and provides what seems to be sensible theoretical guarantees that trade-off between improvements to prediction accuracy and preserving decision-loss. This seems to be a relatively significant improvement to the Reconcile algorithm.\",\n \"weaknesses\": \"The paper has weaknesses in its presentation as well as results that seem somewhat suspicious/hard to interpret precisely.\\n\\nThe following items could be addressed and improved for presentation:\\n- In the paper's introduction, the authors mention calibration several times, but for someone not immediately familiar with the literature it's hard to understand what it is formally. It becomes a little better defined at Lemma 2.6, but having extra background or explanation in the introduction would be helpful in understanding the high-level concepts used to construction the paper's proposed algorithm. \\n- The paper could provide better background on the Reconcile algorithm which seems to be a key motivation for pursuing the work. There is no high level description of how the algorithm works, so Theorem 2.7 for example feels impossible to verify. Additionally, the Reconcile algorithm reproduced in the appendix seems to be missing key elements, like defining $h$ and $\\\\mu$. The former seems to be an important update step.\\n- The paper does not define notation well. For example $\\\\mu$ and $A$ are not defined or are defined offhandedly, making it hard to decipher when it shows up later in the text.\\n- In Theorem 2.7 the distribution $\\\\mathcal{D}$ is not defined which makes the last expectation in the proof hard to verify. I could not immediately come up with a distribution where the expectation is 1/2. \\n- Some notation seems to be incorrect or have a typo. For example the Definition 2.4 of $E_{\\\\ell,a}$ is an indicator outside a set. I don't think this is standard notation and is not precise. My guess is it's a typo.\\n- Is the notation $\\\\ell_a$ necessary? If so how do you obtain $\\\\ell_a$ given $\\\\ell$?\\n- The descriptions of Decision-Calibration, ReDCal, and Decision-Calibration + ReDCal are confusing to me. What exactly is the algorithm for the first and last algorithms. When you count the number of time steps for ReDCal, which loop are you counting? Does it include the loops in decision-calibration? \\n\\nThe following items can be addressed to help improve the results of the paper:\\n\\n**Questions related to Theorem 3.4**\\n- In Theorem 3.4, does item 1. have a typo? \\n- Define $A$ the number of actions in the theorem statement for part 3. since it is off-handedly defined somewhere else in the paper. \\n- How do I choose $\\\\beta$, $\\\\alpha$, and $\\\\eta$ if I wanted to guarantee a certain decision-making loss level while optimizing the Brier scores? Managing the trade-off between prediction accuracy and decision-loss would help improve the impact of this theoretical result. Right now it seems hard since the term $T_i \\\\beta$ seems hard to control.\\n- The decision gap in the numerical results don't seem to correspond to the bound in 3. Shouldn't it be close to 0 if you choose $\\\\beta = 0.00001$?\",\n \"questions\": \"1. One thing I didn't understand was if you are given two models at the end of ReDCal, how do you choose which model to use? Why is it important that the two models agree in the decisions they make if you already control the decision-loss gap?\\n2. If you don't want to affect the decision-loss, it makes sense to me that you just wouldn't change the two predictors. Is that the case and is it reflected in your algorithm if you choose the correct parameters?\\n3. Why does the algorithm converge so fast in the numerics? The upper bound on the time steps seems to suggest the number of time steps that should be run before convergence should be large given the choices of $\\\\beta$ and $\\\\alpha$. Does this imply the bound is is potentially very loose and is it possible for it to be made tighter?\"\n },\n {\n \"id\": \"TbmGeckPjy\",\n \"initial_rating\": 6,\n \"confidence\": 4,\n \"soundness\": 3,\n \"contribution\": 3,\n \"presentation\": 3,\n \"summary\": \"The paper presents a new algorithm for reconciling two or more models in a supervised setting; the goal is to address the model multiplicity problem (which refers to the phenomenon that there may be > 1 model with similar accuracy but substantial disagreements among predictions) in a decision-aware way. Specifically, the paper shows how to perform model reconciliation in a way that (1) makes the models agree in the prediction space, (2) does not hurt (and possibly even improves) the squared loss of either model, and (3) does not hurt (and possibly even improves) the decision loss of a given downstream decision-maker. (The decision loss is an arbitrary linear loss that the decision maker uses to map predictions to actions.) In addition, the paper performs preliminary experiments showing that the method outperforms (with respect to decision loss) a prior reconciliation algorithm on some vision datasets.\",\n \"strengths\": \"This paper is overall a good and novel contribution to the predictive multiplicity literature. Specifically:\\n\\n+ The contribution of this paper appears new in the model multiplicity literature: it gives a rigorous method for reconciliation of any two models with provable guarantees in terms of the resulting model loss (for which only one algorithm exists in the literature), while at the same time ensuring in a rigorous way that downstream decision making is not affected negatively (which is new); and it moreover extends this method to more than 2 models in a natural way. \\n\\n+ In fact, from what I can tell, it has an even broader scope than typically considered in the model multiplicity context, as it does not require the to-be-reconciled models to start off having similar or equal accuracy. When the to-be-reconciled models do have similar accuracy to begin with, then the paper's contribution intuitively appears to offer the interesting and thought-provoking semantics of: \\\"given two different models with similar performance, you can bypass the multiplicity issue by finding an even higher-performing model (which ensembles/combines the two initial models), with similar or better downstream decision making guarantees\\\".\",\n \"weaknesses\": \"While overall I believe this to be a good-quality paper, there is the following (relatively non-major) consideration that I would call a weakness:\\n\\n- The paper currently appears written with a primarily theoretical audience in mind, but I think it could still do a better/more thorough job coming up with/describing experiments. It currently gives two semi-synthetic ones. In the first one, linear decision losses are generated in a Gaussian manner --- so that the two vision models are essentially being calibrated to realizations of random noise (granted, the experiment does illustrate the point that the new method achieves better decision-making properties than the old one, but the setup still sounds strange). In the second one, the decision loss is a \\\"loss function motivated by medical domain knowledge in Zhao et al (2021) and additional random noise\\\" --- and in this case, the paper should at the very least clearly state what the loss function in Zhao et al (2021) is, and perhaps provide more principled perturbations of it than Gaussian noise ones.\\n\\nNeither of these experiments, however, connects in any way to the examples given earlier in the paper on how disregarding decision loss can lead to its deterioration during the reconciliation procedure. These examples are in some sense prototypical of how reconciliation could hurt downstream decision making; and given how easy it is to come up with a synthetic task/models + synthetic loss clearly showcasing them (i.e. making it to where some significant mass of items will flip across the decision boundary as a result of the reconciliation), I would like to urge the authors to do just that as it will make the paper more logically coherent.\",\n \"questions\": \"The paper currently appears mostly in good shape; my main small improvement suggestions are about the experimental section (at least by lightly tweaking existing experiments/adding another simple but conceptual one). \\n\\nIn addition, even though the writing of the paper is mostly quite good, it could still be improved in a few places. For instance, the subsection about handling more than 2 models is somewhat confusingly written: e.g. consider the sentence about exponentially many output predictors in the number of models k, followed by the sentence that union and Chernoff-Hoeffding bounds \\\"suggest\\\" that the sample complexity scales linearly in k --- a reader who may not have checked the details in the appendix may easily confuse this for saying that the method may not necessarily be scalable in k, even though it actually is. Also, the event collection setup in that case, which involves reconciling a single base model with every other model, is not described that well in prose --- so consider e.g. including a quick diagram which could e.g. have the base model with \\\"reconciliation arrows\\\" pointing to the other models, or something similar.\"\n },\n {\n \"id\": \"xM6sILiTmP\",\n \"initial_rating\": 6,\n \"confidence\": 3,\n \"soundness\": 3,\n \"contribution\": 3,\n \"presentation\": 3,\n \"summary\": \"The paper provides a framework using multi-calibration to reconcile model multiplicity in downstream decisions. It seeks to develop a framework to address the inherent discrepancies between predictive models, which often result in varying decision recommendations despite equivalent accuracy levels. The authors propose an algorithm that aligns predictive models with decision-making, improving agreement on best-response actions and losses. Empirical validation shows enhancements in downstream decision performance and addressing prediction disagreements, relevant to multi-calibration advancements. Proposed approach improves the consistency and utility of models in decision-based applications, including cases where only empirical data is available.\",\n \"strengths\": \"The paper presents a novel framework to address the issue of model multiplicity in predictive modeling for decision-making, using multi-calibration techniques. It identifies specific hyperparameters, such as loss margin and decision-calibration tolerance as the driving parameters of key results. In practice, these parameters will need to be adjusted to mediate trade-offs between model fairness and computational efficiency. The authors provide both theoretical insights and empirical validations of their approach. The paper makes a reasonable contribution by integrating theoretical multi-calibration concepts into empirical settings and testing them across multiple datasets.\",\n \"weaknesses\": \"Comments/Questions to Authors: \\n•\\tThe initial concept of multi-calibration was introduced by Hebert-Johnson et al. in 2017 which focused on ensuring fairness across overlapping subpopulations. Recent advancement includes extension of multi-calibration to game dynamics and multi-objective learning – for example, work by Nika Haghtalab and Eric Zhao, which utilize game dynamics to connect fairness and optimization. Another line of work by Roth et al deals with a wide array of issues from online multicalibrated learning to omnipredictors. While the article does a good job at comparing and showing differences with respect to work of Roth, the conceptual difference with respect to work of Haghtalab et al. would be much appreciated. \\n\\n•\\tThe algorithm’s performance depends on several hyperparameters (such as loss margin and decision-calibration tolerance). It seems techniques to fine-tune parameters such as loss margin and decision-calibration tolerance are crucial for achieving the results and trade-offs that is important in practice (e.g., between prediction fairness and accuracy). These are critical for adapting multi-calibration models to robustly function across heterogeneous data distributions and complex decision environments, the paper should clearly discuss how the choice of these hyperparameters impact the performance of proposed algorithm. I am also curious about the third result of Theorem 3.2 (remark 3.3). How does the impact of this result show up in practice/experiments? I believe a major limitation of the paper is lack of explicit focus on loss margin and decision-calibration tolerance hyperparameters.\\n\\n•\\tThe paper could better motivate the empirical tests conducted and discuss how they confirm the theoretical claims and show improvements in decision-making outcomes and model agreement. Avenues for future work on validating the proposed algorithm's effectiveness in real-world data scenarios should be discussed. \\n\\n•\\tAdditional literature – please clarify the relationships / novelty of this paper with respect to work of these authors. I would be especially interested in knowing the relevance of this paper with issues of fairness-accuracy tradeoff. \\n\\n“Inference for an Algorithmic Fairness-Accuracy Frontier,” Yiqi Liu and Francesca Molinari (2024). This work provides a consistent estimator for a fairness-accuracy frontier. Method for testing fairness-related hypotheses in algorithmic decision-making seems relevant. \\n\\n“Fair Representation: Guaranteeing Approximate Multiple Group Fairness for Unknown Tasks,” Xudong Shen and Yongkang Wong and Mohan S. Kankanhalli, IEEE Transactions on Pattern Analysis and Machine Intelligence (2021). Explores approximate fairness using fair representation across multiple fairness notions. \\n\\n“Multigroup Robustness,” Lunjia Hu and Charlotte Peale and Judy Hanwen Shen (2024). This work establishes a connection between multigroup fairness and robustness and discusses robustness algorithms tailored for subpopulations under data corruption.\\n\\n“Inherent Trade-Offs in the Fair Determination of Risk Scores,” J. Kleinberg and S. Mullainathan and Manish Raghavan (2016). Shows inherent trade-offs in algorithmic fairness in risk scores and formalizes three fairness conditions and proves their incompatibility in most cases.\\n\\n“Multi-group Agnostic PAC Learnability,” G. Rothblum and G. Yona, International Conference on Machine Learning (2021). Provides a framework for multi-group agnostic PAC learning.\\n\\nAlso important is work by Michael P. Kim et al. on Multiaccuracy: Black-Box Post-Processing for Fairness in Classification (2018): It introduces a foundational framework for ensuring fairness in classification, with a focus on multi-group fairness using multiaccuracy post-processing techniques. This paper sets a critical baseline for future research on fairness in predictions.\",\n \"questions\": \"Please see my questions above. Additionally: \\n\\n1) How does the framework perform across different domains or datasets that exhibit varying levels of complexity and noise?\\n2) As mentioned, determining optimal ranges and conditions for key hyperparameters parameters across varying contexts can be challenging. \\n3) Although the paper addresses computational efficiency, the scalability of the algorithm for very large datasets remains an open question. How will the computational complexity grow with increased data size?\"\n }\n]","isConversational":false},"scores":{"kind":"string","value":"{\n \"rating\": \"3;6;6;6\",\n \"rating_avg\": 5.25,\n \"confidence\": \"4;3;4;3\",\n \"confidence_avg\": 3.5,\n \"soundness\": \"2;3;3;3\",\n \"soundness_avg\": 2.75,\n \"contribution\": \"2;3;3;2\",\n \"contribution_avg\": 2.5,\n \"presentation\": \"1;3;3;3\",\n \"presentation_avg\": 2.5\n}"},"metadata":{"kind":"string","value":"{\n \"primary_area\": \"\",\n \"track\": \"main\",\n \"venue\": \"ICLR 2025 Poster\",\n \"venueid\": \"ICLR.cc/2025/Conference\",\n \"processed_at\": \"2026-01-14T22:16:03.999401\"\n}"},"meta_review":{"kind":"string","value":"{\n \"id\": \"XEzGmTRo7r\",\n \"metareview\": \"This a paper which just sits above the borderline area. The paper studies the problem of model multiplicity in downstream decision-making.\\nAll reviews have been cautiously positive with various suggestions for improving and strengthening the paper. In particular, the paper studies an important problem and provides some theoretical guarantees. I believe the paper is good enough for publication but the authors should incorporate the improvements suggested by the reviewers.\",\n \"additional_comments\": \"A good discussion among reviewers and authors took place which resulted in some scores being updated.\"\n}"},"decision":{"kind":"string","value":"{\n \"decision\": \"Accept (Poster)\"\n}"}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":36,"numItemsPerPage":100,"numTotalItems":4011,"offset":3600,"length":100}},"jwt":"eyJhbGciOiJFZERTQSIsImtpZCI6Im1sR1pFYktIVHRVSGRhX1RRdEczQ0N1S3k5ME9Qd25DNHY5elRjM3FVYzgifQ.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4MzU4MjMyMiwianRpIjoiMjNjNzk2MjEtZjMwNS00OTMwLTgzNGUtYWFlMGFhZTY5ZWJmIiwic3ViIjoiL2RhdGFzZXRzL1VLUExhYi9Qcm9SZXZpZXdlci1EYXRhc2V0IiwiZXhwIjoxNzgzNTg1OTIyLCJpc3MiOiJodHRwczovL2h1Z2dpbmdmYWNlLmNvIn0.KxISCAg19EQt2J_0lkE2WORoLcb3nXUe2H4U2S-a9KdbhYNLyjXzpmj5dzlFo93NCp9GG8xIMVIyvsOnKXO5DQ","displayUrls":true,"splitSizeSummaries":[{"config":"default","split":"train","numRows":4011,"numBytesParquet":144344024},{"config":"default","split":"test","numRows":1000,"numBytesParquet":45705060}]},"discussionsStats":{"closed":0,"open":1,"total":1},"fullWidth":true,"hasGatedAccess":true,"hasFullAccess":true,"isEmbedded":false,"savedQueries":{"community":[],"user":[]}}">
paper_id string | arxiv_id string | title string | markdown dict | reviews list | scores dict | metadata dict | meta_review dict | decision dict |
|---|---|---|---|---|---|---|---|---|
tdfHABLdxR | 2410.07877v1 | Constrained Skill Discovery: Quadruped Locomotion with Unsupervised Reinforcement Learning | {
"content": "## Abstract\n\nAbstract Representation learning and unsupervised skill discovery can allow robots to acquire diverse and reusable behaviors without the need for task-specific rewards.\nIn this work, we use unsupervised reinforcement learning to learn a latent representation by maximizing the mutual info... | [
{
"id": "Fxn6Zrmsf9",
"initial_rating": 5,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper extends the Lipschitz-constrained Skill Discovery (LSD) framework by Park et al. (2022) to enable the learning of stable and controllable locomotion sk... | {
"rating": "5;5;5;5",
"rating_avg": 5,
"confidence": "5;3;3;3",
"confidence_avg": 3.5,
"soundness": "3;3;4;2",
"soundness_avg": 3,
"contribution": "3;2;2;2",
"contribution_avg": 2.25,
"presentation": "4;3;2;2",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.905980"
} | {
"id": "UAphjzNrCh",
"metareview": "The AC read the paper, reviewers and discussions. As suggested by the general opinions from reviewers, the quality of this paper is below the acceptance threshold.\nThe paper investigates unsupervised RL for constrained skill discovery, with applications to quadrupedal robot loc... | {
"decision": "Reject"
} |
te30nmLaFf | 2407.07612v1 | Teaching Transformers Causal Reasoning through Axiomatic Training | {
"content": "## Abstract\n\nAbstract For text-based AI systems to interact in the real world, causal reasoning is an essential skill. Since interventional data is costly to generate, we study to what extent an agent can learn causal reasoning from passive data. Specifically, we consider an axiomatic training setup w... | [
{
"id": "mpF2U6Ph52",
"initial_rating": 3,
"confidence": 5,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "This paper introduces a way of training Transformers from scratch to learn causal reasoning axioms. The idea, called axiomatic training, is to demonstrate axioms ... | {
"rating": "3;3;5;6",
"rating_avg": 4.25,
"confidence": "3;5;3;3",
"confidence_avg": 3.5,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "3;2;2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.907018"
} | {
"id": "oHqduYIrKx",
"metareview": "The paper investigates the ability of transformers to learn the axioms of causal reasoning. Instead of directly structuring a model to follow the axioms of causal reasoning, the model is trained with data to learn the axioms. The strengths of the paper are the interesting inve... | {
"decision": "Reject"
} |
tfO07iz0b9 | 2408.08192v1 | Stochastic Semi-Gradient Descent for Learning Mean Field Games with Population-Aware Function Approximation | {
"content": "## Abstract\n\nAbstract Mean field games (MFGs) model the interactions within a large-population multi-agent system using the population distribution.\nTraditional learning methods for MFGs are based on fixed-point iteration (FPI), which calculates best responses and induced population distribution sepa... | [
{
"id": "pteuaBmYsz",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper primarily concentrates on the development of a stochastic gradient method for learning in mean-field games.\nThe proposed algorithm significantly reduc... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "3;4;3;4",
"confidence_avg": 3.5,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "3;2;2;3",
"contribution_avg": 2.5,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.908355"
} | {
"id": "UknYzolVGu",
"metareview": "This paper treats the problem of learning in mean-field games (MFGs) and, in particular, analyzes a stochastic approximation method dubbed SemiSGD (stochastic semi-gradient descent) in which agents simultaneously update their policy and population estimates. The authors provide ... | {
"decision": "Accept (Poster)"
} |
tfemquulED | 2410.05760v1 | Training-Free Diffusion Model Alignment with Sampling Demons | {
"content": "## Abstract\n\nAbstract Aligning diffusion models with user preferences has been a key challenge.\nExisting methods for aligning diffusion models either require retraining or are limited to differentiable reward functions.\nTo address these limitations, we propose a stochastic optimization approach, dub... | [
{
"id": "D6ze4D2r3N",
"initial_rating": 8,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 4,
"summary": "This paper studies the problem of aligning diffusion models without additional training or backpropagation. The main idea is to optimize the sampling at inference... | {
"rating": "5;5;6;6;8",
"rating_avg": 6,
"confidence": "3;3;3;3;3",
"confidence_avg": 3,
"soundness": "3;3;3;2;3",
"soundness_avg": 2.8,
"contribution": "3;2;3;2;3",
"contribution_avg": 2.6,
"presentation": "2;3;3;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.909576"
} | {
"id": "NkMTJCDjfr",
"metareview": "This paper proposes a backpropagation-free, inference-time method for preference alignment in diffusion models. Specifically, the authors optimize for the noise distribution to search for an ''optimal'' noise that yields high rewards. Since this method is training-free, it is qu... | {
"decision": "Accept (Poster)"
} |
tfyHbvFZ0K | 2405.14117v1 | Knowledge Localization: Mission Not Accomplished? Enter Query Localization! | {
"content": "## Abstract\n\nAbstract Large language models (LLMs) store extensive factual knowledge, but the mechanisms behind how they store and express this knowledge remain unclear.\nThe Knowledge Neuron (KN) thesis is a prominent theory for explaining these mechanisms. This theory is based on the knowledge local... | [
{
"id": "TJZSeaB2RE",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper investigates the widely accepted knowledge localization (KL) assumption within related research fields by conducting a series of experiments. Through s... | {
"rating": "5;6;8;8",
"rating_avg": 6.75,
"confidence": "3;3;3;2",
"confidence_avg": 2.75,
"soundness": "3;3;4;3",
"soundness_avg": 3.25,
"contribution": "2;3;3;4",
"contribution_avg": 3,
"presentation": "3;3;4;3",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.910555"
} | {
"id": "MmF7KKb7cH",
"metareview": "The authors investigate the mechanisms behind factual knowledge storage and expression in large language models. It re-evaluates the Knowledge Localization (KL) assumption, which posits that specific knowledge neurons can store distinct facts. Through experiments, the authors id... | {
"decision": "Accept (Spotlight)"
} |
tidibw8Xdm | 2402.02651v3 | Vision-Language Models Provide Promptable Representations for Reinforcement Learning | {
"content": "## Abstract\n\nAbstract Humans can quickly learn new behaviors by leveraging background world knowledge. In contrast, agents trained with reinforcement learning (RL) typically learn behaviors from scratch. We thus propose a novel approach that uses the vast amounts of general and indexable world knowled... | [
{
"id": "fJfRUVm0kI",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "The main contributions of this paper are: (1) introducing Promptable Representations for Reinforcement Learning (PR2L), a method that leverages vision-language mo... | {
"rating": "3;6;6",
"rating_avg": 5,
"confidence": "4;5;3",
"confidence_avg": 4,
"soundness": "2;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.911451"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
tj40W2HAKN | 2406.03464v1 | Node-wise Filtering in Graph Neural Networks: A Mixture of Experts Approach | {
"content": "## Abstract\n\nAbstract Graph Neural Networks (GNNs) have proven to be highly effective for node classification tasks across diverse graph structural patterns. Traditionally, GNNs employ a uniform global filter—typically a low-pass filter for homophilic graphs and a high-pass filter for heterophilic gra... | [
{
"id": "LAUUK8UHyp",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "The paper introduces NODE-MOE, a GNN framework that uses a mixture of experts (MoE) model to apply node-specific filters, addressing the limitations of traditiona... | {
"rating": "3;3;5;5;6",
"rating_avg": 4.4,
"confidence": "4;3;4;3;4",
"confidence_avg": 3.6,
"soundness": "2;3;3;3;3",
"soundness_avg": 2.8,
"contribution": "2;2;2;2;2",
"contribution_avg": 2,
"presentation": "3;2;3;2;3",
"presentation_avg": 2.6
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.912292"
} | {
"id": "8TXrf8Nw7P",
"metareview": "This paper proposes NODE-MOE, a Graph Neural Network (GNN) framework that combines a mixture of experts (MoE) with node-specific filters to address the limitations of traditional GNNs relying on uniform global filters. The architecture features a gating model that dynamically as... | {
"decision": "Reject"
} |
tjNf0L8QjR | 2406.09415v1 | An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels | {
"content": "## Abstract\n\nAbstract This work does not introduce a new method. Instead, we present an interesting finding that questions the necessity of the inductive bias – locality in modern computer vision architectures. Concretely, we find that vanilla Transformers can operate by directly treating each individ... | [
{
"id": "2GjNh2VMa3",
"initial_rating": 6,
"confidence": 5,
"soundness": 3,
"contribution": 2,
"presentation": 4,
"summary": "The study investigates whether the inductive bias of locality is necessary for modern computer vision models. It explores the effectiveness of using vanilla Trans... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "3;5;4;5",
"confidence_avg": 4.25,
"soundness": "2;4;3;3",
"soundness_avg": 3,
"contribution": "2;2;2;2",
"contribution_avg": 2,
"presentation": "3;4;3;4",
"presentation_avg": 3.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.912974"
} | {
"id": "C2rEF4WdKf",
"metareview": "This paper explores whether inductive bias towards locality, a hallmark of convolutional networks and Vision Transformers (ViTs), is truly necessary for computer vision tasks. The study replaces patch-based tokenization (e.g., 16x16 patches) with a \"pixels-as-tokens\" approach,... | {
"decision": "Accept (Poster)"
} |
tjlTczcnPz | 2405.14917v1 | SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models | {
"content": "## Abstract\n\nAbstract Large language models (LLMs) achieve remarkable performance in natural language understanding but require substantial computation resources and memory footprint. Post-training quantization (PTQ) is a powerful compression technique extensively investigated for its effectiveness in... | [
{
"id": "KUjGUGaRDx",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "The paper presents \"SliM-LLM,\" a quantization method for large language models (LLMs), focusing on salience-driven, mixed-precision quantization to enhance memo... | {
"rating": "3;5;5;5;6",
"rating_avg": 4.8,
"confidence": "4;4;4;4;3",
"confidence_avg": 3.8,
"soundness": "3;2;3;2;3",
"soundness_avg": 2.6,
"contribution": "2;2;1;2;3",
"contribution_avg": 2,
"presentation": "1;2;3;2;3",
"presentation_avg": 2.2
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.913618"
} | {
"id": "rMOIq9x7rI",
"metareview": "This paper introduces SliM-LLM, a novel salience-driven mixed-precision quantization technique for large language models (LLMs). The method aims to optimize memory efficiency while maintaining model performance by adaptively assigning bit-widths to weight groups based on their s... | {
"decision": "Reject"
} |
tkg9XMFo0H | 2410.03577v1 | Look Twice Before You Answer: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models | {
"content": "## Abstract\n\nAbstract Despite their impressive capabilities, Multimodal Large Language Models (MLLMs) are susceptible to hallucinations, especially assertively fabricating content not present in the visual inputs. To address the aforementioned challenge, we follow a common cognitive process - when one... | [
{
"id": "zc4HnFyDwc",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes MEMVR, a training-free hallucination mitigation paradigm for MLLMs. MEMVR mimics human patterns of image understanding by revisiting image fea... | {
"rating": "3;5;5;6",
"rating_avg": 4.75,
"confidence": "5;4;3;3",
"confidence_avg": 3.75,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "2;3;3;2",
"contribution_avg": 2.5,
"presentation": "1;1;2;3",
"presentation_avg": 1.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.915362"
} | {
"id": "R6KTUPG9Ax",
"metareview": "The paper received mixed feedback, with reviewers acknowledging its innovative \"look twice\" mechanism for mitigating hallucinations in MLLMs but raising concerns about its clarity and broader impact. The proposed method offers a training-free approach that dynamically injects ... | {
"decision": "Reject"
} |
tkiZQlL04w | 2407.15891v1 | RazorAttention: Efficient KV Cache Compression Through Retrieval Heads | {
"content": "## Abstract\n\nAbstract The memory and computational demands of Key-Value (KV) cache present significant challenges for deploying long-context language models. Previous approaches attempt to mitigate this issue by selectively dropping tokens, which irreversibly erases critical information that might be ... | [
{
"id": "ZvHZVXehAS",
"initial_rating": 8,
"confidence": 5,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This work proposed a method advocating keeping a full KV cache for retrieval/induction heads while adopting a StreamingLLM-like cache pattern for the rests. The p... | {
"rating": "3;3;5;5;5;5",
"rating_avg": 4.333333333333333,
"confidence": "4;4;4;4;4;5",
"confidence_avg": 4.166666666666667,
"soundness": "2;3;3;3;2;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;2;2;3;2",
"contribution_avg": 2.1666666666666665,
"presentation": "3;2;3;2;3;3",
"presen... | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.916076"
} | {
"id": "XsJfmBCMTK",
"metareview": "The paper introduces RazorAttention, a novel KV cache compression algorithm that selectively retains full information in retrieval heads while condensing remote tokens in non-retrieval heads using a compensation token. This method achieves over 70% reduction in KV cache size wit... | {
"decision": "Accept (Poster)"
} |
tkqNDbukWW | 2410.18860v1 | DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations | {
"content": "## Abstract\n\nAbstract Large Language Models (LLMs) often hallucinate, producing unfaithful or factually incorrect outputs by misrepresenting the provided context or incorrectly recalling internal knowledge.\nRecent studies have identified specific attention heads within the Transformer architecture, k... | [
{
"id": "v0nhF1VUXE",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "The paper addresses the issue of hallucinations in Large Language Models (LLMs), where outputs may be unfaithful or factually incorrect. It identifies specific at... | {
"rating": "5;5;5;6",
"rating_avg": 5.25,
"confidence": "4;4;3;3",
"confidence_avg": 3.5,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "3;3;2;3",
"contribution_avg": 2.75,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.917108"
} | {
"id": "KUPkYW4ePH",
"metareview": "The paper introduces a variant of contrastive decoding to mitigate hallucinations in language models. The proposed DeCoRe identifies specific retrieval heads that influence the generation of grounded content and uses a masking strategy to induce hallucinations. It then applies c... | {
"decision": "Reject"
} |
tmSWFGpBb8 | 2303.17813v3 | Learning the Complexity of Weakly Noisy Quantum States | {
"content": "## Abstract\n\nAbstract Quantum computers capable of fault-tolerant operation are expected to provide provable advantages over classical computational models. However, the question of whether quantum advantages exist in the noisy intermediate-scale quantum era remains a fundamental and challenging probl... | [
{
"id": "q2jjCwO8IM",
"initial_rating": 6,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 3,
"summary": "This paper investigates a fundamental question in quantum state complexity: how to predict/learn the complexity of weakly noisy quantum states. The authors develo... | {
"rating": "3;5;6",
"rating_avg": 4.666666666666667,
"confidence": "2;3;4",
"confidence_avg": 3,
"soundness": "2;3;2",
"soundness_avg": 2.3333333333333335,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "2;2;3",
"presentation_avg": 2.3333333333333335
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.918323"
} | {
"id": "rYnnqKpzaA",
"metareview": "Summary:\nThis paper introduces an efficient quantum learning algorithm for predicting the complexity of weakly noisy quantum states. The authors develop a method that leverages classical shadow representation of quantum states and provides optimal sample complexity with polynom... | {
"decision": "Accept (Poster)"
} |
tn2mjzjSyR | 2410.03864v1 | DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search | {
"content": "## Abstract\n\nAbstract Enhancing the capability of large language models (LLMs) in reasoning has gained significant attention in recent years. Previous studies have demonstrated the effectiveness of various prompting strategies in aiding LLMs in reasoning (called “reasoning actions”), such as step-by-s... | [
{
"id": "IeBTpUwhin",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper proposes DOTS, an approach enabling LLMs to reason dynamically via optimal reasoning trajectories search, tailored to the specific characteristics of e... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "4;3;4;4",
"confidence_avg": 3.75,
"soundness": "4;3;3;3",
"soundness_avg": 3.25,
"contribution": "2;3;2;3",
"contribution_avg": 2.5,
"presentation": "4;3;2;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.919353"
} | {
"id": "TAz4Wnct7n",
"metareview": "This paper proposes to enhance LLM reasoning via dynamically tailoring trajectory search. The proposed method should be generally applicable. The experiments show that its inference cost could potentially be lower than non-CoT methods. The proposed method is intuitive. However, ... | {
"decision": "Accept (Poster)"
} |
to4PdiiILF | 2410.06491v1 | Honesty to Subterfuge: In-Context Reinforcement Learning Can Make Honest Models Reward Hack | {
"content": "## Abstract\n\nAbstract Previous work has shown that training “helpful-only” LLMs with reinforcement learning on a curriculum of gameable environments can lead models to generalize to egregious specification gaming, such as editing their own reward function or modifying task checklists to appear more su... | [
{
"id": "D0mIlOaydj",
"initial_rating": 3,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper directly builds upon prior work that shows that training on a curriculum of environments that encourage reward specification gaming can lead to general... | {
"rating": "3;3;3",
"rating_avg": 3,
"confidence": "4;4;3",
"confidence_avg": 3.6666666666666665,
"soundness": "2;2;3",
"soundness_avg": 2.3333333333333335,
"contribution": "2;2;2",
"contribution_avg": 2,
"presentation": "3;2;3",
"presentation_avg": 2.6666666666666665
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.920031"
} | {
"id": "eAnzVfBe3B",
"metareview": "The paper addresses the problem of specification gaming (especially, but not exclusively, reward hacking) in in-context reinforcement learning. The authors describe cases where the model edits its own reward function. The paper discusses the risks inherent in replying on current... | {
"decision": "Reject"
} |
tozlOEN4qp | 2410.14171v2 | Heavy-Tailed Diffusion Models | {
"content": "## Abstract\n\nAbstract Diffusion models achieve state-of-the-art generation quality across many applications, but their ability to capture rare or extreme events in heavy-tailed distributions remains unclear. In this work, we show that traditional diffusion and flow-matching models with standard Gaussi... | [
{
"id": "eUUax4RXdS",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The authors provide a framework, termed t-EDM, for diffusion models where the added noise is heavy-tailed, following a Student-t distribution. They build upon the... | {
"rating": "3;6;6;6;8;8",
"rating_avg": 6.166666666666667,
"confidence": "4;4;3;3;4;4",
"confidence_avg": 3.6666666666666665,
"soundness": "2;3;3;3;4;4",
"soundness_avg": 3.1666666666666665,
"contribution": "2;3;2;3;3;3",
"contribution_avg": 2.6666666666666665,
"presentation": "2;3;2;3;4;3",
"prese... | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.920995"
} | {
"id": "NKCimm2qyn",
"metareview": "This paper introduces t-EDM, a diffusion model framework that enhances the modeling of heavy-tailed distributions, particularly relevant for applications in extreme-weather prediction. Building upon the existing EDM framework, t-EDM replaces the Gaussian noise typically used in... | {
"decision": "Accept (Poster)"
} |
tpD1rs25Uu | 2409.10262v1 | Hydra-SGG: Hybrid Relation Assignment for One-stage Scene Graph Generation | {
"content": "## Abstract\n\nAbstract DETR introduces a simplified one-stage framework for scene graph generation (SGG). However, DETR-based SGG models face two challenges: i) Sparse supervision , as each image typically contains fewer than 10 relation annotations, while the models employ over 100 relation queries. T... | [
{
"id": "z0bBc1OgI3",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "The manuscript introduces Hydra-SGG, a novel one-stage Scene Graph Generation (SGG) method designed to address the challenges of sparse supervision and false nega... | {
"rating": "5;6;8",
"rating_avg": 6.333333333333333,
"confidence": "4;2;4",
"confidence_avg": 3.3333333333333335,
"soundness": "2;3;4",
"soundness_avg": 3,
"contribution": "2;2;4",
"contribution_avg": 2.6666666666666665,
"presentation": "3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.922411"
} | {
"id": "nuVYFpEBCc",
"metareview": "This paper introduces Hydra-SGG, a novel method for scene graph generation that addresses sparse supervision and false negatives by employing a hybrid relation assignment strategy and a Hydra branch decoder. These innovations contribute to improved training and competitive resul... | {
"decision": "Accept (Poster)"
} |
tpVQHb4pea | 2410.02229v1 | CodePMP: Scalable Preference Model Pretraining for Large Language Model Reasoning | {
"content": "## Abstract\n\nAbstract Large language models (LLMs) have made significant progress in natural language understanding and generation, driven by scalable pretraining and advanced finetuning. However, enhancing reasoning abilities in LLMs, particularly via reinforcement learning from human feedback (RLHF)... | [
{
"id": "AeBdh62FKP",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "The paper presents CodePMP, a novel and scalable preference model pretraining (PMP) pipeline aimed at enhancing the reasoning capabilities of Large Language Model... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "4;4;3;3",
"confidence_avg": 3.5,
"soundness": "2;2;2;2",
"soundness_avg": 2,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;3;3;2",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.923072"
} | {
"id": "IuYolw71ej",
"metareview": "This paper introduces CodePMP, a scalable preference model pretraining pipeline that leverages synthesized code-preference pairs from publicly available, high-quality source code to improve reward model fine-tuning. I appreciate the authors added additional results during the re... | {
"decision": "Reject"
} |
tpqMR73GzS | 2409.18768v2 | Learning from Demonstration with Implicit Nonlinear Dynamics Models | {
"content": "## Abstract\n\nAbstract Learning from Demonstration (LfD) is a useful paradigm for training policies that solve tasks involving complex motions, such as those encountered in robotic manipulation. In practice, the successful application of LfD requires overcoming error accumulation during policy executio... | [
{
"id": "ScsvvlwqDy",
"initial_rating": 3,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "The paper proposes a method for incorporating non-linear dynamic systems in a policy representation for learning from demonstration (LfD). The approach extends th... | {
"rating": "3;3;5;6",
"rating_avg": 4.25,
"confidence": "4;4;2;2",
"confidence_avg": 3,
"soundness": "1;3;2;3",
"soundness_avg": 2.25,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.924043"
} | {
"id": "lHb41wBiRu",
"metareview": "This paper proposes a method that integrates fixed non linear dynamical systems with learned embeddings to address compounding errors in learning from demonstrations. Experiments on the LASA dataset shows improved spatial performance, robustness to noise and lower latency compar... | {
"decision": "Reject"
} |
trKNi4IUiP | 2406.09836v1 | Robustness Inspired Graph Backdoor Defense | {
"content": "## Abstract\n\nAbstract Graph Neural Networks (GNNs) have achieved promising results in tasks such as node classification and graph classification. However, recent studies reveal that GNNs are vulnerable to backdoor attacks, posing a significant threat to their real-world adoption. Despite initial effor... | [
{
"id": "YU1OwRFrWF",
"initial_rating": 8,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper proposes a novel defense method against graph backdoor attacks, which is composed of poisoned nodes detection and robust training. In poisoned node det... | {
"rating": "5;6;6;8",
"rating_avg": 6.25,
"confidence": "3;3;4;2",
"confidence_avg": 3,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Oral",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.924804"
} | {
"id": "S03XxtW7Ny",
"metareview": "This paper presents a defense method against graph backdoor attacks, combining poisoned node detection and robust training. It observes that edge dropping significantly affects the prediction of poisoned nodes with theoretical verification. The method uses random edge dropping t... | {
"decision": "Accept (Oral)"
} |
trKee5pIFv | 2410.04203v1 | RainbowPO: A Unified Framework for Combining Improvements in Preference Optimization | {
"content": "## Abstract\n\nAbstract Recently, numerous preference optimization algorithms have been introduced as extensions to the Direct Preference Optimization (DPO) family. While these methods have successfully aligned models with human preferences, there is a lack of understanding regarding the contributions o... | [
{
"id": "CaGfgSKCUj",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This paper aims to provide a unified framework for direct preference optimization (DPO) methods which are becoming a popular and successful method to train or fin... | {
"rating": "5;5;5;6",
"rating_avg": 5.25,
"confidence": "4;2;3;4",
"confidence_avg": 3.25,
"soundness": "3;2;2;3",
"soundness_avg": 2.5,
"contribution": "3;2;3;3",
"contribution_avg": 2.75,
"presentation": "2;3;2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.925587"
} | {
"id": "HQm0dLPQTi",
"metareview": "This work reviews several improvements over DPO and identifies \"orthogonal\" complementary components whose contribution to task alignment has not been thoroighly examined. Their proposed training procedure, RainbowDPO, combines aspects of multiple approaches, which can be par... | {
"decision": "Accept (Poster)"
} |
trj2Jq8riA | 2409.09369v3 | Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology | {
"content": "## Abstract\n\nAbstract Histopathology Whole-Slide Images (WSIs) provide an important tool to assess cancer prognosis in computational pathology (CPATH).\nWhile existing survival analysis (SA) approaches have made exciting progress, they are generally limited to adopting highly-expressive architectures ... | [
{
"id": "OzBwg4fP7z",
"initial_rating": 3,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper introduces a Vision-Language-based Survival Analysis (VLSA) paradigm to improve cancer prognosis prediction from histopathology whole-slide images by l... | {
"rating": "3;5;5",
"rating_avg": 4.333333333333333,
"confidence": "3;5;4",
"confidence_avg": 4,
"soundness": "3;2;2",
"soundness_avg": 2.3333333333333335,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "3;3;2",
"presentation_avg": 2.6666666666666665
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.926350"
} | {
"id": "FQXT8GZwQb",
"metareview": "First, there appears to be a misalignment between your paper's stated contribution and its actual focus.\nSecond, critical methodological choices lack sufficient justification. \nThird, the experimental validation needs substantial strengthening. \nFourth, the ordinal survival p... | {
"decision": "Accept (Poster)"
} |
tsfR7JCwTf | 2403.11981v1 | Certified Robustness to Clean-label Poisoning Using Diffusion Denoising | {
"content": "## Abstract\n\nAbstract We present a certified defense\nto clean-label poisoning attacks.\nThese attacks work by injecting\na small number of poisoning samples (e.g., 1%)\nthat contain p 𝑝 p italic_p -norm bounded adversarial perturbations\ninto the training data\nto induce a targeted misclassification... | [
{
"id": "HttUiQniun",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper introduces a defense mechanism against clean-label poisoning attacks by denoising training samples with a diffusion model prior to training. The goal i... | {
"rating": "3;5;5;5",
"rating_avg": 4.5,
"confidence": "5;3;5;4",
"confidence_avg": 4.25,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "2;3;2;2",
"contribution_avg": 2.25,
"presentation": "2;3;2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.927267"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
tu3qwNjrtw | 2407.06483v1 | Composable Interventions for Language Models | {
"content": "## Abstract\n\nAbstract Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining.\nBut despite a flood of new methods, different types of interventions are largely developing independently.\nIn practice, mu... | [
{
"id": "Aw2LhbXoEe",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper experimentally demonstrates the effect of composing various inference time setups for LLMs and suggests possible setups for better performance from its... | {
"rating": "3;6;6;6;6",
"rating_avg": 5.4,
"confidence": "4;4;4;3;4",
"confidence_avg": 3.8,
"soundness": "2;3;2;2;2",
"soundness_avg": 2.2,
"contribution": "2;3;3;3;2",
"contribution_avg": 2.6,
"presentation": "3;3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.928147"
} | {
"id": "1Y5KpvalaK",
"metareview": "This paper proposes a new framework to assess the effects of combining different test-time interventions on language models to improve factual accuracy, efficiency, and mitigate harmful outputs. The authors explore how these interventions interact when used together, analyzing t... | {
"decision": "Accept (Poster)"
} |
tuu4de7HL1 | 2406.14337v1 | Improving Convergence Guarantees of Random Subspace Second-order Algorithm for Nonconvex Optimization | {
"content": "## Abstract\n\nAbstract We propose the Random Subspace Homogenized Trust Region (RSHTR) method,\nwhich efficiently solves high-dimensional non-convex optimization\nproblems by identifying descent directions within randomly selected\nsubspaces. RSHTR provides the strongest theoretical guarantees among ra... | [
{
"id": "XLMTUtG9u7",
"initial_rating": 5,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This paper is primarily inspired by the HSODM method developed by Zhang et al. (2022) and extends their ideas to the domain of random subspace methods. The author... | {
"rating": "5;6;8",
"rating_avg": 6.333333333333333,
"confidence": "2;3;2",
"confidence_avg": 2.3333333333333335,
"soundness": "3;3;3",
"soundness_avg": 3,
"contribution": "3;3;3",
"contribution_avg": 3,
"presentation": "2;3;3",
"presentation_avg": 2.6666666666666665
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.929114"
} | {
"id": "7py8v1slCn",
"metareview": "This paper presents a random subspace trust-region-based method for addressing nonconvex, unconstrained optimization problems. The paper is well-structured and makes a clear contribution to the nonconvex optimization literature. The authors demonstrate, for the first time, local... | {
"decision": "Accept (Spotlight)"
} |
tvQNysCP7C | 2410.12876v2 | In-context KV-Cache Eviction for LLMs via Attention-Gate | {
"content": "## Abstract\n\nAbstract The KV-Cache technique has become the standard for the inference of large language models (LLMs).\nIt caches states of self-attention to avoid recomputation.\nYet, it is widely criticized that KV-Cache can become a bottleneck of the LLM inference system, especially when confronte... | [
{
"id": "FvUOkEaE4i",
"initial_rating": 3,
"confidence": 5,
"soundness": 2,
"contribution": 1,
"presentation": 3,
"summary": "This paper aims to reduce the memory overhead of the KV cache in LLM inference. The high-level goal is to ``evict’ ’certain tokens from the attention computation ... | {
"rating": "3;3;3;5;5",
"rating_avg": 3.8,
"confidence": "5;4;5;4;4",
"confidence_avg": 4.4,
"soundness": "2;3;2;2;2",
"soundness_avg": 2.2,
"contribution": "2;2;1;3;3",
"contribution_avg": 2.2,
"presentation": "2;3;3;3;3",
"presentation_avg": 2.8
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.930025"
} | {
"id": "ksiEkTbYkf",
"metareview": "This paper seeks to address the issue of kv-cache sizes growing fast with the size of models and input context. They do so by learning a model component that tries to determine which tokens in the past kv states need to be stored.\n\nWhile the approach is simple and the eviction... | {
"decision": "Reject"
} |
twIPSx9qHn | 2410.02479v1 | Cross-Embodiment Dexterous Grasping with Reinforcement Learning | {
"content": "## Abstract\n\nAbstract Dexterous hands exhibit significant potential for complex real-world grasping tasks. While recent studies have primarily focused on learning policies for specific robotic hands, the development of a universal policy that controls diverse dexterous hands remains largely unexplored... | [
{
"id": "oHRanXQAom",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 4,
"summary": "This paper learns a unified vision-based reinforcement learning policy via teacher-student learning to control various dexterous hands for grasping different obje... | {
"rating": "3;5;6",
"rating_avg": 4.666666666666667,
"confidence": "3;4;4",
"confidence_avg": 3.6666666666666665,
"soundness": "2;2;3",
"soundness_avg": 2.3333333333333335,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "2;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.930640"
} | {
"id": "B9N6H9Dy4g",
"metareview": "The authors study the important problem of cross-platform dexterous grasping. Their goal is to learn a grasping policy via RL that works on multiple hands, which they propose to do using a unique universal action space based on the human hand eigengrasps. These can be predicted ... | {
"decision": "Accept (Poster)"
} |
twtTLZnG0B | 2311.05589v1 | A Coefficient Makes SVRG Effective | {
"content": "## Abstract\n\nAbstract Stochastic Variance Reduced Gradient (SVRG), introduced by Johnson & Zhang ( 2013 ) , is a theoretically compelling optimization method. However, as Defazio & Bottou ( 2019 ) highlights, its effectiveness in deep learning is yet to be proven. In this work, we demonstrate the pote... | [
{
"id": "AjdnUrW5QH",
"initial_rating": 6,
"confidence": 5,
"soundness": 4,
"contribution": 3,
"presentation": 4,
"summary": "This paper revisit SVRG in deep learning, showing how reasoning more carefully about control variates. The paper starts by a nice recap of what is believed to be ... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "4;4;4;5",
"confidence_avg": 4.25,
"soundness": "3;3;3;4",
"soundness_avg": 3.25,
"contribution": "2;2;2;2",
"contribution_avg": 2,
"presentation": "3;4;3;4",
"presentation_avg": 3.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.931316"
} | {
"id": "DEE4Qe6zvU",
"metareview": "The paper introduces α-SVRG, a variant of SVRG that integrates a coefficient to dynamically control the strength of variance reduction, decaying it over time. The reviewers appreciated the novelty of this approach and its demonstrated empirical effectiveness on various deep-lear... | {
"decision": "Accept (Poster)"
} |
txD9llAYn9 | 2408.08994v3 | Model-based RL as a Minimalist Approach to Horizon-Free and Second-Order Bounds | {
"content": "## Abstract\n\nAbstract Learning a transition model via Maximum Likelihood Estimation (MLE) followed by planning inside the learned model is perhaps the most standard and simplest Model-based Reinforcement Learning (RL) framework. In this work, we show that such a simple Model-based RL scheme, when equi... | [
{
"id": "UpeVLi1xp5",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This paper proposes a minimalist approach to Model-Based Reinforcement Learning (RL) that leverages Maximum Likelihood Estimation (MLE) for learning transition mo... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "3;4;3;4",
"confidence_avg": 3.5,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "3;2;3;3",
"contribution_avg": 2.75,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.932553"
} | {
"id": "NbyFDrm43S",
"metareview": "Model-based RL as a Minimalist Approach to Horizon-Free and Second-Order Bounds \n\nSummary: The paper introduces a minimalist approach to model-based reinforcement learning (MBRL) that combines Maximum Likelihood Estimation (MLE) for transition modeling with optimistic and pess... | {
"decision": "Accept (Poster)"
} |
txV4dNeusx | 2410.06266v1 | Near-Exact Privacy Amplification for Matrix Mechanisms | {
"content": "## Abstract\n\nAbstract We study the problem of computing the privacy parameters for DP machine learning when using privacy amplification via random batching and noise correlated across rounds via a correlation matrix 𝐂 𝐂 \\mathbf{C} bold_C (i.e., the matrix mechanism). Past work on this problem eithe... | [
{
"id": "6iQisobYLN",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper investigates the computation of privacy parameters for differentially private (DP) machine learning, specifically when using privacy amplification thro... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "4;3;3;3",
"confidence_avg": 3.25,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.934237"
} | {
"id": "zQEP1xWgdb",
"metareview": "Reviewers agreed that the paper provides a novel and practical sampling scheme for DP-SGD, and appreciated the good empirical performance. Some technical concerns were addressed during the rebuttal phase. A common concern is that the paper lacks theoretical analysis/guarantee, t... | {
"decision": "Accept (Poster)"
} |
tyFGIjNzlj | 2407.04899v1 | Algorithmic Language Models with Neurally Compiled Libraries | {
"content": "## Abstract\n\nAbstract Important tasks such as reasoning and planning are fundamentally algorithmic,\nmeaning that solving them robustly requires acquiring true reasoning or planning algorithms, rather than shortcuts.\nLarge Language Models lack true algorithmic ability primarily because of the limitat... | [
{
"id": "XcX1Y6WV9k",
"initial_rating": 3,
"confidence": 3,
"soundness": 1,
"contribution": 2,
"presentation": 2,
"summary": "The paper addresses LLMs’ problem with performing symbolic operations. To this end, they investigate one way to incorporate a differentiable interpreter into LLMs... | {
"rating": "3;3;3;6",
"rating_avg": 3.75,
"confidence": "3;5;3;3",
"confidence_avg": 3.5,
"soundness": "2;3;1;2",
"soundness_avg": 2,
"contribution": "1;1;2;3",
"contribution_avg": 1.75,
"presentation": "1;3;2;2",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.934956"
} | {
"id": "2WLXGm63sU",
"metareview": "This is a very interesting paper leveraging pre-compiled differentiable neural libraries as part of a language model architecture. In my opinion, this is definitely an idea worth pursuing further. However, it is also evident that the paper's evaluation is still preliminary, with... | {
"decision": "Reject"
} |
u1EPPYkbgA | 2312.06315v1 | GPTBIAS: A Comprehensive Framework for Evaluating Bias in Large Language Models | {
"content": "## Abstract\n\nAbstract Warning : This paper contains content that may be offensive or upsetting. There has been a significant increase in the usage of large language models (LLMs) in various applications, both in their original form and through fine-tuned adaptations.\nAs a result, LLMs have gained pop... | [
{
"id": "IvAqnzBEyG",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "The paper introduces a bias evaluation framework that leverages advanced LLMs, such as GPT-4, to assess bias across nine distinct bias types. The framework, GPTBI... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "2;2;2;2",
"soundness_avg": 2,
"contribution": "2;2;1;2",
"contribution_avg": 1.75,
"presentation": "2;2;4;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.935499"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
u3TL0qxLWf | 2410.10714v2 | SeedLM: Compressing LLM Weights into Seeds of Pseudo-Random Generators | {
"content": "## Abstract\n\nAbstract Large Language Models (LLMs) have transformed natural language processing, but face significant challenges in widespread deployment due to their high runtime cost. In this paper, we introduce SeedLM, a novel post-training compression method that uses seeds of pseudo-random genera... | [
{
"id": "LxwVSFoxXf",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 4,
"summary": "This is an interesting method of quantization, using pseudo-random generator to point to almost evenly distributed codebook items and fast adjustments.\nThe paper... | {
"rating": "5;6;6",
"rating_avg": 5.666666666666667,
"confidence": "4;5;3",
"confidence_avg": 4,
"soundness": "3;3;3",
"soundness_avg": 3,
"contribution": "2;3;3",
"contribution_avg": 2.6666666666666665,
"presentation": "2;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.936079"
} | {
"id": "W5dEqYEicv",
"metareview": "This paper presents SeedLM, a novel post-training compression method for LLMs that finds input seeds from model weights for pseudo-random generators, specifically using Linear Feedback Shift Registers (LFSRs). By reconstructing weights at runtime from compact seeds, SeedLM signi... | {
"decision": "Accept (Poster)"
} |
u63OVngeSp | 2405.18314v2 | Deriving Causal Order from Single-Variable Interventions: Guarantees & Algorithm | {
"content": "## Abstract\n\nAbstract Targeted and uniform interventions to a system are crucial for unveiling causal relationships. While several methods have been developed to leverage interventional data for causal structure learning, their practical application in real-world scenarios often remains challenging. R... | [
{
"id": "5GbnAwgIKo",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "The authors considered the problem of recovering a causal order in structural causal models (SCM) under the causal sufficiency assumption (it seems that is the ca... | {
"rating": "3;5;6;8;8",
"rating_avg": 6,
"confidence": "4;3;3;3;4",
"confidence_avg": 3.4,
"soundness": "2;2;3;4;3",
"soundness_avg": 2.8,
"contribution": "2;2;3;3;4",
"contribution_avg": 2.8,
"presentation": "3;3;3;3;4",
"presentation_avg": 3.2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.937485"
} | {
"id": "tzwfvifgF0",
"metareview": "The paper introduces an approach for learning the order of a causal DAG model from single variable interventions. Reviewers found it useful, and I'm generally favourable to accept based on the framing of the algorithms being novel and for providing results others may want to bui... | {
"decision": "Accept (Poster)"
} |
u6Y0GdTEYp | 2402.06568v1 | Constrained Multi-Objective Optimization | {
"content": "## Abstract\n\nAbstract Over the last decade, developments in Unmanned Aerial Vehicles (UAVs) has greatly increased, and they are being used in many fields including surveillance, crisis management or automated mission planning. This last field implies the search of plans for missions with multiple task... | [
{
"id": "x5oPxsS6iF",
"initial_rating": 3,
"confidence": 3,
"soundness": 1,
"contribution": 2,
"presentation": 1,
"summary": "This work proposes a gradient-based optimization algorithm, MLM-CMOO, to solve constrained multi-objective optimization (CMOO) problems. The authors conduct a con... | {
"rating": "1;3;3;3",
"rating_avg": 2.5,
"confidence": "4;4;3;3",
"confidence_avg": 3.5,
"soundness": "2;2;3;1",
"soundness_avg": 2,
"contribution": "1;1;2;2",
"contribution_avg": 1.5,
"presentation": "2;2;2;1",
"presentation_avg": 1.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.938305"
} | {
"id": "4bkm7XH4Nd",
"metareview": "The paper proposes a new gradient-based method, MLM-CMOO, for constrained multi-objective optimization (CMOO). It offers a convergence guarantee to Pareto stationary solutions with a rate of O(1/sqrt(T)). However, the paper's title is misleading as it focuses solely on a single ... | {
"decision": "Reject"
} |
u8SYRtXDsZ | 2408.01708v1 | AVESFormer: Efficient Transformer Design for Real-Time Audio-Visual Segmentation | {
"content": "## Abstract\n\nAbstract Recently, transformer-based models have demonstrated remarkable performance on audio-visual segmentation (AVS) tasks.\nHowever, their expensive computational cost makes real-time inference impractical.\nBy characterizing attention maps of the network, we identify two key obstacle... | [
{
"id": "MF1biY1moJ",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper finds two primary challenges in existing audio-visual segmentation models, namely attention dissipation caused by anomalous attention weights after Sof... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "4;4;5;4",
"confidence_avg": 4.25,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "2;2;3;3",
"contribution_avg": 2.5,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.939030"
} | {
"id": "83QHV3jIwa",
"metareview": "This work received three negative ratings (borderline reject) and one positive rating (borderline accept). The reviewers rate negative scores due to the limited novelty, lack of high-level explanation of the motivation and contradictory experimental results raised by reviewer jh... | {
"decision": "Reject"
} |
u8VOQVzduP | 2405.14744v2 | Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View | {
"content": "## Abstract\n\nAbstract Large language models (LLMs) have been shown to face hallucination issues due to the data they trained on often containing human bias; whether this is reflected in the decision-making process of LLM agents remains under-explored. As LLM Agents are increasingly employed in intrica... | [
{
"id": "aDMntMRkEa",
"initial_rating": 6,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper introduces CogMir, a novel open-ended framework designed to evaluate and interpret social intelligence in Large Language Model (LLM) agents through cog... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "4;2;3;3",
"confidence_avg": 3,
"soundness": "2;3;3;2",
"soundness_avg": 2.5,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;3;2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.939734"
} | {
"id": "FyOBDHhW49",
"metareview": "This paper proposed a benchmark called CogMir for evaluating cognitive biases in LLM agents. While reviewers had concerns, especially with the connection to theoretical lines of work in social and cognitive science, these were not insurmountable, and the reviewers' response to t... | {
"decision": "Accept (Poster)"
} |
uAFHCZRmXk | 2404.07983v2 | Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models | {
"content": "## Abstract\n\nAbstract Contrastive vision-language models (VLMs), like CLIP, have gained popularity for their versatile applicability to various downstream tasks. Despite their successes in some tasks, like zero-shot object recognition, they perform surprisingly poor on other tasks, like attribute reco... | [
{
"id": "MyCbWVhJg0",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper presents a comprehensive study of contrastive VLMs, particularly examining why they excel at object recognition but struggle with attribute detection. ... | {
"rating": "5;6;8;8",
"rating_avg": 6.75,
"confidence": "4;3;3;4",
"confidence_avg": 3.5,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "3;3;4;3",
"contribution_avg": 3.25,
"presentation": "1;3;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Oral",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.940547"
} | {
"id": "LkLe4cVXns",
"metareview": "This paper studies the phenomena of modality gap and object bias in contrastive VLMs, and shows that they stem from an information imbalance between modalities, limiting alignment in the embedding space, with the modality gap driven by few dimensions, linked to higher logit entr... | {
"decision": "Accept (Oral)"
} |
uBai0ukstY | 2410.04209v1 | Equivariant Neural Functional Networks for Transformers | {
"content": "## Abstract\n\nAbstract This paper systematically explores neural functional networks (NFN) for transformer architectures. NFN are specialized neural networks that treat the weights, gradients, or sparsity patterns of a deep neural network (DNN) as input data and have proven valuable for tasks such as l... | [
{
"id": "14HGfs6tlo",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper introduces a novel concept of Transformer Neural Functional Networks (Transformer-NFN), designed as equivariant neural functional networks specifically ... | {
"rating": "6;6;8;8",
"rating_avg": 7,
"confidence": "2;4;2;4",
"confidence_avg": 3,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "3;3;3;3",
"contribution_avg": 3,
"presentation": "2;3;3;2",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.942109"
} | {
"id": "PsxvXkKmxM",
"metareview": "The paper proposes an algorithm for equivariant transformer architectures of neural fields. Reviewers are unanimously in favor of accepting the paper, and initially had questions about the computational cost, as well as the significance of improvements. Authors answered with add... | {
"decision": "Accept (Poster)"
} |
uBcx1aFpXy | 2410.04932v1 | OmniBooth: Learning Latent Control for Image Synthesis with Multi-modal Instruction | {
"content": "## Abstract\n\nAbstract We present OmniBooth , an image generation framework that enables spatial control with instance-level multi-modal customization. For all instances, the multi-modal instruction can be described through text prompts or image references. Given a set of user-defined masks and associa... | [
{
"id": "yKpN1fpwi9",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "The paper introduces OmniBooth, a novel framework for image generation that leverages multimodal instructions to enable spatial and instance-level control. The me... | {
"rating": "5;6;6;6",
"rating_avg": 5.75,
"confidence": "4;5;4;2",
"confidence_avg": 3.75,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.948215"
} | {
"id": "aUj78epK6N",
"metareview": "This paper addresses the problem of generating images with precise instance-level spatial control using multi-modal instructions. The proposed approach combines semantic information (from text or image references) and spatial information (from segmentation masks) into a unified ... | {
"decision": "Reject"
} |
uBnM3EFovQ | 2406.14393v3 | Jailbreaking as a Reward Misspecification Problem | {
"content": "## Abstract\n\nAbstract The widespread adoption of large language models (LLMs) has raised concerns about their safety and reliability, particularly regarding their vulnerability to adversarial attacks. In this paper, we propose a novel perspective that attributes this vulnerability to reward misspecifi... | [
{
"id": "RPV8XpyA7n",
"initial_rating": 3,
"confidence": 3,
"soundness": 2,
"contribution": 3,
"presentation": 2,
"summary": "The authors propose that vulnerabilities in large language models (LLMs) stem from reward misspecification during the alignment process. They introduce a metric, ... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "3;4;2;3",
"confidence_avg": 3,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "3;2;3;3",
"contribution_avg": 2.75,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.948967"
} | {
"id": "mP59vqgAn3",
"metareview": "The recommendation is based on the reviewers' comments, the area chair's evaluation, and the author-reviewer discussion. \n\nThis paper studies the reward misspecification problem for LLMs and use the results to propose an automated red-teaming approach. All reviewers find the s... | {
"decision": "Accept (Poster)"
} |
uClUUJk05H | 2411.02728v1 | Compositional simulation-based inference for time series | {
"content": "## Abstract\n\nabstract Amortized simulation-based inference (SBI) methods train neural networks on simulated data to perform Bayesian inference. While this approach avoids the need for tractable likelihoods, it often requires a large number of simulations and has been challenging to scale to time-serie... | [
{
"id": "KgLgnv38ci",
"initial_rating": 5,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes a simulation-based inference (SBI) method for state-space models where the transition dynamics is Markovian. The core idea is to simulate many... | {
"rating": "3;3;5;5;6;8",
"rating_avg": 5,
"confidence": "4;5;2;3;4;4",
"confidence_avg": 3.6666666666666665,
"soundness": "3;2;3;2;3;4",
"soundness_avg": 2.8333333333333335,
"contribution": "2;2;2;2;2;3",
"contribution_avg": 2.1666666666666665,
"presentation": "3;2;2;2;2;4",
"presentation_avg": 2.... | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.949763"
} | {
"id": "mS8PjOxSih",
"metareview": "The reviewers recommend acceptance (6-5-8-6-6-6). The paper presents a simulation-based inference approach for Markovian state-space models, leveraging the structure of the forward model to reduce the computational cost of the inference. The approach is well-motivated and the re... | {
"decision": "Accept (Poster)"
} |
uCqxDfLYrB | 2410.12360v1 | Towards Neural Scaling Laws for Time Series Foundation Models | {
"content": "## Abstract\n\nAbstract Scaling laws offer valuable insights into the design of time series foundation models (TSFMs). However, previous research has largely focused on the scaling laws of TSFMs for in-distribution (ID) data, leaving their out-of-distribution (OOD) scaling behavior and the influence of ... | [
{
"id": "dDxv0xN62P",
"initial_rating": 5,
"confidence": 5,
"soundness": 3,
"contribution": 4,
"presentation": 3,
"summary": "The paper analyzes two common TSFM architectures -- encoder-only and decoder-only transformers -- in terms of both in-distribution and out-of-distribution data. T... | {
"rating": "5;5;6;8",
"rating_avg": 6,
"confidence": "4;5;3;4",
"confidence_avg": 4,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "3;4;3;3",
"contribution_avg": 3.25,
"presentation": "4;3;3;4",
"presentation_avg": 3.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.950609"
} | {
"id": "zYJ1elTdGu",
"metareview": "(a) Summary of Scientific Claims and Findings\nThis paper investigates scaling laws for Time Series Foundation Models (TSFMs), focusing on both in-distribution (ID) and out-of-distribution (OOD) settings. The key contributions include:\nEmpirical Findings on Scaling Laws: The st... | {
"decision": "Accept (Poster)"
} |
uDZ9d4UAUh | 2406.10834v1 | Exposing the Achilles' Heel: Evaluating LLMs Ability to Handle Mistakes in Mathematical Reasoning | {
"content": "## Abstract\n\nAbstract Large Language Models (LLMs) have been applied to Math Word Problems (MWPs) with transformative impacts, revolutionizing how these complex problems are approached and solved in various domains including educational settings. However, the evaluation of these models often prioritiz... | [
{
"id": "6s6qEy4dlX",
"initial_rating": 5,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes a dataset for evaluating LLMs’ abilities in detecting and correcting mistakes. The dataset contains math word problems with both correct and w... | {
"rating": "3;5;5;5",
"rating_avg": 4.5,
"confidence": "4;5;4;3",
"confidence_avg": 4,
"soundness": "2;3;3;2",
"soundness_avg": 2.5,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;3;2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.951770"
} | {
"id": "jMQ98PpAfz",
"metareview": "The paper presents a dataset for evaluating LLMs’ abilities in detecting and correcting mathematical reasoning mistakes. The dataset contains math word problems with both correct and wrong reasoning chains. The proposed pipeline includes perturbations to the original reasoning c... | {
"decision": "Reject"
} |
uDjuCpQH5N | 2410.08827v2 | Do Unlearning Methods Remove Information from Language Model Weights? | {
"content": "## Abstract\n\nAbstract Large Language Models’ knowledge of how to perform cyber-security attacks, create bioweapons, and manipulate humans poses risks of misuse. Previous work has proposed methods to unlearn this knowledge. Historically, it has been unclear whether unlearning techniques are removing in... | [
{
"id": "Qir2JL4Nzy",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper addresses the challenge of unlearning in LLMs and investigates whether unlearning methods effectively remove knowledge or simply make it more difficult... | {
"rating": "3;3;6;8",
"rating_avg": 5,
"confidence": "5;4;5;4",
"confidence_avg": 4.5,
"soundness": "2;2;4;4",
"soundness_avg": 3,
"contribution": "1;2;2;4",
"contribution_avg": 2.25,
"presentation": "3;2;4;4",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.952474"
} | {
"id": "0iuSy6e9vh",
"metareview": "The paper shows that finetuning on a few unlearned facts can recover a large portion of unlearned knowledge.\nThis vulnerability had been demonstrated by prior work, and thus multiple reviewers questioned the novelty of this paper.\nWhile this paper does evaluate this attack vec... | {
"decision": "Reject"
} |
uE84MGbKD7 | 2411.07127v1 | Benchmarking LLMs' Judgments with No Gold Standard | {
"content": "## Abstract\n\nAbstract. We introduce the GEM (Generative Estimator for Mutual Information), an evaluation metric for assessing language generation by Large Language Models (LLMs), particularly in generating informative judgments, without the need for a gold standard reference. GEM broadens the scenario... | [
{
"id": "zBLr573tyf",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper introduces GEM, a novel metric for evaluating the performance of large language models in generating informative judgments using given references but no... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "3;3;4;4",
"confidence_avg": 3.5,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "3;3;3;3",
"contribution_avg": 3,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.953220"
} | {
"id": "i385HcZkhV",
"metareview": "The paper introduces GEM, a metric for evaluating large language models (LLMs) based on the informativeness of their generated responses without needing gold-standard references. GEM estimates mutual information between candidate and reference responses, making it ideal for subj... | {
"decision": "Accept (Poster)"
} |
uEPRY2XAEs | 2410.13835v2 | Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMs | {
"content": "## Abstract\n\nAbstract Practitioners have consistently observed three puzzling phenomena in transformer-based large language models (LLMs): attention sinks , value-state drains , and residual-state peaks , collectively referred to as extreme-token phenomena . These phenomena are characterized by certai... | [
{
"id": "Es2useBInG",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "The paper investigates the underlying mechanisms of \"extreme-token phenomena\" in large language models (LLMs), characterized by three main effects: attention si... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "3;3;3;3",
"confidence_avg": 3,
"soundness": "2;3;3;2",
"soundness_avg": 2.5,
"contribution": "2;2;3;3",
"contribution_avg": 2.5,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.954190"
} | {
"id": "LN6TgHsXwS",
"metareview": "This paper investigated the phenomena called \"extreme-token phenomena\" observed in transformer-based large language models that include attention sinks, value-state drains, and residual-state peaks. Based on a toy task with a small model, the authors identify an active-dormant... | {
"decision": "Reject"
} |
uHgVrGF2Wn | 2406.08035v2 | LVBench: An Extreme Long Video Understanding Benchmark | {
"content": "## Abstract\n\nAbstract Recent progress in multimodal large language models has markedly enhanced the understanding of short videos (typically under one minute), and several evaluation datasets have emerged accordingly. However, these advancements fall short of meeting the demands of real-world applicat... | [
{
"id": "syMZDxRKrm",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "In this paper, the authors introduce LVBench, a new MLLM benchmark designed to evaluate performance on long-term videos. LVBench comprises 103 high-quality videos... | {
"rating": "3;5;5;5",
"rating_avg": 4.5,
"confidence": "5;4;4;4",
"confidence_avg": 4.25,
"soundness": "3;3;3;2",
"soundness_avg": 2.75,
"contribution": "2;3;3;2",
"contribution_avg": 2.5,
"presentation": "4;3;2;2",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.955071"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
uIg9Vcw2CY | 2404.17789v3 | BiLO: Bilevel Local Operator Learning for PDE inverse problems | {
"content": "## Abstract\n\nAbstract We propose a new neural network based method for solving inverse problems for partial differential equations (PDEs) by formulating the PDE inverse problem as a bilevel optimization problem. At the upper level, we minimize the data loss with respect to the PDE parameters. At the l... | [
{
"id": "FhSa7EdCkh",
"initial_rating": 6,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "The paper investigates a method for parameter identification in PDEs (similar to optimal control). The idea is to replace the PDE constraint with novel functional... | {
"rating": "3;6;6;6",
"rating_avg": 5.25,
"confidence": "4;5;4;4",
"confidence_avg": 4.25,
"soundness": "2;3;3;2",
"soundness_avg": 2.5,
"contribution": "1;3;3;2",
"contribution_avg": 2.25,
"presentation": "3;4;2;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.955721"
} | {
"id": "jwTE0gXnnF",
"metareview": "This paper introduces a new approach for solving inverse problems for partial differential equations (PDEs). Specifically, the approach formulates the problem as a bilevel optimization problem. This efficient approach is also able to enforce PDE constraints. Results demonstrate... | {
"decision": "Reject"
} |
uL1H29dM0c | 2405.16305v2 | Efficiently Parameterized Neural Metriplectic Systems | {
"content": "## Abstract\n\nAbstract Metriplectic systems are learned from data in a way that scales quadratically in both the size of the state and the rank of the metriplectic data. Besides being provably energy conserving and entropy stable, the proposed approach comes with approximation results demonstrating its... | [
{
"id": "GbKRiotZf1",
"initial_rating": 5,
"confidence": 2,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "In this work, the authors present a neural network model (NMS) for learning the dynamics of metriplectic systems from trajectory data. It is based on a parameteri... | {
"rating": "5;5;6;8",
"rating_avg": 6,
"confidence": "4;2;2;4",
"confidence_avg": 3,
"soundness": "4;3;3;3",
"soundness_avg": 3.25,
"contribution": "3;2;3;3",
"contribution_avg": 2.75,
"presentation": "4;2;2;2",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.956530"
} | {
"id": "IEmxNAEtye",
"metareview": "This paper proposes a method for learning the dynamics of metriplectic systems from trajectory data. It hardwires structural information in the learning model using geometric techniques, so that the physical constraint is strictly satisfied. Reviewers and I found the method to b... | {
"decision": "Accept (Poster)"
} |
uLAAVg0ymc | 2402.03819v3 | Do we need rebalancing strategies? A theoretical and empirical study around SMOTE and its variants | {
"content": "## Abstract\n\nAbstract Synthetic Minority Oversampling Technique (SMOTE) is a common rebalancing strategy for handling imbalanced tabular data sets. However, few works analyze SMOTE theoretically. In this paper, we prove that SMOTE (with default parameter) tends to copy the original minority samples as... | [
{
"id": "hsHVuhcKeI",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper makes a theretical analysis of the well-known SMOTE method for imbalance classification, and proposes two simplest variants.",
"strengths": "This p... | {
"rating": "3;5;5;8",
"rating_avg": 5.25,
"confidence": "4;4;3;4",
"confidence_avg": 3.75,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "2;2;3;4",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.957543"
} | {
"id": "o4aw6BF7Jp",
"metareview": "The Authors discuss the problem of imbalance classification. Their focus is on analyzing one of the most popular oversampling strategies, known as SMOTE. They prove, for example, that SMOTE under the default parameters tends to copy the original minority samples asymptotically. ... | {
"decision": "Reject"
} |
uM2IDdivyC | 2406.04709v1 | ConDiff: A Challenging Dataset for Neural Solvers of Partial Differential Equations | {
"content": "## Abstract\n\nAbstract We present ConDiff, a novel dataset for scientific machine learning. ConDiff focuses on the diffusion equation with varying coefficients, a fundamental problem in many applications of parametric partial differential equations (PDEs). The main novelty of the proposed dataset is th... | [
{
"id": "Bv5ApYLP6X",
"initial_rating": 3,
"confidence": 3,
"soundness": 3,
"contribution": 1,
"presentation": 3,
"summary": "This paper proposes a challenging dataset as a benchmark for models that use neural networks as PDE solvers. The proposed dataset aims to be more representative o... | {
"rating": "3;3;3;3",
"rating_avg": 3,
"confidence": "4;4;3;3",
"confidence_avg": 3.5,
"soundness": "3;2;2;3",
"soundness_avg": 2.5,
"contribution": "1;1;2;1",
"contribution_avg": 1.25,
"presentation": "3;2;2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.958434"
} | {
"id": "VOgSCppGro",
"metareview": "This paper introduces ConDiff, a new benchmark dataset designed to evaluate machine learning models, particularly neural network-based PDE solvers like FNOs and DeepONets, in solving the 2D diffusion equation with spatially dependent, discontinuous coefficients. It provides a la... | {
"decision": "Reject"
} |
uMEsKEiB7J | 2403.12766v2 | NovelQA: Benchmarking Question Answering on Documents Exceeding 200K Tokens | {
"content": "## Abstract\n\nAbstract The rapid advancement of Large Language Models (LLMs) has introduced a new frontier in natural language processing, particularly in understanding and processing long-context information. However, the evaluation of these models’ long-context abilities remains a challenge due to th... | [
{
"id": "wSqc7ocvlh",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper introduces a new benchmark called NovelQA, which is designed to evaluate the performance of LLMs on extremely long and complex texts. NovelQA uses Engl... | {
"rating": "5;6;6;6;6",
"rating_avg": 5.8,
"confidence": "3;4;4;3;4",
"confidence_avg": 3.6,
"soundness": "3;3;3;3;3",
"soundness_avg": 3,
"contribution": "2;3;3;3;3",
"contribution_avg": 2.8,
"presentation": "3;3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.959070"
} | {
"id": "9tceZzcFCm",
"metareview": "The paper presents NovelQA, a benchmark of question answering where the context can be of length up to than 200k tokens. The benchmark is quite thorough, the paper well written and the reviewers are all positive wrt the paper. The paper also benchmarks many LLMs on this task w... | {
"decision": "Accept (Poster)"
} |
uMLeOlzlZ2 | 2401.17244v3 | LLaMP: Large Language Model Made Powerful for High-fidelity Materials Knowledge Retrieval | {
"content": "## Abstract\n\nAbstract Reducing hallucination of Large Language Models (LLMs) is imperative for use in the sciences where reproducibility is crucial. However, LLMs inherently lack long-term memory, making it a nontrivial, ad hoc , and inevitably biased task to fine-tune them on domain-specific literatu... | [
{
"id": "GAYBUJqaWM",
"initial_rating": 3,
"confidence": 3,
"soundness": 3,
"contribution": 1,
"presentation": 3,
"summary": "The manuscript describes the introduction and test of LLaMP, an LLM fine tuned to interact with the materials science data from the Materials Project database and... | {
"rating": "1;3;6;8",
"rating_avg": 4.5,
"confidence": "4;3;4;4",
"confidence_avg": 3.75,
"soundness": "1;3;3;3",
"soundness_avg": 2.5,
"contribution": "2;1;3;3",
"contribution_avg": 2.25,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.959732"
} | {
"id": "spEWEualnz",
"metareview": "In this work, authors present a multimodal retrieval-augmented generation (RAG) framework that uses hierarchical reasoning-and-acting (ReAct) agents to interact with Materials Project database and run atomistic simulations. The paper claims improved accuracy in material property... | {
"decision": "Reject"
} |
uNomADvF3s | 2406.10513v1 | Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space | {
"content": "## Abstract\n\nAbstract We introduce a new framework for molecular graph generation with 3D molecular generative models. Our Synthetic Coordinate Embedding ( SyCo ) framework maps molecular graphs to Euclidean point clouds via synthetic conformer coordinates and learns the inverse map using an E( n n )-... | [
{
"id": "uG811jE6oO",
"initial_rating": 6,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper introduces a novel 3D latent embedding scheme for 2D molecular graph generation. It presents a set of experiments which try to show that this inductive ... | {
"rating": "5;6;6;6",
"rating_avg": 5.75,
"confidence": "4;4;3;2",
"confidence_avg": 3.25,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "3;3;4;3",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.960799"
} | {
"id": "eERYf9rjow",
"metareview": "**Summary**\n\nThis paper introduces SyCO, a novel method for 2D molecular graph generation that leverages 3D latent embeddings. The approach involves mapping 2D molecular graphs to 3D point clouds using synthetic coordinates generated via the ETKDG algorithm, and then learning ... | {
"decision": "Accept (Poster)"
} |
uOrfve3prk | 2411.04430v1 | Towards Unifying Interpretability and Control: Evaluation via Intervention | {
"content": "## Abstract\n\nAbstract With the growing complexity and capability of large language models (LLMs), a need to understand model reasoning has emerged, often motivated by an underlying goal of controlling and aligning models.\nWhile numerous interpretability and steering methods have been proposed as solu... | [
{
"id": "ZspyOqdOxT",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper addresses the disconnect between interpretability and control in LLMs, proposing intervention as a core goal of interpretability. The authors introduce... | {
"rating": "3;5;5;6",
"rating_avg": 4.75,
"confidence": "4;3;4;2",
"confidence_avg": 3.25,
"soundness": "1;2;3;3",
"soundness_avg": 2.25,
"contribution": "1;3;3;3",
"contribution_avg": 2.5,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.961657"
} | {
"id": "P8y5ZGCMKv",
"metareview": "This paper presents a framework for evaluating the effectiveness of interpretability methods in large language models (LLMs) through the lens of \"intervention,\" aiming to bridge the gap between interpretability and model control. The authors introduce two novel evaluation metr... | {
"decision": "Reject"
} |
uOxoje4Sa9 | 2410.15027v1 | Group Diffusion Transformers are Unsupervised Multitask Learners | {
"content": "## Abstract\n\nAbstract While large language models (LLMs) have revolutionized natural language processing with their task-agnostic capabilities, visual generation tasks such as image translation, style transfer, and character customization still rely heavily on supervised, task-specific datasets. In th... | [
{
"id": "rm1yJXJ3Yw",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 3,
"summary": "This paper presents group diffusion transformer (GDT), targeting the problem of multi-image generation for multi-task learning. The key idea is to bind the self-a... | {
"rating": "3;5;5;6",
"rating_avg": 4.75,
"confidence": "4;3;4;3",
"confidence_avg": 3.5,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "2;2;3;3",
"contribution_avg": 2.5,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.962462"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
uPwe2w78Wx | 2405.19425v2 | Adaptive In-conversation Team Building for Language Model Agents | {
"content": "## Abstract\n\nAbstract Leveraging multiple large language model (LLM) agents has shown to be a promising approach for tackling complex tasks, while the effective design of multiple agents for a particular application remains an art. It is thus intriguing to answer a critical question: Given a task, how... | [
{
"id": "R1Dwr4GjIc",
"initial_rating": 6,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "Multi-agent systems (MAS) has shown to be superior to single agent system if constructed properly. However, designing a perfect MAS requires carefully designing t... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "3;4;4;4",
"confidence_avg": 3.75,
"soundness": "3;2;3;2",
"soundness_avg": 2.5,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "3;3;2;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.963353"
} | {
"id": "eR1RmcoSnx",
"metareview": "The authors address the problem of how to effectively design and manage a team of LLM-based agents to solve complex tasks and propose \"Captain Agent,\" a novel LLM agent that dynamically forms and manages teams of agents for complex tasks, using nested conversations and reflect... | {
"decision": "Reject"
} |
uQEsLZU15E | 2410.07167v2 | Deciphering Cross-Modal Alignment in Large Vision-Language Models with Modality Integration Rate | {
"content": "## Abstract\n\nAbstract We present the Modality Integration Rate (MIR), an effective, robust, and generalized metric to indicate the multi-modal pre-training quality of Large Vision Language Models (LVLMs).\nLarge-scale pre-training plays a critical role in building capable LVLMs, while evaluating its t... | [
{
"id": "pLpwzAOuA1",
"initial_rating": 8,
"confidence": 3,
"soundness": 4,
"contribution": 3,
"presentation": 4,
"summary": "This work proposes Modality Intergration Rate (MIR) to evaluate the pre-training quality of LVLMs, which measures the distance between vision and text modalities.... | {
"rating": "3;5;5;8",
"rating_avg": 5.25,
"confidence": "4;3;2;3",
"confidence_avg": 3,
"soundness": "3;3;3;4",
"soundness_avg": 3.25,
"contribution": "2;3;2;3",
"contribution_avg": 2.5,
"presentation": "3;3;3;4",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.964217"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
uQjySppU9x | 2411.04989v1 | SG-I2V: Self-Guided Trajectory Control in Image-to-Video Generation | {
"content": "## Abstract\n\nAbstract Methods for image-to-video generation have achieved impressive, photo-realistic quality.\nHowever, adjusting specific elements in generated videos, such as object motion or camera movement, is often a tedious process of trial and error, e.g., involving re-generating videos with d... | [
{
"id": "x7zaXbHFg5",
"initial_rating": 5,
"confidence": 5,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "The paper introduces SG-I2V, a self-guided method designed to achieve zero-shot controllable image-to-video (I2V) generation. Technically, SG-I2V first extracts f... | {
"rating": "5;5;6;6;6",
"rating_avg": 5.6,
"confidence": "5;5;4;5;4",
"confidence_avg": 4.6,
"soundness": "3;2;3;3;3",
"soundness_avg": 2.8,
"contribution": "2;2;2;3;3",
"contribution_avg": 2.4,
"presentation": "3;2;3;3;3",
"presentation_avg": 2.8
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.964950"
} | {
"id": "31c5FZUmex",
"metareview": "This work receives mixed reviews. While reviewers initially raised concerns regarding performance discrepancies, limited evaluation, technical novelty, and computational cost, the authors' detailed response and clarification have largely addressed these concerns. The paper's con... | {
"decision": "Accept (Poster)"
} |
uSiyu6CLPh | 2401.13212v1 | AdCorDA: Classifier Refinement via Adversarial Correction and Domain Adaptation | {
"content": "## Abstract\n\nAbstract This paper describes a simple yet effective technique for refining a pretrained classifier network. The proposed AdCorDA method is based on modification of the training set and making use of the duality between network weights and layer inputs. We call this input space training. ... | [
{
"id": "rlfBO2bT8I",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes a method for refining pre-trained classifiers through adversarial correction and domain adaptation. Adversarial correction adds adversarial pe... | {
"rating": "3;3;6;6",
"rating_avg": 4.5,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "2;2;3;4",
"soundness_avg": 2.75,
"contribution": "2;2;3;4",
"contribution_avg": 2.75,
"presentation": "3;3;3;4",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.965549"
} | {
"id": "rMTmhz5Id0",
"metareview": "This paper introduces AdCorDA, a two-stage method for refining pretrained classifiers through adversarial correction and domain adaptation. The authors conducted experiments on standard and weight-quantized neural networks to demonstrate the effectiveness of AdCorDA. The main id... | {
"decision": "Reject"
} |
uTqnyF0JNR | 2406.09870v2 | IGL-Bench: Establishing the Comprehensive Benchmark for Imbalanced Graph Learning | {
"content": "## Abstract\n\nAbstract Deep graph learning has gained grand popularity over the past years due to its versatility and success in representing graph data across a wide range of domains.\nHowever, the pervasive issue of imbalanced graph data distributions, where certain parts exhibit disproportionally ab... | [
{
"id": "plAfsMgV1E",
"initial_rating": 8,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper addresses the gap in graph learning by introducing a benchmark named IGL-Bench for imbalanced graph learning. The benchmark covers 24 algorithms across... | {
"rating": "5;6;6",
"rating_avg": 5.666666666666667,
"confidence": "3;3;3",
"confidence_avg": 3,
"soundness": "3;3;3",
"soundness_avg": 3,
"contribution": "3;3;3",
"contribution_avg": 3,
"presentation": "3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.966358"
} | {
"id": "H7f0QZoHzk",
"metareview": "This paper introduces IGL-Bench, a benchmark for Imbalanced Graph Learning (IGL) that evaluates 24 algorithms across 17 datasets, addressing both class and topology imbalances in node- and graph-level tasks. The benchmark provides a unified data processing pipeline, evaluation p... | {
"decision": "Accept (Spotlight)"
} |
uVDwunWsLz | 2410.07746v1 | Benign Overfitting in Single-Head Attention | {
"content": "## Abstract\n\nAbstract The phenomenon of benign overfitting , where a trained neural network perfectly fits noisy training data but still achieves near-optimal test performance, has been extensively studied in recent years for linear models and fully-connected/convolutional networks. In this work, we s... | [
{
"id": "UkHxibY1wM",
"initial_rating": 3,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper studies benign overfitting in a single-head attention model, proving that under certain signal-to-noise ratio (SNR) conditions, the model achieves beni... | {
"rating": "3;3;5;6",
"rating_avg": 4.25,
"confidence": "4;3;4;2",
"confidence_avg": 3.25,
"soundness": "3;2;3;3",
"soundness_avg": 2.75,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "3;2;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.967900"
} | {
"id": "3Y5hsgbQoA",
"metareview": "This paper theoretically investigates benign overfitting for a single head attention model. \n\nAlthough benign overfitting of transformers is an important research topic, this paper has several drawbacks: \n(i) The biggest concern of this paper is its novelty. First, the an... | {
"decision": "Reject"
} |
uVMZgtw2pf | 2406.11730v2 | CHG Shapley: Efficient Data Valuation and Selection towards Trustworthy Machine Learning | {
"content": "## Abstract\n\nAbstract Understanding the decision-making process of machine learning models is crucial for ensuring trustworthy machine learning. Data Shapley, a landmark study on data valuation, advances this understanding by assessing the contribution of each datum to model accuracy. However, the res... | [
{
"id": "teJbotRbye",
"initial_rating": 3,
"confidence": 4,
"soundness": 4,
"contribution": 1,
"presentation": 4,
"summary": "The paper proposed a gradient-based method to reduce the computational cost of Data Shapley: the CHG (compound of Hardness and Gradient) utility function, which a... | {
"rating": "3;3;5",
"rating_avg": 3.6666666666666665,
"confidence": "3;4;4",
"confidence_avg": 3.6666666666666665,
"soundness": "3;4;2",
"soundness_avg": 3,
"contribution": "2;1;2",
"contribution_avg": 1.6666666666666667,
"presentation": "3;4;2",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.969469"
} | {
"id": "0aX8BmUp1L",
"metareview": "The paper proposed a gradient-based method to reduce the computation of the Shapley value of training data. The proposed CHG (compound of Hardness and Gradient) utility function approximates the utility of each data subset on model performance. Using the CHG utility function, t... | {
"decision": "Reject"
} |
uWtLOy35WD | 2408.15881v3 | LLaVA-MoD: Making LLaVA Tiny via MoE-Knowledge Distillation | {
"content": "## Abstract\n\nAbstract We introduce LLaVA-MoD, a novel framework designed to enable the efficient training of small-scale Multimodal Language Models ( s -MLLM) distilling knowledge from large-scale MLLM ( l -MLLM). Our approach tackles two fundamental challenges in MLLM distillation. First, we optimize... | [
{
"id": "3O8t85Vrcr",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "LLaVA-MoD introduces a framework for creating efficient small-scale multimodal language models through knowledge distillation from larger models. The approach tac... | {
"rating": "5;5;6;6;6;8",
"rating_avg": 6,
"confidence": "4;4;3;3;3;3",
"confidence_avg": 3.3333333333333335,
"soundness": "2;2;3;3;3;4",
"soundness_avg": 2.8333333333333335,
"contribution": "2;2;3;3;2;3",
"contribution_avg": 2.5,
"presentation": "3;3;3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.970105"
} | {
"id": "286SqxaM8T",
"metareview": "This work introduces LLaVA-MoD, a framework for training compact multimodal language models through knowledge distillation from larger models. The study addresses two main challenges: (1) optimizing the network structure using a sparse Mixture of Experts (MoE) architecture, and ... | {
"decision": "Accept (Poster)"
} |
uZFXpPrwSh | 2410.11711v1 | Zero-shot Model-based Reinforcement Learning using Large Language Models | {
"content": "## Abstract\n\nAbstract The emerging zero-shot capabilities of Large Language Models (LLMs) have led to their applications in areas extending well beyond natural language processing tasks.\nIn reinforcement learning, while LLMs have been extensively used in text-based environments, their integration wit... | [
{
"id": "j06EBQCa5I",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This paper proposes a new approach Disentangled In-Context Learning (DICL) to generalize LLM-based in-context learning to the domain of continuous-state-space rei... | {
"rating": "5;6;8;8",
"rating_avg": 6.75,
"confidence": "3;3;2;3",
"confidence_avg": 2.75,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "3;3;3;3",
"contribution_avg": 3,
"presentation": "2;2;3;2",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.970933"
} | {
"id": "qm9PwkOvuE",
"metareview": "This work explores using pre-trained LLMs to predict the dynamics of continuous MDPs in-context, addressing challenges related to multivariate data and control signal integration. The reviewers appreciate the originality of the approach, the theoretical analysis, and the thoroug... | {
"decision": "Accept (Poster)"
} |
uZgK0tcPqd | 2410.01532v1 | Seeing Eye to AI: Human Alignment via Gaze-Based Response Rewards for Large Language Models | {
"content": "## Abstract\n\nAbstract Advancements in Natural Language Processing (NLP), have led to the emergence of Large Language Models (LLMs) such as GPT, Llama, Claude, and Gemini, which excel across a range of tasks but require extensive fine-tuning to align their outputs with human expectations. A widely used... | [
{
"id": "zHx1Gezzt6",
"initial_rating": 8,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 4,
"summary": "The paper examines the possibility of using eye-tracking data to train language reward models. The authors integrate synthetic eye tracking data from existing mod... | {
"rating": "3;6;8",
"rating_avg": 5.666666666666667,
"confidence": "3;4;3",
"confidence_avg": 3.3333333333333335,
"soundness": "2;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "3;3;3",
"contribution_avg": 3,
"presentation": "2;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.971831"
} | {
"id": "KQMpkX8e2e",
"metareview": "This paper introduces a novel framework that integrates eye-tracking data as implicit feedback into the reward model for aligning LLMs with human expectations, called GazeReward. The reviewers appreciate the novelty of the approach and the performance gains it can achieve by inc... | {
"decision": "Accept (Poster)"
} |
uZmmgHY1mD | 2409.20370v1 | The Perfect Blend: Redefining RLHF with Mixture of Judges | {
"content": "## Abstract\n\nAbstract Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off ... | [
{
"id": "vl258B18c8",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper introduces a novel training paradigm called Constrained Generative Policy Optimization (CGPO), to tackle the challenge of multi-objective optimization ... | {
"rating": "3;6;6",
"rating_avg": 5,
"confidence": "4;4;3",
"confidence_avg": 3.6666666666666665,
"soundness": "2;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;3;3",
"contribution_avg": 2.6666666666666665,
"presentation": "2;4;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.972849"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
uaMSBJDnRv | 2410.08847v2 | Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization | {
"content": "## Abstract\n\nAbstract Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences.\nAlthough these methods are designed to teach a model to generate preferred responses more frequently relative to dispreferred responses, prior work ha... | [
{
"id": "ctJKwH01kl",
"initial_rating": 8,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper discusses a curious and interesting phenomenon - likelihood displacement in LLMs where alignment methods actually can lead the model to be unaligned by ... | {
"rating": "5;6;6;8",
"rating_avg": 6.25,
"confidence": "3;3;3;3",
"confidence_avg": 3,
"soundness": "3;3;4;3",
"soundness_avg": 3.25,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "4;3;4;3",
"presentation_avg": 3.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.973990"
} | {
"id": "UCXR1SFYPs",
"metareview": "The paper points out a curious phenomenon in direct preference optimization (DPO) where the likelihood of preferred responses decreases during learning. The authors characterize this as driven by preferences that induce similar embeddings and propose the use of the centered hid... | {
"decision": "Accept (Poster)"
} |
ubIxE93FLM | 2404.06479v4 | Visually Descriptive Language Model for Vector Graphics Reasoning | {
"content": "## Abstract\n\nAbstract Despite significant advancements, current large multimodal models (LMMs) struggle to bridge the gap between low-level visual perception—focusing on shapes, sizes and layouts—and high-level language reasoning involving semantics, events and logic.\nThis limitation becomes evident ... | [
{
"id": "J2ti0tkuUE",
"initial_rating": 5,
"confidence": 2,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "The paper introduces VDLM that aims to decrease inaccuracies associated with modern VLMs by USING an intermediate representation Primal Visual Description (PVD), ... | {
"rating": "3;5;5;5",
"rating_avg": 4.5,
"confidence": "4;4;4;2",
"confidence_avg": 3.5,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "3;2;3;2",
"contribution_avg": 2.5,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.975222"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
ubuGgIPVD0 | 2309.01539v3 | TSTTC: A Large-Scale Dataset for Time-to-Contact Estimation in Driving Scenarios | {
"content": "## Abstract\n\nAbstract Time-to-Contact (TTC) estimation is a critical task for assessing collision risk and is widely used in various driver assistance and autonomous driving systems. The past few decades have witnessed development of related theories and algorithms. The prevalent learning-based method... | [
{
"id": "BFI3DLB03l",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper introduces a time-to-contact dataset for the safety requirements of the redundancy system in ADAS. In addition to being collected from real-world scena... | {
"rating": "5;5;5;5",
"rating_avg": 5,
"confidence": "3;5;4;4",
"confidence_avg": 4,
"soundness": "3;2;2;3",
"soundness_avg": 2.5,
"contribution": "3;2;3;2",
"contribution_avg": 2.5,
"presentation": "3;2;2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.976058"
} | {
"id": "WJxxtmySoZ",
"metareview": "This paper's main contribution is a dataset for evaluating time-to-contact that assesses collision risk in ADAS (driver assistance) systems. The proposed dataset provides monocular videos and consists of manually selected (200) sequences sampled from thousands of hours of drivin... | {
"decision": "Reject"
} |
udfjje2xXb | 2406.18354v1 | Kolmogorov–Arnold Graph Neural Networks | {
"content": "## Abstract\n\nAbstract Graph neural networks (GNNs) excel in learning from network-like data but often\nlack interpretability, making their application challenging in domains requiring\ntransparent decision-making. We propose the Graph Kolmogorov–Arnold Network\n(GKAN), a novel GNN model leveraging spl... | [
{
"id": "HkyyfG57Ap",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "The paper introduces the Kolmogorov-Arnold Network for Graphs (KANG), a new Graph Neural Network (GNN) model that enhances interpretability and accuracy in tasks ... | {
"rating": "1;3;3;3;5",
"rating_avg": 3,
"confidence": "5;3;4;4;4",
"confidence_avg": 4,
"soundness": "2;2;1;3;2",
"soundness_avg": 2,
"contribution": "2;2;2;3;2",
"contribution_avg": 2.2,
"presentation": "2;2;4;3;2",
"presentation_avg": 2.6
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.976965"
} | {
"id": "bi8YNmatmR",
"metareview": "This submission presents KANG, a novel Graph Neural Network (GNN) model that integrates Kolmogorov-Arnold Networks with GNNs and utilizes spline-based activation functions. The primary goal of KANG is to enhance the accuracy and interpretability of GNNs, enabling transparent dec... | {
"decision": "Reject"
} |
ue1Tt3h1VC | 2405.16869v2 | Multiple Heads are Better than One: Mixture of Modality Knowledge Experts for Entity Representation Learning | {
"content": "## Abstract\n\nAbstract Learning high-quality multi-modal entity representations is an important goal of multi-modal knowledge graph (MMKG) representation learning, which can enhance reasoning tasks within the MMKGs, such as MMKG completion (MMKGC). The main challenge is to collaboratively model the str... | [
{
"id": "nEGNmEUNWW",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper tackles MKGE problems by proposing a Mixture of Modality Knowledge Experts framework. The authors develop relation-guided modality knowledge experts to... | {
"rating": "5;6;6;6;6",
"rating_avg": 5.8,
"confidence": "4;4;4;2;4",
"confidence_avg": 3.6,
"soundness": "2;3;3;3;3",
"soundness_avg": 2.8,
"contribution": "2;3;2;3;3",
"contribution_avg": 2.6,
"presentation": "2;3;3;4;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.977791"
} | {
"id": "zR8I0669wl",
"metareview": "This paper proposes a multi-modal knowledge graph (MMKG) representation learning model, a Mixture of Modality Knowledge experts (MoMoK for short), to learn adaptive multi-modal entity representations for better MMKG completion. A key design idea is to learn representations throu... | {
"decision": "Accept (Poster)"
} |
ufi0WPTgWp | 2410.06682v2 | Enhancing Multimodal LLM for Detailed and Accurate Video Captioning using Multi-Round Preference Optimization | {
"content": "## Abstract\n\nAbstract Videos contain a wealth of information, and generating detailed and accurate descriptions in natural language is a key aspect of video understanding. In this paper, we present video-SALMONN 2, an advanced audio-visual large language model (LLM) with low-rank adaptation (LoRA) des... | [
{
"id": "8hVDkx4Xbt",
"initial_rating": 3,
"confidence": 5,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "- The paper introduces video-SALMONN 2, an advanced audio-visual large language model (LLM) for generating detailed and accurate video captions, surpassing models... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "4;5;5;4",
"confidence_avg": 4.5,
"soundness": "2;3;3;2",
"soundness_avg": 2.5,
"contribution": "2;3;2;2",
"contribution_avg": 2.25,
"presentation": "1;3;2;3",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.978484"
} | {
"id": "8S9u7Vlvg9",
"metareview": "# Summary and Recommendation for Rejection\n\n---\n\n## Strengths:\n1. **Innovative Training Techniques**:\n - Introduces **multi-round Directed Preference Optimization (mrDPO)** and **rebirth tuning** for optimizing detailed and accurate video captioning.\n - mrDPO stabiliz... | {
"decision": "Reject"
} |
ugXGFCS6HK | 2410.15433v1 | Discriminating image representations with principal distortions | {
"content": "## Abstract\n\nAbstract Image representations (artificial or biological) are often compared in terms of their global geometry; however, representations with similar global structure can have strikingly different local geometries. Here, we propose a framework for comparing a set of image representations ... | [
{
"id": "ZWvclA96cJ",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 4,
"summary": "This paper introduces a novel metric on image representations to measure differences in local geometry. The authors then leverage this metric to generate \"princi... | {
"rating": "5;5;6;6;8",
"rating_avg": 6,
"confidence": "4;3;3;3;4",
"confidence_avg": 3.4,
"soundness": "3;3;4;3;4",
"soundness_avg": 3.4,
"contribution": "2;3;3;3;3",
"contribution_avg": 2.8,
"presentation": "3;4;3;3;4",
"presentation_avg": 3.4
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.979220"
} | {
"id": "JDoNTqZA2i",
"metareview": "The current paper applies Fisher information w.r.t. multiple neural networks to find local distortions in the input space of image representations. The reviews highlighted the novelty and clear presentation. Most reviewers acknowledged that the authors' rebuttal has sufficiently... | {
"decision": "Accept (Poster)"
} |
uhaLuZcCjH | 2410.04234v1 | Functional Homotopy: Smoothing Discrete Optimization via Continuous Parameters for LLM Jailbreak Attacks | {
"content": "## Abstract\n\nAbstract Warning: This paper contains potentially offensive and harmful text. Optimization methods are widely employed in deep learning to identify and mitigate undesired model responses. While gradient-based techniques have proven effective for image models, their application to language... | [
{
"id": "oaIHaF7NFZ",
"initial_rating": 6,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper proposes the functional homotopy method, a novel optimization approach for LLM attacks. FH fintunes the LLM and solves a number of easy-to-hard problems... | {
"rating": "3;3;6;6",
"rating_avg": 4.5,
"confidence": "4;3;2;2",
"confidence_avg": 2.75,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "2;1;3;3",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.979879"
} | {
"id": "J5pY3K3OTy",
"metareview": "This paper proposes a new method called the Functional Homotopy (FH) method to solve discrete optimization problems that appear in jailbreak attacks on LLMs. The proposed method achieves 20%-30% improvement in success rates. \n\nThe proposed method is based on an interesting id... | {
"decision": "Accept (Poster)"
} |
ujNe7sybJu | 2410.04511v1 | Realizing Video Summarization from the Path of Language-based Semantic Understanding | {
"content": "## Abstract\n\nAbstract The recent development of Video-based Large Language Models (VideoLLMs), has significantly advanced video summarization by aligning video features—and, in some cases, audio features—with Large Language Models (LLMs). Each of these VideoLLMs possesses unique strengths and weakness... | [
{
"id": "UKc7leB360",
"initial_rating": 3,
"confidence": 5,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "* This work introduces a new framework inspired by the Mixture of Experts (MoE) approach, allowing VideoLLMs to complement each other’s strengths without fine-tun... | {
"rating": "1;3;3;3",
"rating_avg": 2.5,
"confidence": "5;4;4;5",
"confidence_avg": 4.5,
"soundness": "3;3;2;3",
"soundness_avg": 2.75,
"contribution": "1;2;1;2",
"contribution_avg": 1.5,
"presentation": "1;3;2;2",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.980529"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
ujpAYpFDEA | 2410.03168v1 | Can Watermarked LLMs be Identified by Users via Crafted Prompts? | {
"content": "## Abstract\n\nAbstract Text watermarking for Large Language Models (LLMs) has made significant progress in detecting LLM outputs and preventing misuse. Current watermarking techniques offer high detectability, minimal impact on text quality, and robustness to text editing.\nHowever, current researches ... | [
{
"id": "D6U1irciZl",
"initial_rating": 8,
"confidence": 4,
"soundness": 4,
"contribution": 4,
"presentation": 4,
"summary": "Considering that “even if individual watermarked texts are imperceptible, the distribution of numerous watermarked texts may reveal whether the LLM is watermarked... | {
"rating": "5;5;6;8",
"rating_avg": 6,
"confidence": "3;4;3;4",
"confidence_avg": 3.5,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "2;2;3;4",
"contribution_avg": 2.75,
"presentation": "3;3;3;4",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.981196"
} | {
"id": "sGaBsh29Od",
"metareview": "This paper focuses on the imperceptibility of watermarking techniques in Large Language Models (LLMs). While existing methods ensure detectability, robustness, and minimal text quality degradation, their imperceptibility to users remains underexplored. The authors introduce Wate... | {
"decision": "Accept (Spotlight)"
} |
ulGwcj1egv | 2410.12513v1 | FiRST: Finetuning Router-Selective Transformers for Input-Adaptive Latency Reduction | {
"content": "## Abstract\n\nAbstract Auto-regressive Large Language Models (LLMs)\ndemonstrate remarkable performance across domanins such as vision and language processing. However, due to sequential processing through a stack of transformer layers, autoregressive decoding faces significant computation/latency chal... | [
{
"id": "BfHJ0AyUvP",
"initial_rating": 3,
"confidence": 5,
"soundness": 2,
"contribution": 3,
"presentation": 2,
"summary": "This paper provides an algorithm FIRST that reduces inference latency using layer selection corresponding to input sequences and tasks. \nThe authors evaluate FIR... | {
"rating": "3;3;3;3",
"rating_avg": 3,
"confidence": "5;4;4;5",
"confidence_avg": 4.5,
"soundness": "2;2;2;2",
"soundness_avg": 2,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "2;3;3;2",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.981859"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
ulJNq6FQrw | 2408.02599v2 | Progressively Label Enhancement for Large Language Model Alignment | {
"content": "## Abstract\n\nAbstract Large Language Models (LLM) alignment aims to prevent models from producing content that misaligns with human expectations, which can lead to ethical and legal concerns.\nIn the last few years, Reinforcement Learning from Human Feedback (RLHF) has been the most prominent method f... | [
{
"id": "VrV3zp5C7n",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 2,
"summary": "The paper proposes a new method for aligning large language models. \n\nThe method proposed in the paper is PLE, i.e. Progressively Label Enhancement for LLM Alig... | {
"rating": "3;5;5;5;5",
"rating_avg": 4.6,
"confidence": "3;4;3;3;4",
"confidence_avg": 3.4,
"soundness": "2;2;2;2;2",
"soundness_avg": 2,
"contribution": "2;2;2;2;3",
"contribution_avg": 2.2,
"presentation": "2;3;3;2;2",
"presentation_avg": 2.4
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.982475"
} | {
"id": "GmrUDtk128",
"metareview": "This paper proposes Progressively Label Enhancement (PLE) for LLM Alignment, which dynamically adjusts the model's training process based on the quality of the generated data. The reviewers raised significant concerns including the novelty of the method and the soundness of the ... | {
"decision": "Reject"
} |
umggmAFhRD | 2407.00805v2 | Towards shutdownable agents via stochastic choice | {
"content": "## Abstract\n\nAbstract Some worry that advanced artificial agents may resist being shut down. The Incomplete Preferences Proposal (IPP) is an idea for ensuring that doesn’t happen. A key part of the IPP is using a novel ‘ D iscounted RE ward for S ame-Length T rajectories (DREST)’ reward function to tr... | [
{
"id": "NUxkgakZQV",
"initial_rating": 3,
"confidence": 3,
"soundness": 1,
"contribution": 1,
"presentation": 3,
"summary": "The paper focuses on the shutdown problem in AI safety: that advanced AIs might take actions that impede humans from shutting them down. It proposes a method for ... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "4;3;3;2",
"confidence_avg": 3,
"soundness": "2;1;2;2",
"soundness_avg": 1.75,
"contribution": "1;1;2;3",
"contribution_avg": 1.75,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.983265"
} | {
"id": "72hpLzHp5h",
"metareview": "This paper proposes a novel reward function that aims to have agents learn to achieve their tasks, while being indifferent to being \"shutdown\". This is achieved through the proposed DREST reward formulation, which results in agents that are both USEFUL and NEUTRAL (metrics def... | {
"decision": "Reject"
} |
unDQOUah0F | 2410.19100v1 | VideoWebArena: Evaluating Long Context Multimodal Agents with Video Understanding Web Tasks | {
"content": "## Abstract\n\nAbstract Videos are often used to learn or extract the necessary information to complete tasks in ways different than what text and static imagery alone can provide. However, many existing agent benchmarks neglect long-context video understanding, instead focusing on text or static image ... | [
{
"id": "Y7sLOjGa77",
"initial_rating": 5,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper introduces VideoWebArena, a benchmark designed to evaluate multimodal AI models’ abilities to process and understand long video sequences alongside tex... | {
"rating": "5;6;6;6",
"rating_avg": 5.75,
"confidence": "3;4;3;3",
"confidence_avg": 3.25,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "3;2;3;3",
"contribution_avg": 2.75,
"presentation": "3;3;2;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.984250"
} | {
"id": "cpQrUylmIe",
"metareview": "This paper was reviewed by five experts in the field. Reviewer pi6N disappeared and didn't participate in the discussion despite multiple reminders from the AC and authors. The authors' rebuttal resolved most concerns, and the other four reviewers unanimously agreed to accept th... | {
"decision": "Accept (Poster)"
} |
upzyG4wRBr | 2406.11334v1 | Program Synthesis Benchmark for Visual Programming in XLogoOnline Environment | {
"content": "## Abstract\n\nAbstract Large language and multimodal models have shown remarkable successes on various benchmarks focused on specific skills such as general-purpose programming, natural language understanding, math word problem-solving, and visual question answering. However, it is unclear how well the... | [
{
"id": "GcBcoU2pSQ",
"initial_rating": 3,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 4,
"summary": "This paper\n1. curates a small (85 tasks) program synthesis benchmark based on real-world tasks in the XLogoOnline visual programming environment, which requires ... | {
"rating": "3;3;3;8;8",
"rating_avg": 5,
"confidence": "5;3;4;4;4",
"confidence_avg": 4,
"soundness": "2;2;3;3;4",
"soundness_avg": 2.8,
"contribution": "2;2;2;3;3",
"contribution_avg": 2.4,
"presentation": "2;3;4;4;4",
"presentation_avg": 3.4
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.984898"
} | {
"id": "N55sca168r",
"metareview": "The paper introduces a new benchmark to evaluate the performance of LLM on program synthesis tasks within the XLogoOnline visual environment. The primary concerns raised include:\n\n1. Limited technical novelty\n\n2. Insufficient experimental evaluation\n\n While the paper has n... | {
"decision": "Reject"
} |
uq9TLFT7tF | 2405.18132v1 | EG4D: Explicit Generation of 4D Object without Score Distillation | {
"content": "## Abstract\n\nAbstract In recent years, the increasing demand for dynamic 3D assets in design and gaming applications has given rise to powerful generative pipelines capable of synthesizing high-quality 4D objects.\nPrevious methods generally rely on score distillation sampling (SDS) algorithm to infer... | [
{
"id": "0V8Pobf7Zt",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 3,
"summary": "The paper introduces EG4D, a multi-stage framework designed for generating 4D objects from a single image input without relying on score distillation sampling (SD... | {
"rating": "5;5;6",
"rating_avg": 5.333333333333333,
"confidence": "5;4;4",
"confidence_avg": 4.333333333333333,
"soundness": "2;2;3",
"soundness_avg": 2.3333333333333335,
"contribution": "2;3;3",
"contribution_avg": 2.6666666666666665,
"presentation": "3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.985478"
} | {
"id": "GiAqi78DCs",
"metareview": "This paper receives 3x ratings of 6s. The AC follows the recommendations of the reviewers to accept the paper. The reviewers think that the paper is well-written, the proposed method is effective and the experimental results are good. The weaknesses are well-addressed by the aut... | {
"decision": "Accept (Poster)"
} |
uqWM9hBDAE | 2402.05835v1 | How Much is Unseen Depends Chiefly on Information About the Seen | {
"content": "## Abstract\n\nAbstract It might seem counter-intuitive at first: We find that, in expectation , the proportion of data points in an unknown population—that belong to classes that do not appear in the training data—is almost entirely determined by the number f k subscript 𝑓 𝑘 f_{k} of classes that do ... | [
{
"id": "MNoUIvzBYC",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper makes a contribution to the estimation of the missing mass probability by providing a distribution free estimator that minimizes the mean-squared error ... | {
"rating": "6;8;8",
"rating_avg": 7.333333333333333,
"confidence": "4;3;2",
"confidence_avg": 3,
"soundness": "3;4;3",
"soundness_avg": 3.3333333333333335,
"contribution": "3;2;3",
"contribution_avg": 2.6666666666666665,
"presentation": "3;4;3",
"presentation_avg": 3.3333333333333335
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.986450"
} | {
"id": "KQRWgXUrl0",
"metareview": "The reviewers reach the consensus that the paper is well-written, tackling challenging missing class probability estimation and provide MSE studies based on a class of estimators. The reviewers has reached a consensus that the contribution is substantial and generally agreed tha... | {
"decision": "Accept (Spotlight)"
} |
urQi0TgXFY | 2410.03768v1 | Hidden in Plain Text: Emergence & Mitigation of Steganographic Collusion in LLMs | {
"content": "## Abstract\n\nAbstract The rapid proliferation of frontier model agents promises significant societal advances but also raises concerns about systemic risks arising from unsafe interactions.\nCollusion to the disadvantage of others has been identified as a central form of undesirable agent cooperation.... | [
{
"id": "aW7fhkHipA",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper presents results for testing the ability of models to embed steganographic information into text to allow for collaboration between models. The threat ... | {
"rating": "3;3;6;6",
"rating_avg": 4.5,
"confidence": "5;4;4;3",
"confidence_avg": 4,
"soundness": "3;2;3;3",
"soundness_avg": 2.75,
"contribution": "1;2;3;3",
"contribution_avg": 2.25,
"presentation": "3;2;4;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.987323"
} | {
"id": "qzM1oDF4jl",
"metareview": "This paper uncovers a phenomenon called steganographic collusion, which arises indirectly from optimization in LLMs. While the phenomenon itself is intriguing, reviewers raised concerns about the experimental setting, where the ground-truth labels of the training dataset are sol... | {
"decision": "Reject"
} |
urcEYsZOBz | 2409.00127v3 | Latent-EnSF: A Latent Ensemble Score Filter for High-Dimensional Data Assimilation with Sparse Observation Data | {
"content": "## Abstract\n\nAbstract Accurate modeling and prediction of complex physical systems often rely on data assimilation techniques to correct errors inherent in model simulations. Traditional methods like the Ensemble Kalman Filter (EnKF) and its variants as well as the recently developed Ensemble Score Fi... | [
{
"id": "dNQEdTuXDP",
"initial_rating": 3,
"confidence": 3,
"soundness": 2,
"contribution": 1,
"presentation": 2,
"summary": "The Ensemble Kalman Filter (EnKF) enables data assimilation for Nonlinear State-Space Models, but it encounters the problem of vanishing gradients in tasks where ... | {
"rating": "3;3;5;5;6;6",
"rating_avg": 4.666666666666667,
"confidence": "4;3;4;4;3;4",
"confidence_avg": 3.6666666666666665,
"soundness": "2;2;2;3;3;3",
"soundness_avg": 2.5,
"contribution": "2;1;2;2;3;3",
"contribution_avg": 2.1666666666666665,
"presentation": "1;2;2;3;3;3",
"presentation_avg": 2... | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.988136"
} | {
"id": "EuKgxIdNNL",
"metareview": "The reviewers recommend acceptance (5-6-6-8-6-6). The paper proposes Latent Ensemble Score Filter for improving data assimilation in high-dimensional and nonlinear systems. The approach is sound and the results are good. The author-reviewer discussion has been very constructive ... | {
"decision": "Accept (Poster)"
} |
urf8a5G59f | 2404.19604v1 | X-Diffusion: Generating Detailed 3D MRI Volumes From a Single Image Using Cross-Sectional Diffusion Models | {
"content": "## Abstract\n\nAbstract In this work, we present X-Diffusion , a cross-sectional diffusion model tailored for Magnetic Resonance Imaging (MRI) data.\nX-Diffusion is capable of generating the entire MRI volume from just a single MRI slice or optionally from few multiple slices, setting new benchmarks in ... | [
{
"id": "hLbRXXSuhm",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The authors introduces a new model for generating detailed 3D-MRI volumes from sparsified spatial-domain inputs, resulting in accurate 2D-to-3D scans reconstructi... | {
"rating": "3;3;6;6",
"rating_avg": 4.5,
"confidence": "5;4;4;3",
"confidence_avg": 4,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "2;1;4;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.988856"
} | {
"id": "Ga3qHRGlu2",
"metareview": "This manuscript proposes diffusion model based volumetric conditional generation (inpainting with extremely sparse inputs) intended for MRI (structural contrasts, presumably). They provide results on a variety of datasets, including UK Biobank (whole body MR), BRATS (brain tumor... | {
"decision": "Reject"
} |
us5riDkeBW | 2312.11441v2 | Social Learning: Towards Collaborative Learning with Large Language Models | {
"content": "## Abstract\n\nAbstract We introduce the framework of \"social learning\" in the context of large language models (LLMs), whereby models share knowledge with each other in a privacy-aware manner using natural language. We present and evaluate two approaches for knowledge transfer between LLMs. In the fi... | [
{
"id": "qNVXuEJneO",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "This paper proposed a new way to transfer knowledge in a collaborative learning setting where multiple teachers want to collaborate to teach a good student. A new... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "3;4;3;4",
"confidence_avg": 3.5,
"soundness": "1;2;3;3",
"soundness_avg": 2.25,
"contribution": "3;2;2;2",
"contribution_avg": 2.25,
"presentation": "3;2;2;2",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.989646"
} | {
"id": "qDpzibBw5s",
"metareview": "The paper proposes a social learning method for LLMs, which aims to achieve information sharing among LLMs.\n\nThe reviewers think that the idea is novel and clever and hence has good potential.\n\nHowever, the reviewers also identified some notable weaknesses: insufficient disc... | {
"decision": "Reject"
} |
usFdPd4Ghs | 2410.01284v1 | Deep Kernel Posterior Learning under Infinite Variance Prior Weights | {
"content": "## Abstract\n\nAbstract Neal ( 1996 ) proved that infinitely wide shallow Bayesian neural networks (BNN) converge to Gaussian processes (GP), when the network weights have bounded prior variance. Cho & Saul ( 2009 ) provided a useful recursive formula for deep kernel processes for relating the covarianc... | [
{
"id": "o7pc81civr",
"initial_rating": 5,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "The authors consider Bayesian neural networks in the infinite-width limit under infinite variance priors on the weights. They show that the resulting covariance k... | {
"rating": "5;5;6;6;8",
"rating_avg": 6,
"confidence": "3;2;2;3;4",
"confidence_avg": 2.8,
"soundness": "2;3;3;3;3",
"soundness_avg": 2.8,
"contribution": "2;3;3;3;3",
"contribution_avg": 2.8,
"presentation": "2;2;3;2;3",
"presentation_avg": 2.4
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.990382"
} | {
"id": "6BM0bCE4HY",
"metareview": "This paper explores infinite-width Bayesian neural networks with elliptically-distributed weights with infinite variance, demonstrating convergence to a process with α-stable marginals and conditionally Gaussian representations. This results in a stochastic, data-dependent covar... | {
"decision": "Accept (Poster)"
} |
uswS6tUCN2 | 2410.09771v1 | Magnituder Layers for Implicit Neural Representations in 3D | {
"content": "## Abstract\n\nAbstract Improving the efficiency and performance of implicit neural representations in 3D, particularly Neural Radiance Fields (NeRF) and Signed Distance Fields (SDF) is crucial for enabling their use in real-time applications. These models, while capable of generating photo-realistic no... | [
{
"id": "xXe57TXIsv",
"initial_rating": 3,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper proposes a novel neural network layer called the \"Magnituder,\" designed to reduce the number of training parameters in implicit 3D representation mod... | {
"rating": "3;3;5",
"rating_avg": 3.6666666666666665,
"confidence": "4;3;3",
"confidence_avg": 3.3333333333333335,
"soundness": "2;2;2",
"soundness_avg": 2,
"contribution": "2;2;2",
"contribution_avg": 2,
"presentation": "2;2;3",
"presentation_avg": 2.3333333333333335
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.991301"
} | {
"id": "C637nxc78V",
"metareview": "This submission proposes a new network layer, towards enabling efficiency improvements for implicit neural representation (INR) tasks. The paper draws on inspiration from kernel method approximations, where input magnitudes are considered in order to disentangle them from MLP la... | {
"decision": "Reject"
} |
utkGLDSNOk | 2411.02442v1 | TODO: Enhancing LLM Alignment with Ternary Preferences | {
"content": "## Abstract\n\nAbstract Aligning large language models (LLMs) with human intent is critical for enhancing their performance across a variety of tasks. Standard alignment techniques, such as Direct Preference Optimization (DPO), often rely on the binary Bradley-Terry (BT) model, which can struggle to cap... | [
{
"id": "gsg1QBPsYK",
"initial_rating": 5,
"confidence": 5,
"soundness": 2,
"contribution": 3,
"presentation": 3,
"summary": "The authors present an interesting approach to enhancing the alignment of LLMs through the use of ternary preferences. A limitation of binary models is their inab... | {
"rating": "3;5;6;8",
"rating_avg": 5.5,
"confidence": "3;5;4;3",
"confidence_avg": 3.75,
"soundness": "3;2;3;3",
"soundness_avg": 2.75,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.992052"
} | {
"id": "Qa4ADqH561",
"metareview": "The authors introduce an extension of the Bradley-Terry (BT) model, called the Tie-rank Oriented Bradley-Terry (TOBT) model. The TOBT model incorporates ties, allowing for a more nuanced representation of preferences. Building on this foundation, the Tie-rank Oriented Direct Pre... | {
"decision": "Accept (Poster)"
} |
uuEQsqb0GH | 2402.08062v3 | Avoiding Catastrophe in Online Learning by Asking for Help | {
"content": "## Abstract\n\nAbstract Most learning algorithms with formal regret guarantees assume that no mistake is irreparable and essentially rely on trying all possible behaviors. This approach is problematic when some mistakes are catastrophic , i.e., irreparable. We propose an online learning problem where th... | [
{
"id": "k01gUIABZS",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "The submission studies an online disaster avoidance problem. The learning agent is allowed to query the mentor to acquire information to avoid disaster. To circum... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "3;3;2;3",
"confidence_avg": 2.75,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "3;3;3;2",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.993056"
} | {
"id": "r6QqcMp3iW",
"metareview": "This is a borderline paper with potential. However, there is also criticism and there is a feeling that the paper is not quite there yet. This concerns, in particular, links to the AI risk literature and, more importantly, the local generalization assumption. I believe that the ... | {
"decision": "Reject"
} |
uuPkll6i7m | 2405.13922v1 | Towards Certification of Uncertainty Calibration under Adversarial Attacks | {
"content": "## Abstract\n\nAbstract Since neural classifiers are known to be sensitive to adversarial perturbations that alter their accuracy, certification methods have been developed to provide provable guarantees on the insensitivity of their predictions to such perturbations. Furthermore, in safety-critical app... | [
{
"id": "qLpBzm56xw",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This work addresses a critical vulnerability in neural network classifiers by examining how adversarial attacks can significantly harm model calibration (the reli... | {
"rating": "5;5;6;8",
"rating_avg": 6,
"confidence": "4;3;4;4",
"confidence_avg": 3.75,
"soundness": "2;3;4;3",
"soundness_avg": 3,
"contribution": "2;3;4;3",
"contribution_avg": 3,
"presentation": "3;3;4;3",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.994296"
} | {
"id": "b8KyvjXNGQ",
"metareview": "This paper studies a seemingly new and interesting problem: certifying uncertainty calibration of predictive models under adversarial attacks. The authors demonstrate that attacks can be devised that drastically change the calibration of preditors without necessarily changing th... | {
"decision": "Accept (Poster)"
} |
uuXPWRtwvK | 2411.02454v1 | Graph-based Confidence Calibration for Large Language Models | {
"content": "## Abstract\n\nAbstract One important approach to improving the reliability of large language models (LLMs) is to provide accurate confidence estimations regarding the correctness of their answers. However, developing a well-calibrated confidence estimation model is challenging, as mistakes made by LLMs... | [
{
"id": "UxhTQOZP6K",
"initial_rating": 5,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "To enhance the calibration performance of LLMs, this paper proposes to combine the LLM’s self-consistency with labeled data and train an auxiliary model to estima... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "4;5;5;3",
"confidence_avg": 4.25,
"soundness": "2;3;3;2",
"soundness_avg": 2.5,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;3;3;2",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:03.995226"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
uvHmnahyp1 | 2405.01155v2 | SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints | {
"content": "## Abstract\n\nAbstract Generative models see increasing use in computer-aided drug design. However,\nwhile performing well at capturing distributions of molecular motifs, they often produce synthetically inaccessible molecules.\nTo address this, we introduce SynFlowNet, a GFlowNet model whose action sp... | [
{
"id": "coVA6I4MJC",
"initial_rating": 8,
"confidence": 4,
"soundness": 4,
"contribution": 3,
"presentation": 4,
"summary": "It was a pleasure to review this paper on synthesizability-constrained molecular generation using GFlowNets. Overall, I felt the paper was strong and focuses on a... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "4;4;4;4",
"confidence_avg": 4,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "3;3;2;4",
"contribution_avg": 3,
"presentation": "3;3;2;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.996000"
} | {
"id": "KEv7hQrpvI",
"metareview": "In this work, authors introduce SynFlowNet, a generative model that combines GFlowNets with reaction-based Markov Decision Processes to generate synthetically accessible drug-like molecules. The key technical innovation is incorporating forward synthesis constraints directly int... | {
"decision": "Accept (Spotlight)"
} |
uwzyMFwyOO | 2405.19933v1 | Learning Latent Graph Structures and their Uncertainty | {
"content": "## Abstract\n\nAbstract Within a prediction task, Graph Neural Networks (GNNs) use relational information as an inductive bias to enhance the model’s accuracy. As task-relevant relations might be unknown, graph structure learning approaches have been proposed to learn them while solving the downstream p... | [
{
"id": "QxVoCp22Su",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper tackles the latent structure learning on graph structured data and demonstrate that minimizing prediction function does not guarantee a calibrated mode... | {
"rating": "3;5;5;5;5",
"rating_avg": 4.6,
"confidence": "4;3;3;3;3",
"confidence_avg": 3.2,
"soundness": "2;4;3;3;2",
"soundness_avg": 2.8,
"contribution": "2;2;3;3;3",
"contribution_avg": 2.6,
"presentation": "3;3;3;3;2",
"presentation_avg": 2.8
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:03.996897"
} | {
"id": "g5oyotrfAi",
"metareview": "The paper addresses problem of graph structure learning problem. It mentions limitations of current point-prediction methods which cannot guarantee the calibration of the distribution of adjacency matrix. A sampling-based optimization method using Maximum Mean Discrepancy (MMD) ... | {
"decision": "Reject"
} |
uxVBbSlKQ4 | 2410.03024v1 | Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting | {
"content": "## Abstract\n\nAbstract Recent advancements in generative modeling, particularly diffusion models, have opened new directions for time series modeling, achieving state-of-the-art performance in forecasting and synthesis.\nHowever, the reliance of diffusion-based models on a simple, fixed prior complicat... | [
{
"id": "tBf42IhNBQ",
"initial_rating": 6,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The existing diffusion models have problems in the time series generation since the data and prior distributions differ. The authors handle this problem by utiliz... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "3;2;4;2",
"confidence_avg": 2.75,
"soundness": "3;3;4;3",
"soundness_avg": 3.25,
"contribution": "3;2;2;3",
"contribution_avg": 2.5,
"presentation": "2;2;4;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.997821"
} | {
"id": "g83aQncGrU",
"metareview": "This paper introduces a new generative framework for time series forecasting, based on conditional flow matching and leveraging conditional Gaussian Processes as informed priors. By aligning the prior distribution more closely with the data distribution, this approach simplifies... | {
"decision": "Accept (Poster)"
} |
uy31tqVuNo | 2410.18975v2 | Unbounded: A Generative Infinite Game of Character Life Simulation | {
"content": "## Abstract\n\nAbstract We introduce the concept of a generative infinite game , a video game that transcends the traditional boundaries of finite, hard-coded systems by using generative models. Inspired by James P. Carse’s distinction between finite and infinite games (Carse, 1986 ) , we leverage recen... | [
{
"id": "B5fG2n5Q8V",
"initial_rating": 6,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "The authors present \"UNBOUNDED\" a virtual pet game based on LLMs. The authors present two novel technical elements to support this game: a regional IP-Adapter f... | {
"rating": "5;6",
"rating_avg": 5.5,
"confidence": "4;4",
"confidence_avg": 4,
"soundness": "2;3",
"soundness_avg": 2.5,
"contribution": "2;2",
"contribution_avg": 2,
"presentation": "2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.998593"
} | {
"id": "VB2N4zOWA8",
"metareview": "The paper presents an interesting experiment in creating a generative game where the core \"game engine\" is based on LLMs and diffusion models. While obviously there is a long way to go, there are interesting things to learn from the paper. The \"core technical\" contribution o... | {
"decision": "Accept (Poster)"
} |
uy4EavBEwl | 2405.19667v1 | Reconciling Model Multiplicity for Downstream Decision Making | {
"content": "## Abstract\n\nAbstract We consider the problem of model multiplicity in downstream decision-making, a setting where two predictive models of equivalent accuracy cannot agree on the best-response action for a downstream loss function. We show that even when the two predictive models approximately agree ... | [
{
"id": "FSZcgEYDki",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper studied the problem of model multiplicity in downstream decision-making. In this setting, two predictive models of equivalent accuracy do not agree on ... | {
"rating": "3;6;6;6",
"rating_avg": 5.25,
"confidence": "4;3;4;3",
"confidence_avg": 3.5,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "2;3;3;2",
"contribution_avg": 2.5,
"presentation": "1;3;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:03.999401"
} | {
"id": "XEzGmTRo7r",
"metareview": "This a paper which just sits above the borderline area. The paper studies the problem of model multiplicity in downstream decision-making.\nAll reviews have been cautiously positive with various suggestions for improving and strengthening the paper. In particular, the paper stud... | {
"decision": "Accept (Poster)"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.