Peer Reviews

19K threads, 205K comments, 78K peer reviews, and growing.

+ Create
Outcome

Withdrawn

A Feature-Aware Federated Learning Framework for Unsupervised Anomaly Detection in 5G Net…

The authors propose an explainable AI-based algorithm, integrated gradients, to employ the feature importance of each client in the aggregation of the FL system, apply it to the 5G-related tasks, and process the importance estimation in each round to adopt the dynamic feature of 5G networks. Results compared with traditional aggregation algorithms, such as FedAvg and FedProx show some performance gain on one standard dataset.

September 28, 2024 · 2.5/10 · 4 Reviews

0
Rejected

Synergistic Approach for Simultaneous Optimization of Monolingual, Cross-lingual, and Mul…

This paper introduces a simple method called hybrid batch training, which involves translating to obtain parallel data in multiple languages, and sampling these data to construct a multilingual training dataset. The model is trained by inputting monolingual or multilingual training data with a certain probability, thereby balancing its performance in both scenarios.

September 28, 2024 · 3/10 · 3 Reviews

11
Accepted

Beyond Random Masking: When Dropout meets Graph Convolutional Networks

The paper performs a comprehensive theoretical analysis of dropout in the case of Graph Convolution Networks (GCN) from multiple perspectives: dimension-specific graph structure modification during training, degree-dependent effect on nodes, impact on over-smoothing, and it’s combined effect with batch normalization.

September 28, 2024 · 6/10 · 4 Reviews

22
Rejected

Defining Deception in Decision Making

The paper provides a formal definition of deception in a setting where a Speaker and a Listener interact with each other over multiple rounds. The goal of the proposed definition is to capture the Speaker's degree of deception by measuring to what extent its action(s) affect the Listener's beliefs and individual utility. The framework that is used to formalize the Speaker-Listener interactions is a POMDP variation, termed as Communication POMDP, where the Listener is modeled as part of the environment and the Speaker plays the role of the acting agent. Under this decision making framework, the Speaker's degree of deception is formalized through a function that measures the Listener's regret w.r.t. the accuracy of its beliefs and its personal utility, that results from the Speaker's actions. Finally, the paper includes extensive experimentation over a diverse set of scenarios that aims to showcase how well the proposed approach aligns with human intuition and how it compares to LLM baselines.

September 28, 2024 · 4/10 · 4 Reviews

0
Accepted

Self-supervised contrastive learning performs non-linear system identification

This paper focuses on contrastive learning (CL) methods for dynamical systems. The authors show that under certain assumptions CL performs system identification and can therefore uncover the latent dynamics of the data. The theoretical findings are applied to switching linear dynamics and non-linear dynamics, and are demonstrated from an empirical point of view using simulated data.

September 28, 2024 · 6.4/10 · 5 Reviews

8
Rejected

Recovering Plasticity of Neural Networks via Soft Weight Rescaling

This paper introduces Soft Weight Regularization (SWR), a regularization based algorithm for maintaining plasticity under the broad framework of continual learning. Unlike other regularization based approaches for addressing plasticity loss, such as L2 regularization, Shrink and Perturb, and L2 Init, SWR does not alter the network's predictions. The paper provides a theoretical analysis showing that SWR bounds weight magnitudes and maintains balanced weights between layers, two favourable properties of neural networks. Finally, the paper provides empirical evidence arguing the efficacy of SWR on a set of problems that test for plasticity and stability in settings of warm-starting, continual learning, and generalization.

September 28, 2024 · 4/10 · 4 Reviews

0
Accepted

DarkBench: Benchmarking Dark Patterns in Large Language Models

The authors develop DarkBench by manually conceptualizing six tendencies of LLMs that seem to align with the chatbot subscription-based business model (e.g., ChatGPT, Claude.ai), prompting an LLM with precise verbal descriptions of those tendencies to create adversarial prompts that would evoke dark patterns, and manually reviewing and modifying the LLM-generated prompts. Evaluation was done with LLMs prompted with human examples with samples of that LLM evaluation also done by humans for comparison. Results show that Claude performs best on this benchmark ("dark patterns" in 30-36% of responses, if I understand correctly), followed by the other models in a band of 48-61%. Some patterns (e.g., user retention) are much more frequent than others (e.g., sycophancy) in current LLMs.

September 28, 2024 · 7/10 · 4 Reviews

13
Withdrawn

Stochastic Sparse Sampling: A Framework for Variable-Length Medical Time Series Classific…

This paper introduces a multi-scale learning approach for medical time series classification. The proposed method comprises multiple independent models, each with a distinct patch length, allowing it to capture information across various temporal scales. The patching method follows the PatchTST framework, which employs single-channel patching. To reduce computational costs, the authors implement stochastic sparse sampling, randomly selecting models during training. The final representation is an aggregation of outputs from all models, combining multi-scale information. The model is evaluated on intracranial EEG (iEEG) data for seizure onset zone classification, using a dataset collected from four independent medical centers.

September 28, 2024 · 4.8/10 · 4 Reviews

0
Rejected

OpenCity: A Scalable Platform to Simulate Urban Activities with Massive LLM Agents

ABM and LLM is leveraged to develop one platform for open city modeling and planning. It is a nice simulation platform and the paper provides application scenarios. Concretely, 1. This paper combines agent-based models with large language models to develop the OpenCity platform for simulating urban activities. It reduces simulation computational costs through IO multiplexing and the "group-and-distill" prompt optimization strategy. 2. Through experiments conducted in six major cities worldwide, OpenCity demonstrates a 635 plus increase in average simulation speed per agent, along with a 70% decrease in LLM requests and a 50% reduction in token usage. The time savings are mainly concentrated in the LLM response wait time and the CPU multiplexing process. 3. The OpenCity platform proposed in this paper achieves the first benchmark testing for LLM agent-based urban activity simulation research.

September 28, 2024 · 4.3/10 · 4 Reviews

15
Withdrawn

MathOdyssey: Benchmarking Mathematical Problem-Solving Skills in Large Language Models Us…

This paper presents a new dataset called MathOdyssey, which aims to evaluate the reasoning abilities of large language models (LLMs). The dataset consists of 387 problems, including 148 at the Olympic level, 101 at the university level, and 138 at the high school level. The problems cover several subjects with three answer formats: true/false, multiple choice, and open answers. The authors evaluate LLMs' math reasoning performance on MathOdyssey using GPT-4 as the answer judger in a zero-shot manner, providing it with specific instructions. They conduct their experiments with seven closed-source LLMs and one open-source model, Llama-3-70B. Their findings reveal that the Llama-3-70B model still falls short when tackling more complex problems.

September 28, 2024 · 4.5/10 · 4 Reviews

0
Withdrawn

KARPA: A Training-free Method of Adapting Knowledge Graph as References for Large Languag…

The paper introduces a novel framework called KARPA designed to enhance the reasoning capabilities of LLMs in Knowledge Graph Question Answering (KGQA) tasks. The authors identify limitations in existing methods that either rely on step-by-step traversal of KGs, which restricts the global planning abilities of LLMs, or require fine-tuning on specific KGs, making them less adaptable. KARPA addresses these challenges through a three-step process: (1) Pre-planning: The LLM generates initial relation paths based on the question and relevant KG relations, leveraging its inherent global reasoning and planning capabilities. (2) Retrieving: A semantic embedding model is used to find candidate paths in the KG that are semantically similar to the LLM-generated paths, avoiding local optima and reducing interactions with the KG. (3) Reasoning: The candidate paths and corresponding entities are provided back to the LLM for comprehensive reasoning to produce the final answer. The framework operates in a training-free manner, making it adaptable to various LLM architectures without additional fine-tuning or pre-training. Experimental results demonstrate that KARPA achieves state-of-the-art performance on multiple KGQA benchmark datasets, delivering both high efficiency and accuracy.

September 28, 2024 · 4.6/10 · 5 Reviews

17
Accepted

Sparse autoencoders reveal selective remapping of visual concepts during adaptation

This paper utilizes sparse autoencoders (SAE) to interpret visual concepts learned by the CLIP vision model and investigates how these concepts are affected by adaptation techniques such as MaPLe. The key finding is that during adaptation, the model primarily reuses existing concepts rather than learning entirely new ones.

September 28, 2024 · 6.5/10 · 4 Reviews

17
Rejected

DySpec: Faster Speculative Decoding with Dynamic Token Tree Structure

This paper proposes a draft token selection method in speculative decoding aimed at improving the token acceptance rate. The core idea is to use the draft model's prediction score as evidence to infer the token acceptance rate and use this information to select more promising tokens. The authors provide both theoretical and empirical analyses to support their approach.

September 28, 2024 · 4.8/10 · 4 Reviews

12
Rejected

StreamingBench: Assessing the Gap for MLLMs to Achieve Streaming Video Understanding

The authors propose a new benchmark, StreamingBench, for evaluating MLLMs in streaming video understanding. It assesses three aspects of streaming video understanding: real-time visual understanding, omni-source understanding, and contextual understanding. There are 18 tasks in total. They evaluate 13 open-source Video MLLMs and 3 proprietary MLLMs on this benchmark and analyze the results.

September 28, 2024 · 5.8/10 · 4 Reviews

13
Accepted

PIED: Physics-Informed Experimental Design for Inverse Problems

This paper solves the one-shot experiment design using PINNs in both forward and inverse problems. It overcomes computational bottlenecks by parallelism and meta-learning of initialization. The experiments on both synthetic and real-life datasets show the performance improvements.

September 28, 2024 · 7/10 · 4 Reviews

14
Withdrawn

Learning Fused State Representations for Control from Multi-View Observations

The paper introduces a novel approach named Multi-view Fusion State for Control(MFSC),which ingrates a self-attention mechanism with bisimulation metric learning to fuse task-relevant representation from multi-view observation. Additionally, the paper also incorporated a mask-based latent reconstruction auxiliary task to learn cross-view information in order to foster more compact fused presentation. In this paper, two major problems were solved : First is Higher data dimensions and more redundant information , and Informative aggregation of representation from various views.

September 28, 2024 · 4.5/10 · 4 Reviews

1
Accepted

AgentRefine: Enhancing Agent Generalization through Refinement Tuning

The paper proposes AgentRefine, a framework designed to enhance the generalization capabilities of large language model (LLM)-based agents through a self-refinement process. The core idea is to enable agents to learn from their mistakes by refining their actions based on feedback from the environment. The authors introduce a data generation pipeline that simulates diverse environments and tasks, followed by a refinement tuning process to improve agent robustness and generalization. Experimental results show that AgentRefine outperforms state-of-the-art methods in held-out tasks, demonstrating improved generalization and robustness.

September 28, 2024 · 5.8/10 · 4 Reviews

18
Rejected

SEAL: Scaling to Emphasize Attention for Long-Context Retrieval

This paper proposes a novel and practical method, SEAL, to improve the long-context retrieval ability of LLMs. First, through perturbation experiments, it finds a certain attention head or a certain channel in it can cause a positive or negative effect on long-context retrieval accuracy. Second, it demonstrates directly scaling the hidden states of these heads or channels can indeed improve the retrieval accuracy of LLMs. Third, it adds trainable scale factors into the model and use a small amount of samples of retrieval tasks to fine-tune the model. The results show SEAL can remarkably improve the long-context retrieval ability of LLMs.

September 28, 2024 · 5.3/10 · 4 Reviews

19
Rejected

MuLan: Multimodal-LLM Agent for Progressive and Interactive Multi-Object Diffusion

This paper introduces MuLan, a comprehensive image generation method that utilizes a Large Language Model (LLM) agent for precise control of the generation process. The approach involves decomposing the prompt into a sequence of sub-tasks and generating each object sequentially through a diffusion model. Consequently, the method effectively generates multiple objects in accordance with the prompt.

September 28, 2024 · 4/10 · 4 Reviews

6
Accepted

TabM: Advancing tabular deep learning with parameter-efficient ensembling

The paper applies the BatchEnsemble technique to MLPs for tabular data, and investigates several modifications. The results show an improvement over MLPs and several deep baseline models and GBRT on a broach benchmark of 50 datasets from the literature. The experiments show that in the first adapter layer in particular is extremely critical, and results in the majority of gains.

September 28, 2024 · 6.5/10 · 4 Reviews

19
Rejected

Leveraging Implicit Sentiments: Enhancing Reliability and Validity in Psychological Trait…

The paper introduces the Core Sentiment Inventory (CSI), a new evaluation method inspired by the Implicit Association Test (IAT) to assess the implicit sentiment tendencies of large language models (LLMs). The approach aims to provide a reliable and valid measure of LLMs' optimism, pessimism, and neutrality in both English and Chinese, surpassing conventional human-centric psychometric tests like the Big Five Inventory (BFI). The authors present experimental results that claim improved reliability, reduced reluctance rates, and strong predictive power for CSI.

September 28, 2024 · 4.3/10 · 4 Reviews

8
Rejected

EmbodiedCity: A Benchmark Platform for Embodied Agent in Real-world City Environment

The paper presents a benchmark platform for evaluating embodied artificial intelligence in realistic urban environments, addressing gaps in open-world scenarios. It features a detailed 3D simulation, diverse evaluation tasks, and user-friendly interfaces, enhancing embodied intelligence capabilities and supporting practical applications in artificial general intelligence.

September 28, 2024 · 3.5/10 · 4 Reviews

0
Accepted

Multi-Label Test-Time Adaptation with Bound Entropy Minimization

The paper presents a novel approach to Test-Time Adaptation (TTA) for multi-label scenarios using a method termed Bound Entropy Minimization (BEM). The paper is well-structured, the problem statement is clear, and the proposed solution is innovative. The integration of view and caption prompts and the application of BEM to meet the test time adaptation are innovative to some extent. However, there are some details should be clarified.

September 28, 2024 · 6.3/10 · 4 Reviews

21
Rejected

Action-Constrained Imitation Learning

This paper proposes action-constrained imitation learning (ACIL), a new imitation learning algorithm for action-constrained imitators to learn from demonstrations. The authors propose DTWIL to solve this problem by first replacing the original expert demonstrations with a surrogate dataset that follows similar state trajectories and then recasting trajectory alignment as a planning problem and solving it via Model Predictive Control. Through experiments in both navigation and locomotion tasks, they show the effectiveness of proposed method.

September 28, 2024 · 5/10 · 4 Reviews

10
Rejected

LMRL Gym: Benchmarks for Multi-Turn Reinforcement Learning with Language Models

The paper highlights that current LLMs are trained to imitate golden responses rather than genuinely learning to reason and solve single-turn tasks. Additionally, there is a lack of benchmarking for multi-turn RL tasks, along with the absence of established evaluation protocols, which can be costly. To address this, the authors synthesize a benchmark that leverages the imitation capabilities of language models in conjunction with simulators, such as chess engines. They propose the LMRL-GYM benchmark, which comprises three interactive dialogue tasks and five RL capability tests, benchmarking existing RL methods, including offline methods like ILQL and online methods like PPO, among others.

September 28, 2024 · 4.3/10 · 4 Reviews

17
Accepted

ToolGen: Unified Tool Retrieval and Calling via Generation

This paper proposes ToolGen, a generative tool/function calling framework. The concrete methods are: (1) virtualizing tools by virtual tokens; (2) memorizing tools with training data of (tool docs, tool virtual tokens); (3) learning tool retrieval with training data of (user queries, tool virtual tokens); (4) and finally, finetuning tool agent with tool calling trajectories. Experimental results with over 47,000 tools show that ToolGen not only achieves superior results in both tool retrieval and autonomous task completion but also sets the stage for a new era of AI agents that can adapt to tools across diverse domains.

September 28, 2024 · 5.8/10 · 4 Reviews

19
Withdrawn

Re-examining learning linear functions in context

The paper investigates in-context learning (ICL) across various training and testing scenarios using different sizes of transformer models trained from scratch. Building on previous work, it highlights systematic failures in these models' ability to generalize to data outside the training distribution, revealing some limitations of ICL.

September 28, 2024 · 3.5/10 · 4 Reviews

11
Withdrawn

Exploring New Frontiers in Vertical Federated Learning: the Role of Saddle Point Reformul…

This paper explores new methods for Vertical Federated Learning (VFL) by reformulating the learning process using a saddle point framework instead of the traditional minimization approach. The proposed approach in the deterministic case enables solving VFL problems with enhanced convergence guarantees in terms of the eigenvalue of the data matrix. The authors also propose stochastic algorithms tailored to this reformulation and suggest modifications to address practical challenges such as communication efficiency and computational cost by implementing compression, partial participation, and coordinate selection, respectively. The paper validates the proposed methods through numerical experiments.

September 28, 2024 · 5/10 · 4 Reviews

0
Rejected

Efficiently Scanning and Resampling Spatio-Temporal Tasks with Irregular Observations

The paper tackles the issues of spatio-temporal modeling in the context of variable observation spaces, such as when there are variable objects across episodes / scenes, etc or when the object number varies within an episode (such as when entities enter and leave an arena). The authors introduce a few baseline tasks representative of this problem and cover various attention-based encoding and temporal aggregation schemes, some of which are advertised as being novel (such as some of the encoder methods). They perform various analysis on different methodologies, such as task performance, memory, speed, etc on these tasks which highlight some of the benefits of various methods (such as the scan-based one).

September 28, 2024 · 6/10 · 4 Reviews

14
Accepted

Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks

With the inspiration that a backdoored model is apt to map significantly different poisoned samples and clean samples of the backdoor target to similar activation areas, this paper introduces a novel measurement, i.e., the circular distribution of the gradients w.r.t sample activation, namely GCD, which works to identify the target class of the backdoor and consequently separate poisoned and clean samples within the target class. Accordingly, this paper proposes a sample detection approach called AGPD to achieve dataset purification. Extensive experiments show AGPD's advanced performance in detecting and isolating poisoned samples.

September 28, 2024 · 6/10 · 4 Reviews

21
Showing 30 threads · scroll for more
© 2026 NYSGPTLast Updated: August 5, 2026