Leiden University
AI Insights
- It characterizes ranking behavior through three steerable directions β decision, evidence, and role β and shows that ranking performance can be effectively calibrated at inference time through simple projection-based interventions. (ML: 0.99)
- The method can be applied to other ranking paradigms and retrieval settings, making it a promising area of research. (ML: 0.96)
- Steerable directions: The three dimensions through which the activation steering is performed β decision, evidence, and role. (ML: 0.95)
- The method outperforms state-of-the-art pointwise LLM ranking methods on TREC-DL 2020 and the BEIR benchmarks with three backbone LLMs. (ML: 0.93)
- RankSteer is a post-hoc activation steering method for pointwise Large Language Model (LLM) rankers. (ML: 0.90)
- RankSteer preserves the inference complexity of standard pointwise ranking, making it a computationally efficient post-hoc calibration mechanism for pointwise LLM rankers. (ML: 0.90)
- Activation steering: A method that directly manipulates the activation geometry of an LLM to improve its performance on specific tasks. (ML: 0.90)
- RankSteer provides an efficient post-hoc calibration mechanism for pointwise LLM rankers, recovering under-utilized ranking capacity with far fewer labeled queries than fine-tuning. (ML: 0.89)
- The method requires a large amount of labeled data to train the anchor queries. (ML: 0.85)
- Pointwise LLM ranker: An LLM-based ranking system that generates a ranked list of documents based on their relevance to a given query. (ML: 0.83)
Abstract
Large language models (LLMs) have recently shown strong performance as zero-shot rankers, yet their effectiveness is highly sensitive to prompt formulation, particularly role-play instructions. Prior analyses suggest that role-related signals are encoded along activation channels that are largely separate from query-document representations, raising the possibility of steering ranking behavior directly at the activation level rather than through brittle prompt engineering. In this work, we propose RankSteer, a post-hoc activation steering framework for zero-shot pointwise LLM ranking. We characterize ranking behavior through three disentangled and steerable directions in representation space: a \textbf{decision direction} that maps hidden states to relevance scores, an \textbf{evidence direction} that captures relevance signals not directly exploited by the decision head, and a \textbf{role direction} that modulates model behavior without injecting relevance information. Using projection-based interventions at inference time, RankSteer jointly controls these directions to calibrate ranking behavior without modifying model weights or introducing explicit cross-document comparisons. Experiments on TREC DL 20 and multiple BEIR benchmarks show that RankSteer consistently improves ranking quality using only a small number of anchor queries, demonstrating that substantial ranking capacity remains under-utilized in pointwise LLM rankers. We further provide a geometric analysis revealing that steering improves ranking by stabilizing ranking geometry and reducing dispersion, offering new insight into how LLMs internally represent and calibrate relevance judgments.
Large language models (LLMs) have recently shown strong performance as zero-shot rankers, yet their effectiveness is highly sensitive to prompt formulation, particularly role-play instructions. Prior analyses suggest that role-related signals are encoded along activation channels that are largely separate from query-document representations, raising the possibility of steering ranking behavior directly at the activation level rather than through brittle prompt engineering. In this work, we propose RankSteer, a post-hoc activation steering framework for zero-shot pointwise LLM ranking. We characterize ranking behavior through three disentangled and steerable directions in representation space: a \textbf{decision direction} that maps hidden states to relevance scores, an \textbf{evidence direction} that captures relevance signals not directly exploited by the decision head, and a \textbf{role direction} that modulates model behavior without injecting relevance information. Using projection-based interventions at inference time, RankSteer jointly controls these directions to calibrate ranking behavior without modifying model weights or introducing explicit cross-document comparisons. Experiments on TREC DL 20 and multiple BEIR benchmarks show that RankSteer consistently improves ranking quality using only a small number of anchor queries, demonstrating that substantial ranking capacity remains under-utilized in pointwise LLM rankers. We further provide a geometric analysis revealing that steering improves ranking by stabilizing ranking geometry and reducing dispersion, offering new insight into how LLMs internally represent and calibrate relevance judgments.
Why we are recommending this paper?
Due to your Interest in Ranking
This paper directly addresses ranking, a core interest, by exploring how LLMs can be steered for improved ranking performance. The focus on activation steering within LLMs aligns strongly with the user's interest in personalization and search techniques.
Due to your Interest in Ranking
This paper directly addresses ranking, a core interest, by exploring how LLMs can be steered for improved ranking performance. The focus on activation steering within LLMs aligns strongly with the user's interest in personalization and search techniques.
Palo Alto Networks
AI Insights
- While the RL agent learns from engagement signals, incorporating cognitive load and user satisfaction as explicit reward components could further improve adaptation quality. (ML: 0.96)
- The current framework assumes stable user behavior distributions and relies on anonymized interaction data, which may not capture all contextual factors. (ML: 0.96)
- Predictive modeling: A technique used to forecast user behavior based on past interactions and contextual factors. (ML: 0.95)
- Results indicate that AI-driven personalization increases engagement metrics by up to 30% while reducing average interaction latency. (ML: 0.95)
- Reinforcement learning: An AI approach that enables systems to learn from rewards or penalties received after taking actions in an environment. (ML: 0.94)
- A key insight from this work is the synergistic relationship between sequence prediction and reward-based optimization, which balances immediacy and sustained engagement. (ML: 0.90)
- The proposed system integrates predictive modeling and reinforcement learning within a unified personalization loop to enhance front-end adaptability. (ML: 0.90)
- The study presents a concrete step toward practical, real-time AI-driven personalization in production-scale front-end systems. (ML: 0.86)
- Unified personalization loop: A framework that combines predictive modeling, reinforcement learning, and adaptive rendering to continuously improve UI adaptation. (ML: 0.86)
- By coupling prediction, optimization, and adaptive rendering within a single feedback architecture, it establishes a foundation for next-generation user interfaces that evolve intelligently alongside their users. (ML: 0.82)
Abstract
Front-end personalization has traditionally relied on static designs or rule-based adaptations, which fail to fully capture user behavior patterns. This paper presents an AI driven approach for dynamic front-end personalization, where UI layouts, content, and features adapt in real-time based on predicted user behavior. We propose three strategies: dynamic layout adaptation using user path prediction, content prioritization through reinforcement learning, and a comparative analysis of AI-driven vs. rule-based personalization. Technical implementation details, algorithms, system architecture, and evaluation methods are provided to illustrate feasibility and performance gains.
Front-end personalization has traditionally relied on static designs or rule-based adaptations, which fail to fully capture user behavior patterns. This paper presents an AI driven approach for dynamic front-end personalization, where UI layouts, content, and features adapt in real-time based on predicted user behavior. We propose three strategies: dynamic layout adaptation using user path prediction, content prioritization through reinforcement learning, and a comparative analysis of AI-driven vs. rule-based personalization. Technical implementation details, algorithms, system architecture, and evaluation methods are provided to illustrate feasibility and performance gains.
Why we are recommending this paper?
Due to your Interest in Personalization
Given the user's interest in personalization, this paper's exploration of AI-driven front-end adaptation is highly relevant. The approach of dynamically adapting UI layouts based on user behavior directly addresses the need for tailored experiences.
Due to your Interest in Personalization
Given the user's interest in personalization, this paper's exploration of AI-driven front-end adaptation is highly relevant. The approach of dynamically adapting UI layouts based on user behavior directly addresses the need for tailored experiences.
Karlsruhe Institute of Technology
AI Insights
- Arithmetic encoding: This is a method for compressing binary data by representing it as a single number. (ML: 0.92)
- Compact representation: This refers to a data structure that uses less memory than the original data. (ML: 0.89)
- Rank/select queries: These are operations that count the number of 1s (or 0s) before a given position in a bit vector, or select the first 1 (or 0) after a given position. (ML: 0.86)
- The proposed method is compared with existing methods, including Spider, Sux, and Movi, in terms of memory usage and query performance. (ML: 0.85)
- Bit vectors: These are arrays of bits used to represent binary data. (ML: 0.84)
- The paper discusses the development of a new data structure for rank/select queries on bit vectors. (ML: 0.79)
- The proposed method is suitable for applications where memory usage is limited, such as in embedded systems or mobile devices. (ML: 0.72)
- The authors propose a compact representation of the data structure using a combination of bitwise operations and arithmetic encoding. (ML: 0.72)
- The results show that the new data structure achieves better memory efficiency and faster query times than the existing methods. (ML: 0.72)
- The new data structure achieves better memory efficiency and faster query times than existing methods. (ML: 0.72)
Abstract
Given a text, a query $\mathsf{rank}(q, c)$ counts the number of occurrences of character $c$ among the first $q$ characters of the text. Space-efficient methods to answer these rank queries form an important building block in many succinct data structures. For example, the FM-index is a widely used data structure that uses rank queries to locate all occurrences of a pattern in a text. In bioinformatics applications, the goal is usually to process a given input as fast as possible. Thus, data structures should have high throughput when used with many threads. Contributions. For the binary alphabet, we develop BiRank with 3.28% space overhead. It merges the central ideas of two recent papers: (1) we interleave (inline) offsets in each cache line of the underlying bit vector [Laws et al., 2024], reducing cache-misses, and (2) these offsets are to the middle of each block so that only half of them need popcounting [Gottlieb and Reinert, 2025]. In QuadRank (14.4% space overhead), we extend these techniques to the $Ο=4$ (DNA) alphabet. Both data structures require only a single cache miss per query, making them highly suitable for high-throughput and memory-bound settings. To enable efficient batch-processing, we support prefetching the cache lines required to answer upcoming queries. Results. BiRank and QuadRank are around $1.5\times$ and $2\times$ faster than similar-overhead methods that do not use inlining. Prefetching gives an additional $2\times$ speedup, at which point the dual-channel DDR4 RAM bandwidth becomes a hard limit on the total throughput. With prefetching, both methods outperform all other methods apart from SPIDER [Laws et al., 2024] by $2\times$. When using QuadRank with prefetching in a toy count-only FM-index, QuadFm, this results in a smaller size and up to $4\times$ speedup over Genedex, a state-of-the-art batching FM-index implementation.
Given a text, a query $\mathsf{rank}(q, c)$ counts the number of occurrences of character $c$ among the first $q$ characters of the text. Space-efficient methods to answer these rank queries form an important building block in many succinct data structures. For example, the FM-index is a widely used data structure that uses rank queries to locate all occurrences of a pattern in a text. In bioinformatics applications, the goal is usually to process a given input as fast as possible. Thus, data structures should have high throughput when used with many threads. Contributions. For the binary alphabet, we develop BiRank with 3.28% space overhead. It merges the central ideas of two recent papers: (1) we interleave (inline) offsets in each cache line of the underlying bit vector [Laws et al., 2024], reducing cache-misses, and (2) these offsets are to the middle of each block so that only half of them need popcounting [Gottlieb and Reinert, 2025]. In QuadRank (14.4% space overhead), we extend these techniques to the $Ο=4$ (DNA) alphabet. Both data structures require only a single cache miss per query, making them highly suitable for high-throughput and memory-bound settings. To enable efficient batch-processing, we support prefetching the cache lines required to answer upcoming queries. Results. BiRank and QuadRank are around $1.5\times$ and $2\times$ faster than similar-overhead methods that do not use inlining. Prefetching gives an additional $2\times$ speedup, at which point the dual-channel DDR4 RAM bandwidth becomes a hard limit on the total throughput. With prefetching, both methods outperform all other methods apart from SPIDER [Laws et al., 2024] by $2\times$. When using QuadRank with prefetching in a toy count-only FM-index, QuadFm, this results in a smaller size and up to $4\times$ speedup over Genedex, a state-of-the-art batching FM-index implementation.
Why we are recommending this paper?
Due to your Interest in Ranking
This paper's focus on efficient ranking methods, particularly using character counts, aligns with the user's interest in search and information retrieval. The emphasis on high-throughput rank queries is a key aspect of effective ranking systems.
Due to your Interest in Ranking
This paper's focus on efficient ranking methods, particularly using character counts, aligns with the user's interest in search and information retrieval. The emphasis on high-throughput rank queries is a key aspect of effective ranking systems.
Seoul National University
AI Insights
- Vision-Language Models (VLMs): A type of artificial intelligence model that combines natural language processing and computer vision capabilities. (ML: 0.97)
- The paper discusses the concept of contextualized visual personalization in vision-language models (VLMs), which involves tailoring the model's responses to a user's specific context and preferences. (ML: 0.97)
- MCQA Accuracy: The accuracy of a VLM's ability to answer multiple-choice questions accurately. (ML: 0.97)
- Contextualized Visual Personalization: The ability of a VLM to tailor its responses to a user's specific context and preferences. (ML: 0.96)
- The paper concludes by emphasizing the need for more research on contextualized visual personalization and the development of new evaluation metrics and benchmarks to assess VLMs' performance in this area. (ML: 0.96)
- The authors highlight the importance of addressing issues such as object hallucination, where the model generates objects that are not present in the input image, and entity name inclusion, where the model fails to include relevant entities in its response. (ML: 0.93)
- The authors propose a new benchmark for evaluating VLMs' ability to personalize their responses, which includes tasks such as object hallucination, entity name inclusion, and positive MCQA accuracy. (ML: 0.93)
- The paper also discusses various techniques for improving VLMs' personalization capabilities, including retrieval-augmented personalization, knowledge-augmented large language models, and multi-concept customization of text-to-image diffusion. (ML: 0.92)
- Entity Name Inclusion: When a VLM fails to include relevant entities in its response. (ML: 0.89)
- Object Hallucination: When a VLM generates objects that are not present in the input image. (ML: 0.83)
Abstract
Despite recent progress in vision-language models (VLMs), existing approaches often fail to generate personalized responses based on the user's specific experiences, as they lack the ability to associate visual inputs with a user's accumulated visual-textual context. We newly formalize this challenge as contextualized visual personalization, which requires the visual recognition and textual retrieval of personalized visual experiences by VLMs when interpreting new images. To address this issue, we propose CoViP, a unified framework that treats personalized image captioning as a core task for contextualized visual personalization and improves this capability through reinforcement-learning-based post-training and caption-augmented generation. We further introduce diagnostic evaluations that explicitly rule out textual shortcut solutions and verify whether VLMs truly leverage visual context. Extensive experiments demonstrate that existing open-source and proprietary VLMs exhibit substantial limitations, while CoViP not only improves personalized image captioning but also yields holistic gains across downstream personalization tasks. These results highlight CoViP as a crucial stage for enabling robust and generalizable contextualized visual personalization.
Despite recent progress in vision-language models (VLMs), existing approaches often fail to generate personalized responses based on the user's specific experiences, as they lack the ability to associate visual inputs with a user's accumulated visual-textual context. We newly formalize this challenge as contextualized visual personalization, which requires the visual recognition and textual retrieval of personalized visual experiences by VLMs when interpreting new images. To address this issue, we propose CoViP, a unified framework that treats personalized image captioning as a core task for contextualized visual personalization and improves this capability through reinforcement-learning-based post-training and caption-augmented generation. We further introduce diagnostic evaluations that explicitly rule out textual shortcut solutions and verify whether VLMs truly leverage visual context. Extensive experiments demonstrate that existing open-source and proprietary VLMs exhibit substantial limitations, while CoViP not only improves personalized image captioning but also yields holistic gains across downstream personalization tasks. These results highlight CoViP as a crucial stage for enabling robust and generalizable contextualized visual personalization.
Why we are recommending this paper?
Due to your Interest in Personalization
Considering the user's interest in deep learning and information retrieval, this paper's exploration of personalized responses using VLMs is a strong match. The focus on associating visual inputs with user context is a valuable area of research.
Due to your Interest in Personalization
Considering the user's interest in deep learning and information retrieval, this paper's exploration of personalized responses using VLMs is a strong match. The focus on associating visual inputs with user context is a valuable area of research.
University of Essen
AI Insights
- The use of large language models as annotators has shown potential in improving the quality of annotated data. (ML: 0.98)
- AI-augmented annotation: The use of artificial intelligence to assist in annotating data, such as labeling or categorizing it. (ML: 0.98)
- The system is designed to allocate tasks between humans and AI models based on their respective strengths and weaknesses. (ML: 0.97)
- Human-AI collaboration: The process of working together between humans and AI models to accomplish a task, such as data annotation. (ML: 0.97)
- The authors propose a novel approach to human-AI collaboration in data annotation, using large language models as annotators. (ML: 0.97)
- The proposed system has shown promising results in improving efficiency and accuracy in data annotation tasks. (ML: 0.94)
- The evaluation framework provides a comprehensive set of metrics for measuring the quality of generated text, which can be used to evaluate the performance of retrieval-augmented generation systems. (ML: 0.92)
- The paper presents an evaluation framework for retrieval-augmented generation systems, which includes metrics for measuring the quality of generated text. (ML: 0.91)
- The paper discusses the development of AI-augmented annotation systems for information retrieval tasks, with a focus on improving efficiency and accuracy. (ML: 0.90)
- Retrieval-augmented generation: A type of natural language processing that uses retrieval techniques to generate text based on the input provided. (ML: 0.87)
Abstract
The rise of Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) has rapidly increased the need for high-quality, curated information retrieval datasets. These datasets, however, are currently created with off-the-shelf annotation tools that make the annotation process complex and inefficient. To streamline this process, we developed a specialized annotation tool - AIANO. By adopting an AI-augmented annotation workflow that tightly integrates human expertise with LLM assistance, AIANO enables annotators to leverage AI suggestions while retaining full control over annotation decisions. In a within-subject user study ($n = 15$), participants created question-answering datasets using both a baseline tool and AIANO. AIANO nearly doubled annotation speed compared to the baseline while being easier to use and improving retrieval accuracy. These results demonstrate that AIANO's AI-augmented approach accelerates and enhances dataset creation for information retrieval tasks, advancing annotation capabilities in retrieval-intensive domains.
The rise of Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) has rapidly increased the need for high-quality, curated information retrieval datasets. These datasets, however, are currently created with off-the-shelf annotation tools that make the annotation process complex and inefficient. To streamline this process, we developed a specialized annotation tool - AIANO. By adopting an AI-augmented annotation workflow that tightly integrates human expertise with LLM assistance, AIANO enables annotators to leverage AI suggestions while retaining full control over annotation decisions. In a within-subject user study ($n = 15$), participants created question-answering datasets using both a baseline tool and AIANO. AIANO nearly doubled annotation speed compared to the baseline while being easier to use and improving retrieval accuracy. These results demonstrate that AIANO's AI-augmented approach accelerates and enhances dataset creation for information retrieval tasks, advancing annotation capabilities in retrieval-intensive domains.
Why we are recommending this paper?
Due to your Interest in Information Retrieval
This paper directly addresses information retrieval, a key interest for the user, by proposing a method for enhancing datasets with AI-augmented annotation. The use of LLMs and RAG aligns with the user's interest in modern retrieval techniques.
Due to your Interest in Information Retrieval
This paper directly addresses information retrieval, a key interest for the user, by proposing a method for enhancing datasets with AI-augmented annotation. The use of LLMs and RAG aligns with the user's interest in modern retrieval techniques.
Northwestern University
AI Insights
- Gradual resolution of uncertainty: the process by which consumers gradually learn about the quality and prices of products over time. (ML: 0.98)
- Benchmark model: a model without hidden fees used for comparison purposes. (ML: 0.95)
- The paper concludes that gradual resolution of uncertainty leads to higher prices and lower consumer welfare, but regulation can mitigate these effects. (ML: 0.95)
- It also highlights the importance of considering the impact of regulatory policies on equilibrium prices and consumer welfare in markets with hidden fees. (ML: 0.94)
- It finds that gradual resolution of uncertainty leads to higher prices and lower consumer welfare compared to a benchmark model without hidden fees. (ML: 0.94)
- However, regulation that prohibits hidden fees can lead to lower prices and higher consumer welfare. (ML: 0.93)
- The paper studies the effect of gradual resolution of uncertainty on prices and consumer welfare in a monopolistic competition model with hidden fees. (ML: 0.91)
- Hidden fees: fees charged by firms to consumers that are not explicitly disclosed. (ML: 0.91)
- Monopolistic competition: a market structure where firms compete with each other but also have some degree of market power. (ML: 0.88)
- The paper also examines the effect of regulation on equilibrium prices in unregulated and regulated markets and finds that it increases consumer surplus. (ML: 0.83)
Abstract
I introduce and study a nested search problem modeled as a tree structure that generalizes Weitzman (1979) in two ways: (1) search progresses incrementally, reflecting real-life scenarios where agents gradually acquire information about the prizes; and (2) the realization of prizes can be correlated, capturing similarities among them. I derive the optimal policy, which takes the form of an index solution. I apply this result to study monopolistic competition in a market with two stages of product inspection. My application illustrates that regulations on drip pricing lower equilibrium price and raise consumer surplus.
I introduce and study a nested search problem modeled as a tree structure that generalizes Weitzman (1979) in two ways: (1) search progresses incrementally, reflecting real-life scenarios where agents gradually acquire information about the prizes; and (2) the realization of prizes can be correlated, capturing similarities among them. I derive the optimal policy, which takes the form of an index solution. I apply this result to study monopolistic competition in a market with two stages of product inspection. My application illustrates that regulations on drip pricing lower equilibrium price and raise consumer surplus.
Why we are recommending this paper?
Due to your Interest in Search
Due to your Interest in Search
The George Washington University
AI Insights
- The analysis shows that the reflection mechanism reduces the effective variance of the value estimation, leading to a lower regret bound. (ML: 0.90)
- Corollary 2 provides an upper bound on the order of the regret bound, showing that it grows at most logarithmically with T. (ML: 0.88)
- The proof of Theorem 1, which establishes the regret bound for Agent Alpha, relies on Freedman's Inequality and the unconfirmed knowledge framework introduced by Kato. (ML: 0.85)
- K: set of actions (arms) at a given expansion step Β΅a: true expected value of action a aβ: optimal action with mean Β΅β βa: optimality gap for suboptimal action a Ο2res,a: conditional residual variance Cn,a(Ξ΄): confidence radius for arm a at time n (ML: 0.83)
- The final bound is given by RTβ€XaΜΈ=aβ 8Ο2res,alnT βa+16 lnT/3+ 2βa!. (ML: 0.53)
Abstract
While scaling test-time compute through trajectory-level sampling has significantly improved Graphical User Interface (GUI) agents, the lack of regressive ability prevents the reuse of partial successes and the recovery from early missteps. In this paper, we introduce Agent Alpha, a unified framework that synergizes generation, exploration, and evaluation through step-level Monte Carlo Tree Search (MCTS). It enables active modeling or exploiting structures of the planning space. By integrating alpha-UCT guided search into the interaction loop, Agent Alpha enables deliberate planning, facilitating early pruning of suboptimal branches and efficient prefix reuse. We also employ comparison-driven evaluation to mitigate absolute scoring biases and diversity-constrained expansion to maintain a compact, informative search space. Regret bound of alpha-UCT is analyzed. On the OSWorld benchmark, Agent Alpha achieves a state-of-the-art success rate of $\sim 77\%$, significantly outperforming trajectory-level baselines under equivalent compute.
While scaling test-time compute through trajectory-level sampling has significantly improved Graphical User Interface (GUI) agents, the lack of regressive ability prevents the reuse of partial successes and the recovery from early missteps. In this paper, we introduce Agent Alpha, a unified framework that synergizes generation, exploration, and evaluation through step-level Monte Carlo Tree Search (MCTS). It enables active modeling or exploiting structures of the planning space. By integrating alpha-UCT guided search into the interaction loop, Agent Alpha enables deliberate planning, facilitating early pruning of suboptimal branches and efficient prefix reuse. We also employ comparison-driven evaluation to mitigate absolute scoring biases and diversity-constrained expansion to maintain a compact, informative search space. Regret bound of alpha-UCT is analyzed. On the OSWorld benchmark, Agent Alpha achieves a state-of-the-art success rate of $\sim 77\%$, significantly outperforming trajectory-level baselines under equivalent compute.
Why we are recommending this paper?
Due to your Interest in Search
Due to your Interest in Search
Luxembourg Institute of Science and Technology
AI Insights
- Additionally, the generated neural network architectures may not always outperform state-of-the-art models in various tasks. (ML: 0.98)
- They also discuss the limitations and challenges associated with this approach. (ML: 0.98)
- This can help improve the performance of various machine learning tasks such as image classification, object detection, and natural language processing. (ML: 0.97)
- However, it relies heavily on the capabilities of LLMs, which may not be available to all researchers or practitioners. (ML: 0.97)
- The proposed method relies heavily on the capabilities of LLMs, which may not be available to all researchers or practitioners. (ML: 0.97)
- The paper proposes a method for generating neural network architectures using large language models (LLMs). (ML: 0.92)
- The authors cite several papers that demonstrate the effectiveness of using LLMs for generating neural network architectures. (ML: 0.92)
- The authors demonstrate the effectiveness of their approach by generating neural network architectures that outperform state-of-the-art models in several tasks. (ML: 0.92)
- The authors demonstrate the effectiveness of their approach by generating neural network architectures that outperform state-of-the-art models in several tasks. (ML: 0.92)
- The paper presents a novel approach to generating neural network architectures using large language models (LLMs). (ML: 0.91)
- The paper proposes a new way to generate neural network architectures using large language models (LLMs). (ML: 0.91)
- The authors propose a method that leverages the capabilities of LLMs to generate neural network architectures, which can be used for various tasks such as image classification, object detection, and natural language processing. (ML: 0.91)
- LLM: Large Language Model The proposed method for generating neural network architectures using LLMs is a promising approach that can be used to improve the performance of various machine learning tasks. (ML: 0.89)
- The proposed method is based on a combination of two techniques: instruction-guided autoregressive neural network parameter generation and tabular data generation using agentic LLM methods. (ML: 0.85)
Abstract
Neural networks are increasingly used to support decision-making. To verify their reliability and adaptability, researchers and practitioners have proposed a variety of tools and methods for tasks such as NN code verification, refactoring, and migration. These tools play a crucial role in guaranteeing both the correctness and maintainability of neural network architectures, helping to prevent implementation errors, simplify model updates, and ensure that complex networks can be reliably extended and reused. Yet, assessing their effectiveness remains challenging due to the lack of publicly diverse datasets of neural networks that would allow systematic evaluation. To address this gap, we leverage large language models (LLMs) to automatically generate a dataset of neural networks that can serve as a benchmark for validation. The dataset is designed to cover diverse architectural components and to handle multiple input data types and tasks. In total, 608 samples are generated, each conforming to a set of precise design choices. To further ensure their consistency, we validate the correctness of the generated networks using static analysis and symbolic tracing. We make the dataset publicly available to support the community in advancing research on neural network reliability and adaptability.
Neural networks are increasingly used to support decision-making. To verify their reliability and adaptability, researchers and practitioners have proposed a variety of tools and methods for tasks such as NN code verification, refactoring, and migration. These tools play a crucial role in guaranteeing both the correctness and maintainability of neural network architectures, helping to prevent implementation errors, simplify model updates, and ensure that complex networks can be reliably extended and reused. Yet, assessing their effectiveness remains challenging due to the lack of publicly diverse datasets of neural networks that would allow systematic evaluation. To address this gap, we leverage large language models (LLMs) to automatically generate a dataset of neural networks that can serve as a benchmark for validation. The dataset is designed to cover diverse architectural components and to handle multiple input data types and tasks. In total, 608 samples are generated, each conforming to a set of precise design choices. To further ensure their consistency, we validate the correctness of the generated networks using static analysis and symbolic tracing. We make the dataset publicly available to support the community in advancing research on neural network reliability and adaptability.
Why we are recommending this paper?
Due to your Interest in Deep Learning
Due to your Interest in Deep Learning
Beijing University of Posts and Telecommunications
AI Insights
- Scale-invariant components: Components in neural networks that are invariant to scaling, such as BatchNorm layers. (ML: 0.92)
- Architecture-aware updates/projections: Updates and projections that are aware of the architecture of the neural network, such as BatchNorm layers. (ML: 0.92)
- Curvature-adaptive radial step sizing: An adaptive learning rate scheme that adjusts the step size based on the curvature of the loss function. (ML: 0.89)
- Its performance is superior to other algorithms, including AdamW and AdamP, on both CIFAR-100 and modular-arithmetic Grokking tasks. (ML: 0.88)
- The paper evaluates AdamO on CIFAR-100 and modular-arithmetic Grokking tasks, showing that it outperforms other optimization algorithms, including AdamW and AdamP. (ML: 0.84)
- Orthogonal dynamics: A method for optimizing neural networks by decoupling the update rules for different dimensions. (ML: 0.81)
- AdamO is designed to handle scale-invariant components in neural networks, such as BatchNorm, by using projections to suppress ineffective updates. (ML: 0.81)
- AdamO is a robust and effective optimization algorithm for deep learning tasks, particularly those involving scale-invariant components. (ML: 0.71)
- AdamO's performance is robust across a wide range of hyperparameters, making it easier to tune and use in practice. (ML: 0.67)
- The paper proposes a new adaptive optimization algorithm called AdamO, which is fully decoupled orthogonal dynamics with curvature-adaptive radial step sizing and architecture-aware updates/projections. (ML: 0.61)
Abstract
Is the standard weight decay in AdamW truly optimal? Although AdamW decouples weight decay from adaptive gradient scaling, a fundamental conflict remains: the Radial Tug-of-War. In deep learning, gradients tend to increase parameter norms to expand effective capacity while steering directions to learn features, whereas weight decay indiscriminately suppresses norm growth. This push--pull interaction induces radial oscillations, injecting noise into Adam's second-moment estimates and potentially degrading delicate tangential feature learning. We argue that magnitude and direction play distinct roles and should be decoupled in optimizer dynamics. We propose Orthogonal Dynamics Decoupling and instantiate it as AdamO: an SGD-style update handles the one-dimensional norm control, while Adam's adaptive preconditioning is confined to the tangential subspace. AdamO further incorporates curvature-adaptive radial step sizing and architecture-aware rules and projections for scale-invariant layers and low-dimensional parameters. Experiments on vision and language tasks show that AdamO improves generalization and stability over AdamW without introducing additional complex constraints.
Is the standard weight decay in AdamW truly optimal? Although AdamW decouples weight decay from adaptive gradient scaling, a fundamental conflict remains: the Radial Tug-of-War. In deep learning, gradients tend to increase parameter norms to expand effective capacity while steering directions to learn features, whereas weight decay indiscriminately suppresses norm growth. This push--pull interaction induces radial oscillations, injecting noise into Adam's second-moment estimates and potentially degrading delicate tangential feature learning. We argue that magnitude and direction play distinct roles and should be decoupled in optimizer dynamics. We propose Orthogonal Dynamics Decoupling and instantiate it as AdamO: an SGD-style update handles the one-dimensional norm control, while Adam's adaptive preconditioning is confined to the tangential subspace. AdamO further incorporates curvature-adaptive radial step sizing and architecture-aware rules and projections for scale-invariant layers and low-dimensional parameters. Experiments on vision and language tasks show that AdamO improves generalization and stability over AdamW without introducing additional complex constraints.
Why we are recommending this paper?
Due to your Interest in Deep Learning
Due to your Interest in Deep Learning
Beihang University
AI Insights
- The performance of the model on SciFact and NF-Corpus with qwen-8b and gritlm shows marginal variability and overall robustness. (ML: 0.96)
- The performance of the model on SciFact and NF-Corpus with qwen-8b and gritlm shows marginal variability and overall robustness, which may indicate some sensitivity to hyperparameters. (ML: 0.94)
- The method's performance on SciFact and NF-Corpus with qwen-8b and gritlm shows marginal variability and overall robustness. (ML: 0.94)
- Hard-negative pool size (K): The number of hard negatives in the pool, which is used to compute the loss function. (ML: 0.91)
- Temperature: The temperature parameter is used to control the softening or hardening of the negative sampling distribution. (ML: 0.89)
- In-batch negatives (M): The number of negative samples generated for each positive sample in a batch, which are used to compute the loss function. (ML: 0.86)
- It is set to a default value of 0.01. (ML: 0.80)
- It is set to a default value of 1000. (ML: 0.79)
- Hyperparameter search ranges used in the experiments are provided in Table 6. (ML: 0.78)
- Hyperparameter search ranges used in the experiments are provided in Table 6. (ML: 0.78)
- The method is largely insensitive to initialization randomness and other stochastic factors. (ML: 0.75)
- The method is largely insensitive to initialization randomness and other stochastic factors. (ML: 0.75)
- It is set to a default value of 64. (ML: 0.71)
Abstract
Dense retrieval represents queries and docu-002 ments as high-dimensional embeddings, but003 these representations can be redundant at the004 query level: for a given information need, only005 a subset of dimensions is consistently help-006 ful for ranking. Prior work addresses this via007 pseudo-relevance feedback (PRF) based dimen-008 sion importance estimation, which can produce009 query-aware masks without labeled data but010 often relies on noisy pseudo signals and heuris-011 tic test-time procedures. In contrast, super-012 vised adapter methods leverage relevance labels013 to improve embedding quality, yet they learn014 global transformations shared across queries015 and do not explicitly model query-aware di-016 mension importance. We propose a Query-017 Aware Adaptive Dimension Selection frame-018 work that learns to predict per-dimension im-019 portance directly from query embedding. We020 first construct oracle dimension importance dis-021 tributions over embedding dimensions using022 supervised relevance labels, and then train a023 predictor to map a query embedding to these024 label-distilled importance scores. At inference,025 the predictor selects a query-aware subset of026 dimensions for similarity computation based027 solely on the query embedding, without pseudo-028 relevance feedback. Experiments across multi-029 ple dense retrievers and benchmarks show that030 our learned dimension selector improves re-031 trieval effectiveness over the full-dimensional032 baseline as well as PRF-based masking and033 supervised adapter baselines.
Dense retrieval represents queries and docu-002 ments as high-dimensional embeddings, but003 these representations can be redundant at the004 query level: for a given information need, only005 a subset of dimensions is consistently help-006 ful for ranking. Prior work addresses this via007 pseudo-relevance feedback (PRF) based dimen-008 sion importance estimation, which can produce009 query-aware masks without labeled data but010 often relies on noisy pseudo signals and heuris-011 tic test-time procedures. In contrast, super-012 vised adapter methods leverage relevance labels013 to improve embedding quality, yet they learn014 global transformations shared across queries015 and do not explicitly model query-aware di-016 mension importance. We propose a Query-017 Aware Adaptive Dimension Selection frame-018 work that learns to predict per-dimension im-019 portance directly from query embedding. We020 first construct oracle dimension importance dis-021 tributions over embedding dimensions using022 supervised relevance labels, and then train a023 predictor to map a query embedding to these024 label-distilled importance scores. At inference,025 the predictor selects a query-aware subset of026 dimensions for similarity computation based027 solely on the query embedding, without pseudo-028 relevance feedback. Experiments across multi-029 ple dense retrievers and benchmarks show that030 our learned dimension selector improves re-031 trieval effectiveness over the full-dimensional032 baseline as well as PRF-based masking and033 supervised adapter baselines.
Why we are recommending this paper?
Due to your Interest in Information Retrieval
Due to your Interest in Information Retrieval
π¬ Help Shape Our Pricing
We're exploring pricing options to make this project sustainable. Take 3 minutes to share what you'd be willing to pay (if anything). Your input guides our future investment.
Help us improve your experience!
This project is on its early stages your feedback can be pivotal on the future of the project.
Let us know what you think about this week's papers and suggestions!
Give Feedback