Alibaba Group
AI Insights
- Regression: A type of prediction problem where the goal is to predict a continuous value. (ML: 0.98)
- The ablation studies reveal that ranking-centric training alone achieves robust ordinal performance, challenging conventional reliance on auxiliary regression supervision. (ML: 0.96)
- Two-stage training strategy: A training approach where the model is trained in two stages, with different objectives and rewards in each stage. (ML: 0.96)
- Ordinal ranking: A type of classification problem where the goal is to predict a rank or order for each sample. (ML: 0.96)
- RARL synergistically enhances regression accuracy and ranking performance through bidirectional regularization. (ML: 0.95)
- The work demonstrates the effectiveness of RARL in achieving state-of-the-art results on ordinal ranking tasks. (ML: 0.95)
- Ranking-aware verifiable rewards: Rewards that are designed to encourage the model to produce accurate rankings. (ML: 0.95)
- Extensive experiments demonstrate state-of-the-art results across three benchmarks, with ablation studies revealing that ranking-centric training alone achieves robust ordinal performance. (ML: 0.92)
- The work introduces RARL, an efficient and scalable framework for ordinal ranking. (ML: 0.91)
- The work relies heavily on the Qwen2.5-VL model, which may not be applicable to other models or tasks. (ML: 0.71)
Abstract
Ordinal regression and ranking are challenging due to inherent ordinal dependencies that conventional methods struggle to model. We propose Ranking-Aware Reinforcement Learning (RARL), a novel RL framework that explicitly learns these relationships. At its core, RARL features a unified objective that synergistically integrates regression and Learning-to-Rank (L2R), enabling mutual improvement between the two tasks. This is driven by a ranking-aware verifiable reward that jointly assesses regression precision and ranking accuracy, facilitating direct model updates via policy optimization. To further enhance training, we introduce Response Mutation Operations (RMO), which inject controlled noise to improve exploration and prevent stagnation at saddle points. The effectiveness of RARL is validated through extensive experiments on three distinct benchmarks.
Ordinal regression and ranking are challenging due to inherent ordinal dependencies that conventional methods struggle to model. We propose Ranking-Aware Reinforcement Learning (RARL), a novel RL framework that explicitly learns these relationships. At its core, RARL features a unified objective that synergistically integrates regression and Learning-to-Rank (L2R), enabling mutual improvement between the two tasks. This is driven by a ranking-aware verifiable reward that jointly assesses regression precision and ranking accuracy, facilitating direct model updates via policy optimization. To further enhance training, we introduce Response Mutation Operations (RMO), which inject controlled noise to improve exploration and prevent stagnation at saddle points. The effectiveness of RARL is validated through extensive experiments on three distinct benchmarks.
Why we are recommending this paper?
Due to your Interest in Ranking
This paper directly addresses ranking challenges, aligning with the user's interest in ranking and personalization. The use of reinforcement learning for ordinal regression is a key area of interest within deep learning and information retrieval.
Due to your Interest in Ranking
This paper directly addresses ranking challenges, aligning with the user's interest in ranking and personalization. The use of reinforcement learning for ordinal regression is a key area of interest within deep learning and information retrieval.
University of Liverpool
AI Insights
- Instance-dependent complexity: The complexity of an algorithm depends on the specific instance it is applied to, rather than just its worst-case performance. (ML: 0.96)
- Future directions include closing the constant-factor gap between the upper and lower bounds on instance-dependent complexity. (ML: 0.90)
- The paper assumes the existence of weak and strong oracles, which may not be feasible in all scenarios. (ML: 0.89)
- The algorithms ACE and ACE-W adaptively focus strong evaluations on critical items, yielding instance-dependent complexity governed by the near-tie mass. (ML: 0.88)
- PAC guarantees: Probably Approximately Correct Two-oracle framework: A framework that uses two types of oracles (weak and strong) to certify the exact top-k set. (ML: 0.81)
- The paper introduces a two-oracle framework for certifying the exact top-k set under PAC guarantees. (ML: 0.81)
- The algorithms ACE and ACE-W are shown to be effective in reducing strong oracle usage, making them suitable for applications where computational resources are limited. (ML: 0.80)
- The paper provides a new approach to certifying the exact top-k set under PAC guarantees, with significant improvements over existing methods. (ML: 0.78)
- The experiments show that ACE and ACE-W achieve significant reductions in strong oracle usage, with speedups of 2.4x and 2.8x over TA and STC respectively. (ML: 0.67)
Abstract
Identifying the top-$k$ items is fundamental but often prohibitive when exact valuations are expensive. We study a two-oracle setting with a fast, noisy weak oracle and a scarce, high-fidelity strong oracle (e.g., human expert verification or expensive simulation). We first analyze a simple screen-then-certify baseline (STC) and prove it makes at most $m(4\varepsilon_{\max})$ strong calls given jointly valid weak confidence intervals with maximum radius $\varepsilon_{\max}$, where $m(\cdot)$ denotes the near-tie mass around the top-$k$ threshold. We establish a conditional lower bound of $Ī©(m(\varepsilon_{\max}))$ for any algorithm given the same weak uncertainty. Our main contribution is ACE, an adaptive certification algorithm that focuses strong queries on critical boundary items, achieving the same $O(m(4\varepsilon_{\max}))$ bound while reducing strong calls in practice. We then introduce ACE-W, a fully adaptive two-phase method that allocates weak budget adaptively before running ACE, further reducing strong costs.
Identifying the top-$k$ items is fundamental but often prohibitive when exact valuations are expensive. We study a two-oracle setting with a fast, noisy weak oracle and a scarce, high-fidelity strong oracle (e.g., human expert verification or expensive simulation). We first analyze a simple screen-then-certify baseline (STC) and prove it makes at most $m(4\varepsilon_{\max})$ strong calls given jointly valid weak confidence intervals with maximum radius $\varepsilon_{\max}$, where $m(\cdot)$ denotes the near-tie mass around the top-$k$ threshold. We establish a conditional lower bound of $Ī©(m(\varepsilon_{\max}))$ for any algorithm given the same weak uncertainty. Our main contribution is ACE, an adaptive certification algorithm that focuses strong queries on critical boundary items, achieving the same $O(m(4\varepsilon_{\max}))$ bound while reducing strong calls in practice. We then introduce ACE-W, a fully adaptive two-phase method that allocates weak budget adaptively before running ACE, further reducing strong costs.
Why we are recommending this paper?
Due to your Interest in Ranking
The focus on adaptive ranking and utilizing both weak and strong oracles is highly relevant to the user's interest in personalization and search strategies. This work tackles a fundamental problem in information retrieval.
Due to your Interest in Ranking
The focus on adaptive ranking and utilizing both weak and strong oracles is highly relevant to the user's interest in personalization and search strategies. This work tackles a fundamental problem in information retrieval.
York University
AI Insights
- AUPRC: Area Under the Precision-Recall Curve Lack of spread: A measure of model performance that is not well-defined in this context. (ML: 0.97)
- Calibration slope: A measure of model calibration. (ML: 0.96)
- When using the second loss function (Lāā), which includes the AUPRC as a measure of discrimination, the optimal Mis proportion is much higher than under Lā. (ML: 0.96)
- CITL: Calibration-In-The-Large, a measure of model calibration. (ML: 0.96)
- The best performing model under Lā has an AUPRC value of 0.475 and a lack of spread value of 0.063 when Ī± = 0.1. (ML: 0.95)
- The results suggest that the choice of loss function and the value of Ī± have a significant impact on the performance of the model. (ML: 0.95)
- The proposed algorithm for tuning the size of subpopulation (Mis) is applied to a real-world dataset from the eICU cardiac database. (ML: 0.93)
- When using the first loss function (Lā), the optimal Mis proportion is 0.29 under all values of Ī± except Ī± = 0.9, where it is 0.20. (ML: 0.90)
- The results show that the optimal Mis value varies depending on the choice of alpha (Ī±), which controls the emphasis on discrimination and calibration in the loss function. (ML: 0.89)
Abstract
Advances in precision medicine increasingly drive methodological innovation in health research. A key development is the use of personalized prediction models (PPMs), which are fit using a similar subpopulation tailored to a specific index patient, and have been shown to outperform one-size-fits-all models, particularly in terms of model discrimination performance. We propose a generalized loss function that enables tuning of the subpopulation size used to fit a PPM. This loss function allows joint optimization of discrimination and calibration, allowing both the performance measures and their relative weights to be specified by the user. To reduce computational burden, we conducted extensive simulation studies to identify practical bounds for the grid of subpopulation sizes. Based on these results, we recommend using a lower bound of 20\% and an upper bound of 70\% of the entire training dataset. We apply the proposed method to both simulated and real-world datasets and demonstrate that previously observed relationships between subpopulation size and model performance are robust. Furthermore, we show that the choice of performance measures in the loss function influences the optimal subpopulation size selected. These findings support the flexible and computationally efficient implementation of PPMs in precision health research.
Advances in precision medicine increasingly drive methodological innovation in health research. A key development is the use of personalized prediction models (PPMs), which are fit using a similar subpopulation tailored to a specific index patient, and have been shown to outperform one-size-fits-all models, particularly in terms of model discrimination performance. We propose a generalized loss function that enables tuning of the subpopulation size used to fit a PPM. This loss function allows joint optimization of discrimination and calibration, allowing both the performance measures and their relative weights to be specified by the user. To reduce computational burden, we conducted extensive simulation studies to identify practical bounds for the grid of subpopulation sizes. Based on these results, we recommend using a lower bound of 20\% and an upper bound of 70\% of the entire training dataset. We apply the proposed method to both simulated and real-world datasets and demonstrate that previously observed relationships between subpopulation size and model performance are robust. Furthermore, we show that the choice of performance measures in the loss function influences the optimal subpopulation size selected. These findings support the flexible and computationally efficient implementation of PPMs in precision health research.
Why we are recommending this paper?
Due to your Interest in Personalization
This paper's exploration of personalized predictive models aligns with the user's interest in personalization and deep learning. The use of a mixture loss function is a sophisticated approach to model optimization.
Due to your Interest in Personalization
This paper's exploration of personalized predictive models aligns with the user's interest in personalization and deep learning. The use of a mixture loss function is a sophisticated approach to model optimization.
Plaksha University
AI Insights
- Further research is needed to refine the granularity of real-time difficulty calibration and explore the long-term effects of using such a system. (ML: 0.98)
- The study suggests that GuideAI's use of physiological data to inform adaptive interventions can lead to better learning outcomes and increased user engagement. (ML: 0.98)
- Further research is needed to refine the granularity of real-time difficulty calibration. (ML: 0.97)
- Previous studies have shown that personalized learning systems can improve learning outcomes, but few have explored the use of physiological data to inform adaptive interventions. (ML: 0.97)
- The study found that GuideAI, a real-time personalized learning solution with adaptive interventions, significantly improved learning outcomes and user experience compared to a control group. (ML: 0.96)
- The study demonstrates the potential of GuideAI's biosensor-driven approach to improve learning outcomes and user experience. (ML: 0.96)
- The study's sample size was relatively small. (ML: 0.95)
- GuideAI: A real-time personalized learning solution with adaptive interventions. (ML: 0.95)
- GuideAI's biosensor-driven interventions were rated positively by participants, who appreciated the system's ability to detect and respond to cognitive-affective shifts in real time. (ML: 0.94)
- Biosensor-driven interventions: Adaptive interventions informed by physiological data, such as heart rate or skin conductance, to adjust the learning experience in real time. (ML: 0.94)
Abstract
Large Language Models (LLMs) have emerged as powerful learning tools, but they lack awareness of learners' cognitive and physiological states, limiting their adaptability to the user's learning style. Contemporary learning techniques primarily focus on structured learning paths, knowledge tracing, and generic adaptive testing but fail to address real-time learning challenges driven by cognitive load, attention fluctuations, and engagement levels. Building on findings from a formative user study (N=66), we introduce GuideAI, a multi-modal framework that enhances LLM-driven learning by integrating real-time biosensory feedback including eye gaze tracking, heart rate variability, posture detection, and digital note-taking behavior. GuideAI dynamically adapts learning content and pacing through cognitive optimizations (adjusting complexity based on learning progress markers), physiological interventions (breathing guidance and posture correction), and attention-aware strategies (redirecting focus using gaze analysis). Additionally, GuideAI supports diverse learning modalities, including text-based, image-based, audio-based, and video-based instruction, across varied knowledge domains. A preliminary study (N = 25) assessed GuideAI's impact on knowledge retention and cognitive load through standardized assessments. The results show statistically significant improvements in both problem-solving capability and recall-based knowledge assessments. Participants also experienced notable reductions in key NASA-TLX measures including mental demand, frustration levels, and effort, while simultaneously reporting enhanced perceived performance. These findings demonstrate GuideAI's potential to bridge the gap between current LLM-based learning systems and individualized learner needs, paving the way for adaptive, cognition-aware education at scale.
Large Language Models (LLMs) have emerged as powerful learning tools, but they lack awareness of learners' cognitive and physiological states, limiting their adaptability to the user's learning style. Contemporary learning techniques primarily focus on structured learning paths, knowledge tracing, and generic adaptive testing but fail to address real-time learning challenges driven by cognitive load, attention fluctuations, and engagement levels. Building on findings from a formative user study (N=66), we introduce GuideAI, a multi-modal framework that enhances LLM-driven learning by integrating real-time biosensory feedback including eye gaze tracking, heart rate variability, posture detection, and digital note-taking behavior. GuideAI dynamically adapts learning content and pacing through cognitive optimizations (adjusting complexity based on learning progress markers), physiological interventions (breathing guidance and posture correction), and attention-aware strategies (redirecting focus using gaze analysis). Additionally, GuideAI supports diverse learning modalities, including text-based, image-based, audio-based, and video-based instruction, across varied knowledge domains. A preliminary study (N = 25) assessed GuideAI's impact on knowledge retention and cognitive load through standardized assessments. The results show statistically significant improvements in both problem-solving capability and recall-based knowledge assessments. Participants also experienced notable reductions in key NASA-TLX measures including mental demand, frustration levels, and effort, while simultaneously reporting enhanced perceived performance. These findings demonstrate GuideAI's potential to bridge the gap between current LLM-based learning systems and individualized learner needs, paving the way for adaptive, cognition-aware education at scale.
Why we are recommending this paper?
Due to your Interest in Personalization
Given the user's interest in personalization and deep learning, this paper's exploration of LLMs for adaptive learning is a strong match. The focus on real-time interventions and learner states is particularly relevant.
Due to your Interest in Personalization
Given the user's interest in personalization and deep learning, this paper's exploration of LLMs for adaptive learning is a strong match. The focus on real-time interventions and learner states is particularly relevant.
Tencent Youtu Lab
AI Insights
- The model is trained on two datasets: TopiOCQA and HotpotQA, with the goal of improving performance in CQA. (ML: 0.94)
- The authors use a two-stage CRL framework, where Stage I focuses on generating sub-queries and Stage II refines the generated sub-queries. (ML: 0.87)
- The paper presents an approach to conversational question answering (CQA) using reinforcement learning (RL). (ML: 0.86)
- The paper evaluates the performance of the model under both sparse and dense retrievers, using metrics such as MRR@K, NDCG@K, and MAP@10. (ML: 0.82)
- CRL: Constrained Reinforcement Learning BM25: A popular information retrieval algorithm ANCE: An efficient neural network-based retriever DAPO: A deep reinforcement learning framework for optimization ACQO: A model that generates concise sub-queries The proposed approach improves performance in CQA by using a two-stage CRL framework. (ML: 0.79)
Abstract
Query optimization is a crucial component for the efficacy of Retrieval-Augmented Generation (RAG) systems. While reinforcement learning (RL)-based agentic and reasoning methods have recently emerged as a promising direction on query optimization, most existing approaches focus on the expansion and abstraction of a single query. However, complex user queries are prevalent in real-world scenarios, often requiring multiple parallel and sequential search strategies to handle disambiguation and decomposition. Directly applying RL to these complex cases introduces significant hurdles. Determining the optimal number of sub-queries and effectively re-ranking and merging retrieved documents vastly expands the search space and complicates reward design, frequently leading to training instability. To address these challenges, we propose a novel RL framework called Adaptive Complex Query Optimization (ACQO). Our framework is designed to adaptively determine when and how to expand the search process. It features two core components: an Adaptive Query Reformulation (AQR) module that dynamically decides when to decompose a query into multiple sub-queries, and a Rank-Score Fusion (RSF) module that ensures robust result aggregation and provides stable reward signals for the learning agent. To mitigate training instabilities, we adopt a Curriculum Reinforcement Learning (CRL) approach, which stabilizes the training process by progressively introducing more challenging queries through a two-stage strategy. Our comprehensive experiments demonstrate that ACQO achieves state-of-the-art performance on three complex query benchmarks, significantly outperforming established baselines. The framework also showcases improved computational efficiency and broad compatibility with different retrieval architectures, establishing it as a powerful and generalizable solution for next-generation RAG systems.
Query optimization is a crucial component for the efficacy of Retrieval-Augmented Generation (RAG) systems. While reinforcement learning (RL)-based agentic and reasoning methods have recently emerged as a promising direction on query optimization, most existing approaches focus on the expansion and abstraction of a single query. However, complex user queries are prevalent in real-world scenarios, often requiring multiple parallel and sequential search strategies to handle disambiguation and decomposition. Directly applying RL to these complex cases introduces significant hurdles. Determining the optimal number of sub-queries and effectively re-ranking and merging retrieved documents vastly expands the search space and complicates reward design, frequently leading to training instability. To address these challenges, we propose a novel RL framework called Adaptive Complex Query Optimization (ACQO). Our framework is designed to adaptively determine when and how to expand the search process. It features two core components: an Adaptive Query Reformulation (AQR) module that dynamically decides when to decompose a query into multiple sub-queries, and a Rank-Score Fusion (RSF) module that ensures robust result aggregation and provides stable reward signals for the learning agent. To mitigate training instabilities, we adopt a Curriculum Reinforcement Learning (CRL) approach, which stabilizes the training process by progressively introducing more challenging queries through a two-stage strategy. Our comprehensive experiments demonstrate that ACQO achieves state-of-the-art performance on three complex query benchmarks, significantly outperforming established baselines. The framework also showcases improved computational efficiency and broad compatibility with different retrieval architectures, establishing it as a powerful and generalizable solution for next-generation RAG systems.
Why we are recommending this paper?
Due to your Interest in Search
This paper's exploration of query optimization using reinforcement learning aligns with the user's interests in search and deep learning. The focus on RAG systems and adaptive query strategies is highly relevant.
Due to your Interest in Search
This paper's exploration of query optimization using reinforcement learning aligns with the user's interests in search and deep learning. The focus on RAG systems and adaptive query strategies is highly relevant.
Karlsruhe Institute of Technology
AI Insights
- Large language model (LLM): a type of artificial intelligence model that is trained on large amounts of text data and can generate human-like language. (ML: 0.94)
- Heuristic: an approximate algorithm used to find a good solution to a complex problem, often used when exact algorithms are too slow or impractical. (ML: 0.94)
- The proposed method relies on the quality of the initial heuristic set, which may not be optimal for all problem instances. (ML: 0.94)
- Combinatorial optimization problem: a type of mathematical problem that involves finding the optimal solution among a finite set of possible solutions. (ML: 0.93)
- The paper presents a novel approach to automatic heuristic design using large language models (LLMs). (ML: 0.92)
- The paper demonstrates the potential of LLMs in automatic heuristic design, showing improved solution quality and computational efficiency compared to existing methods. (ML: 0.84)
- EOH-S requires significant computational resources to train and evaluate the LLMs. (ML: 0.83)
- The proposed method, called EOH-S, uses LLMs to evolve and improve heuristics for combinatorial optimization problems. (ML: 0.81)
- EOH-S provides a promising approach for solving complex combinatorial optimization problems, with applications in various fields such as logistics, finance, and engineering. (ML: 0.75)
- EOH-S is shown to outperform existing methods in terms of solution quality and computational efficiency. (ML: 0.69)
Abstract
Heuristic functions are essential to the performance of tree search algorithms such as A*, where their accuracy and efficiency directly impact search outcomes. Traditionally, such heuristics are handcrafted, requiring significant expertise. Recent advances in large language models (LLMs) and evolutionary frameworks have opened the door to automating heuristic design. In this paper, we extend the Evolution of Heuristics (EoH) framework to investigate the automated generation of guiding heuristics for A* search. We introduce a novel domain-agnostic prompt augmentation strategy that includes the A* code into the prompt to leverage in-context learning, named Algorithmic - Contextual EoH (A-CEoH). To evaluate the effectiveness of A-CeoH, we study two problem domains: the Unit-Load Pre-Marshalling Problem (UPMP), a niche problem from warehouse logistics, and the classical sliding puzzle problem (SPP). Our computational experiments show that A-CEoH can significantly improve the quality of the generated heuristics and even outperform expert-designed heuristics.
Heuristic functions are essential to the performance of tree search algorithms such as A*, where their accuracy and efficiency directly impact search outcomes. Traditionally, such heuristics are handcrafted, requiring significant expertise. Recent advances in large language models (LLMs) and evolutionary frameworks have opened the door to automating heuristic design. In this paper, we extend the Evolution of Heuristics (EoH) framework to investigate the automated generation of guiding heuristics for A* search. We introduce a novel domain-agnostic prompt augmentation strategy that includes the A* code into the prompt to leverage in-context learning, named Algorithmic - Contextual EoH (A-CEoH). To evaluate the effectiveness of A-CeoH, we study two problem domains: the Unit-Load Pre-Marshalling Problem (UPMP), a niche problem from warehouse logistics, and the classical sliding puzzle problem (SPP). Our computational experiments show that A-CEoH can significantly improve the quality of the generated heuristics and even outperform expert-designed heuristics.
Why we are recommending this paper?
Due to your Interest in Search
Due to your Interest in Search
Universit Paris Cit , CNRS
AI Insights
- The authors show that any feedforward network with ReLU activations can be viewed as a place-independent IFS, and they extend this result to other types of neural networks, including residual blocks and MoE models. (ML: 0.92)
- The paper discusses the interpretation of deep neural networks as iterated function systems (IFSs) and provides a general framework for analyzing their convergence properties. (ML: 0.89)
- The paper provides several examples of neural network architectures that can be interpreted as IFSs, including ResNet with Softplus activation, Transformer block, and MoE model. (ML: 0.88)
- The authors use the Hutchinson operator to analyze the convergence properties of IFSs and show that they can be used to bound the Wasserstein distance between the output of a neural network and its fixed point. (ML: 0.87)
- Definition 1: A Markov recursion is a sequence of random variables {Xn} defined by X0 = x and Xt+1 = w(Xt, Ī), where w is a function that depends on the current state Xt and the parameter Ī. (ML: 0.82)
- Definition 3: A place-dependent IFS (P-IFS) is an IFS {wĪ¾} where each wĪ¾ depends on the current state x and the parameter Ī. (ML: 0.80)
- Definition 2: An iterated function system (IFS) is a collection of functions {wĪ¾} indexed by Ī¾ ā I, where each wĪ¾ is a Lipschitz map from X to itself. (ML: 0.80)
- They also introduce the concept of strong average Lipschitz contractivity for place-dependent IFSs and provide conditions under which it holds. (ML: 0.75)
- Definition 5: A P-IFS {wĪ¾} is strongly average-contractive if sup_xāX ā_{Ī¾āI} pĪ¾(x)cĪ¾ ā¤ c < 1. (ML: 0.67)
- Definition 4: The Hutchinson operator T is a contraction on the space of probability measures PP(X) with respect to the Wasserstein distance W2 if there exists a constant c < 1 such that W2(T(Āµ), T(Ī½)) ā¤ cW2(Āµ, Ī½) for all Āµ, Ī½ ā PP(X). (ML: 0.67)
- The Hutchinson operator T is defined as T(Āµ) = ā_{Ī¾āI} pwĪ¾#Āµq. (ML: 0.49)
Abstract
Deep neural networks (DNNs) achieve remarkable performance on a wide range of tasks, yet their mathematical analysis remains fragmented: stability and generalization are typically studied in disparate frameworks and on a case-by-case basis. Architecturally, DNNs rely on the recursive application of parametrized functions, a mechanism that can be unstable and difficult to train, making stability a primary concern. Even when training succeeds, there are few rigorous results on how well such models generalize beyond the observed data, especially in the generative setting. In this work, we leverage the theory of stochastic Iterated Function Systems (IFS) and show that two important deep architectures can be viewed as, or canonically associated with, place-dependent IFS. This connection allows us to import results from random dynamical systems to (i) establish the existence and uniqueness of invariant measures under suitable contractivity assumptions, and (ii) derive a Wasserstein generalization bound for generative modeling. The bound naturally leads to a new training objective that directly controls the collage-type approximation error between the data distribution and its image under the learned transfer operator. We illustrate the theory on a controlled 2D example and empirically evaluate the proposed objective on standard image datasets (MNIST, CelebA, CIFAR-10).
Deep neural networks (DNNs) achieve remarkable performance on a wide range of tasks, yet their mathematical analysis remains fragmented: stability and generalization are typically studied in disparate frameworks and on a case-by-case basis. Architecturally, DNNs rely on the recursive application of parametrized functions, a mechanism that can be unstable and difficult to train, making stability a primary concern. Even when training succeeds, there are few rigorous results on how well such models generalize beyond the observed data, especially in the generative setting. In this work, we leverage the theory of stochastic Iterated Function Systems (IFS) and show that two important deep architectures can be viewed as, or canonically associated with, place-dependent IFS. This connection allows us to import results from random dynamical systems to (i) establish the existence and uniqueness of invariant measures under suitable contractivity assumptions, and (ii) derive a Wasserstein generalization bound for generative modeling. The bound naturally leads to a new training objective that directly controls the collage-type approximation error between the data distribution and its image under the learned transfer operator. We illustrate the theory on a controlled 2D example and empirically evaluate the proposed objective on standard image datasets (MNIST, CelebA, CIFAR-10).
Why we are recommending this paper?
Due to your Interest in Deep Learning
Due to your Interest in Deep Learning
Harvard University
AI Insights
- Milestones serve dual pedagogical and validation purposes, providing motivation through historical framing and demonstrating implementation correctness through real-world task performance. (ML: 0.98)
- Each module concludes with systems reasoning prompts measuring conceptual understanding beyond syntactic correctness. (ML: 0.97)
- Milestones are designed to be challenging but achievable, allowing students to demonstrate their understanding of complex concepts through real-world tasks. (ML: 0.96)
- Assessment validates both isolated correctness and cross-module integration. (ML: 0.96)
- The TinyTorch framework is designed for teaching machine learning concepts through hands-on implementation and analysis. (ML: 0.95)
- Reflect: Systems Analysis Questions. (ML: 0.94)
- TinyTorch follows a consistent Build-Use-Reflect cycle, integrating implementation, application, and systems reasoning to address multiple learning objectives. (ML: 0.94)
- It's a pedagogical tool aimed at bridging the gap between theoretical understanding and practical application. (ML: 0.94)
- Students implement components in Jupyter notebooks with scaffolded guidance. (ML: 0.91)
- TinyTorch's design emphasizes systems thinking, encouraging students to analyze and understand the relationships between components, rather than just focusing on individual functions. (ML: 0.87)
- The framework includes six historical milestones that recreate actual breakthroughs using exclusively student code, validating success through task-appropriate performance. (ML: 0.85)
- The framework is built with a focus on explicit dependencies, making it easier for students to understand where each module fits in the larger architecture. (ML: 0.83)
- Use: Integration Testing Beyond Unit Tests. (ML: 0.77)
- Build: Implementation with Explicit Dependencies. (ML: 0.66)
Abstract
Machine learning education faces a fundamental gap: students learn algorithms without understanding the systems that execute them. They study gradient descent without measuring memory, attention mechanisms without analyzing O(N^2) scaling, optimizer theory without knowing why Adam requires 3x the memory of SGD. This "algorithm-systems divide" produces practitioners who can train models but cannot debug memory failures, optimize inference latency, or reason about deployment trade-offs--the very skills industry demands as "ML systems engineering." We present TinyTorch, a 20-module curriculum that closes this gap through "implementation-based systems pedagogy": students construct PyTorch's core components (tensors, autograd, optimizers, CNNs, transformers) in pure Python, building a complete framework where every operation they invoke is code they wrote. The design employs three patterns: "progressive disclosure" of complexity, "systems-first integration" of profiling from the first module, and "build-to-validate milestones" recreating 67 years of ML breakthroughs--from Perceptron (1958) through Transformers (2017) to MLPerf-style benchmarking. Requiring only 4GB RAM and no GPU, TinyTorch demonstrates that deep ML systems understanding is achievable without specialized hardware. The curriculum is available open-source at mlsysbook.ai/tinytorch.
Machine learning education faces a fundamental gap: students learn algorithms without understanding the systems that execute them. They study gradient descent without measuring memory, attention mechanisms without analyzing O(N^2) scaling, optimizer theory without knowing why Adam requires 3x the memory of SGD. This "algorithm-systems divide" produces practitioners who can train models but cannot debug memory failures, optimize inference latency, or reason about deployment trade-offs--the very skills industry demands as "ML systems engineering." We present TinyTorch, a 20-module curriculum that closes this gap through "implementation-based systems pedagogy": students construct PyTorch's core components (tensors, autograd, optimizers, CNNs, transformers) in pure Python, building a complete framework where every operation they invoke is code they wrote. The design employs three patterns: "progressive disclosure" of complexity, "systems-first integration" of profiling from the first module, and "build-to-validate milestones" recreating 67 years of ML breakthroughs--from Perceptron (1958) through Transformers (2017) to MLPerf-style benchmarking. Requiring only 4GB RAM and no GPU, TinyTorch demonstrates that deep ML systems understanding is achievable without specialized hardware. The curriculum is available open-source at mlsysbook.ai/tinytorch.
Why we are recommending this paper?
Due to your Interest in Deep Learning
Due to your Interest in Deep Learning
Google LLC
AI Insights
- Domain Generalization: The ability of a model to generalize to new domains without significant performance degradation. (ML: 0.98)
- It emphasizes the importance of understanding query distributions during inference when designing retrieval systems. (ML: 0.98)
- Distributional overfitting refers to the model's inability to generalize to new domains, while structural blindness refers to the loss of fine-grained entity distinctions and term matching due to geometric compression. (ML: 0.98)
- The authors identify two primary failure modes: distributional overfitting and structural blindness. (ML: 0.96)
- The paper highlights the importance of understanding query distributions during inference when designing retrieval systems. (ML: 0.96)
- Structural Blindness: The loss of fine-grained entity distinctions and term matching due to geometric compression in dense retrievers. (ML: 0.95)
- The paper highlights the limitations of dense retrieval architectures, particularly in terms of domain generalization and structural blindness. (ML: 0.93)
- It also discusses the limitations of dense retrievers in resolving queries contingent on rare entities, serial numbers, or specific proper nouns. (ML: 0.90)
- Dense Retrieval: A method of information retrieval that uses a neural network to map queries and documents into a shared vector space. (ML: 0.90)
- The paper discusses the challenges and limitations of dense retrieval architectures in information retrieval systems. (ML: 0.85)
Abstract
Designing an embedding retrieval system requires navigating a complex design space of conflicting trade-offs between efficiency and effectiveness. This work structures these decisions as a vertical traversal of the system design stack. We begin with the Representation Layer by examining how loss functions and architectures, specifically Bi-encoders and Cross-encoders, define semantic relevance and geometric projection. Next, we analyze the Granularity Layer and evaluate how segmentation strategies like Atomic and Hierarchical chunking mitigate information bottlenecks in long-context documents. Moving to the Orchestration Layer, we discuss methods that transcend the single-vector paradigm, including hierarchical retrieval, agentic decomposition, and multi-stage reranking pipelines to resolve capacity limitations. Finally, we address the Robustness Layer by identifying architectural mitigations for domain generalization failures, lexical blind spots, and the silent degradation of retrieval quality due to temporal drift. By categorizing these limitations and design choices, we provide a comprehensive framework for practitioners to optimize the efficiency-effectiveness frontier in modern neural search systems.
Designing an embedding retrieval system requires navigating a complex design space of conflicting trade-offs between efficiency and effectiveness. This work structures these decisions as a vertical traversal of the system design stack. We begin with the Representation Layer by examining how loss functions and architectures, specifically Bi-encoders and Cross-encoders, define semantic relevance and geometric projection. Next, we analyze the Granularity Layer and evaluate how segmentation strategies like Atomic and Hierarchical chunking mitigate information bottlenecks in long-context documents. Moving to the Orchestration Layer, we discuss methods that transcend the single-vector paradigm, including hierarchical retrieval, agentic decomposition, and multi-stage reranking pipelines to resolve capacity limitations. Finally, we address the Robustness Layer by identifying architectural mitigations for domain generalization failures, lexical blind spots, and the silent degradation of retrieval quality due to temporal drift. By categorizing these limitations and design choices, we provide a comprehensive framework for practitioners to optimize the efficiency-effectiveness frontier in modern neural search systems.
Why we are recommending this paper?
Due to your Interest in Information Retrieval
Due to your Interest in Information Retrieval
University of Helsinki
AI Insights
- Using a separate sample of actual queries for training LEMUR yields a consistent performance improvement on HotpotQA and ViDoRe. (ML: 0.92)
- BEIR: A benchmark for evaluating information retrieval models. (ML: 0.91)
- The embedding dimension ablation study shows that LEMUR performs better with higher embedding dimensions, especially for k= 100. (ML: 0.88)
- MuVERA: Another multi-vector embedding model compared to LEMUR in the paper. (ML: 0.87)
- The model is robust to hyperparameters and uses the same hyperparameters for all experiments in this paper. (ML: 0.85)
- On all six datasets, LEMUR significantly outperforms the baseline methods, including ColBERTv2. (ML: 0.85)
- LEMUR is a learned multi-vector retrieval model that outperforms baseline methods on six BEIR datasets. (ML: 0.80)
- LEMUR is a robust and effective learned multi-vector retrieval model that outperforms baseline methods on six BEIR datasets. (ML: 0.78)
- LEMR: Learned Multi-Vector Retrieval, a model that outperforms baseline methods on six BEIR datasets. (ML: 0.76)
- ColBERTv2: A baseline method used in the experiments. (ML: 0.76)
Abstract
Multi-vector representations generated by late interaction models, such as ColBERT, enable superior retrieval quality compared to single-vector representations in information retrieval applications. In multi-vector retrieval systems, both queries and documents are encoded using one embedding for each token, and similarity between queries and documents is measured by the MaxSim similarity measure. However, the improved recall of multi-vector retrieval comes at the expense of significantly increased latency. This necessitates designing efficient approximate nearest neighbor search (ANNS) algorithms for multi-vector search. In this work, we introduce LEMUR, a simple-yet-efficient framework for multi-vector similarity search. LEMUR consists of two consecutive problem reductions: We first formulate multi-vector similarity search as a supervised learning problem that can be solved using a one-hidden-layer neural network. Second, we reduce inference under this model to single-vector similarity search in its latent space, which enables the use of existing single-vector ANNS methods for speeding up retrieval. In addition to performance evaluation on ColBERTv2 embeddings, we evaluate LEMUR on embeddings generated by modern multi-vector text models and multi-vector visual document retrieval models. LEMUR is an order of magnitude faster than earlier multi-vector similarity search methods.
Multi-vector representations generated by late interaction models, such as ColBERT, enable superior retrieval quality compared to single-vector representations in information retrieval applications. In multi-vector retrieval systems, both queries and documents are encoded using one embedding for each token, and similarity between queries and documents is measured by the MaxSim similarity measure. However, the improved recall of multi-vector retrieval comes at the expense of significantly increased latency. This necessitates designing efficient approximate nearest neighbor search (ANNS) algorithms for multi-vector search. In this work, we introduce LEMUR, a simple-yet-efficient framework for multi-vector similarity search. LEMUR consists of two consecutive problem reductions: We first formulate multi-vector similarity search as a supervised learning problem that can be solved using a one-hidden-layer neural network. Second, we reduce inference under this model to single-vector similarity search in its latent space, which enables the use of existing single-vector ANNS methods for speeding up retrieval. In addition to performance evaluation on ColBERTv2 embeddings, we evaluate LEMUR on embeddings generated by modern multi-vector text models and multi-vector visual document retrieval models. LEMUR is an order of magnitude faster than earlier multi-vector similarity search methods.
Why we are recommending this paper?
Due to your Interest in Information Retrieval
Due to your Interest in Information Retrieval
š¬ 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