Personalized Daily Arxiv Papers 02/26/2025

Total ArXiv papers: 625

Total scanned papers: 384

Total relevant papers: 37

Table of contents with paper titles:

1. Reversal Blessing: Thinking Backward May Outpace Thinking Forward in Multi-choice Questions Authors: Yizhe Zhang, Richard Bai, Zijin Gu, Ruixiang Zhang, Jiatao Gu, Emmanuel Abbe, Samy Bengio, Navdeep Jaitly

2. Mechanistic PDE Networks for Discovery of Governing Equations Authors: Adeel Pervez, Efstratios Gavves, Francesco Locatello

3. How Do Large Language Monkeys Get Their Power (Laws)? Authors: Rylan Schaeffer, Joshua Kazdan, John Hughes, Jordan Juravsky, Sara Price, Aengus Lynch, Erik Jones, Robert Kirk, Azalia Mirhoseini, Sanmi Koyejo

4. AMPO: Active Multi-Preference Optimization Authors: Taneesh Gupta, Rahul Madhavan, Xuchao Zhang, Chetan Bansal, Saravan Rajmohan

5. Global law of conjugate kernel random matrices with heavy-tailed weights Authors: Alice Guionnet, Vanessa Piccolo

6. Graded Neural Networks Authors: Tony Shaska

7. PICASO: Permutation-Invariant Context Composition with State Space Models Authors: Tian Yu Liu, Alessandro Achille, Matthew Trager, Aditya Golatkar, Luca Zancato, Stefano Soatto

8. SpargeAttn: Accurate Sparse Attention Accelerating Any Model Inference Authors: Jintao Zhang, Chendong Xiang, Haofeng Huang, Jia Wei, Haocheng Xi, Jun Zhu, Jianfei Chen

9. Unveiling and Causalizing CoT: A Causal Pespective Authors: Jiarun Fu, Lizhong Ding, Hao Li, Pengqi Li, Qiuning Wei, Xu Chen

10. The Lottery LLM Hypothesis, Rethinking What Abilities Should LLM Compression Preserve? Authors: Zhenheng Tang, Xiang Liu, Qian Wang, Peijie Dong, Bingsheng He, Xiaowen Chu, Bo Li

11. Optimal Brain Apoptosis Authors: Mingyuan Sun, Zheng Fang, Jiaxu Wang, Junjie Jiang, Delei Kong, Chenming Hu, Yuetong Fang, Renjing Xu

12. CoKV: Optimizing KV Cache Allocation via Cooperative Game Authors: Qiheng Sun, Hongwei Zhang, Haocheng Xia, Jiayao Zhang, Jinfei Liu, Kui Ren

13. Jacobian Sparse Autoencoders: Sparsify Computations, Not Just Activations Authors: Lucy Farnik, Tim Lawson, Conor Houghton, Laurence Aitchison

14. C-LoRA: Continual Low-Rank Adaptation for Pre-trained Models Authors: Xin Zhang, Liang Bai, Xian Yang, Jiye Liang

15. Near-Optimal Approximations for Bayesian Inference in Function Space Authors: Veit Wild, James Wu, Dino Sejdinovic, Jeremias Knoblauch

16. The Gradient of Algebraic Model Counting Authors: Jaron Maene, Luc De Raedt

17. Effective Field Neural Network Authors: Xi Liu, Yujun Zhao, Chun Yu Wan, Yang Zhang, Junwei Liu

18. Scalable Equilibrium Sampling with Sequential Boltzmann Generators Authors: Charlie B. Tan, Avishek Joey Bose, Chen Lin, Leon Klein, Michael M. Bronstein, Alexander Tong

19. Golden Ratio Mixing of Real and Synthetic Data for Stabilizing Generative Model Training Authors: Hengzhi He, Shirong Xu, Guang Cheng

20. Aligning Compound AI Systems via System-level DPO Authors: Xiangwen Wang, Yibo Jacky Zhang, Zhoujie Ding, Katherine Tsai, Sanmi Koyejo

21. C-3DPO: Constrained Controlled Classification for Direct Preference Optimization Authors: Kavosh Asadi, Julien Han, Xingzi Xu, Dominique Perrault-Joncas, Shoham Sabach, Karim Bouyarmane, Mohammad Ghavamzadeh

22. A Fokker-Planck-Based Loss Function that Bridges Dynamics with Density Estimation Authors: Zhixin Lu, {\L}ukasz Ku\'smierz, Stefan Mihalas

23. Knowledge Distillation with Training Wheels Authors: Guanlin Liu, Anand Ramachandran, Tanmay Gangwani, Yan Fu, Abhinav Sethy

24. Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning Authors: Wenkai Yang, Shuming Ma, Yankai Lin, Furu Wei

25. LeanKAN: A Parameter-Lean Kolmogorov-Arnold Network Layer with Improved Memory Efficiency and Convergence Behavior Authors: Benjamin C. Koenig, Suyong Kim, Sili Deng

26. An Overview of Large Language Models for Statisticians Authors: Wenlong Ji, Weizhe Yuan, Emily Getzen, Kyunghyun Cho, Michael I. Jordan, Song Mei, Jason E Weston, Weijie J. Su, Jing Xu, Linjun Zhang

27. A General Framework to Enhance Fine-tuning-based LLM Unlearning Authors: Jie Ren, Zhenwei Dai, Xianfeng Tang, Hui Liu, Jingying Zeng, Zhen Li, Rahul Goutam, Suhang Wang, Yue Xing, Qi He, Hui Liu

28. Synthetic Text Generation for Training Large Language Models via Gradient Matching Authors: Dang Nguyen, Zeman Li, Mohammadhossein Bateni, Vahab Mirrokni, Meisam Razaviyayn, Baharan Mirzasoleiman

29. A Priori Generalizability Estimate for a CNN Authors: Cito Balsells, Beatrice Riviere, David Fuentes

30. Hallucination Detection in LLMs Using Spectral Features of Attention Maps Authors: Jakub Binkowski, Denis Janiak, Albert Sawczyn, Bogdan Gabrys, Tomasz Kajdanowicz

31. Representation Engineering for Large-Language Models: Survey and Research Challenges Authors: Lukasz Bartoszcze, Sarthak Munshi, Bryan Sukidi, Jennifer Yen, Zejia Yang, David Williams-King, Linh Le, Kosi Asuzu, Carsten Maple

32. Model-Free Adversarial Purification via Coarse-To-Fine Tensor Network Representation Authors: Guang Lin, Duc Thien Nguyen, Zerui Tao, Konstantinos Slavakis, Toshihisa Tanaka, Qibin Zhao

33. Learning Backbones: Sparsifying Graphs through Zero Forcing for Effective Graph-Based Learning Authors: Obaid Ullah Ahmad, Anwar Said, Mudassir Shabbir, Xenofon Koutsoukos, Waseem Abbas

34. An Improved Privacy and Utility Analysis of Differentially Private SGD with Bounded Domain and Smooth Losses Authors: Hao Liang, Wanrong Zhang, Xinlei He, Kaishun He, Hong Xing

35. Scaling LLM Pre-training with Vocabulary Curriculum Authors: Fangyuan Yu

36. Generalized Exponentiated Gradient Algorithms Using the Euler Two-Parameter Logarithm Authors: Andrzej Cichocki

37. A Survey on Mechanistic Interpretability for Multi-Modal Foundation Models Authors: Zihao Lin, Samyadeep Basu, Mohammad Beigi, Varun Manjunatha, Ryan A. Rossi, Zichao Wang, Yufan Zhou, Sriram Balasubramanian, Arman Zarei, Keivan Rezaei, Ying Shen, Barry Menglong Yao, Zhiyang Xu, Qin Liu, Yuxiang Zhang, Yan Sun, Shilong Liu, Li Shen, Hongxuan Li, Soheil Feizi, Lifu Huang

1. Reversal Blessing: Thinking Backward May Outpace Thinking Forward in Multi-choice Questions

ArXiv ID: 2502.18435

Authors: Yizhe Zhang, Richard Bai, Zijin Gu, Ruixiang Zhang, Jiatao Gu, Emmanuel Abbe, Samy Bengio, Navdeep Jaitly

Abstract: Language models usually use left-to-right (L2R) autoregressive factorization. However, L2R factorization may not always be the best inductive bias. Therefore, we investigate whether alternative factorizations of the text distribution could be beneficial in some tasks. We investigate right-to-left (R2L) training as a compelling alternative, focusing on multiple-choice questions (MCQs) as a test bed for knowledge extraction and reasoning. Through extensive experiments across various model sizes (2B-8B parameters) and training datasets, we find that R2L models can significantly outperform L2R models on several MCQ benchmarks, including logical reasoning, commonsense understanding, and truthfulness assessment tasks. Our analysis reveals that this performance difference may be fundamentally linked to multiple factors including calibration, computability and directional conditional entropy. We ablate the impact of these factors through controlled simulation studies using arithmetic tasks, where the impacting factors can be better disentangled. Our work demonstrates that exploring alternative factorizations of the text distribution can lead to improvements in LLM capabilities and provides theoretical insights into optimal factorization towards approximating human language distribution, and when each reasoning order might be more advantageous.

Comment: Author match

2. Mechanistic PDE Networks for Discovery of Governing Equations

ArXiv ID: 2502.18377

Authors: Adeel Pervez, Efstratios Gavves, Francesco Locatello

Abstract: We present Mechanistic PDE Networks -- a model for discovery of governing partial differential equations from data. Mechanistic PDE Networks represent spatiotemporal data as space-time dependent linear partial differential equations in neural network hidden representations. The represented PDEs are then solved and decoded for specific tasks. The learned PDE representations naturally express the spatiotemporal dynamics in data in neural network hidden space, enabling increased power for dynamical modeling. Solving the PDE representations in a compute and memory-efficient way, however, is a significant challenge. We develop a native, GPU-capable, parallel, sparse, and differentiable multigrid solver specialized for linear partial differential equations that acts as a module in Mechanistic PDE Networks. Leveraging the PDE solver, we propose a discovery architecture that can discover nonlinear PDEs in complex settings while also being robust to noise. We validate PDE discovery on a number of PDEs, including reaction-diffusion and Navier-Stokes equations.

Comment: The paper proposes Mechanistic PDE Networks for discovering governing equations, which aligns with foundational research in AI for Science and introduces a novel architecture.

Relevance: 9 Novelty: 9

3. How Do Large Language Monkeys Get Their Power (Laws)?

ArXiv ID: 2502.17578

Authors: Rylan Schaeffer, Joshua Kazdan, John Hughes, Jordan Juravsky, Sara Price, Aengus Lynch, Erik Jones, Robert Kirk, Azalia Mirhoseini, Sanmi Koyejo

Abstract: Recent research across mathematical problem solving, proof assistant programming and multimodal jailbreaking documents a striking finding: when (multimodal) language model tackle a suite of tasks with multiple attempts per task -- succeeding if any attempt is correct -- then the negative log of the average success rate scales a power law in the number of attempts. In this work, we identify an apparent puzzle: a simple mathematical calculation predicts that on each problem, the failure rate should fall exponentially with the number of attempts. We confirm this prediction empirically, raising a question: from where does aggregate polynomial scaling emerge? We then answer this question by demonstrating per-problem exponential scaling can be made consistent with aggregate polynomial scaling if the distribution of single-attempt success probabilities is heavy tailed such that a small fraction of tasks with extremely low success probabilities collectively warp the aggregate success trend into a power law - even as each problem scales exponentially on its own. We further demonstrate that this distributional perspective explains previously observed deviations from power law scaling, and provides a simple method for forecasting the power law exponent with an order of magnitude lower relative error, or equivalently, ${\sim}2-4$ orders of magnitude less inference compute. Overall, our work contributes to a better understanding of how neural language model performance improves with scaling inference compute and the development of scaling-predictable evaluations of (multimodal) language models.

Comment: The paper provides theoretical insights into power law scaling in large language models, which aligns with foundational research in LLM behavior and interpretability.

Relevance: 9 Novelty: 8

4. AMPO: Active Multi-Preference Optimization

ArXiv ID: 2502.18293

Authors: Taneesh Gupta, Rahul Madhavan, Xuchao Zhang, Chetan Bansal, Saravan Rajmohan

Abstract: Multi-preference optimization enriches language-model alignment beyond pairwise preferences by contrasting entire sets of helpful and undesired responses, thereby enabling richer training signals for large language models. During self-play alignment, these models often produce numerous candidate answers per query, rendering it computationally infeasible to include all responses in the training objective. In this work, we propose $\textit{Active Multi-Preference Optimization}$ (AMPO), a novel approach that combines on-policy generation, a multi-preference group-contrastive loss, and active subset selection. Specifically, we score and embed large candidate pools of responses and then select a small, yet informative, subset that covers reward extremes and distinct semantic clusters for preference optimization. Our contrastive training scheme is capable of identifying not only the best and worst answers but also subtle, underexplored modes that are crucial for robust alignment. Theoretically, we provide guarantees for expected reward maximization using our active selection method, and empirically, AMPO achieves state-of-the-art results on $\textit{AlpacaEval}$ using Llama 8B.

Comment: The paper proposes a novel multi-preference optimization framework for LLM alignment, which aligns with foundational research in LLM training and optimization techniques.

Relevance: 9 Novelty: 8

5. Global law of conjugate kernel random matrices with heavy-tailed weights

ArXiv ID: 2502.18428

Authors: Alice Guionnet, Vanessa Piccolo

Abstract: We study the asymptotic spectral behavior of the conjugate kernel random matrix $YY^\top$, where $Y= f(WX)$ arises from a two-layer neural network model. We consider the setting where $W$ and $X$ are both random rectangular matrices with i.i.d. entries, where the entries of $W$ follow a heavy-tailed distribution, while those of $X$ have light tails. Our assumptions on $W$ include a broad class of heavy-tailed distributions, such as symmetric $\alpha$-stable laws with $\alpha \in (0,2)$ and sparse matrices with $\mathcal{O}(1)$ nonzero entries per row. The activation function $f$, applied entrywise, is nonlinear, smooth, and odd. By computing the eigenvalue distribution of $YY^\top$ through its moments, we show that heavy-tailed weights induce strong correlations between the entries of $Y$, leading to richer and fundamentally different spectral behavior compared to models with light-tailed weights.

Comment: The paper studies the spectral behavior of kernel random matrices with heavy-tailed weights, which provides theoretical insights into neural network training dynamics and aligns with foundational research.

Relevance: 9 Novelty: 8

6. Graded Neural Networks

ArXiv ID: 2502.17751

Authors: Tony Shaska

Abstract: This paper presents a novel framework for graded neural networks (GNNs) built over graded vector spaces $\V_\w^n$, extending classical neural architectures by incorporating algebraic grading. Leveraging a coordinate-wise grading structure with scalar action $\lambda \star \x = (\lambda^{q_i} x_i)$, defined by a tuple $\w = (q_0, \ldots, q_{n-1})$, we introduce graded neurons, layers, activation functions, and loss functions that adapt to feature significance. Theoretical properties of graded spaces are established, followed by a comprehensive GNN design, addressing computational challenges like numerical stability and gradient scaling. Potential applications span machine learning and photonic systems, exemplified by high-speed laser-based implementations. This work offers a foundational step toward graded computation, unifying mathematical rigor with practical potential, with avenues for future empirical and hardware exploration.

Comment: The paper introduces graded neural networks, which propose a novel architectural framework with theoretical underpinnings, aligning with model architecture innovations.

Relevance: 9 Novelty: 8

7. PICASO: Permutation-Invariant Context Composition with State Space Models

ArXiv ID: 2502.17605

Authors: Tian Yu Liu, Alessandro Achille, Matthew Trager, Aditya Golatkar, Luca Zancato, Stefano Soatto

Abstract: Providing Large Language Models with relevant contextual knowledge at inference time has been shown to greatly improve the quality of their generations. This is often achieved by prepending informative passages of text, or 'contexts', retrieved from external knowledge bases to their input. However, processing additional contexts online incurs significant computation costs that scale with their length. State Space Models (SSMs) offer a promising solution by allowing a database of contexts to be mapped onto fixed-dimensional states from which to start the generation. A key challenge arises when attempting to leverage information present across multiple contexts, since there is no straightforward way to condition generation on multiple independent states in existing SSMs. To address this, we leverage a simple mathematical relation derived from SSM dynamics to compose multiple states into one that efficiently approximates the effect of concatenating textual contexts. Since the temporal ordering of contexts can often be uninformative, we enforce permutation-invariance by efficiently averaging states obtained via our composition algorithm across all possible context orderings. We evaluate our resulting method on WikiText and MSMARCO in both zero-shot and fine-tuned settings, and show that we can match the strongest performing baseline while enjoying on average 5.4x speedup.

Comment: The paper introduces a novel method for permutation-invariant context composition using state space models, which aligns with foundational research in efficient context representation for LLMs.

Relevance: 9 Novelty: 8

8. SpargeAttn: Accurate Sparse Attention Accelerating Any Model Inference

ArXiv ID: 2502.18137

Authors: Jintao Zhang, Chendong Xiang, Haofeng Huang, Jia Wei, Haocheng Xi, Jun Zhu, Jianfei Chen

Abstract: An efficient attention implementation is essential for large models due to its quadratic time complexity. Fortunately, attention commonly exhibits sparsity, i.e., many values in the attention map are near zero, allowing for the omission of corresponding computations. Many studies have utilized the sparse pattern to accelerate attention. However, most existing works focus on optimizing attention within specific models by exploiting certain sparse patterns of the attention map. A universal sparse attention that guarantees both the speedup and end-to-end performance of diverse models remains elusive. In this paper, we propose SpargeAttn, a universal sparse and quantized attention for any model. Our method uses a two-stage online filter: in the first stage, we rapidly and accurately predict the attention map, enabling the skip of some matrix multiplications in attention. In the second stage, we design an online softmax-aware filter that incurs no extra overhead and further skips some matrix multiplications. Experiments show that our method significantly accelerates diverse models, including language, image, and video generation, without sacrificing end-to-end metrics. The codes are available at https://github.com/thu-ml/SpargeAttn.

Comment: The paper introduces SpargeAttn, a universal sparse attention mechanism, which aligns with foundational research in model compression and sparse attention techniques.

Relevance: 9 Novelty: 8

9. Unveiling and Causalizing CoT: A Causal Pespective

ArXiv ID: 2502.18239

Authors: Jiarun Fu, Lizhong Ding, Hao Li, Pengqi Li, Qiuning Wei, Xu Chen

Abstract: Although Chain-of-Thought (CoT) has achieved remarkable success in enhancing the reasoning ability of large language models (LLMs), the mechanism of CoT remains a ``black box''. Even if the correct answers can frequently be obtained, existing CoTs struggle to make the reasoning understandable to human. In this paper, we unveil and causalize CoT from a causal perspective to ensure both correctness and understandability of all reasoning steps (to the best of our knowledge, the first such). We model causality of CoT via structural causal models (SCM) to unveil the reasoning mechanism of CoT. To measure the causality of CoT, we define the CoT Average Causal Effect (CACE) to test the causal relations between steps. For those steps without causality (wrong or unintelligible steps), we design a role-playing causal query algorithm to causalize these steps, resulting a causalized CoT with all steps correct and understandable. Experimental results on both open-source and closed-source LLMs demonstrate that the causal errors commonly in steps are effectively corrected and the reasoning ability of LLMs is significantly improved.

Comment: The paper explores causal perspectives on Chain-of-Thought reasoning in LLMs, providing theoretical insights into reasoning mechanisms. This aligns with foundational research in LLM behavior and interpretability.

Relevance: 9 Novelty: 8

10. The Lottery LLM Hypothesis, Rethinking What Abilities Should LLM Compression Preserve?

ArXiv ID: 2502.17535

Authors: Zhenheng Tang, Xiang Liu, Qian Wang, Peijie Dong, Bingsheng He, Xiaowen Chu, Bo Li

Abstract: Motivated by reducing the computational and storage costs of LLMs, model compression and KV cache compression have attracted much attention from researchers. However, current methods predominantly emphasize maintaining the performance of compressed LLMs, as measured by perplexity or simple accuracy on tasks of common sense knowledge QA and basic arithmetic reasoning. In this blog, we present a brief review of recent advancements in LLMs related to retrieval-augmented generation, multi-step reasoning, external tools, and computational expressivity, all of which substantially enhance LLM performance. Then, we propose a lottery LLM hypothesis suggesting that for a given LLM and task, there exists a smaller lottery LLM capable of producing the same performance as the original LLM with the assistance of multi-step reasoning and external tools. Based on the review of current progress in LLMs, we discuss and summarize the essential capabilities that the lottery LLM and KV cache compression must possess, which are currently overlooked in existing methods.

Comment: The paper discusses the Lottery LLM Hypothesis, which is highly relevant to model compression and foundational insights into LLM capabilities.

Relevance: 9 Novelty: 8

11. Optimal Brain Apoptosis

ArXiv ID: 2502.17941

Authors: Mingyuan Sun, Zheng Fang, Jiaxu Wang, Junjie Jiang, Delei Kong, Chenming Hu, Yuetong Fang, Renjing Xu

Abstract: The increasing complexity and parameter count of Convolutional Neural Networks (CNNs) and Transformers pose challenges in terms of computational efficiency and resource demands. Pruning has been identified as an effective strategy to address these challenges by removing redundant elements such as neurons, channels, or connections, thereby enhancing computational efficiency without heavily compromising performance. This paper builds on the foundational work of Optimal Brain Damage (OBD) by advancing the methodology of parameter importance estimation using the Hessian matrix. Unlike previous approaches that rely on approximations, we introduce Optimal Brain Apoptosis (OBA), a novel pruning method that calculates the Hessian-vector product value directly for each parameter. By decomposing the Hessian matrix across network layers and identifying conditions under which inter-layer Hessian submatrices are non-zero, we propose a highly efficient technique for computing the second-order Taylor expansion of parameters. This approach allows for a more precise pruning process, particularly in the context of CNNs and Transformers, as validated in our experiments including VGG19, ResNet32, ResNet50, and ViT-B/16 on CIFAR10, CIFAR100 and Imagenet datasets. Our code is available at https://github.com/NEU-REAL/OBA.

Comment: The paper introduces a novel pruning method, Optimal Brain Apoptosis (OBA), which advances parameter importance estimation using the Hessian matrix. This aligns closely with the model compression criterion, particularly in pruning and efficiency improvements.

Relevance: 9 Novelty: 8

12. CoKV: Optimizing KV Cache Allocation via Cooperative Game

ArXiv ID: 2502.17501

Authors: Qiheng Sun, Hongwei Zhang, Haocheng Xia, Jiayao Zhang, Jinfei Liu, Kui Ren

Abstract: Large language models (LLMs) have achieved remarkable success on various aspects of human life. However, one of the major challenges in deploying these models is the substantial memory consumption required to store key-value pairs (KV), which imposes significant resource demands. Recent research has focused on KV cache budget allocation, with several approaches proposing head-level budget distribution by evaluating the importance of individual attention heads. These methods, however, assess the importance of heads independently, overlooking their cooperative contributions within the model, which may result in a deviation from their true impact on model performance. In light of this limitation, we propose CoKV, a novel method that models the cooperation between heads in model inference as a cooperative game. By evaluating the contribution of each head within the cooperative game, CoKV can allocate the cache budget more effectively. Extensive experiments show that CoKV achieves state-of-the-art performance on the LongBench benchmark using LLama-3-8B-Instruct and Mistral-7B models.

Comment: The paper proposes CoKV, a novel method for optimizing KV cache allocation in LLMs using cooperative game theory. This directly addresses efficiency and memory challenges in LLMs, aligning with the model compression criterion.

Relevance: 9 Novelty: 8

13. Jacobian Sparse Autoencoders: Sparsify Computations, Not Just Activations

ArXiv ID: 2502.18147

Authors: Lucy Farnik, Tim Lawson, Conor Houghton, Laurence Aitchison

Abstract: Sparse autoencoders (SAEs) have been successfully used to discover sparse and human-interpretable representations of the latent activations of LLMs. However, we would ultimately like to understand the computations performed by LLMs and not just their representations. The extent to which SAEs can help us understand computations is unclear because they are not designed to "sparsify" computations in any sense, only latent activations. To solve this, we propose Jacobian SAEs (JSAEs), which yield not only sparsity in the input and output activations of a given model component but also sparsity in the computation (formally, the Jacobian) connecting them. With a na\"ive implementation, the Jacobians in LLMs would be computationally intractable due to their size. One key technical contribution is thus finding an efficient way of computing Jacobians in this setup. We find that JSAEs extract a relatively large degree of computational sparsity while preserving downstream LLM performance approximately as well as traditional SAEs. We also show that Jacobians are a reasonable proxy for computational sparsity because MLPs are approximately linear when rewritten in the JSAE basis. Lastly, we show that JSAEs achieve a greater degree of computational sparsity on pre-trained LLMs than on the equivalent randomized LLM. This shows that the sparsity of the computational graph appears to be a property that LLMs learn through training, and suggests that JSAEs might be more suitable for understanding learned transformer computations than standard SAEs.

Comment: The paper introduces Jacobian Sparse Autoencoders (JSAEs) to sparsify computations in LLMs, which aligns with foundational research in representation learning and sparsity. It also provides efficient methods for computing Jacobians.

Relevance: 9 Novelty: 8

14. C-LoRA: Continual Low-Rank Adaptation for Pre-trained Models

ArXiv ID: 2502.17920

Authors: Xin Zhang, Liang Bai, Xian Yang, Jiye Liang

Abstract: Low-Rank Adaptation (LoRA) is an efficient fine-tuning method that has been extensively applied in areas such as natural language processing and computer vision. Existing LoRA fine-tuning approaches excel in static environments but struggle in dynamic learning due to reliance on multiple adapter modules, increasing overhead and complicating inference. We propose Continual Low-Rank Adaptation (C-LoRA), a novel extension of LoRA for continual learning. C-LoRA uses a learnable routing matrix to dynamically manage parameter updates across tasks, ensuring efficient reuse of learned subspaces while enforcing orthogonality to minimize interference and forgetting. Unlike existing approaches that require separate adapters for each task, C-LoRA enables a integrated approach for task adaptation, achieving both scalability and parameter efficiency in sequential learning scenarios. C-LoRA achieves state-of-the-art accuracy and parameter efficiency on benchmarks while providing theoretical insights into its routing matrix's role in retaining and transferring knowledge, establishing a scalable framework for continual learning.

Comment: The paper proposes C-LoRA, a novel extension of Low-Rank Adaptation (LoRA) for continual learning, which aligns with the model compression and efficiency criteria. The use of a learnable routing matrix for task adaptation is a significant methodological contribution.

Relevance: 9 Novelty: 8

15. Near-Optimal Approximations for Bayesian Inference in Function Space

ArXiv ID: 2502.18279

Authors: Veit Wild, James Wu, Dino Sejdinovic, Jeremias Knoblauch

Abstract: We propose a scalable inference algorithm for Bayes posteriors defined on a reproducing kernel Hilbert space (RKHS). Given a likelihood function and a Gaussian random element representing the prior, the corresponding Bayes posterior measure $\Pi_{\text{B}}$ can be obtained as the stationary distribution of an RKHS-valued Langevin diffusion. We approximate the infinite-dimensional Langevin diffusion via a projection onto the first $M$ components of the Kosambi-Karhunen-Lo`eve expansion. Exploiting the thus obtained approximate posterior for these $M$ components, we perform inference for $\Pi_{\text{B}}$ by relying on the law of total probability and a sufficiency assumption. The resulting method scales as $O(M^3+JM^2)$, where $J$ is the number of samples produced from the posterior measure $\Pi_{\text{B}}$. Interestingly, the algorithm recovers the posterior arising from the sparse variational Gaussian process (SVGP) (see Titsias, 2009) as a special case, owed to the fact that the sufficiency assumption underlies both methods. However, whereas the SVGP is parametrically constrained to be a Gaussian process, our method is based on a non-parametric variational family $\mathcal{P}(\mathbb{R}^M)$ consisting of all probability measures on $\mathbb{R}^M$. As a result, our method is provably close to the optimal $M$-dimensional variational approximation of the Bayes posterior $\Pi_{\text{B}}$ in $\mathcal{P}(\mathbb{R}^M)$ for convex and Lipschitz continuous negative log likelihoods, and coincides with SVGP for the special case of a Gaussian error likelihood.

Comment: The paper proposes a scalable Bayesian inference algorithm in function space, which is a foundational contribution to probabilistic modeling and aligns with emerging trends in theoretical research.

Relevance: 8 Novelty: 9

16. The Gradient of Algebraic Model Counting

ArXiv ID: 2502.18406

Authors: Jaron Maene, Luc De Raedt

Abstract: Algebraic model counting unifies many inference tasks on logic formulas by exploiting semirings. Rather than focusing on inference, we consider learning, especially in statistical-relational and neurosymbolic AI, which combine logical, probabilistic and neural representations. Concretely, we show that the very same semiring perspective of algebraic model counting also applies to learning. This allows us to unify various learning algorithms by generalizing gradients and backpropagation to different semirings. Furthermore, we show how cancellation and ordering properties of a semiring can be exploited for more memory-efficient backpropagation. This allows us to obtain some interesting variations of state-of-the-art gradient-based optimisation methods for probabilistic logical models. We also discuss why algebraic model counting on tractable circuits does not lead to more efficient second-order optimization. Empirically, our algebraic backpropagation exhibits considerable speed-ups as compared to existing approaches.

Comment: The paper introduces a novel loss function derived from the Fokker-Planck equation, bridging dynamics with density estimation. This aligns with emerging trends in foundational research.

Relevance: 8 Novelty: 9

17. Effective Field Neural Network

ArXiv ID: 2502.17665

Authors: Xi Liu, Yujun Zhao, Chun Yu Wan, Yang Zhang, Junwei Liu

Abstract: In recent years, with the rapid development of machine learning, physicists have been exploring its new applications in solving or alleviating the curse of dimensionality in many-body problems. In order to accurately reflect the underlying physics of the problem, domain knowledge must be encoded into the machine learning algorithms. In this work, inspired by field theory, we propose a new set of machine learning models called effective field neural networks (EFNNs) that can automatically and efficiently capture important many-body interactions through multiple self-refining processes. Taking the classical $3$-spin infinite-range model and the quantum double exchange model as case studies, we explicitly demonstrate that EFNNs significantly outperform fully-connected deep neural networks (DNNs) and the effective model. Furthermore, with the help of convolution operations, the EFNNs learned in a small system can be seamlessly used in a larger system without additional training and the relative errors even decrease, which further demonstrates the efficacy of EFNNs in representing core physical behaviors.

Comment: The paper introduces Effective Field Neural Networks (EFNNs), inspired by field theory, to model many-body interactions. This aligns with foundational research in representation learning and emerging trends, as it proposes a novel paradigm for encoding domain knowledge into neural networks.

Relevance: 8 Novelty: 9

18. Scalable Equilibrium Sampling with Sequential Boltzmann Generators

ArXiv ID: 2502.18462

Authors: Charlie B. Tan, Avishek Joey Bose, Chen Lin, Leon Klein, Michael M. Bronstein, Alexander Tong

Abstract: Scalable sampling of molecular states in thermodynamic equilibrium is a long-standing challenge in statistical physics. Boltzmann generators tackle this problem by pairing powerful normalizing flows with importance sampling to obtain statistically independent samples under the target distribution. In this paper, we extend the Boltzmann generator framework and introduce Sequential Boltzmann generators (SBG) with two key improvements. The first is a highly efficient non-equivariant Transformer-based normalizing flow operating directly on all-atom Cartesian coordinates. In contrast to equivariant continuous flows of prior methods, we leverage exactly invertible non-equivariant architectures which are highly efficient both during sample generation and likelihood computation. As a result, this unlocks more sophisticated inference strategies beyond standard importance sampling. More precisely, as a second key improvement we perform inference-time scaling of flow samples using annealed Langevin dynamics which transports samples toward the target distribution leading to lower variance (annealed) importance weights which enable higher fidelity resampling with sequential Monte Carlo. SBG achieves state-of-the-art performance w.r.t. all metrics on molecular systems, demonstrating the first equilibrium sampling in Cartesian coordinates of tri, tetra, and hexapeptides that were so far intractable for prior Boltzmann generators.

Comment: The paper introduces Sequential Boltzmann Generators with a Transformer-based normalizing flow, which is relevant to foundational research in generative modeling and architecture-level innovations.

Relevance: 8 Novelty: 8

19. Golden Ratio Mixing of Real and Synthetic Data for Stabilizing Generative Model Training

ArXiv ID: 2502.18049

Authors: Hengzhi He, Shirong Xu, Guang Cheng

Abstract: Recent studies identified an intriguing phenomenon in recursive generative model training known as model collapse, where models trained on data generated by previous models exhibit severe performance degradation. Addressing this issue and developing more effective training strategies have become central challenges in generative model research. In this paper, we investigate this phenomenon theoretically within a novel framework, where generative models are iteratively trained on a combination of newly collected real data and synthetic data from the previous training step. To develop an optimal training strategy for integrating real and synthetic data, we evaluate the performance of a weighted training scheme in various scenarios, including Gaussian distribution estimation and linear regression. We theoretically characterize the impact of the mixing proportion and weighting scheme of synthetic data on the final model's performance. Our key finding is that, across different settings, the optimal weighting scheme under different proportions of synthetic data asymptotically follows a unified expression, revealing a fundamental trade-off between leveraging synthetic data and generative model performance. Notably, in some cases, the optimal weight assigned to real data corresponds precisely to the reciprocal of the golden ratio. Finally, we validate our theoretical results on extensive simulated datasets and a real tabular dataset.

Comment: The paper investigates generative model training stability and proposes a novel weighting scheme, which aligns with foundational research in representation learning and training dynamics.

Relevance: 8 Novelty: 8

20. Aligning Compound AI Systems via System-level DPO

ArXiv ID: 2502.17721

Authors: Xiangwen Wang, Yibo Jacky Zhang, Zhoujie Ding, Katherine Tsai, Sanmi Koyejo

Abstract: Compound AI systems, comprising multiple interacting components such as LLM agents and external tools, demonstrate state-of-the-art results across diverse tasks. It is hence crucial to align components within the system to produce consistent results that match human expectations. However, conventional alignment methods, such as Direct Preference Optimization (DPO), are not directly applicable to compound AI systems. These challenges include the non-differentiable interactions between components, making end-to-end gradient optimization infeasible. Additionally, system-level preferences cannot be directly translated into component-level preferences, further complicating alignment. We address the issues by formulating compound AI systems as Directed Acyclic Graphs (DAGs), capturing the connections between agents and the data generation processes. We propose a system-level DPO (SysDPO) to jointly align compound systems by adapting the DPO to operate on these DAGs. We study the joint alignment of an LLM and a diffusion model to demonstrate the effectiveness of our approach. Our exploration provides insights into the alignment of compound AI systems and lays a foundation for future advancements.

Comment: The paper introduces a system-level alignment method for compound AI systems, which is relevant to emerging trends in foundational AI research.

Relevance: 8 Novelty: 8

21. C-3DPO: Constrained Controlled Classification for Direct Preference Optimization

ArXiv ID: 2502.17507

Authors: Kavosh Asadi, Julien Han, Xingzi Xu, Dominique Perrault-Joncas, Shoham Sabach, Karim Bouyarmane, Mohammad Ghavamzadeh

Abstract: Direct preference optimization (DPO)-style algorithms have emerged as a promising approach for solving the alignment problem in AI. We present a novel perspective that formulates these algorithms as implicit classification algorithms. This classification framework enables us to recover many variants of DPO-style algorithms by choosing appropriate classification labels and loss functions. We then leverage this classification framework to demonstrate that the underlying problem solved in these algorithms is under-specified, making them susceptible to probability collapse of the winner-loser responses. We address this by proposing a set of constraints designed to control the movement of probability mass between the winner and loser in the reference and target policies. Our resulting algorithm, which we call Constrained Controlled Classification DPO (\texttt{C-3DPO}), has a meaningful RLHF interpretation. By hedging against probability collapse, \texttt{C-3DPO} provides practical improvements over vanilla \texttt{DPO} when aligning several large language models using standard preference datasets.

Comment: The paper proposes a novel constrained classification framework for preference optimization, which provides theoretical insights into DPO-style algorithms. This aligns with foundational research in representation learning.

Relevance: 8 Novelty: 8

22. A Fokker-Planck-Based Loss Function that Bridges Dynamics with Density Estimation

ArXiv ID: 2502.17690

Authors: Zhixin Lu, {\L}ukasz Ku\'smierz, Stefan Mihalas

Abstract: We have derived a novel loss function from the Fokker-Planck equation that links dynamical system models with their probability density functions, demonstrating its utility in model identification and density estimation. In the first application, we show that this loss function can enable the extraction of dynamical parameters from non-temporal datasets, including timestamp-free measurements from steady non-equilibrium systems such as noisy Lorenz systems and gene regulatory networks. In the second application, when coupled with a density estimator, this loss facilitates density estimation when the dynamic equations are known. For density estimation, we propose a density estimator that integrates a Gaussian Mixture Model with a normalizing flow model. It simultaneously estimates normalized density, energy, and score functions from both empirical data and dynamics. It is compatible with a variety of data-based training methodologies, including maximum likelihood and score matching. It features a latent space akin to a modern Hopfield network, where the inherent Hopfield energy effectively assigns low densities to sparsely populated data regions, addressing common challenges in neural density estimators. Additionally, this Hopfield-like energy enables direct and rapid data manipulation through the Concave-Convex Procedure (CCCP) rule, facilitating tasks such as denoising and clustering. Our work demonstrates a principled framework for leveraging the complex interdependencies between dynamics and density estimation, as illustrated through synthetic examples that clarify the underlying theoretical intuitions.

Comment: The paper introduces a scalable graph condensation method with evolving capabilities, which aligns with foundational research in model compression and efficiency.

Relevance: 8 Novelty: 8

23. Knowledge Distillation with Training Wheels

ArXiv ID: 2502.17717

Authors: Guanlin Liu, Anand Ramachandran, Tanmay Gangwani, Yan Fu, Abhinav Sethy

Abstract: Knowledge distillation is used, in generative language modeling, to train a smaller student model using the help of a larger teacher model, resulting in improved capabilities for the student model. In this paper, we formulate a more general framework for knowledge distillation where the student learns from the teacher during training, and also learns to ask for the teacher's help at test-time following rules specifying test-time restrictions. Towards this, we first formulate knowledge distillation as an entropy-regularized value optimization problem. Adopting Path Consistency Learning to solve this, leads to a new knowledge distillation algorithm using on-policy and off-policy demonstrations. We extend this using constrained reinforcement learning to a framework that incorporates the use of the teacher model as a test-time reference, within constraints. In this situation, akin to a human learner, the model needs to learn not only the learning material, but also the relative difficulty of different sections to prioritize for seeking teacher help. We examine the efficacy of our method through experiments in translation and summarization tasks, observing trends in accuracy and teacher use, noting that our approach unlocks operating points not available to the popular Speculative Decoding approach.

Comment: The paper proposes a novel framework for knowledge distillation with test-time teacher assistance, which aligns with foundational research in model compression and training dynamics.

Relevance: 8 Novelty: 8

24. Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning

ArXiv ID: 2502.18080

Authors: Wenkai Yang, Shuming Ma, Yankai Lin, Furu Wei

Abstract: Recent studies have shown that making a model spend more time thinking through longer Chain of Thoughts (CoTs) enables it to gain significant improvements in complex reasoning tasks. While current researches continue to explore the benefits of increasing test-time compute by extending the CoT lengths of Large Language Models (LLMs), we are concerned about a potential issue hidden behind the current pursuit of test-time scaling: Would excessively scaling the CoT length actually bring adverse effects to a model's reasoning performance? Our explorations on mathematical reasoning tasks reveal an unexpected finding that scaling with longer CoTs can indeed impair the reasoning performance of LLMs in certain domains. Moreover, we discover that there exists an optimal scaled length distribution that differs across different domains. Based on these insights, we propose a Thinking-Optimal Scaling strategy. Our method first uses a small set of seed data with varying response length distributions to teach the model to adopt different reasoning efforts for deep thinking. Then, the model selects its shortest correct response under different reasoning efforts on additional problems for self-improvement. Our self-improved models built upon Qwen2.5-32B-Instruct outperform other distillation-based 32B o1-like models across various math benchmarks, and achieve performance on par with QwQ-32B-Preview.

Comment: The paper explores the optimal scaling of test-time compute for LLM reasoning, introducing a Thinking-Optimal Scaling strategy. This aligns with foundational research on LLM behavior and efficiency.

Relevance: 8 Novelty: 8

25. LeanKAN: A Parameter-Lean Kolmogorov-Arnold Network Layer with Improved Memory Efficiency and Convergence Behavior

ArXiv ID: 2502.17844

Authors: Benjamin C. Koenig, Suyong Kim, Sili Deng

Abstract: The recently proposed Kolmogorov-Arnold network (KAN) is a promising alternative to multi-layer perceptrons (MLPs) for data-driven modeling. While original KAN layers were only capable of representing the addition operator, the recently-proposed MultKAN layer combines addition and multiplication subnodes in an effort to improve representation performance. Here, we find that MultKAN layers suffer from a few key drawbacks including limited applicability in output layers, bulky parameterizations with extraneous activations, and the inclusion of complex hyperparameters. To address these issues, we propose LeanKANs, a direct and modular replacement for MultKAN and traditional AddKAN layers. LeanKANs address these three drawbacks of MultKAN through general applicability as output layers, significantly reduced parameter counts for a given network structure, and a smaller set of hyperparameters. As a one-to-one layer replacement for standard AddKAN and MultKAN layers, LeanKAN is able to provide these benefits to traditional KAN learning problems as well as augmented KAN structures in which it serves as the backbone, such as KAN Ordinary Differential Equations (KAN-ODEs) or Deep Operator KANs (DeepOKAN). We demonstrate LeanKAN's simplicity and efficiency in a series of demonstrations carried out across both a standard KAN toy problem and a KAN-ODE dynamical system modeling problem, where we find that its sparser parameterization and compact structure serve to increase its expressivity and learning capability, leading it to outperform similar and even much larger MultKANs in various tasks.

Comment: The paper introduces LeanKAN, a parameter-efficient alternative to Kolmogorov-Arnold networks, with improved memory efficiency and convergence. It aligns with foundational research in model architecture and efficiency.

Relevance: 8 Novelty: 8

26. An Overview of Large Language Models for Statisticians

ArXiv ID: 2502.17814

Authors: Wenlong Ji, Weizhe Yuan, Emily Getzen, Kyunghyun Cho, Michael I. Jordan, Song Mei, Jason E Weston, Weijie J. Su, Jing Xu, Linjun Zhang

Abstract: Large Language Models (LLMs) have emerged as transformative tools in artificial intelligence (AI), exhibiting remarkable capabilities across diverse tasks such as text generation, reasoning, and decision-making. While their success has primarily been driven by advances in computational power and deep learning architectures, emerging problems -- in areas such as uncertainty quantification, decision-making, causal inference, and distribution shift -- require a deeper engagement with the field of statistics. This paper explores potential areas where statisticians can make important contributions to the development of LLMs, particularly those that aim to engender trustworthiness and transparency for human users. Thus, we focus on issues such as uncertainty quantification, interpretability, fairness, privacy, watermarking and model adaptation. We also consider possible roles for LLMs in statistical analysis. By bridging AI and statistics, we aim to foster a deeper collaboration that advances both the theoretical foundations and practical applications of LLMs, ultimately shaping their role in addressing complex societal challenges.

Comment: The paper explores the intersection of statistics and LLMs, focusing on uncertainty quantification, interpretability, and fairness. This aligns with foundational research in LLM behavior and interpretability.

Relevance: 8 Novelty: 7

27. A General Framework to Enhance Fine-tuning-based LLM Unlearning

ArXiv ID: 2502.17823

Authors: Jie Ren, Zhenwei Dai, Xianfeng Tang, Hui Liu, Jingying Zeng, Zhen Li, Rahul Goutam, Suhang Wang, Yue Xing, Qi He, Hui Liu

Abstract: Unlearning has been proposed to remove copyrighted and privacy-sensitive data from Large Language Models (LLMs). Existing approaches primarily rely on fine-tuning-based methods, which can be categorized into gradient ascent-based (GA-based) and suppression-based methods. However, they often degrade model utility (the ability to respond to normal prompts). In this work, we aim to develop a general framework that enhances the utility of fine-tuning-based unlearning methods. To achieve this goal, we first investigate the common property between GA-based and suppression-based methods. We unveil that GA-based methods unlearn by distinguishing the target data (i.e., the data to be removed) and suppressing related generations, which is essentially the same strategy employed by suppression-based methods. Inspired by this finding, we introduce Gated Representation UNlearning (GRUN) which has two components: a soft gate function for distinguishing target data and a suppression module using Representation Fine-tuning (ReFT) to adjust representations rather than model parameters. Experiments show that GRUN significantly improves the unlearning and utility. Meanwhile, it is general for fine-tuning-based methods, efficient and promising for sequential unlearning.

Comment: The paper proposes a framework for enhancing fine-tuning-based LLM unlearning, which aligns with foundational research in LLM training and optimization.

Relevance: 8 Novelty: 7

28. Synthetic Text Generation for Training Large Language Models via Gradient Matching

ArXiv ID: 2502.17607

Authors: Dang Nguyen, Zeman Li, Mohammadhossein Bateni, Vahab Mirrokni, Meisam Razaviyayn, Baharan Mirzasoleiman

Abstract: Synthetic data has the potential to improve the performance, training efficiency, and privacy of real training examples. Nevertheless, existing approaches for synthetic text generation are mostly heuristics and cannot generate human-readable text without compromising the privacy of real data or provide performance guarantees for training Large Language Models (LLMs). In this work, we propose the first theoretically rigorous approach for generating synthetic human-readable text that guarantees the convergence and performance of LLMs during fine-tuning on a target task. To do so, we leverage Alternating Direction Method of Multipliers (ADMM) that iteratively optimizes the embeddings of synthetic examples to match the gradient of the target training or validation data, and maps them to a sequence of text tokens with low perplexity. In doing so, the generated synthetic text can guarantee convergence of the model to a close neighborhood of the solution obtained by fine-tuning on real data. Experiments on various classification tasks confirm the effectiveness of our proposed approach.

Comment: The paper proposes a theoretically rigorous approach for generating synthetic text for LLM training, which aligns with foundational research in representation learning and LLM training dynamics.

Relevance: 8 Novelty: 7

29. A Priori Generalizability Estimate for a CNN

ArXiv ID: 2502.17622

Authors: Cito Balsells, Beatrice Riviere, David Fuentes

Abstract: We formulate truncated singular value decompositions of entire convolutional neural networks. We demonstrate the computed left and right singular vectors are useful in identifying which images the convolutional neural network is likely to perform poorly on. To create this diagnostic tool, we define two metrics: the Right Projection Ratio and the Left Projection Ratio. The Right (Left) Projection Ratio evaluates the fidelity of the projection of an image (label) onto the computed right (left) singular vectors. We observe that both ratios are able to identify the presence of class imbalance for an image classification problem. Additionally, the Right Projection Ratio, which only requires unlabeled data, is found to be correlated to the model's performance when applied to image segmentation. This suggests the Right Projection Ratio could be a useful metric to estimate how likely the model is to perform well on a sample.

Comment: The paper introduces a novel diagnostic tool using singular value decomposition for CNNs, which aligns with foundational research in model architecture and generalization analysis.

Relevance: 8 Novelty: 7

30. Hallucination Detection in LLMs Using Spectral Features of Attention Maps

ArXiv ID: 2502.17598

Authors: Jakub Binkowski, Denis Janiak, Albert Sawczyn, Bogdan Gabrys, Tomasz Kajdanowicz

Abstract: Large Language Models (LLMs) have demonstrated remarkable performance across various tasks but remain prone to hallucinations. Detecting hallucinations is essential for safety-critical applications, and recent methods leverage attention map properties to this end, though their effectiveness remains limited. In this work, we investigate the spectral features of attention maps by interpreting them as adjacency matrices of graph structures. We propose the $\text{LapEigvals}$ method, which utilises the top-$k$ eigenvalues of the Laplacian matrix derived from the attention maps as an input to hallucination detection probes. Empirical evaluations demonstrate that our approach achieves state-of-the-art hallucination detection performance among attention-based methods. Extensive ablation studies further highlight the robustness and generalisation of $\text{LapEigvals}$, paving the way for future advancements in the hallucination detection domain.

Comment: The paper introduces a novel method for hallucination detection in LLMs using spectral features of attention maps. This aligns with the criteria for theoretical insights into LLM behavior and interpretability, making it relevant to foundational research.

Relevance: 8 Novelty: 7

31. Representation Engineering for Large-Language Models: Survey and Research Challenges

ArXiv ID: 2502.17601

Authors: Lukasz Bartoszcze, Sarthak Munshi, Bryan Sukidi, Jennifer Yen, Zejia Yang, David Williams-King, Linh Le, Kosi Asuzu, Carsten Maple

Abstract: Large-language models are capable of completing a variety of tasks, but remain unpredictable and intractable. Representation engineering seeks to resolve this problem through a new approach utilizing samples of contrasting inputs to detect and edit high-level representations of concepts such as honesty, harmfulness or power-seeking. We formalize the goals and methods of representation engineering to present a cohesive picture of work in this emerging discipline. We compare it with alternative approaches, such as mechanistic interpretability, prompt-engineering and fine-tuning. We outline risks such as performance decrease, compute time increases and steerability issues. We present a clear agenda for future research to build predictable, dynamic, safe and personalizable LLMs.

Comment: The paper surveys representation engineering for large language models, which aligns with foundational research in representation learning and interpretability.

Relevance: 8 Novelty: 6

32. Model-Free Adversarial Purification via Coarse-To-Fine Tensor Network Representation

ArXiv ID: 2502.17972

Authors: Guang Lin, Duc Thien Nguyen, Zerui Tao, Konstantinos Slavakis, Toshihisa Tanaka, Qibin Zhao

Abstract: Deep neural networks are known to be vulnerable to well-designed adversarial attacks. Although numerous defense strategies have been proposed, many are tailored to the specific attacks or tasks and often fail to generalize across diverse scenarios. In this paper, we propose Tensor Network Purification (TNP), a novel model-free adversarial purification method by a specially designed tensor network decomposition algorithm. TNP depends neither on the pre-trained generative model nor the specific dataset, resulting in strong robustness across diverse adversarial scenarios. To this end, the key challenge lies in relaxing Gaussian-noise assumptions of classical decompositions and accommodating the unknown distribution of adversarial perturbations. Unlike the low-rank representation of classical decompositions, TNP aims to reconstruct the unobserved clean examples from an adversarial example. Specifically, TNP leverages progressive downsampling and introduces a novel adversarial optimization objective to address the challenge of minimizing reconstruction error but without inadvertently restoring adversarial perturbations. Extensive experiments conducted on CIFAR-10, CIFAR-100, and ImageNet demonstrate that our method generalizes effectively across various norm threats, attack types, and tasks, providing a versatile and promising adversarial purification technique.

Comment: The paper proposes a novel tensor network decomposition for adversarial purification, which aligns with foundational research in model robustness and efficiency.

Relevance: 7 Novelty: 7

33. Learning Backbones: Sparsifying Graphs through Zero Forcing for Effective Graph-Based Learning

ArXiv ID: 2502.17713

Authors: Obaid Ullah Ahmad, Anwar Said, Mudassir Shabbir, Xenofon Koutsoukos, Waseem Abbas

Abstract: This paper introduces a novel framework for graph sparsification that preserves the essential learning attributes of original graphs, improving computational efficiency and reducing complexity in learning algorithms. We refer to these sparse graphs as "learning backbones". Our approach leverages the zero-forcing (ZF) phenomenon, a dynamic process on graphs with applications in network control. The key idea is to generate a tree from the original graph that retains critical dynamical properties. By correlating these properties with learning attributes, we construct effective learning backbones. We evaluate the performance of our ZF-based backbones in graph classification tasks across eight datasets and six baseline models. The results demonstrate that our method outperforms existing techniques. Additionally, we explore extensions using node distance metrics to further enhance the framework's utility.

Comment: The paper introduces a novel graph sparsification framework using zero-forcing, which is relevant to model compression and sparsity.

Relevance: 7 Novelty: 7

34. An Improved Privacy and Utility Analysis of Differentially Private SGD with Bounded Domain and Smooth Losses

ArXiv ID: 2502.17772

Authors: Hao Liang, Wanrong Zhang, Xinlei He, Kaishun He, Hong Xing

Abstract: Differentially Private Stochastic Gradient Descent (DPSGD) is widely used to protect sensitive data during the training of machine learning models, but its privacy guarantees often come at the cost of model performance, largely due to the inherent challenge of accurately quantifying privacy loss. While recent efforts have strengthened privacy guarantees by focusing solely on the final output and bounded domain cases, they still impose restrictive assumptions, such as convexity and other parameter limitations, and often lack a thorough analysis of utility. In this paper, we provide rigorous privacy and utility characterization for DPSGD for smooth loss functions in both bounded and unbounded domains. We track the privacy loss over multiple iterations by exploiting the noisy smooth-reduction property and establish the utility analysis by leveraging the projection's non-expansiveness and clipped SGD properties. In particular, we show that for DPSGD with a bounded domain, (i) the privacy loss can still converge without the convexity assumption, and (ii) a smaller bounded diameter can improve both privacy and utility simultaneously under certain conditions. Numerical results validate our results.

Comment: The paper provides an improved privacy and utility analysis for DPSGD, which is relevant to foundational research in model training dynamics and efficiency.

Relevance: 7 Novelty: 7

35. Scaling LLM Pre-training with Vocabulary Curriculum

ArXiv ID: 2502.17910

Authors: Fangyuan Yu

Abstract: Modern language models rely on static vocabularies, fixed before pretraining, in contrast to the adaptive vocabulary acquisition observed in human language learning. To bridge this gap, we introduce vocabulary curriculum learning, an approach that improves pretraining efficiency with log-linear scaling gains relative to vocabulary size. Our method alternates between entropy-guided vocabulary expansion and model optimization, enabling models to learn transferable representations across diverse tokenization granularities. This approach naturally gives rise to an optimal computation allocation pattern: longer tokens capture predictable content, while shorter tokens focus on more complex, harder-to-predict contexts. Experiments on small-scale GPT models demonstrate improved scaling efficiency, reinforcing the effectiveness of dynamic tokenization. We release our code to support further research and plan to extend our experiments to larger models and diverse domains.

Comment: The paper proposes a vocabulary curriculum for LLM pretraining, which offers insights into training efficiency and tokenization but does not introduce a major architectural or theoretical breakthrough.

Relevance: 7 Novelty: 7

36. Generalized Exponentiated Gradient Algorithms Using the Euler Two-Parameter Logarithm

ArXiv ID: 2502.17500

Authors: Andrzej Cichocki

Abstract: In this paper we propose and investigate a new class of Generalized Exponentiated Gradient (GEG) algorithms using Mirror Descent (MD) approaches, and applying as a regularization function the Bregman divergence with two-parameter deformation of logarithm as a link function. This link function (referred to as the Euler logarithm) is associated with a wide class of generalized entropies. In order to derive novel GEG/MD updates, we estimate generalized exponential function, which closely approximates the inverse of the Euler two-parameter logarithm. The characteristic/shape and properties of the Euler logarithm and its inverse -- deformed exponential functions are tuned by two or even more hyperparameters. By learning these hyperparameters, we can adapt to distribution of training data, and we can adjust them to achieve desired properties of gradient descent algorithms. The concept of generalized entropies and associated deformed logarithms provide deeper insight into novel gradient descent updates. In literature, there exist nowadays over fifty mathematically well-defined entropic functionals and associated deformed logarithms, so impossible to investigate all of them in one research paper. Therefore, we focus here on a wide-class of trace-form entropies and associated generalized logarithm. We applied the developed algorithms for Online Portfolio Selection (OPLS) in order to improve its performance and robustness.

Comment: The paper introduces a new class of gradient algorithms using generalized entropies and deformed logarithms. It provides theoretical insights into optimization methods, which could have implications for representation learning.

Relevance: 7 Novelty: 7

37. A Survey on Mechanistic Interpretability for Multi-Modal Foundation Models

ArXiv ID: 2502.17516

Authors: Zihao Lin, Samyadeep Basu, Mohammad Beigi, Varun Manjunatha, Ryan A. Rossi, Zichao Wang, Yufan Zhou, Sriram Balasubramanian, Arman Zarei, Keivan Rezaei, Ying Shen, Barry Menglong Yao, Zhiyang Xu, Qin Liu, Yuxiang Zhang, Yan Sun, Shilong Liu, Li Shen, Hongxuan Li, Soheil Feizi, Lifu Huang

Abstract: The rise of foundation models has transformed machine learning research, prompting efforts to uncover their inner workings and develop more efficient and reliable applications for better control. While significant progress has been made in interpreting Large Language Models (LLMs), multimodal foundation models (MMFMs) - such as contrastive vision-language models, generative vision-language models, and text-to-image models - pose unique interpretability challenges beyond unimodal frameworks. Despite initial studies, a substantial gap remains between the interpretability of LLMs and MMFMs. This survey explores two key aspects: (1) the adaptation of LLM interpretability methods to multimodal models and (2) understanding the mechanistic differences between unimodal language models and crossmodal systems. By systematically reviewing current MMFM analysis techniques, we propose a structured taxonomy of interpretability methods, compare insights across unimodal and multimodal architectures, and highlight critical research gaps.

Comment: The survey focuses on interpretability for multimodal foundation models, which aligns with foundational research in understanding model behavior but lacks direct methodological contributions.

Relevance: 7 Novelty: 6

Paper Selection Prompt

You are a helpful paper reading assistant whose job is to read daily posts from ArXiv and identify a few papers that your friend will enjoy reading. Your job is to carefully read the paper titles and abstracts below and find the ones that match the criteria below.

Relevant Topics

Use the following relevance criteria to focus on foundational research. Keep relevant papers and filter out irrelevant ones. Avoid purely application-driven work.

1. Representation Learning - Relevant: Insights into how deep networks encode information, feature/dictionary learning, sparse/contrastive methods, training dynamics in neural networks. - Irrelevant: Standard applications of known techniques lacking new theoretical or methodological contributions.

2. Model Architecture - Relevant: Mixture-of-Experts (MoE), Transformers, Conditional/Dynamic Networks, Autoencoders, analysis on existing architectures (like encoder-decoder), or other architectural innovations. - Irrelevant: Merely using existing architectures for a certain task without insights into the structure themselves.

3. Model Compression - Relevant: Sparsity, pruning, quantization, low-rank approaches, KV cache, or other algorithmic/theoretical efficiency breakthroughs. - Irrelevant: Straightforward applications of existing compression methods to new tasks.

4. Large Language Models (LLMs) - Relevant: Major breakthroughs in training or architecture, theoretical insights into LLM behavior/interpretability. - Irrelevant: Domain-specific usage (e.g., translation, jail-breaking), minor tweaks (e.g., instruction tuning, CoT, data mixing), or purely empirical dataset/benchmark studies and text-level analysis (e.g. hallucination, reasoning, safety).

5. AI for Science - Relevant: Foundational research in molecular/protein modeling, new generative paradigms, or significant architecture-level innovations. - Irrelevant: Conventional, domain-specific applications without new theoretical perspectives.

6. Emerging Trends - Relevant: Cutting-edge theoretical work challenging established assumptions or introducing broad new paradigms. - Irrelevant: Incremental improvements or trend-following without novel insights.

Keywords:

• Relevant: Mixture of Experts (MoE), Representation Learning, Compression/Efficiency, Sparse/Sparsity, Pruning, Quantization, Low-rank, Foundation Model, etc.
• Irrelevant: Reinforcement Learning, Transfer Learning, Federated Learning, Online Learning, Diffusion Models, etc.
• Application Work: Image Segmentation, Medical Imaging, 3D Vision, Video Understanding, Information Retrieval, Summarization, Recommendation Systems, Machine Translation, Speech Recognition, Signal Processing, Spatial/Temporal Modeling, Time Series, etc.

Scoring Criteria

The "Relevance" score measures how closely the paper aligns with the core topics of the prompt. The "Novelty" score assesses the originality and impact of the paper. They are two ORTHONORMAL axes and SHOULD NOT be confused with each other.

Relevance Scoring

• Relevance 9-10 (Completely Relevant)
• Focus: Fully aligned with core topics with no deviation, score the highest if contains relevant keywords in it.

• Examples: Papers focused on foundational methods or theoretical research, whose titles contain topic keywords like "MoE".

• Relevance 7-8 (Relevant)

• Focus: Retain a solid link to the main research area, though may touch on peripheral elements.

• Examples: Papers research on the fundamental part of MoE through a less critical aspect like its behavior in GNN.

• Relevance 5-6 (Borderline)

• Focus: Maintains a link to the core topic but also extends into at least one other domain/area beyond the primary focus.

• Examples: Work referencing MoE centered on reinforcement learning.

• Relevance 3-4 (Irrelevant)

• Focus: Largely outside our interests with no association to our topics.

• Examples: Application-focused papers like using MoE to solve a problem in the real world.

• Relevance 1-2 (Ignore)

• Focus: Purely unrelated to our topics. Completely a different domain.
• Exception: If the paper hints at a cutting-edge, radically new direction that could eventually transform the primary domain, consider a score of 9–10 despite initial appearances. (Usually a very rare concept that belongs to the fundamental research)

Novelty Scoring

• Novelty 9-10 (Breakthrough)
• Definition: Groundbreaking methods/theory introducing new directions or solving major challenges.

• Examples: Entirely new paradigm for foundational models; a novel theory transforming representation learning.

• Novelty 7-8 (Improvements)

• Definition: Substantial insights/enhancements, though not a full paradigm shift.

• Examples: Modifications on existing methods yielding significantly better results.

• Novelty 5-6 (Borderline)

• Definition: Incremental contributions with possible long-term benefits, not immediately transformative.

• Examples: Moderately novel extension to an existing architecture; refining current methods without fundamentally altering them.

• Novelty 3-4 (Tangential)

• Definition: Minor or domain-specific improvements with limited broader impact.

• Examples: Slight modifications to known methods with strange motivation; purely engineering jobs like a new benchmark/dataset.

• Novelty 1-2 (Low)

• Definition: Minimal originality, applying standard approaches without real innovation.
• Examples: Using an off-the-shelf model without adding new insights; purely application-driven studies like finetuning a pretrained model using existing methods.

Papers

[PAPER LIST HERE]

Instructions

Write the response in JSONL format with {ARXIVID, COMMENT, RELEVANCE, NOVELTY} on each line, one for each paper.

• ARXIVID: should be the ArXiv ID.
• COMMENT: should identify whether there is a criteria that match the paper very closely. These matches should not be based on general terms like "language modeling" or "advancements" and should specifically refer to a criterion. No need to mention the non-matching criteria.
• RELEVANCE: should be a score from 1-10.
• NOVELTY: should be a score from 1-10.