论文检索

Deep Learning · Large Language Models

Yash Akhauri, Mohamed Abdelfattah

Efficient LLM inference research has largely focused on reducing the cost of each decoding step (e.g., using quantization, pruning, or sparse attention), typically applying a uniform computation budget to every generated token. In practice, token difficulty varies widely, so static compression can over-compute on easy steps and under-compute on hard ones. We study **dynamic budget allocation** for autoregressive decoding: learning how much computation to spend **per token** from within a single model. Self-Optimizing Language Models (SOL) pair a frozen LLM with a lightweight policy network that reads the LLM hidden state and selects a discrete **efficiency action** at each decode step. Actions can jointly control (i) token-level attention sparsity, (ii) structured activation pruning in the MLP, and (iii) activation quantization bit-width, while leaving the base model weights unchanged. We train the policy with group-relative policy optimization on teacher-forced episodes: the token sequence is fixed, while we sample multiple compute schedules (i.e., “counterfactual” schedules that vary only the efficiency actions for the same token path) and compare their likelihoods under the same supervision. Our reward trades off language‑model quality against soft penalties that encourage episode‑average budget usage to match a requested target. Across model variants and compute regimes, SOL improves quality at matched budget over static allocation and strong random schedule search, offering a complementary axis for inference‑efficiency optimization. SOL discovers a better quality-efficiency pareto-front across all our experiments and improves MMLU accuracy by upto 7.3\% over uniform budget allocation strategies.

Deep Learning · Large Language Models

Yash Akhauri, Xingyou Song, Arissa Wongpanich, Bryan Lewandowski, Mohamed Abdelfattah

We study \textbf{code-to-metric regression}: predicting numeric outcomes of code executions, a challenging task due to the open-ended nature of programming languages. While prior methods have resorted to heavy and domain-specific feature engineering, we show that a single unified Regression Language Model (RLM) can simultaneously predict directly from text, (i) the memory footprint of code across multiple high-level languages such as Python and C++, (ii) the latency of Triton GPU kernels, and (iii) the accuracy and speed of trained neural networks represented in ONNX. In particular, a relatively small 300M parameter RLM initialized from T5Gemma, obtains $>$0.9 Spearman-rank on competitive programming submissions from APPS, and a single unified model achieves $>$0.5 average Spearman-rank across 17 separate languages from CodeNet. Furthermore, the RLM can obtain the highest average Kendall-Tau of 0.46 on five classic NAS design spaces previously dominated by graph neural networks, and simultaneously predict architecture latencies on numerous hardware platforms.

Deep Learning · Large Language Models

Junnan Ren, Yan Zhang, Qian Chen, Yunhang Shen, Ke Li, Shengchuan Zhang, Liujuan Cao, Rongrong Ji

Recent Large Reasoning Models (LRMs) have demonstrated powerful multi-step problem-solving capabilities but often suffer from inefficiency due to an ``overthinking phenomenon", where they apply complex reasoning to simple tasks, resulting in unnecessary computational cost and latency. While adaptive reasoning models that can switch between generating explicit reasoning and producing direct answers offer a potential solution, their effectiveness is compromised by a critical flaw: they are often misled by superficial linguistic complexity, mistaking verbosely phrased simple problems for complex ones. To address this, we propose a two-stage training framework to create a more robust adaptive reasoner. The first stage uses supervised fine-tuning with augmented data—presenting simple problems in both concise and redundant forms—to teach the model to ignore superficial verbosity. Subsequently, a reinforcement learning phase utilizes Generalized Reward Policy Optimization (GRPO) with a custom reward function to refine the model's adaptive policy, ensuring it selects a reasoning mode based on true task complexity rather than surface-level cues. The resulting model reduces computational overhead without sacrificing accuracy and demonstrates improved robustness to misleading linguistic cues.

Reinforcement Learning · Inverse

Raphaël Baur, Yannick Metz, Maria Gkoulta, Mennatallah El-Assady, Giorgia Ramponi, Thomas Kleine Buening

Reward learning typically relies on a single feedback type or combines multiple feedback types using manually weighted loss terms. Currently, it remains unclear how to jointly learn reward functions from heterogeneous feedback types such as demonstrations, comparisons, ratings, rankings, and stops that provide qualitatively different signals. We address this challenge by formulating reward learning from multiple feedback types as Bayesian inference over a shared latent reward function, where each feedback type contributes information through an explicit likelihood. We introduce a scalable amortized variational inference approach that learns a shared reward encoder and feedback-specific likelihood decoders and is trained by optimizing a single evidence lower bound. Our approach avoids reducing feedback to a common intermediate representation and eliminates the need for manual loss balancing. Across discrete and continuous-control benchmarks, we show that jointly inferred reward posteriors outperform single-type baselines, exploit complementary information across feedback types, and yield policies that are more robust to environment perturbations. The inferred reward uncertainty further provides interpretable signals for analyzing model confidence and consistency across feedback types.

Deep Learning · Large Language Models

Avinandan Bose, Stella Li, Faeze Brahman, Pang Wei Koh, Simon Du, Yulia Tsvetkov, Maryam Fazel, Lin Xiao, Asli Celikyilmaz

Cold-start personalization requires inferring preferences from minimal interaction when no user-specific historical data is available. The space of possible preferences is vast, yet users care about only a sparse subset and rarely articulate them upfront; combined with limited interaction budgets, this makes preference elicitation challenging. Our key insight is that preferences exhibit predictable structure across populations; e.g., users who want detailed explanations often also value worked examples. We propose PEP (Preference Elicitation with Priors), a principled system decomposition framework for cold-start personalization: learning a structured world model of preference correlations offline using latent variables, then performing Bayesian inference online without retraining. Even simple belief model instantiations (e.g., linear regression) substantially outperform end-to-end RL. Across medical, mathematical, social, and commonsense reasoning, PEP achieves 80.8% alignment with ground-truth user preferences versus 68.5% for RL, requires 3-5× fewer interactions, and adapts twice as often. Our contribution is a principled decomposition of cold-start personalization that makes Bayesian preference elicitation practical at scale for LLM systems.

Deep Learning · Robustness

Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Chao Li, Cesar F Caiafa, Junbin Gao, Qibin Zhao

Adversarial purification uses generative models to restore clean data distributions from unseen attacks without retraining classifiers. However, unimodal diffusion-based approaches struggle to preserve semantic consistency, while recent multimodal variants rely on computationally expensive adversarial training or distillation. Both approaches often lack theoretical guarantees. In this work, we propose MultiDAP, a novel framework leveraging multimodal diffusion models for efficient adversarial purification. MultiDAP first learns continuous class-agnostic prompts from clean data to capture rich semantic priors, replacing rigid hand-crafted templates. Guided by these prompts, MultiDAP purifies adversarial inputs by minimizing a regularized DDPM loss for only a few steps (e.g., 5-20). We provide theoretical guarantees for both the likelihood improvement via prompt learning and the convergence of the purification process. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet-1K demonstrate that MultiDAP matches the robustness of state-of-the-art baselines but with improved efficiency.

General Machine Learning · Everything Else

Junhua Zeng, Yuning Qiu, Binghua Li, Chao Li, Qibin Zhao, Guoxu Zhou

Over the years, the unsupervised and supervised learning research directions of tensor networks (TNs) have mainly developed in parallel. In this paper, we provide a view for their cooperative advancement through a novel mixed tensor network learning (MTNL) framework that unifies the two fields. Specifically, inspired by supervised TN learning tasks, multiple TNs are fused in a deep-network style in MTNL, enhancing the expressive power for the unsupervised TN learning tasks. We then develop a more flexible TN structure search prior with theoretical guarantees for learning multiple TN structures, aligning with trends in many supervised learning setups. More interestingly, by combining these components within a Bayesian framework, we show that MTNL induces a lightweight uncertainty quantification mechanism that is theoretically guaranteed by its connection to the dropout-based counterpart problem, making the mechanism a potential alternative for large-scale learning problems. Finally, we demonstrate the effectiveness of the MTNL framework on tensor recovery, parameter-efficient fine-tuning, and tensor regression experiments.

Xianzhen Luo, Jingyuan Zhang, Shiqi Zhou, JinYang Huang, Chuan Xiao, Qingfu Zhu, Zhiyuan Ma, YUE XING, Yang Yue, WencongZeng 等

Evaluating and improving the security capabilities of code agents requires high-quality, executable vulnerability tasks. However, existing works rely on costly, unscalable manual reproduction and suffer from outdated data distributions. To address these, we present CVE-Factory, the first multi-agent framework to achieve expert-level quality in automatically transforming sparse CVE metadata into fully executable agentic tasks. Cross-validation against human expert reproductions shows that CVE-Factory achieves 95\% solution correctness and 96\% environment fidelity, confirming its expert-level quality. It is also evaluated on the latest realistic vulnerabilities and achieves a 66.2\% verified success. This automation enables two downstream contributions. First, we construct LiveCVEBench, a continuously updated benchmark of 190 tasks spanning 14 languages and 153 repositories that captures emerging threats including AI-tooling vulnerabilities. Second, we synthesize over 1,000 executable training environments, the first large-scale scaling of agentic tasks in code security. Fine-tuned Qwen3-32B improves from 5.3\% to 35.8\% on LiveCVEBench, surpassing Claude 4.5 Sonnet, with gains generalizing to Terminal Bench (12.5\% to 31.3\%). We open-source all code, data, and models.

General Machine Learning · Causality

Michael Lindon, Nathan Kallus

Delayed outcomes are ubiquitous in online experimentation. When such a temporal dimension is present, treatment influences not only the outcome value but also the outcome timing, which can move in opposite directions. Motivated by the desire to continuously monitor the performance of treatment arms, we develop an anytime-valid approach to inference in the delayed outcome setting. We adopt a design-based framework where both the outcome timing and value are fixed potential outcomes, and randomness is introduced by treatment assignment only. We target the sample cumulative reward as a function of time, a causal estimand that avoids modeling the unobserved future, which would require strong assumptions violated by the nonstationarity and heterogeneity of our setting. We prove that the estimation error for the Horvitz-Thompson (IPW) estimator forms a martingale with respect to a specific single-arm filtration. Conversely, the estimation error for the AIPW estimator fails to be adapted to this filtration. We prove a fundamental negative result for the treatment effect: the estimation error is not a martingale under any filtration, arising from cross-arm covariance induced by randomized assignment. We resolve this using a union bound, showing it yields tighter intervals than the standard variance upper bound when treatment induces asymmetry in outcome arrival rates.

Deep Learning · Theory

Ruizhe Shi, Minhak Song, Runlong Zhou, Zihan Zhang, Maryam Fazel, Simon Du

We present a fine-grained theoretical analysis of the performance gap between reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) under a representation gap. Our study decomposes this gap into two sources: an explicit representation gap under exact optimization and an implicit representation gap under finite samples. In the exact optimization setting, we characterize how the relative capacities of the reward and policy model classes influence the final policy qualities. We show that RLHF, DPO, or online DPO can outperform one another depending on type of model mis-specifications. Notably, online DPO can outperform both RLHF and standard DPO when the reward and policy model classes are isomorphic and both mis-specified. In the approximate optimization setting, we provide a concrete construction where the ground-truth reward is implicitly sparse and show that RLHF requires significantly fewer samples than DPO to recover an effective reward model, highlighting a statistical advantage of two-stage learning. Together, these results provide a comprehensive understanding of the performance gap between RLHF and DPO under various settings, and offer practical insights into when each method is preferred.

Deep Learning · Large Language Models

Fangru Lin, Valentin Hofmann, Xingchen Wan, Weixing Wang, Zifeng Ding, Anthony Cohn, Janet Pierrehumbert

Large language models (LLMs) are trained and tested extensively on symbolic representations such as code and graphs, yet real-world user tasks are often specified in natural language. To what extent can LLMs generalize across these representations? Here, we approach this question by studying isomorphic tasks involving procedures represented in code, graphs, and natural language (e.g., scheduling steps in planning). We find that training LLMs with popular post-training methods on graphs or code data alone does not reliably generalize to corresponding natural language tasks, while training solely on natural language can lead to inefficient performance gains. To address this gap, we propose a two-stage data curriculum that first trains on symbolic, then natural language data. The curriculum substantially improves model performance across model families and tasks. Remarkably, a 1.5B Qwen model trained by our method can closely match zero-shot GPT-4o in naturalistic planning. Finally, our analysis suggests that successful cross-representation generalization can be interpreted as a form of generative analogy, which our curriculum effectively encourages.

Wenbo Pan, Zhichao Liu, Xianlong Wang, Yu Haining, Xiaohua Jia

Token attribution methods provide intuitive explanations for language model outputs by identifying causally important input tokens. However, as modern LLMs increasingly rely on extended reasoning chains, existing schemes face two critical challenges: (1) efficiency bottleneck, where attributing a sequence of $|\mathbf{S}|$ tokens requires $\mathcal{O}(|\mathbf{S}|^2)$ operations, making long-context attribution prohibitively slow; and (2) faithfulness drop, where intermediate reasoning tokens absorb attribution mass, preventing importance from propagating back to the original input. To address these, we introduce **FlashTrace**, an efficient multi-token attribution method that employs span-wise aggregation to compute attribution over *multi-token targets in a single pass*, reducing complexity to $\mathcal{O}(|\mathbf{S}|)$. Moreover, we design a recursive attribution mechanism that traces importance through intermediate reasoning chains back to source inputs. Extensive experiments on long-context retrieval (RULER) and multi-step reasoning (MATH, MorehopQA) tasks demonstrate that FlashTrace achieves over 130× speedup over existing baselines while maintaining superior faithfulness. We further analyze the dynamics of recursive attribution, showing that even a single recursive hop substantially improves faithfulness by tracing importance through the reasoning chain.

Deep Learning · Large Language Models

Zhiwei Ning, Xuanang Gao, Jiaxi Cao, Gengming Zhang, Shengnan Ma, Wenwen Tong, Hanming Deng, JIE YANG, Wei Liu

Multimodal large language models (MLLMs) have achieved remarkable success in general perception, yet complex multi-step visual reasoning remains a persistent challenge. Although recent agentic approaches incorporate tool use, they often neglect critical execution feedback. Consequently, they suffer from the imagination-action-observer (IAO) bias, a misalignment between prior imagination and observer feedback that undermines reasoning stability and optimality. To bridge this gap, we introduce V-ABS, an action-observer driven beam search framework that enables deliberate reasoning through thinker-actor-observer iterations. We also propose an entropy-based adaptive weighting algorithm to mitigate the IAO bias by dynamically balancing the confidence scores between the policy priors and the observational feedback. Moreover, we construct a large-scale supervised fine-tuning (SFT) dataset comprising over 80k samples to guide the model to assign higher prior confidence to correct action paths. Extensive experiments across eight diverse benchmarks show that V-ABS achieves state-of-the-art performance, delivering an average improvement of 19.7\% on the Qwen3-VL-8B baseline and consistent gains across both open-source and proprietary models.

Vincent Siu, Nicholas Crispino, David Park, Nathan Henry, Zhun Wang, Yang Liu, Dawn Song, Chenguang Wang

We introduce STEERINGSAFETY, a benchmark for evaluating representation steering methods across nine safety perspectives spanning 18 datasets. While prior work highlights general capabilities of representation steering, we focus on safety perspectives including bias, harmfulness, hallucination, social behaviors, reasoning, epistemic integrity, and normative judgment. Our benchmark provides modularized building blocks for state-of-the-art steering methods, enabling unified implementation of DIM, ACE, CAA, PCA, and LAT with recent enhancements like conditional steering. Results on Gemma-2-2B, Llama-3.1-8B, and Qwen-2.5-7B reveal that strong steering performance depends critically on pairing of method, model, and specific perspective. For instance, DIM shows consistent effectiveness, but all methods exhibit substantial entanglement - where improving effectiveness on one perspective changes performance in other safety perspectives. Social behaviors show highest vulnerability (reaching degradation as high as 76%), jailbreaking often compromises normative judgment such as commonsense morality (degradation up to 26%), and hallucination steering unpredictably shifts political views, from 21% shifts right to 19% shifts to the political left. Our findings underscore the critical need for understanding steering methods from various safety angles.

Rajat Vadiraj Dwaraknath, Sungyoon Kim, Mert Pilanci

Sparse sketches such as the sparse Johnson–Lindenstrauss transform are a core primitive in randomized numerical linear algebra because they leverage random sparsity to reduce the arithmetic cost of sketching, while still offering strong approximation guarantees. Their random sparsity, however, is at odds with efficient implementations on modern GPUs, since it leads to irregular memory access patterns that degrade memory bandwidth utilization. Motivated by this tension, we pursue a sketch–kernel co-design approach: we design a new family of sparse sketches, BlockPerm-SJLT, whose sparsity structure is chosen to enable FlashSketch, a corresponding optimized CUDA kernel that implements these sketches efficiently. The design of BlockPerm-SJLT introduces a tunable parameter that explicitly trades off the tension between GPU-efficiency and sketching robustness. We provide theoretical guarantees for BlockPerm-SJLT under the oblivious subspace embedding (OSE) framework, and also analyze the effect of the tunable parameter on sketching quality. We empirically evaluate FlashSketch on standard RandNLA benchmarks, as well as an end-to-end ML data attribution pipeline called GraSS. FlashSketch pushes the Pareto frontier of sketching quality versus speed, across a range of regimes and tasks, and achieves a global geomean speedup of roughly $1.7 \times$ over the prior state-of-the-art GPU sketches.

Deep Learning · Theory

Vicente Mendes, Lorenzo Bardone, Cédric Koller, Jorge Medina Moreira, Vittorio Erba, Emanuele Troiani, Lenka Zdeborova

Many real-world datasets contain hidden structure that cannot be detected by simple linear correlations between input features. For example, latent factors may influence the data in a coordinated way, even though their effect is invisible to covariance-based methods such as PCA. In practice, nonlinear neural networks often succeed in extracting such hidden structure in unsupervised and self-supervised learning. However, constructing a minimal high-dimensional model where this advantage can be rigorously analyzed has remained an open theoretical challenge. We introduce a tractable high-dimensional spiked model with two latent factors: one visible to covariance, and one statistically dependent yet uncorrelated, appearing only in higher-order moments. PCA and linear autoencoders fail to recover the latter, while a minimal nonlinear autoencoder provably extracts both. We analyze both the population risk, and empirical risk minimization. Our model also provides a tractable example where self-supervised test loss is poorly aligned with representation quality: nonlinear autoencoders recover latent structure that linear methods miss, even though their reconstruction loss is higher.

Lorenzo Bardone, Claudia Merger, Sebastian Goldt

While diffusion models have emerged as a powerful class of generative models, their learning dynamics remain poorly understood. We address this issue first by empirically showing that standard diffusion models trained on natural images exhibit a simplicity bias, learning simple, pair-wise input statistics before specializing to higher-order correlations. We reproduce this behaviour in simple denoisers trained on a minimal data model, the mixed cumulant model, where we precisely control both pair-wise and higher-order correlations of the inputs. We identify a scalar invariant of the model that governs the sample complexity of learning pair-wise and higher-order correlations that we call the _diffusion information exponent_, in analogy to related invariants in different learning paradigms. Using this invariant, we prove that the denoiser learns simple, pair-wise statistics of the inputs at linear sample complexity, while more complex higher-order statistics, such as the fourth cumulant, require at least cubic sample complexity. We also prove that the sample complexity of learning the fourth cumulant is linear if pair-wise and higher-order statistics share a correlated latent structure. Our work describes a key mechanism for how diffusion models can learn distributions of increasing complexity.

Deep Learning · Large Language Models

Yue Yu, Qiwei Di, Quanquan Gu, Dongruo Zhou

Test-time compute (TTC) has become an increasingly prominent paradigm for enhancing large language models (LLMs). Despite the empirical success of methods such as best-of-$n$ (BoN) sampling and sequential revision, their fundamental limits remain unclear. We address this gap by analyzing a mixture-of-reference policy model and proving that standard BoN is inherently suboptimal. To move closer to the optimal frontier, we study reward-filtered sequential inference, a simple procedure that selectively incorporates only high-reward generations into the context. This mechanism concentrates computation on superior policy candidates and suppresses inferior ones. On the theoretical side, we show that reward-filtered sequential inference yields strictly stronger guarantees than standard TTC paradigms. On the empirical side, we evaluate such an inference strategy across diverse benchmarks and observe consistent improvements over widely used approaches, demonstrating the practical effectiveness of our framework.

Applications · Chemistry, Physics, and Earth Sciences

Mujie Lin, Yutian Liu, Yudi Guo, Yanzhen Hou, Yiheng Tao, Ruochong Zheng, Kaiwen Cheng, Xin Shan, Youdong Mao, Jie Chen

Generating long-horizon, all-atom molecular dynamics (MD) is difficult due to error accumulation in time-domain autoregressive models (causing drift) and fixed step-size constraints on temporal resolution. We propose **BioDynaSpec**, which reformulates protein dynamics as spatio-spectral generation: **Independent Windowed Fourier Decomposition (IWFD)** decomposes trajectories into independent windowed frequency representations, and a generator combines low-to-high frequency autoregression with diffusion denoising to reconstruct continuous motion. We introduce **Inter-Residue Frequency Coupling (IRFC)** bias, a learnable Gaussian distance bias in attention that injects a resonance-inspired structural prior to stabilize training and improve cross-residue, cross-frequency consistency. On ATLAS, BioDynaSpec improves 250-frame trajectory generation with $R_{250}=1.509$ Å (where $R_s$ denotes the mean per-frame C$\alpha$-RMSE over the first $s$ frames after alignment), reducing error by 60.4\% vs. MDGEN and 57.2\% vs. ProAR, and achieves the best PCA-2D displacement-profile correlation and stepwise distribution matching. For equilibrium conformational sampling, it achieves Root Mean $W_2=1.31$, MD PCA $W_2=0.90$, and Joint PCA $W_2=1.19$ (50.03\%, 35.25\%, and 47.58\% lower than the next best), while ablations show removing IRFC severely degrades RMSE/MAE and correlation.

Deep Learning · Large Language Models

Xiaoling Zhou, Shuaiyu Zhou, Zhemg Lee, Tao Chen, Xirui Li, Peng Chen, Jie Jiang, Wei Ye, Shikun Zhang

Although reinforcement learning (RL) enhances the reasoning capabilities of large language models (LLMs), it is primarily learned from the model's self-generated distribution, limiting its ability to acquire reasoning skills beyond its initial knowledge. To overcome this, we propose a Difficulty-Aware Learning Strategy Allocation (DALSA) framework, which adaptively assigns appropriate learning strategies to samples based on their difficulty signals. DALSA is built on the key insight that samples beyond models' knowledge scope are better addressed through supervised fine-tuning (SFT), while those within the boundary but insufficiently mastered benefit more from RL, and well-learned samples are discarded to avoid redundant updates. To realize this principle, we extract a series of difficulty-aware training characteristics and employ a learnable strategy allocator to dynamically determine the optimal learning strategy for each sample based on its training dynamics. The allocator and the LLM are alternately optimized, enabling adaptive strategy allocation. Furthermore, two regularization techniques, anti-curriculum weighting and adversarial label smoothing, are integrated to alleviate the inherent limitations of RL and SFT, backed by comprehensive theoretical analyses. Extensive experiments on ten LLMs ranging from 1.5B to 70B across various tasks indicate that DALSA consistently outperforms baselines under both full and parameter-efficient fine-tuning settings.