论文检索

Deep Learning · Theory

Themistoklis Haris, Zihan Zhang, Yuichi Yoshida

Understanding simplicity biases in deep learning offers a promising path toward developing reliable AI. A common metric for this, inspired by Boolean function analysis, is average sensitivity, which captures a model's robustness to single-token perturbations. We argue that average sensitivity has two key limitations: it lacks a natural generalization to real-valued domains and fails to explain the "junta-like" input dependence we empirically observe in modern LLMs. To address these limitations, we propose *noise stability* as a more comprehensive simplicity metric. Noise stability expresses a model's robustness to correlated noise applied to *all* input coordinates simultaneously. We provide a theoretical analysis of noise stability for single-layer attention and ReLU MLP layers and tackle the multi-layer propagation problem with a covariance interval propagation approach. Building on this theory, we develop a practical *noise stability regularization* method. Experiments on algorithmic and next-token-prediction tasks show that our regularizer consistently catalyzes grokking and accelerates training by approximately $35$\% and $75$\% respectively. Our results establish noise stability as a powerful tool for understanding and improving modern Transformers.

Applications · Language, Speech and Dialog

Hao Zhu, Phil Cuvin, Xinkai Yu, Charlotte Yan, Jason Zhang, Diyi Yang

Agents are predominantly evaluated and optimized via task success metrics, which are coarse, rely on manual design from experts, and fail to reward intermediate emergent behaviors. We propose AutoLibra, a framework for agent evaluation, that transforms open-ended human feedback e.g. “If you find that the button is disabled, don’t click it again”, or “This agent has too much autonomy to decide what to do on its own” into metrics for evaluating fine-grained behaviors in agent trajectories. AutoLibra accomplishes this by grounding feedback to an agent’s behavior, clustering similar positive and negative behaviors, and creating concrete metrics with clear definitions and concrete examples, which can be used for prompting LLM-as-a-Judge as evaluators. We further propose two meta-metrics to evaluate the alignment of a set of (induced) metrics with open feedback: “coverage” and “redundancy”. Through optimizing these meta-metrics, we experimentally demonstrate AutoLibra’s ability to induce more concrete agent evaluation metrics than the ones proposed in previous agent evaluation benchmarks and discover new metrics to analyze agents. We also present two applications of AutoLibra in agent improvement: First, we show that AutoLibra serve human prompt engineers for diagonalize agent failures and improve prompts iterative. Moreover, we find that AutoLibra can induce metrics for automatic optimization for agents, which makes agents improve through self-regulation. Our results suggest that AutoLibra is a powerful task-agnostic tool for evaluating and improving language agents.

Vaishnavi Shrivastava, Ahmed H Awadallah, Vidhisha Balachandran, Shivam Garg, Harkirat Behl, Dimitris Papailiopoulos

Large language models trained with reinforcement learning on verifiable rewards often inflate response length—trading brevity for accuracy. While longer reasoning can help on hard problems, many extra tokens are filler: verbose text making little progress. We introduce GFPO (Group Filtered Policy Optimization), which curbs this length explosion by sampling larger groups per problem and only training on responses filtered by (1) length and (2) token efficiency (reward per token). By sampling more during training time, GFPO teaches models to think less at inference time. On Phi-4-reasoning, GFPO cuts GRPO’s length inflation by up to 85\% across STEM and coding benchmarks (AIME 24/25, GPQA, Omni-MATH, LiveCodeBench) while preserving accuracy. We find that GFPO also outperforms Dr. GRPO in both accuracy and length reduction and generalizes across model sizes and families. We further propose Adaptive Difficulty GFPO, which allocates more training exploration to harder problems, yielding better efficiency-accuracy trade-offs on challenging questions. With only a 7\% increase in training time, GFPO reduces end-to-end latency by $\sim$30\%, cutting response time on hard queries by 90 seconds. GFPO trades modest training-time increases for lasting gains in inference—an effective recipe for efficient reasoning.

Social Aspects · Fairness, Equity, Justice and Safety

Tri Cao, Bennett Lim, Yue Liu, Yuan Sui, YUEXIN LI, Shumin Deng, Lin Lu, Nay Oo, Shuicheng YAN, Bryan Hooi

Computer-Use Agents (CUAs) with full system access enable powerful task automation but pose significant security and privacy risks due to their ability to manipulate files, access user data, and execute arbitrary commands. While prior work has focused on browser-based agents and HTML-level attacks, the vulnerabilities of CUAs remain underexplored. In this paper, we propose an end-to-end threat model where Visual Prompt Injection (VPI) manipulates CUAs in black-box settings to perform unauthorized actions or leak sensitive information, capturing the entire attack chain from injection to harmful outcomes. Then, we propose VPI-Bench, a benchmark of 306 test cases across five widely used platforms, to evaluate agent robustness under VPI threats. Each test case is a variant of a web platform, designed to be interactive, deployed in a realistic environment, and containing a visually embedded malicious prompt. Our empirical study shows that current CUAs and BUAs can be deceived at rates of up to 51\% and 100\%, respectively, on certain platforms. The experimental results also indicate that existing defense methods offer only limited improvements. These findings highlight the need for robust, context-aware defenses to ensure the safe deployment of multimodal AI agents in real-world environments.

Deep Learning · Theory

Shyam Venkatasubramanian, Sean Moushegian, Ahmed Aloui, VAHID TAROKH

Quantifying the limitations of classical neural network architectures is a critically underexplored area of machine learning research. Deriving lower bounds on the optimal performance of these architectures can facilitate improved neural architecture search and overfitting detection. We present an information-theoretic lower bound on the generalization mean squared error of autoencoders with sigmoid activation functions. Through the Estimation Error and Differential Entropy (EEDE) inequality for continuous random vectors, we derive this lower bound, which provides a new perspective on the inherent limitations and capabilities of autoencoders. Our analysis extends to the examination of how this lower bound is influenced by various architectural features and data distribution characteristics. This study enriches our theoretical understanding of autoencoders and has substantial practical implications for their design, optimization, and application in the field of deep learning.

Social Aspects · Accountability, Transparency and Interpretability

Borui Yang, Mingxuan Ma, Liyao Xiang, Nan Chen, Xin Zhang, Linghe Kong, Xinghao Jiang

Code language models (LMs) represent valuable intellectual property (IP) as their training involves immense investments, including large-scale code corpora, proprietary annotations, extensive computational resources, and specialized designs. Hence the threat of model IP infringements such as unauthorized redistribution or model theft has become increasingly concerning. While neural network watermarking has been widely studied as a measure to support model ownership verification, watermarking code LMs is particularly challenging due to the seemingly conflicting requirements of code generation: adhering to strict syntactic rules and semantic consistency while allowing flexible changes to embed watermarks, keeping high fidelity of the generated content while being robust to extraction attacks, etc. To resolve the issues, we propose CodeGenGuard, a watermarking framework for code LMs. CodeGenGuard leverages semantic-preserving transformations (SPTs) to encode the watermark and incorporates a dead-code-based data augmentation pipeline to diversify SPT patterns. To improve robustness, we incorporate an efficient dual-LoRA shadow training scheme and an optimizable trigger prompt that learns to extract watermark from both the watermarked and the shadow models. As most SPTs take place in specific contexts, we implant auxiliary prompts during verification to encourage the generation of the context, further enhancing the detection rate. Evaluation results on representative code generation models demonstrate that CodeGenGuard achieves superior watermarking performance to the state-of-the-art.

Yao Lu, Hao Cheng, Yujie Fang, Zeyu Wang, Jiaheng Wei, Dongwei Xu, Qi Xuan, Zhaowei Zhu

Although large language models (LLMs) have achieved remarkable success across various domains, their considerable scale necessitates substantial computational resources, posing significant challenges for deployment in resource-constrained environments. Layer pruning, as a simple yet effective compression method, removes layers of a model directly, reducing computational overhead. However, what are the best practices for layer pruning in LLMs? Are sophisticated layer selection metrics truly effective? Does the LoRA (Low-Rank Approximation) family, widely regarded as a leading method for pruned model fine-tuning, truly meet expectations when applied to post-pruning fine-tuning? To answer these questions, we dedicate thousands of GPU hours to benchmarking layer pruning in LLMs and gaining insights across multiple dimensions. Our results demonstrate that a simple approach, i.e., pruning the final layers followed by fine-tuning the lm\_head and the remaining last three layers, yields remarkably strong performance. These pruning strategies are further supported by theoretical analyses based on the gradient flow. Following this guide, our method surpasses existing state-of-the-art pruning methods by $5.62\%$–$17.27\%$ on Llama-3.1-8B-It, by $2.36\%$–$19.45\%$ on Llama-3-8B and by $4.34\%$–$9.59\%$ on Llama-3-70B. The code is available at at https://github.com/yaolu-zjut/Navigation_LLM_layer_pruning.

Reinforcement Learning · Multi-agent

Ziyue Wang, Junde Wu, Linghan Cai, Chang Low, Xihong Yang, Qiaxuan Li, Yueming Jin

Modern clinical diagnosis relies on the comprehensive analysis of multi-modal patient data, drawing on medical expertise to ensure systematic and rigorous reasoning. Recent advances in Vision–Language Models (VLMs) and agent-based methods are reshaping medical diagnosis by effectively integrating multi-modal information. However, they often output direct answers and empirical-driven conclusions without clinical evidence supported by quantitative analysis, which compromises their reliability and hinders clinical usability. Here we propose MedAgent-Pro, an agentic reasoning paradigm that mirrors modern diagnosis principles via a hierarchical diagnostic workflow, consisting of disease-level standardized plan generation and patient-level personalized step-by-step reasoning. To support disease-level planning, a retrieval-augmented generation agent is designed to access medical guidelines for alignment with clinical standards. For patient-level reasoning, MedAgent-Pro leverages professional tools such as visual models to take various actions to analyze multi-modal input, and performs evidence-based reflection to iteratively adjust memory, enforcing rigorous reasoning throughout the process. Extensive experiments across a wide range of anatomical regions, imaging modalities, and diseases demonstrate the superiority of MedAgent-Pro over mainstream VLMs, agentic systems and leading expert models. Ablation studies and expert evaluation further confirm its robustness and clinical relevance. Anonymized code link is available in the reproducibility statement.

General Machine Learning · Representation Learning

Albert Miao, Chenliang Zhou, Jiawei Zhou, Cengiz Oztireli

Sparse Autoencoders (SAEs) have found human-interpretable features in LLM activations, clarifying how LLMs transform input to output. However, they have rarely been applied outside of text, limiting explorations of feature dynamics. We present the first application of SAEs to the 3D domain, analyzing the features found in 53k 3D objects encoded by a state-of-the-art 3D reconstruction VAE. We observe that the model encodes discrete rather than continuous features, leading to our key finding: the model's feature activations approximate a discrete state space, driven by phase-like transitions. Through this state space framework, we address three otherwise unintuitive behaviors — the preference for positional encoding features, the sigmoidal relationship between feature ablation and reconstruction loss, and the bimodal distribution of phase transition points. This final observation suggests the model redistributes superposition interference to prioritize the high-importance features. Our work not only catalogs and explains unexpected feature dynamics, but also provides a framework to explain the model's learning dynamics. The code is available at https://feature3d.github.io/Dora-SAE/.

General Machine Learning · Everything Else

Hao Cheng, Rui Zhang, Ling Li, Na Di, Jiaheng Wei, Zhaowei Zhu, Bo Han

Task-specific fine-tuning of foundation models is critically dependent on the quality and relevance of the instruction data. While prevailing data selection methods rely exclusively on instruction instances X to approximate the target distribution, we argue that selection should align with the joint distribution of instructions and task-specific labels (X,Y). However, task-specific labels Y are typically unavailable in practice. To address this, we reformulate the task-specific data selection problem and present a novel pipeline that leverages the reasoning capabilities of large language models (LLMs) to infer proxy labels, thereby facilitating joint distribution alignment. Our approach begins by propagating proxy labels from a small target set to a large, unlabeled source corpus. A two-stage filtering process then removes instances with label noise and refines the subset through distribution alignment. This strategy produces more semantically meaningful and task-aware selections than conventional similarity measures based on $X$ alone. Experimental results show that fine-tuning on a subset of only 10K samples, selected from a pool of 300K, achieves performance competitive or superior to state-of-the-art methods. Code is available at https://github.com/tmlr-group/TADS.

Ruochen Mao, Yuling Shi, Xiaodong Gu, Jiaheng Wei

Aligning large language models with human preferences is critical for creating reliable and controllable AI systems. A human preference can be visualized as a high-dimensional vector where different directions represent trade-offs between desired attributes (e.g., helpfulness vs. verbosity). Yet, because the training data often reflects dominant, average preferences, LLMs tend to perform well on com- mon requests but falls short in specific, individual needs. This mismatch creates a preference coverage gap. Existing methods often address this through costly retraining, which may not be generalized to the full spectrum of diverse preferences. This brittleness means that when a user’s request reflects a nuanced preference deviating from the training data’s central tendency, model performance can degrade unpredictably. To address this challenge, we introduce Robust Preference Selection (RPS), a post-hoc, training-free method by leveraging directional neighborhood consensus. Instead of forcing a model to generate a response from a single, highly specific preference, RPS samples multiple responses from a local neighborhood of related preferences to create a superior candidate pool. It then selects the re- sponse that best aligns with the user’s original intent. We provide a theoretical framework showing that, under mild conditions where (i) nearby preference direc- tions correspond to better-trained regions of the model and (ii) the reward-model scores change smoothly with small angular changes in the preference vector, our neighborhood generation strategy yields a higher expected best score than a strong baseline that also samples multiple candidates. Comprehensive experiments across three distinct alignment paradigms (DPA, DPO, and SFT) demonstrate that RPS consistently improves robustness against this baseline, achieving win rates of up to 69% on challenging preferences from under-represented regions of the space without any model retraining. Our work presents a practical, theoretically-grounded solution for enhancing the reliability of preference-aligned models.

Computer Vision · Vision Models & Multimodal

Christian Greisinger, Steffen Eger

Large language models (LLMs) are increasingly used to assist scientists across diverse workflows. A key challenge is generating high-quality figures from textual descriptions, often represented as TikZ programs that can be rendered as scientific images. Prior research has proposed a variety of datasets and modeling approaches for this task. However, existing datasets for Text-to-TikZ are too small and noisy to capture the complexity of TikZ, causing mismatches between text and rendered figures. Moreover, prior approaches rely solely on supervised fine-tuning (SFT), which does not expose the model to the rendered semantics of the figure, often resulting in errors such as looping, irrelevant content, and incorrect spatial relations. To address these issues, we construct DaTikZ-V4, a dataset more than four times larger and substantially higher in quality than DaTikZ-V3, enriched with LLM-generated figure descriptions. Using this dataset, we train TikZilla, a family of small open-source Qwen models (3B and 8B) with a two-stage pipeline of SFT followed by reinforcement learning (RL). For RL, we leverage an image encoder trained via inverse graphics to provide semantically faithful reward signals. Extensive human evaluations with over 1,000 judgments show that TikZilla improves by 1.5-2 points over its base models on a 5-point scale, surpasses GPT-4o by 0.5 points, and matches GPT-5 in the image-based evaluation, while operating at much smaller model sizes. Code, data, and models will be made available.

Reinforcement Learning · Multi-agent

Cédric Colas, Tracey Mills, Ben Prystawski, Michael Tessler, Noah Goodman, Jacob Andreas, Joshua B Tenenbaum

The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models human social learning as joint probabilistic inference over structured, executable world models given sensorimotor and linguistic data. We make this possible by turning a pretrained language model into a probabilistic model of how humans share advice conditioned on their beliefs, allowing our agents both to generate advice for others and to interpret linguistic input as evidence during Bayesian inference. Using behavioral experiments and simulations across 10 video games, we show how linguistic guidance can shape exploration and accelerate learning by reducing risky interactions and speeding up key discoveries in both humans and models. We further explore how knowledge can accumulate across generations through iterated learning experiments and demonstrate successful knowledge transfer between humans and models—revealing how structured, language-compatible representations might facilitate human-machine collaborative learning.

Deep Learning · Attention Mechanisms

Pascal Bergsträßer, Ryan Cotterell, Anthony W. Lin

We propose succinctness as a measure of expressive power of a transformer in describing a concept. To this end, we prove that transformers are highly expressive in that they can represent formal languages substantially more succinctly than standard representations of formal languages like finite automata and Linear Temporal Logic (LTL) formulas. As a by-product of this expressivity, verifying even simple properties of transformers is shown to be provably intractable (i.e. EXPSPACE-complete).

Applications · Language, Speech and Dialog

Hongjue Zhao, Haosen Sun, Jiangtao Kong, Xiaochang Li, Qineng Wang, Liwei Jiang, Qi Zhu, Tarek Abdelzaher, Yejin Choi, Manling Li 等

Activation steering, or representation engineering, offers a lightweight approach to align large language models (LLMs) by manipulating their internal activations at inference time. However, current methods suffer from two key limitations: \textit{(i)} the lack of a unified theoretical framework for guiding the design of steering directions, and \textit{(ii)} an over-reliance on \textit{one-step steering} that fail to capture complex patterns of activation distributions. In this work, we propose a unified ordinary differential equations (ODEs)-based \textit{theoretical} framework for activation steering in LLM alignment. We show that conventional activation addition can be interpreted as a first-order approximation to the solution of an ODE. Based on this ODE perspective, identifying a steering direction becomes equivalent to designing a \textit{barrier function} from control theory. Derived from this framework, we introduce ODESteer, a kind of ODE-based steering guided by barrier functions, which shows \textit{empirical} advancement in LLM alignment. ODESteer identifies steering directions by defining the barrier function as the log-density ratio between positive and negative activations, and employs it to construct an ODE for \textit{multi-step and adaptive} steering. Compared to state-of-the-art activation steering methods, ODESteer achieves consistent empirical improvements on diverse LLM alignment benchmarks, a notable $5.7\%$ improvement over TruthfulQA, $2.5\%$ over UltraFeedback, and $2.4\%$ over RealToxicityPrompts. Our work establishes a principled new view of activation steering in LLM alignment by unifying its theoretical foundations via ODEs, and validating it empirically through the proposed ODESteer method.

General Machine Learning · Representation Learning

Zhuochun Li, Yong Zhang, Ming Li, Yuelyu Ji, Yiming Zeng, ning Cheng, Yun Zhu, Yanmeng Wang, Shaojun Wang, Jing Xiao 等

Large language models (LLMs) are widely used as reference-free evaluators via prompting, but this “LLM-as-a-Judge” paradigm is costly, opaque, and sensitive to prompt design. In this work, we investigate whether smaller models can serve as efficient evaluators by leveraging internal representations instead of surface generation. We uncover a consistent empirical pattern: small LMs, despite with weak generative ability, encode rich evaluative signals in their hidden states. This motivates us to propose the Semantic Capacity Asymmetry Hypothesis: evaluation requires significantly less semantic capacity than generation and can be grounded in intermediate representations, suggesting that evaluation does not necessarily need to rely on large-scale generative models but can instead leverage latent features from smaller ones. Our findings motivate a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge, a decoding-free evaluation strategy that probes internal model structure rather than relying on prompted output. We instantiate this paradigm through INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations. Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show that INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while offering a more efficient, reliable, and interpretable alternative for scalable evaluation. The code and data are available at: https://github.com/zhuochunli/Representation-as-a-judge

Deep Learning · Everything Else

Ryma Boumazouza, Raya Elsaleh, Melanie Ducoffe, Shahaf Bassan, Guy Katz

We propose $\textbf{FAME}$ (Formal Abstract Minimal Explanations), a new class of abductive explanations grounded in abstract interpretation. FAME is the first method to scale to large neural networks while reducing explanation size. Our main contribution is the design of dedicated perturbation domains that eliminate the need for traversal order. FAME progressively shrinks these domains and leverages LiRPA-based bounds to discard irrelevant features, ultimately converging to a $\textbf{formal abstract minimal explanation}$. To assess explanation quality, we introduce a procedure that measures the worst-case distance between an abstract minimal explanation and a true minimal explanation. This procedure combines adversarial attacks with an optional $VERI{\large X}+$ refinement step. We benchmark FAME against $VERI{\large X}+$ and demonstrate consistent gains in both explanation size and runtime on medium- to large-scale neural networks.

Deep Learning · Graph Neural Networks

Jun Chen, Ziyue Qiao, Qin Zhang, Kaize Ding, Xiao Luo

The pretrain-finetuning paradigm has achieved notable success in graph learning. Moreover, merging models fine-tuned on different tasks to enable a parameter-efficient model with multi-task capabilities is gaining increasing attention for its practicality. However, existing model merging methods, such as weight averaging and task arithmetic, struggle to generalize well to graph structures and Graph Neural Network (GNN) models due to the unique structural heterogeneity of graph data. In this paper, we propose an innovative graph model merging framework called G-Merging for merging multiple task-specific fine-tuned GNN models. G-Merging first employs task arithmetic to coarsely merge graph models, capturing shared cross-task knowledge. Second, it introduces a Topology-aware Wasserstein Distance (TWD) loss to train lightweight task adapters, preserving domain-specific graph patterns via aligning the embeddings of merged and fine-tuned models. Third, G-Merging integrates the adapters into a training-free, topology-aware router within a mixture-of-experts (MoE) architecture, dynamically routing input graphs to task-specific adapters based on structural similarity, thereby mitigating conflicts and enhancing knowledge sharing. Extensive experiments on 8 graph downstream datasets demonstrate the effectiveness of G-Merging, showing impressive performance close to or exceeding individual finetuned models while improving parameters and training efficiency. Our code is available at https://github.com/cjcj46262/G-Merging.

Reinforcement Learning · Multi-agent

Haocheng Yang, Fengxiang Cheng, Tianjun Yao, Mengyue Yang, Jiajun Chai, Xiaohan Wang, Guojun Yin, Wei Lin, Soummya Kar, Fenrong Liu 等

Large language models (LLMs) struggle with complex logical reasoning. Previous methods can be briefly summarized into two pipelines: (1) translating natural language (NL) to symbolic language (SL) then reasoning via external solvers, and (2) adopting LLMs to reason directly in NL based on prompting or fine-tuning. However, we point out that on the one hand, the translation relying on a specific SL often fails to capture different important features of raw NL, leading to information loss or translation errors. On the other hand, both two pipelines have unignorable limitations. For example, the former (SL-based) methods are highly sensitive to imperfect translation, and the latter (NL-based) methods are prone to hallucinations. Motivated by this, we are the first to propose a multi-agent debate framework to leverage the strengths of different SLs and reasoning methods, achieving better performance in both translation and reasoning stages. Specifically, in the translation stage, multiple agents translate the NL into different SL and refine translations through debate. In the reasoning stage, multiple agents based on SL (obtained by the corresponding solver) and NL debate multiple rounds, with the final answer determined by majority vote. In addition, to address the inefficiency of multi-agent debates, we introduce an adaptive sparse communication strategy that prunes unnecessary interactions based on agent confidence and information gains. Extensive experiments on three datasets show that our method enhances logical QA performance while reducing computational cost.

Reinforcement Learning · Deep RL

Waris Radji, Thomas Michel, Hector Piteau

Reinforcement learning (RL) research requires diverse, challenging environments that are both tractable and scalable. While modern video games may offer rich dynamics, they are computationally expensive and poorly suited for large-scale experimentation due to their CPU-bound execution. We introduce Octax, a high-performance suite of classic arcade game environments implemented in JAX, based on CHIP-8 emulation, a predecessor to Atari, which is widely adopted as a benchmark in RL research. Octax provides the JAX community with a long-awaited end-to-end GPU alternative to Atari games, offering image-based environments, spanning puzzle, action, and strategy genres, all executable at massive scale on modern GPUs. Our JAX-based implementation achieves orders-of-magnitude speedups over traditional CPU emulators. We demonstrate Octax's capabilities by training RL agents across multiple games, showing significant improvements in training speed and scalability compared to existing solutions. The environment's modular design enables researchers to easily extend the suite with new games or generate novel environments using large language models, making it an ideal platform for large-scale RL experimentation. Our open-source framework is available at https://github.com/riiswa/octax/.