论文检索

Deep Learning · Algorithms

Shwai He, Guoheng Sun, Haichao Zhang, Yun Fu, Ang Li

Network pruning, which removes less important parameters or architectures, is often expected to improve efficiency while preserving performance. However, this expectation does not consistently hold across language tasks: pruned models can perform well on non-generative tasks but frequently fail in generative settings. To demystify how such discrepancies arise under pruning, we analyze network pruning from a representation-hierarchy perspective, decomposing the internal computation of language models into three sequential spaces: \textit{embedding} (hidden representations), \textit{logit} (pre-softmax outputs), and \textit{probability} (post-softmax distributions). While representations in the embedding and logit spaces are largely robust to pruning-induced perturbations, the subsequent nonlinear transformation from logits to the probability space amplifies such deviations, whose persistence across time steps leads to substantial degradation during generation. By contrast, the stability of the categorical-token probability subspace, together with the robustness of the embedding space, supports the effectiveness of pruning for non-generative tasks such as retrieval and multiple-choice selection. Our representation-level analysis disentangles the effects of pruning across tasks and offers practical guidance on its application. Code will be released upon acceptance.

Deep Learning · Large Language Models

Ji Zeng, Dayuan Fu, Tiantian Mi, Zhuang Yumin, Yaxing Huang, Xuefeng Li, Lyumanshan Ye, Muhang Xie, Qishuo Hua, Zhen Huang 等

Recently, the frontier of Large Language Model (LLM) capabilities has shifted from single-turn code generation to agentic software engineering—a paradigm where models autonomously navigate, edit, and test complex repositories. While post-training methods have become the de facto approach for code agents, *agentic mid-training*—mid-training (MT) on large-scale data that mirrors authentic agentic workflows—remains critically underexplored due to substantial resource requirements, despite offering a more scalable path to instilling foundational agentic behaviors than relying solely on expensive reinforcement learning. A central challenge in realizing effective agentic mid-training is the distribution mismatch between static training data and the dynamic, feedback-rich environment of real development. To address this, we present a systematic study of agentic mid-training, establishing both the data synthesis principles and training methodology for effective agent development at scale. Central to our approach is *agent-native data*—supervision comprising two complementary types of trajectories: *contextually-native trajectories* that preserve the complete information flow an agent experiences, offering broad coverage and diversity; and *environmentally-native trajectories* collected from executable repositories where observations stem from actual tool invocations and test executions, providing depth and interaction authenticity. We verify the model’s agentic capabilities on `SWE-Bench Verified`. We demonstrate our superiority over the previous open software engineering mid-training recipe `Kimi-Dev` under two post-training settings with an aligned base model and agentic scaffold, while using less than half mid-training tokens (73.1B). Besides relative advantage, our best performing 32B and 72B models achieve **56.1%** and **58.5%** resolution rates, respectively, which are state-of-the-art among open training recipes using agentic scaffolds under their model sizes, despite starting from non-coder `Qwen2.5-Base` base models. Beyond these agentic capabilities, we also observe performance gains on general code generation and scientific benchmarks. We plan to open-source a significant portion of our datasets, recipes, and model checkpoints—resources representing substantial computational investment typically unavailable to the broader community—to facilitate further research in this underexplored paradigm.

Optimization · Non-Convex

Zuqi Zhang, Chenghe Sun, Xiangyi Chu, Wei-Han Yu, Ka-Fai Un, Rui Martins, Pui-In Mak, Jiawei Xu

Adaptive rounding techniques in Post-Training Quantization (PTQ) enable the efficient deployment of Large Language Models (LLMs) with low resource and data dependencies. While learning-based rounding methods are accurate yet costly, compensation-based approaches offer a highly efficient alternative. However, synergizing these two to realize their full potential is hindered by spatiotemporal misalignment in the decoupled paradigm. Key challenges include temporal parameter conflict, the invalidation of the initial Round-to-Nearest (RTN) assumption, and spatially-inconsistent optimization objectives. This paper introduces STLA, a novel rounding-optimized PTQ framework that achieves both fast and accurate LLM quantization. STLA resolves temporal inconsistency through cluster-wise integrated rounding optimization, which collocates the learning and compensation phases. STLA achieves spatial alignment through a unified global objective derived from the Schur Complement, enabling the solver to look ahead and align local rounding decisions with the optimal future compensation of remaining weights. Furthermore, we propose a Hessian-guided clustering strategy that exploits both diagonal and off-diagonal information to maximize intra-cluster error cancellation. Extensive experiments demonstrate that STLA establishes a new state-of-the-art for low-bit PTQ while maintaining high computational efficiency. The code is available at https://anonymous.4open.science/r/STLA.

General Machine Learning · Transfer, Multitask and Meta-learning

Stefan Horoi, Guy Wolf, Eugene Belilovsky, Gintare Karolina Dziugaite

Modern deep learning is increasingly characterized by the use of open-weight foundation models that can be fine-tuned on specialized datasets. This has led to a proliferation of expert models and adapters, often shared via platforms like HuggingFace and AdapterHub. Model merging has recently emerged as an effective way to leverage these existing resources, enabling the composition of capabilities from different model checkpoints. A natural pipeline has thus formed to harness the benefits of transfer learning and amortize sunk training costs: models are pre-trained on general data, fine-tuned on specific tasks, and then multiple checkpoints are merged to obtain a more capable model. A prevailing assumption is that improvements at one stage of this pipeline propagate downstream, leading to gains at subsequent steps. In this work, we challenge that assumption by examining how expert fine-tuning affects model merging. We show that long fine-tuning of experts that optimizes for their individual performance leads to degraded merging performance across vision and language modalities, multiple model scales, and both fully fine-tuned and LoRA-adapted models. We trace this degradation to the memorization of a small set of difficult examples that dominate late fine-tuning steps. This causes negative parameter interference and encodes knowledge that is forgotten during merging. Finally, we demonstrate that task-dependent aggressive early stopping strategies can significantly improve model merging performance.

Deep Learning · Large Language Models

Chao Wang, Bei Li, Jiaqi Zhang, Xinyu Liu, Yuchun Fan, Linkun Lyu, Xin Chen, Jingang Wang, Tong Xiao, Peng Pei 等

The success of Large Language Models (LLMs) hinges on the stable training of deep Transformer architectures. A critical design choice is the placement of normalization layers, leading to a fundamental trade-off: the ''PreNorm'' architecture ensures training stability at the cost of potential performance degradation in deep models, while the ''PostNorm'' architecture offers strong performance but suffers from severe training instability. In this work, we propose SpanNorm, a novel technique designed to resolve this dilemma by integrating the strengths of both paradigms. SpanNorm adopts the clean residual path of PreNorm to stabilize signal propagation while employing a PostNorm-style computation that normalizes the output of the residual connection, thereby enhancing model performance. We provide a theoretical analysis demonstrating that SpanNorm, combined with a principled scaling strategy, maintains bounded signal variance throughout the network, preventing the gradient issues that plague PostNorm models, and alleviating the representation collapse of PreNorm. Empirically, SpanNorm consistently outperforms standard normalization schemes in both dense and Mixture-of-Experts (MoE) scenarios, paving the way for more powerful and stable Transformer architectures.

Social Aspects · Accountability, Transparency, and Interpretability

Xi Chen, Mingyu Jin, Jingcheng (Frank) Niu, Yutong Yin, Jinman Zhao, Bangwei Guo, Dimitris Metaxas, Zhaoran Wang, Yutao Yue, Gerald Penn

In this paper, we present empirical and theoretical evidence against a central but largely implicit assumption in circuit and sheaf discovery (CSD), which we term the *Functional Anisotropy Hypothesis*: the idea that functions in large language models (LLMs) are localised to a unique or near-unique internal mechanism. We show that a single LLM task can instead be supported by multiple, structurally distinct circuits or sheaves that are simultaneously faithful, sparse, and complete. To systematically uncover such competing mechanisms, we introduce Overlap-Aware Sheaf Repulsion, a method that augments the CSD objective with an explicit penalty on structural overlap across multiple discovery runs, enabling the discovery of circuits or sheaves with strong task performance but minimal shared structure across a plethora of common CSD benchmarks. We find that this phenomenon becomes increasingly pronounced as the number of discovered sheaves grows and persists robustly across major CSD methods. We further identify an ultra-sparse three-edge sheaf and show that none of its edges is individually indispensable, undermining even weakened notions of canonical or essential components. To explain these findings, we propose a *Distributive Dense Circuit Hypothesis* and provide a theoretical analysis demonstrating that non-unique, low-overlap circuit explanations arise naturally from high-dimensional superposition under mild assumptions. Together, our results suggest that mechanistic explanations in LLMs are inherently non-canonical and call for a rethinking of how CSD results should be interpreted and evaluated.

Applications · Computer Vision

Weijie Wang, Xiaoxuan He, Youping Gu, Zeyu Zhang, Yefei He, Yanbo Ding, Donny Y. Chen, Xirui Hu, Zhiyuan He, Yuqing Yang 等

Recent video foundation models demonstrate impressive visual synthesis but frequently suffer from geometric inconsistencies. While existing methods attempt to inject 3D priors via architectural modifications, they often incur high computational costs and limit scalability. We propose World-R1, a framework that aligns video generation with 3D constraints through reinforcement learning. To facilitate this alignment, we introduce a specialized pure text dataset tailored for world simulation. Utilizing Flow-GRPO, we optimize the model using feedback from pre-trained 3D foundation models and vision-language models to enforce structural coherence without altering the underlying architecture. We further employ a periodic decoupled training strategy to balance rigid geometric consistency with dynamic scene fluidity. Extensive evaluations reveal that our approach significantly enhances 3D consistency while preserving the original visual quality of the foundation model, effectively bridging the gap between video generation and scalable world simulation.

Applications · Computer Vision

Ziwei Zhou, Zeyuan Lai, Rui Wang, Yifan Yang, Yuqing Yang, Qi Dai, Lili Qiu, Chong Luo

Text-to-Audio-Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing benchmarks largely assess audio and video in isolation or rely on coarse embedding similarity, failing to capture fine-grained joint correctness required by realistic prompts. We introduce AVGen-Bench, a task-driven benchmark for T2AV generation, featuring high-quality prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with Multimodal Large Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering, speech coherence, physical reasoning, and universal breakdown in musical pitch control.

Deep Learning · Large Language Models

Junyi Li, Xiaowei Qian, Yingyi Zhang, Wenlin Zhang, Guojing Li, Sheng Zhang, Xiao Han, Yichao Wang, Xiangyu Zhao

Recent advances in tool-integrated language agents have significantly improved their ability to solve complex reasoning tasks. However, existing alignment methods predominantly focus on maximizing task accuracy, while overlooking auxiliary objectives such as tool-use efficiency, which are essential for practical deployment. To address this gap, we introduce {ParetoPO}, a two-stage multi-objective optimization framework for aligning tool-using large language models (LLMs) under competing objectives. In the first stage, ParetoPO leverages hypervolume-guided dynamic scalarization to adapt reward weights based on global Pareto frontier progress. In the second stage, it replaces scalarized learning signals with Pareto-ranking-based advantage computation, promoting nondominated trajectories through dominance-aware credit assignment. This design enables fine-grained, action-level optimization across multiple conflicting objectives. Experimental results on mathematic reasoning and deep search tasks show that ParetoPO consistently discovers policies with superior accuracy-efficiency trade-offs compared to static and heuristic baselines.

Reinforcement Learning · Deep RL

Rishabh Tiwari, Aditya Tomar, Udbhav Bamba, Monishwaran Maheswaran, Heng Yang, Michael Mahoney, Kurt Keutzer, Amir Gholaminejad

Process Reward Models (PRMs) are rapidly becoming the backbone of LLM reasoning pipelines, yet we demonstrate that state-of-the-art PRMs are systematically exploitable under optimization pressure. We introduce a three-tiered diagnostic framework that applies increasing adversarial pressure to quantify these vulnerabilities. Static perturbation analysis uncovers a fluency-logic dissociation: high invariance to surface-level style changes (reward changes $<$0.1) yet inconsistent detection of logically corrupted reasoning, with different models failing on different attack types. Adversarial optimization demonstrates that gradient-based attacks inflate rewards on invalid trajectories, with reward landscapes exhibiting wide, exploitable peaks. RL-induced reward hacking exposes the critical failure mode: policies trained on AIME problems achieve near-perfect PRM rewards ($>$0.9) while ground-truth accuracy remains below 4\%, with 43\% of reward gains attributable to stylistic shortcuts. These findings reveal that current PRMs function as fluency detectors rather than reasoning verifiers, creating systematic blind spots that undermine their use as training signals. We release PRM-BiasBench and a diagnostic toolkit to enable robustness evaluation before deployment.

Deep Learning · Large Language Models

Udbhav Bamba, Arnav Chavan, Aryamaan Thakur, Deepak Gupta, Steven Teig

The scaling of Large Language Models (LLMs) has driven significant performance gains but created substantial challenges in inference efficiency. While Mixture of Experts (MoEs) architectures address this by decoupling model size from inference cost, training MoEs from scratch is often unstable and compute intensive. Conversion of pre-trained dense models into sparse MoEs has emerged as an alternative solution; however, existing methods typically rely on heuristic neuron clustering or random splitting to partition Feed-Forward Networks (FFNs) into experts. In this work, we propose DOT-MoE, a novel framework that formulates the decomposition of dense layers as a Differentiable Optimal Transport (DOT) problem. Instead of static heuristics, we model neuron assignment as a balanced transport problem, utilizing differentiable Sinkhorn-Knopp iterations to enforce strict expert capacity constraints. Furthermore, we utilize Straight-Through Estimators (STE) to jointly learn the discrete neuron-to-expert assignment and the token-to-expert routing policy end-to-end. Extensive experiments across multiple architectures and benchmarks demonstrate that DOT-MoE significantly outperforms structured pruning, heuristic clustering, and random-split baselines, retaining 90% of the original dense model's performance while reducing active parameters by 50%.

Xin Ding, Jianyu Wei, Yifan Yang, Shiqi Jiang, QIANXI ZHANG, Hao Wu, Fucheng Jia, Liang Mi, Yuxuan Yan, weijun wang 等

Vision-Language Navigation (VLN) requires agents to follow natural language instructions by grounding them in sequential visual observations over long horizons. Explicit reasoning could enhance temporal consistency and perception–action alignment, but reasoning at fixed steps often leads to suboptimal performance and unnecessary computation. To address this, we propose AdaNav, an uncertainty-based adaptive reasoning framework for VLN. At its core is the Uncertainty-Adaptive Reasoning Block (UAR), a lightweight plugin that dynamically triggers reasoning. We introduce Action Entropy as a policy prior for UAR and progressively refine it through a Heuristics-to-RL training method, enabling agents to learn difficulty-aware reasoning policies under the strict data limitations of embodied tasks. Results show that with only 6K training samples, AdaNav achieves substantial gains over closed-source models trained on million-scale data, improving success rate by 20% on R2R val-unseen, 11.7% on RxR-CE, and 11.4% in real-world scenes.

Optimization · Non-Convex

Christopher Harker, Aditya Bhaskara

The DeepWalk algorithm has been widely used for learning node embeddings in graphs. Combined with the idea of _negative sampling_, the DeepWalk algorithm has been shown to be implementable at scale, easily handling graphs with millions of nodes. However, theoretical guarantees on the resulting embeddings are much less understood. Recent results have studied the minimizers of the objective and have shown interesting guarantees for certain graph classes. However, the optimization _trajectory_, i.e., what happens when we start at a random initialization and run gradient descent, remains poorly understood. This is especially true for the implementation of DeepWalk using Skip-gram with negative sampling (SGNS), since the variance of the stochastic updates turns out to be very large. In this work, we make progress on this question. We show that for "small norm" initialization, under a spectral gap assumption on the graph, the DeepWalk embeddings align with the column space of a fixed low-rank matrix. For graphs generated from Stochastic Block Models with certain separation conditions, our results imply that the DeepWalk embeddings recover cluster structure. To the best of our knowledge, our results give the first analysis of the optimization trajectory of DeepWalk with negative sampling on non-trivial graph classes.

Optimization · Non-Convex

Zhiyao Zhang, Myeung Suk Oh, Zhen Qin, Jiaxiang Li, Xin Zhang, Jia (Kevin) Liu

In recent years, bilevel optimization (BLO) has attracted significant attention for its broad applications in machine learning. However, most existing works on BLO remain confined to the single-task setting and rely on the lower-level strong convexity assumption, which significantly restricts their applicability to modern machine learning problems of growing complexity. In this paper, we make the first attempt to extend BLO to the multi-task setting under a relaxed lower-level general convexity (LLGC) assumption. To this end, we reformulate the multi-task bilevel learning (MTBL) problem with LLGC into an equality constrained multi-objective optimization (ECMO) problem. However, ECMO itself is a new problem that has not yet been studied in the literature. To address this gap, we first establish a new Karush–Kuhn–Tucker (KKT)-based Pareto stationarity as the convergence criterion for ECMO algorithm design. Based on this foundation, we propose a weighted Chebyshev (WC)-penalty algorithm that achieves a finite-time convergence rate of $\mathcal{O}(ST^{-\frac{1}{2}})$ to KKT-based Pareto stationarity in both deterministic and stochastic settings, where $S$ denotes the number of objectives, and $T$ is the total iterations. Moreover, by varying the preference vector over the $S$-dimensional simplex, our WC-penalty method systematically explores the Pareto front. Finally, solutions to the ECMO problem translate directly into solutions for the original MTBL problem, thereby closing the loop between these two foundational optimization frameworks.

Theory · Game Theory

Omar Abbadi, Rida Laraki, Panayotis Mertikopoulos

We examine the interplay between ordinal, preference-based solution concepts in games and the outcomes of payoff-driven learning dynamics, asking to what extent the combinatorial data of a game—its preference graph—can predict the long-run behavior of no-regret dynamics such as *follow-the-regularized-leader* (FTRL). In one direction, we show that the skeleton of every *dynamically stable* set, i.e., the set of pure profiles it contains, must be *preferentially stable*, that is, closed under pure profitable deviations. We then ask the converse question: when are preferences sufficient to describe long-run behavior? For *subgames*—subsets of pure profiles obtained by restricting players’ action sets—preferences are enough to fully characterize asymptotic stability. Beyond subgames however, we construct a three-player counterexample with a preferentially stable set whose span is dynamically *unstable*, thus establishing that preferences are *not sufficient* to describe dynamically stable behavior in general. To restore stability, we introduce the notion of *leaklessness*, a measure of aggregate payoff drift away from a set of pure profiles, and use it to identify a payoff-based condition under which the span of a set of pure profiles remains stable and attracting, thereby setting forth a natural cardinal guarantee of dynamic stability.

Reinforcement Learning · Online

Udbhav Bamba, Minghao Fang, Yifan Yu, Haizhong Zheng, Fan Lai

Reinforcement learning algorithms such as GRPO have driven recent advances in large language model (LLM) reasoning. While scaling the number of rollouts stabilizes training, existing approaches suffer from limited exploration on challenging prompts and leave informative feedback signals underexploited, due to context-independent rollout allocation across prompts (e.g., generating 16 rollouts per prompt) and relying heavily on sparse rewards. This paper presents XRPO (eXplore–eXploit GRPO), a unified framework that recasts policy optimization through the principled lens of rollout exploration–exploitation. To enhance exploration, XRPO introduces a mathematically grounded rollout allocator that adaptively prioritizes prompts with higher potential for uncertainty reduction. It further addresses stagnation on zero-reward prompts through an in-context seeding strategy that injects curated exemplars, steering the model into more difficult reasoning trajectories. To strengthen exploitation, XRPO develops a group-relative, novelty-aware advantage sharpening mechanism that leverages sequence likelihoods to amplify low-probability yet correct responses, thereby extending the policy’s reach beyond sparse rewards. Experiments across diverse math and coding benchmarks on both reasoning and non-reasoning models demonstrate that XRPO outperforms existing advances (e.g., GRPO and GSPO) up to 4% pass@1 and 6% cons@32, while accelerating training convergence by up to 2.7x.

Optimization · Non-Convex

Kangke Cheng, Jiawei Huang, Jingni Song, Wanlin Zhang, Bangxian Han, Hu Ding

The Gromov-Wasserstein (GW) distance enables comparison across different spaces but remains fragile to structural noise due to its global quadratic coupling. Existing robust extensions primarily rely on node-centric mass relaxation. However, we argue that this strategy is far from sufficient: it only addresses node-induced structural noise (outliers) while neglecting edge-induced distortions where spurious connections exist between valid nodes. To overcome this limitation, we propose the Structurally Robust Gromov-Wasserstein (SRGW) distance, a novel formulation that adaptively filters geometric distortions during optimization. By introducing a structure-aware dual-mask mechanism, our method effectively isolates these stubborn structural outliers while preserving strict marginal constraints for balanced transport. We solve this objective using a Mask-Guided GW Algorithm, which jointly optimizes the transport plan and the structural noise filters. We provide a rigorous theoretical analysis proving that our algorithm converges to a critical point under the Kurdyka-Łojasiewicz framework. Extensive experiments on synthetic geometric matching and real-world subgraph alignment benchmarks demonstrate that Mask-Guided GW achieves superior alignment quality, particularly under severe structural noise.

Theory · Reinforcement Learning and Planning

Kenta Hoshino, Yashaswi Shashank Aluru, Xiyu Deng, Yorie Nakahira

This paper studies the theoretical foundations of guidance mechanisms in continuous-time score-based diffusion models. We adopt Doob’s h-transform as a principled framework for characterizing ideal guided diffusion processes and analyze the discrepancy between ideal and approximate guidance. Our analysis provides explicit error bounds and yields probabilistic guarantees on satisfying prescribed constraints, which are particularly important for safety-critical planning. We further show that the Doob-based formulation induces a stochastic optimal control problem, enabling practical guidance design without additional model training. We demonstrate the effectiveness of the proposed framework on robotic navigation tasks, including language-conditioned planning.

Omar Abbadi, Rida Laraki, Panayotis Mertikopoulos

We examine the interplay between ordinal, preference-based solution concepts in games and the outcomes of payoff-driven learning dynamics, asking to what extent the combinatorial data of a game—its preference graph—can predict the long-run behavior of no-regret dynamics such as *follow-the-regularized-leader* (FTRL). In one direction, we show that the skeleton of every *dynamically stable* set, i.e., the set of pure profiles it contains, must be *preferentially stable*, that is, closed under pure profitable deviations. We then ask the converse question: when are preferences sufficient to describe long-run behavior? For *subgames*—subsets of pure profiles obtained by restricting players’ action sets—preferences are enough to fully characterize asymptotic stability. Beyond subgames however, we construct a three-player counterexample with a preferentially stable set whose span is dynamically *unstable*, thus establishing that preferences are *not sufficient* to describe dynamically stable behavior in general. To restore stability, we introduce the notion of *leaklessness*, a measure of aggregate payoff drift away from a set of pure profiles, and use it to identify a payoff-based condition under which the span of a set of pure profiles remains stable and attracting, thereby setting forth a natural cardinal guarantee of dynamic stability.

Deep Learning · Large Language Models

LangYu Bian, Jinwu Hu, Zitian Zhang, Dongjin Yang, Yufeng Wang, Qing Du, Qi Chen, Mingkui Tan

Large language models (LLMs) inevitably encounter distribution shifts during real-world deployment, leading to performance degradation. Although test-time learning (TTL) adapts LLMs from unlabeled test streams, applying entropy minimization to autoregressive generation faces two challenges: (i) early decoding errors can steer later tokens off track, and updating on them can push the model further off course, and (ii) updates on unreliable tokens can amplify confident error predictions and trigger model collapse. To address these challenges, we propose Future-Gain Guided Test-Time Learning (FG-TTL) for LLMs, which learns selectively from the model's own generations. Our key idea is to update only on tokens that reduce uncertainty in subsequent generation rather than tokens that are merely uncertain at the current step. Specifically, we develop a Future-Gain Guided Token Selection (FTS) strategy to decide where to learn. We introduce Future-Gain as a token-level metric for this purpose and update the model only on high-gain tokens, concentrating learning on informative positions and mitigating temporal error propagation. In addition, we design a Risk-Aware Adaptation (RAA) mechanism that controls how strongly to update by combining gain-based weighting with adaptive temperature scaling based on intrinsic uncertainty, suppressing unreliable gradients while enabling stronger learning on high-gain tokens. Experiments on six benchmarks with three LLM backbones show that FG-TTL achieves the best average performance.