ICML 2026 论文列表

Optimizing complex systems, ranging from LLM prompts to multi-turn agents, traditionally requires labor-intensive manual iteration. We formalize this as a stochastic generative optimization problem where a language model acts as the optimizer, guided by numerical and text feedback to discover the best system. We introduce **P**rioritized **O**ptimization with **L**ocal **C**ontextual **A**ggregation (POLCA), a scalable framework designed to handle stochasticity in optimization—such as noisy feedback, sampled minibatches, and stochastic system behaviors—while effectively managing the unconstrained expansion of the solution space. POLCA maintains a priority queue to enable exploration–exploitation, tracking candidate solutions and their evaluation histories. To enhance efficiency, we integrate an $\varepsilon$-Net mechanism to maintain parameter diversity and an LLM Summarizer to perform meta-learning across historical trials. We theoretically prove that POLCA converges to near-optimal candidate solutions under stochasticity. We evaluate our framework on diverse benchmarks, including $\tau$-bench, HotpotQA, VeriBench, and KernelBench. Experimental results demonstrate that POLCA achieves robust, sample- and time-efficient performance, consistently outperforming state-of-the-art algorithms in both deterministic and stochastic problems. The codebase for this work is publicly available at https://github.com/rlx-lab/POLCA.

Being probabilistic models, during inference large language models (LLMs) display *rare events*: behaviour that is far from typical but highly significant. By definition all rare events are hard to see, but the enormous scale of LLM usage means that events completely unobserved during development are likely to become prominent in deployment. Here we present an end-to-end framework for the systematic analysis of rare events in LLMs. We provide a practical implementation spanning theory, efficient generation strategies, probability estimation and error analysis, which we illustrate with concrete examples. We outline extensions and applications to other models and contexts, highlighting the generality of the concepts and techniques presented here.

Large Language Models (LLMs) have demonstrated significant potential as autonomous software engineering (SWE) agents. Recent work has further explored augmenting these agents with memory mechanisms to support long-horizon reasoning. However, these approaches typically operate at a coarse instance granularity, treating the entire problem-solving episode as the atomic unit of storage and retrieval. We empirically demonstrate that instance-level memory suffers from a fundamental granularity mismatch, resulting in misguided retrieval when tasks with similar surface descriptions require distinct reasoning logic at specific stages. To address this, we propose Structurally Aligned Subtask-Level Memory, a method that aligns memory storage, retrieval, and updating with the agent’s functional decomposition. Extensive experiments on SWE-bench Verified demonstrate that our method consistently outperforms both vanilla agents and strong instance-level memory baselines across diverse backbones, improving mean Pass@1 over the vanilla agent by +4.7 pp on average (e.g., +6.8 pp on Gemini 2.5 Pro). Performance gains grow with more interaction steps, showing that leveraging past experience benefits long-horizon reasoning in complex software engineering tasks.

Multimodal large language models (MLLMs) have achieved remarkable success across a broad range of vision tasks. However, constrained by the capacity of their internal world knowledge, prior work has proposed augmenting MLLMs by ``reasoning-then-tool-call'' for visual and textual search engines to obtain substantial gains on tasks requiring extensive factual information. However, these approaches typically define multimodal search in a naive setting, assuming that a single full-level or entity-level image query and few text query suffices to retrieve the key evidence needed to answer the question, which is unrealistic in real-world scenarios with substantial visual noise. Moreover, they are often limited in the reasoning depth and search breadth, making it difficult to solve complex questions that require aggregating evidence from diverse visual and textual sources. Building on this, we propose Vision-DeepResearch, which proposes one new multimodal deep-research paradigm, i.e., performs multi-turn, multi-entity and multi-scale visual and textual search to robustly hit real-world search engines under heavy noise. Our Vision-DeepResearch supports dozens of reasoning steps and hundreds of engine interactions, while internalizing deep-research capabilities into the MLLM via cold-start supervision and RL training, resulting in a strong end-to-end multimodal deep-research MLLM. It substantially outperforming existing multimodal deep-research MLLMs, and workflows built on strong closed-source foundation model such as GPT-5, Gemini-2.5-pro and Claude-4-Sonnet.

Vector-quantized variational autoencoders (VQ-VAEs) are discrete autoencoders that compress images into discrete tokens. However, they are difficult to train due to discretization. In this paper, we propose a simple yet effective technique dubbed __Gaussian Quant (GQ)__, which first trains a Gaussian VAE under certain constraints and then converts it into a VQ-VAE without additional training. For conversion, GQ generates random Gaussian noise as a codebook and finds the closest noise vector to the posterior mean. Theoretically, we prove that when the logarithm of the codebook size exceeds the bits-back coding rate of the Gaussian VAE, a small quantization error is guaranteed. Practically, we propose a heuristic to train Gaussian VAEs for effective conversion, named the target divergence constraint (TDC). Empirically, we show that GQ outperforms previous VQ-VAEs, such as VQGAN, FSQ, LFQ, and BSQ, on both UNet and ViT architectures. Furthermore, TDC also improves previous Gaussian VAE discretization methods, such as TokenBridge. The source code is provided in https://github.com/tongdaxu/VQ-VAE-from-Gaussian-VAE.

Instruction tuning is essential for aligning large language models (LLMs) to downstream tasks and commonly relies on large, diverse corpora. However, small, high-quality subsets, known as coresets, can deliver comparable or superior results, though curating them remains challenging. Existing methods often rely on coarse, sample-level signals like gradients, an approach that is computationally expensive and overlooks fine-grained features. To address this, we introduce TRIM (Token Relevance via Interpretable Multi-layer Attention), a forward-only, token-centric framework. Instead of using gradients, TRIM operates by matching underlying representational patterns identified via attention-based "fingerprints" from a handful of target samples. Such an approach makes TRIM highly efficient and uniquely sensitive to the structural features that define a task. Coresets selected by our method consistently outperform state-of-the-art baselines by up to 9% on downstream tasks and even surpass the performance of full-data fine-tuning in some settings. By avoiding expensive backward passes, TRIM achieves this at a fraction of the computational cost. These findings establish TRIM as a scalable and efficient alternative for building high-quality instruction-tuning datasets.

As Multimodal Large Language Models (MLLMs) continue to be trained, the availability of public data diminishes, limiting the possibility for further training and adaptation. However, private data remains an underutilized yet valuable resource. Federated Learning (FL) enables decentralized training on private data, yet extending it to MLLMs is challenging: heterogeneous client modalities induce architectural incompatibility, and full-parameter fine-tuning of billion-scale models incurs prohibitive communication costs. Parameter-efficient methods like LoRA alleviate these issues but introduce aggregation inconsistency, as averaged low-rank updates fail to recover the true global update faithfully. To address these issues, we propose **UniFLoW**(Universal multi-modal Federated LoRA fine-tuning framework With Analytical Aggregation), a unified federated framework that leverages pre-trained large language models and a multi-modal Encoder architecture, and our proposed Federated Aggregating Analytical Low-Rank Adaption$FedA^2$-$LoRA$). UniFLoW effectively utilizes fragmented client-side multi-modal data while $FedA^2$-$LoRA$ ensuring consistent aggregation. And modality-specific encoders and a II stage training strategy ensure effective integration of diverse modalities without overfitting. Experiments on text, image, and speech demonstrate that **UniFLoW** enables scalable, communication-efficient, and aggregation-consistent federated fine-tuning, with $FedA^2$-$LoRA$ achieving state-of-the-art performance compared to existing FedLoRA approaches. We envision UniFLoW as a promising solution to the growing scarcity of public data.

Fine-tuning Multimodal Large Language Models (MLLMs) on task-specific data is an effective way to improve performance on downstream applications. However, such adaptation often leads to a degradation in generalization on pretrained tasks, a phenomenon known as Catastrophic Forgetting. Existing methods that aim to mitigate this issue either become ineffective when fine-tuning deeper layers of the language decoder or scale poorly with increasing model size. To address these limitations, we propose Model-Dowser, a novel sparse fine-tuning approach for MLLMs. Model-Dowser measures a principled importance score for each model parameter with respect to pretrained generalization (prior to downstream adaptation) by jointly considering weight magnitudes, input activations, and output sensitivities. During fine-tuning, Model-Dowser selectively preserves high-importance parameters and updates the remaining. Comprehensive experiments on two representative MLLMs, LLaVA and NVILA, demonstrate that Model-Dowser effectively mitigates catastrophic forgetting and consistently outperforms prior methods, while remaining resource-efficient and scalable to multi-billion-parameter models.

Multimodal contrastive learning (MCL) aims to embed data from two modalities in a shared embedding space. However, in practice, image and text representations occupy completely separated regions of the embedding space, a phenomenon called the modality gap. Meanwhile, empirical findings on how the modality gap affects downstream performance remain inconsistent. These observations motivate two key questions: (1) What causes the modality gap? (2) What determines downstream performance? To address these questions, we develop the first theoretical framework for analyzing the geometry of convergent optimal representations (COR) of MCL when training is optimized. We prove that the modality gap emerges when image and text representations collapse into different subspaces, a phenomenon called \emph{dimension collapse}. Our theory further reveals that although the modality gap prevents direct alignment between image and text representations, their projections onto the shared subspace can align. Moreover, we show that shared space alignment is a dominant factor in downstream performance, while the effect of the modality gap is limited. Inspired by these findings, we propose Shared Space Alignment (SSA) to improve MCL pretraining by enhancing alignment in the shared space without optimizing for modality gap reduction. Extensive empirical results validate our theoretical analysis and the proposed method.

Understanding how transformer components operate in LLMs is important, as it is at the core of recent technological advances in artificial intelligence. In this work, we revisit the challenges associated with interpretability of feed-forward modules (FFNs) and propose MemoryLLM, which aims to decouple FFNs from self-attention and enables us to study the decoupled FFNs as context-free token-wise neural retrieval memory. In detail, we investigate how input tokens access memory locations within FFN parameters and the importance of FFN memory across different downstream tasks. MemoryLLM achieves context-free FFNs by training them in isolation from self-attention directly using the token embeddings. This approach allows FFNs to be pre-computed as token-wise lookups (ToLs), enabling on-demand transfer between VRAM and storage, additionally enhancing inference efficiency. We also introduce Flex-MemoryLLM, positioning it between a conventional transformer design and MemoryLLM. This architecture bridges the performance gap caused by training FFNs with context-free token-wise embeddings.

Direct Preference Optimization (DPO) is a widely used RL-free method for aligning language models from pairwise preferences, but it models preferences over full sequences even though generation is driven by per-token decisions. Existing token-level extensions typically decompose a sequence-level Bradley–Terry objective across timesteps, leaving per-prefix (state-wise) optimality implicit. We study how to recover **token-level** preference optimality using only standard sequence-level pairwise comparisons. We introduce **Token-level Bregman Preference Optimization (TBPO)**, which posits a token-level Bradley--Terry preference model over next-token actions conditioned on the prefix, and derive a Bregman-divergence density-ratio matching objective that generalizes the logistic/DPO loss while preserving the optimal policy induced by the token-level model and maintaining DPO-like simplicity. We introduce two instantiations: TBPO-Q, which explicitly learns a lightweight state baseline, and TBPO-A, which removes the baseline through advantage normalization. Across instruction following, helpfulness/harmlessness, and summarization, TBPO improves alignment quality and training stability and increases output diversity relative to strong sequence-level and token-level baselines.

The increasing complexity of AI tasks has shifted the paradigm from monolithic models toward multi-agent large language model (LLM) systems. However, these collaborative architectures introduce a critical bottleneck: redundant prefill computation for shared content generated by previous agents, which significantly increases KV cache memory usage and time-to-first-token (TTFT). While various KV cache methods have been proposed to mitigate prefill redundancy, they either fail to maintain accuracy on agent-generated outputs or exhibit low reuse rates due to rigid constraints. We present RelayCaching, a training-free inference method that directly reuses decoding phase KV caches from previous agents in subsequent prefill phases. Our key insight is that KV caches for identical content are highly consistent across phases, while prefix-induced deviations are sparse and localized within a limited range of layers and token positions. By selectively recomputing KV caches at these positions, RelayCaching preserves model accuracy with minimal overhead, yielding a superior accuracy–efficiency trade-off over existing methods. Experiments on diverse collaborative LLM tasks spanning mathematical reasoning, general knowledge, and code generation demonstrate that RelayCaching achieves over $80$\% KV cache reuse, reduces TTFT by up to $4.7\times$ compared to the standard pipeline, all with negligible accuracy degradation.

Large Language Models (LLMs) with billions of parameters have transformed AI applications but require immense computational and memory resources during inference. Adaptive sparse activation inference, which activates only a small number of neurons for each token, offers a novel way to accelerate model inference without degrading performance, showing great potential for resource-constrained hardware devices. Yet existing token-level MLP-based methods frequently alter activation maps, reducing efficiency gains. In this paper, we introduce \textbf{SparseInfer}, an MLP-free adaptive sparse activation inference method based on sentence-level prediction. We first propose the concept of core neurons and empirically demonstrate that, for an input sentence, LLMs only need the core neurons to maintain performance. Remarkably, we discovered that core neurons exhibit both stability and similarity in relation to the sentence's semantics—an insight overlooked by previous studies. Building on this finding, we design two semantic-based methods for predicting core neurons to fit different input scenarios, which enables core neurons to be determined during the pre-filling stage and fixed during the encoding stage. Our experiments verify SparseInfer exhibits good performance in various tasks and achieve 10.33$\times$ speed up.

Class-incremental learning (CIL) with pre-trained models (PTMs) faces a critical trade-off between efficient adaptation and long-term stability. While analytic learning enables rapid, recursive closed-form updates, its efficacy is often compromised by accumulated errors and feature incompatibility. In this paper, we first conduct a systematic study to dissect the failure modes of PTM-based analytic CIL, identifying representation rigidity as the primary bottleneck. Motivated by this insight, we propose **VILA**, a novel dual-branch framework that advances analytic CIL via a two-level vision-language calibration strategy. Specifically, we coherently fuse plastic, task-adapted features with a frozen, universal visual anchor at the feature level through geometric calibration, and leverage cross-modal semantic priors at the decision level to rectify prediction bias. This confluence maintains analytic-learning's extreme efficiency while overcoming its inherent brittleness. Extensive experiments across eight benchmarks demonstrate that VILA consistently yields superior performance, particularly in fine-grained and long-sequence scenarios. Our framework harmonizes high-fidelity prediction with the simplicity of analytic learning. Our code is available at https://github.com/byzhaoAI/VILA.

Spatiotemporal neural dynamics and oscillatory synchronization are widely implicated in biological information processing and have been hypothesized to support flexible coordination such as feature binding. By contrast, most deep learning architectures represent and propagate information through activation values, neglecting the joint dynamics of rate and phase. In this work, we introduce Kuramoto oscillatory Phase Encoding (KoPE) as an additional, evolving phase state to Vision Transformers, incorporating a neuro-inspired synchronization mechanism to advance learning efficiency. We show that KoPE can improve training, parameter, and data efficiency of vision models through synchronization-enhanced structure learning. Moreover, KoPE benefits tasks requiring structured understanding, including semantic and panoptic segmentation, representation alignment with language, and few-shot abstract visual reasoning (ARC-AGI). Theoretical analysis and empirical verification further suggest that KoPE can accelerate attention concentration for learning efficiency. These results indicate that synchronization can serve as a scalable, neuro-inspired mechanism for advancing state-of-the-art neural network models.

Just as LEGO pieces can be assembled into an unlimited variety of structures, heterogeneous federated learning (HFL) can be viewed as the assembly of diverse model components. Inspired by this analogy, we reformulate HFL as a LEGO-like assembly game. The central challenge in HFL lies in learning across heterogeneous model architectures, which hinders direct parameter sharing. To address this challenge, we propose to decompose models into a set of modular components—analogous to LEGO pieces and collaboratively learn these components across clients under predefined composition rules. Based on this perspective, we develop a novel federated learning framework, termed LEGO-FL, which enables flexible model construction while preserving collaborative learning. Extensive experiments validate the effectiveness of LEGO-FL under different heterogeneous settings and system scales.

Large language models (LLMs) exhibit persistent miscalibration, especially after instruction tuning and preference alignment. Modified training objectives can improve calibration, but retraining is expensive. Inference-time steering offers a lightweight alternative, yet most existing methods optimize proxies for correctness rather than correctness itself. We introduce CORAL (Correctness-Optimized Residual Activation Lens), a regularized inference-time steering method that captures distributed correctness signals from model internal activations using weight-decay MLP probes. We evaluate CORAL across three 7B-parameter models and find that it consistently improves accuracy by 10\% and expected calibration error (ECE) by 50\% on average. We additionally demonstrate that these gains transfer without retraining to the complete published test sets of four held-out benchmarks (ARC-Challenge, HellaSwag, Math-MC, OpenBookQA), averaging 14\% accuracy improvements and 49\% ECE improvements. Our results support the hypothesis that distributed information in model internals can be extracted using regularized probes when individual neurons are insufficient. CORAL thus provides a compute-efficient, transferable, and calibration-aware approach to improve MCQA performance during inference.

Engineering problem solving is central to real-world decision-making, requiring mathematical formulations that not only represent complex problems but also produce feasible solutions under data and physical constraints. Unlike mathematical problem solving, which operates on predefined formulations, engineering tasks demand open-ended analysis, feasibility-driven modeling, and iterative refinement. Although large language models (LLMs) have shown strong capabilities in reasoning and code generation, they often fail to ensure feasibility, which limits their applicability to engineering problem solving. To address this challenge, we propose EngiAgent, a multi-agent system with a fully connected coordinator that simulates expert workflows through specialized agents for problem analysis, modeling, verification, solving, and solution evaluation. The fully connected coordinator enables flexible feedback routing, overcoming the rigidity of prior pipeline-based reflection methods and ensuring feasibility at every stage of the process. This design not only improves robustness to diverse failure cases such as data extraction errors, constraint inconsistencies, and solver failures, but also enhances the overall quality of problem solving. Empirical results across four representative domains demonstrate that EngiAgent achieves substantial improvements in feasibility compared to prior approaches, establishing a new paradigm for feasibility-oriented engineering problem solving with LLMs. Our source code and data are available at https://github.com/AI4Engi/EngiAgent.

Manifold clustering has demonstrated strong capability in capturing complex data structures and has been widely studied in cluster analysis. However, many existing methods mainly focus on combining K-means with manifold learning, while overlooking the consistency between data structures and clustering labels, and often suffer from high computational cost when handling large scale data. To address these issues, we propose a manifold balanced clustering method based on anchor induced distance(LMBC), grounded in the relationship between K-means clustering and manifold learning. Specifically, the LMBC uses label information to guide the construction of the manifold structure, thereby ensuring consistency between data structures and clustering labels. To enable large scale clustering, we introduce an anchor induced distance representation that models manifold structure in a compact anchor space, significantly reducing computational complexity while preserving essential structural information. Furthermore, to naturally maintain class balance during clustering, we maximize the Schatten-p norm of the label representation and provide theoretical analysis to support its effectiveness. Experimental results on several benchmark datasets demonstrate the effectiveness and scalability of the proposed method.