ICML 2026 Papers

Direct Alignment Algorithms (DAAs) such as DPO have become a common way to post-train and align LLMs with human preferences. However, DAAs have been observed to over-optimize their implicit reward model and decrease the likelihood of preferred responses. This results in a decrease in the total likelihood assigned to responses seen in the preference dataset, potentially resulting in undesirable behavior. To counteract this undesired side-effect of DAAs, we examine the effect of using objectives that add a regularization term to maintain the total length-normalized probabilities of the chosen and rejected responses. To better understand over-optimization, we investigate how response likelihood changes are distributed over the tokens with and without regularization. We find that a significant portion of the likelihood changes are due to a small set of outlier tokens, which explains how DAAs improve generation quality despite decreasing the likelihoods of chosen responses. We apply the proposed regularization to reference-based (DPO) and reference-free (SimPO) methods and find (1) improved trade-offs between generation quality and general benchmark capability and (2) improvements in reward modeling across datasets. For example, on Llama-3.1-8B-Instruct, we see both a >20\% relative increase in AlpacaEval2 scores and >9\% relative performance gains on general benchmarks. Additionally, we find that the added regularization term effectively mitigates the amount of displacement within preferred responses overall, and for the outlier tokens specifically, by utilizing low-likelihood tokens.

Composed image retrieval uses a reference image plus a natural-language edit to retrieve images that apply the requested change while preserving other relevant visual content. Classic fusion pipelines typically rely on supervised triplets and can lose fine-grained cues, while recent zero-shot approaches often caption the reference image and merge the caption with the edit, which may miss implicit user intent and return repetitive results. We present Pix2Key, which represents both queries and candidates as open-vocabulary visual dictionaries, enabling intent-aware constraint matching and diversity-aware reranking in a unified embedding space. A self-supervised pretraining component, V-Dict-AE, further improves the dictionary representation using only images, strengthening fine-grained attribute understanding without CIR-specific supervision. On the extended DeepFashion-MM benchmark, Pix2Key improves Recall@10 up to 3.2 points, and adding V-Dict-AE yields an additional 2.3-point gain while improving intent consistency and maintaining high list diversity.

Recent advancements in large reasoning models (LRMs) have greatly improved their capabilities on complex reasoning tasks through Long Chains of Thought (CoTs). However, this approach often results in substantial redundancy, impairing computational efficiency and causing significant delays in real-time applications. Recent studies show that longer reasoning chains are frequently uncorrelated with correctness and can even be detrimental to accuracy. In a further in-depth analysis of this phenomenon, we surprisingly uncover and empirically verify that LRMs implicitly know the appropriate time to stop thinking, while this capability is obscured by current sampling paradigms. Motivated by this, we introduce SAGE (Self-Aware Guided Efficient Reasoning), a novel sampling paradigm that unleashes this efficient reasoning potential. Furthermore, integrating SAGE as mixed sampling into group-based reinforcement learning (SAGE-RL) enables SAGE-RL to effectively incorporate SAGE-discovered efficient reasoning patterns into standard pass@1 inference, markedly enhancing both the reasoning accuracy and efficiency of LRMs across multiple challenging mathematical benchmarks.

Primal heuristics play a crucial role in quickly finding feasible solutions for NP-hard integer linear programming (ILP). Although $\textit{end-to-end learning}$-based primal heuristics (E2EPH) have recently been proposed, they are typically unable to independently generate feasible solutions. To address this challenge, we propose RL-SPH, a novel reinforcement learning-based start primal heuristic capable of independently generating feasible solutions, even for ILP involving non-binary integers. Empirically, RL-SPH rapidly obtains high-quality feasible solutions with a 100% feasibility rate, achieving on average a 28.6$\times$ lower primal gap and a 2.6$\times$ lower primal integral compared to existing start primal heuristics.

Consistency training encourages a model to produce similar outputs across related inputs or sampling procedures. Such methods are simple, scalable, and largely label-free, but their effects on model alignment remain poorly understood. Could the self-bootstrapping nature of these methods amplify undesired behavior in models? We test seven consistency training methods on 108 model organisms: open-source models (7B--70B) fine-tuned to exhibit various forms of controlled misaligned behavior. We find that outcomes vary significantly: consistency training generally suppresses reward hacking and emergent misalignment but amplifies sycophancy. We present evidence that distribution shifts induced by the consistency labeling process, rather than variation in the selection operators, may be the primary driver of systematic alignment effects. Finally, we present a unifying theoretical framework to derive conditions under which consistency training will amplify or suppress misalignment. In total, our study establishes that consistency training is not alignment-neutral, and that its use in critical systems should be carefully audited.

Large language models (LLMs) can generate syntactically valid optimization programs, yet often struggle to reliably choose an effective modeling strategy, leading to incorrect formulations and inefficient solver behavior. We propose **SAGE**, a strategy-aware framework that makes *Modeling Strategy* explicit in both data construction and post-training. SAGE builds a solver-verified multi-strategy dataset and trains a student model with supervised fine-tuning followed by Segment-Weighted GRPO using a composite reward over format compliance, correctness, and solver efficiency. Across eight benchmarks spanning synthetic and real-world settings, SAGE improves average pass@1 from 72.7 to 80.3 over the strongest open-source baseline. With multiple generations, SAGE discovers more distinct correct formulations and improves component-level diversity at pass@16 by 19-29%. At the largest scale, SAGE produces more compact constraint systems with 14.2% fewer constraints than the baseline, consistent with solver-efficient modeling. Overall, these results show that making *Modeling Strategy* explicit improves automated optimization modeling. The code and data are available at https://github.com/rachhhhing/SAGE.

Probabilistic super-resolution of high-dimensional spatial fields using diffusion models is often computationally prohibitive due to the cost of operating directly in pixel space. We propose PODiff, a structured conditional generative framework that performs diffusion in a fixed, variance-ordered Proper Orthogonal Decomposition (POD) coefficient space, exploiting the orthogonality of POD modes to impose an interpretable, variance-ordered latent geometry. This design enables efficient ensemble generation, preserves dominant spatial structure, and yields spatially interpretable, well-calibrated uncertainty at substantially lower computational cost. We evaluate PODiff on sea surface temperature downscaling over the West Australian coast and on a controlled advection--diffusion benchmark. PODiff achieves reconstruction accuracy comparable to pixel-space and learned-latent diffusion while using a substantially smaller structured latent space, requiring significantly less memory, and producing more reliable uncertainty estimates than deterministic and Monte Carlo Dropout baselines.

While data augmentation is essential for robust point cloud recognition, conventional spatial mixup strategies often compromise geometric integrity by generating physically unrealistic samples. To overcome this limitation, we propose PSMix, which shifts the mixing paradigm to the spectral domain via the Spherical Harmonic Transform. Instead of simple coordinate interpolation, PSMix performs a rotation-aware hierarchical mixing on spectral coefficients. This approach explicitly preserves global structural properties while diversifying local details, achieving a balance that spatial methods struggle to maintain. Complementing this, we introduce an adversarial rotation optimization strategy to enforce invariance against challenging orientations. Extensive experiments on ModelNet-C and ScanObjectNN-C demonstrate that PSMix achieves state-of-the-art robustness, while also serving as an orthogonal plug-in that further boosts the performance of existing spatial strategies.

LoRA and its variants have attracted considerable attention because of their abilities to tune a negligible number of parameters while achieving comparable downstream performance. This success is largely attributed to the intrinsic low-rank structure of model parameter spaces, which allows LoRA to train two projection matrices to project weights into a low-dimensional subspace and then map them back. However, it does not consider how to explore this low-rank subspace sufficiently and may lose the expression ability accordingly. Moreover, when using LoRA to tune convolution layers, a flatten operation is required to convert tensors into matrices. We argue that this will degrade the model's performance. In this paper, we address this issue from a general parameter sub-space perspective: we present a unified **L**anguage **A**nd **V**ision **A**daption finetuning framework (called **LAVA**). Specifically, we verify the existence of low-rank subspaces in convolution layers empirically and propose to parameterize the increment of both convolution kernels and matrices as sum of learnable rank-1 components. To improve training stability, we analyze the optimization dynamics of LoRA and incorporate orthogonal regularization into our parameterization, for which we give theoretical proof that it will help reduce the variance of the gradient. We conduct various experiments on different downstreaming tasks to validate LAVA's superiority. For example, when tuning LLaMA2-7b for commonsense tasks, the performance of our LAVA is **+1.9\%** higher than that of LoRA. For metric depth estimation tasks, LAVA only tunes $\sim$1.5\% of Depth-Anything (335.3M), and achieves **+3.5\%** $\delta_1$ accuracy against that of LoRA and **+5.6\%** $\delta_1$ accuracy against that of SVDiff.

We derive a robust update rule for the online infinite hidden Markov model (iHMM) for when the streaming data contains outliers and the model is misspecified. Leveraging recent advances in generalised Bayesian inference, we define robustness via the posterior influence function (PIF), and provide conditions under which the online iHMM has bounded PIF. Imposing robustness inevitably induces an adaptation lag for regime switching. Our method, which is called Batched Robust iHMM (BR-iHMM), balances adaptivity and robustness with two additional tunable parameters. Across limit order book data, hourly electricity demand, and a synthetic high-dimensional linear system, BR-iHMM reduces one-step-ahead forecasting error by up to 67% relative to competing online Bayesian methods. Together with theoretical guarantees of bounded PIF, our results highlight the practicality of our approach for both forecasting and interpretable online learning.

Diffusion posterior sampling conditions diffusion priors on measurements, but data-consistency updates are typically scaled by hand-tuned guidance weights and can destabilize sampling under stiff, operator-dependent curvature. We replace scalar guidance with a per-noise-level damped Gauss--Newton correction computed in diffusion-state coordinates. The correction pulls likelihood gradients back through the denoiser, uses a one-sided curvature model that avoids forward denoiser Jacobians, and applies diffusion-calibrated rank-one damping aligned with the denoiser residual. Each correction is solved with matrix-free GMRES using automatic differentiation, and sampling proceeds with a variance-preserving Langevin transition with a closed-form drift/noise split. On FFHQ and ImageNet across inverse problems, it achieves competitive PSNR/SSIM/LPIPS while running markedly faster than most of the compared baselines; on accelerated MRI reconstruction, it achieves the best PSNR/SSIM among the compared baselines. Code is available at https://github.com/Seunghyeok0715/CLAMP

In large vision-language models, visual tokens typically constitute the majority of input tokens, leading to substantial computational overhead. To address this, recent studies have explored pruning redundant or less informative visual tokens for image understanding tasks. However, these methods struggle with pixel grounding tasks, where token importance is highly contingent on the input text. Through an in-depth analysis of CLIP, we observe that visual tokens within referent regions often exhibit low similarity to their textual representation. Motivated by this insight, we introduce LiteLVLM, a training-free, text-guided token pruning strategy for efficient pixel grounding inference. By reversing the ranking of CLIP's visual-text similarity, LiteLVLM effectively retains visual tokens covering the referent regions, while recovering context tokens to enable clear foreground-background separation. Extensive experiments demonstrate that LiteLVLM significantly outperforms existing methods by over 5% across diverse token budgets. Without any training or fine-tuning, LiteLVLM maintains 90% of the original performance with a 22% speedup and a 2.3$\times$ memory reduction. Our code is available at https://github.com/sejong-rcv/LiteLVLM.

As Large Multimodal Models (LMMs) scale up and reinforcement learning (RL) methods mature, LMMs have made notable progress in complex reasoning and decision making. Yet training still relies on static data and fixed recipes, making it difficult to diagnose capability blind spots or provide dynamic, targeted reinforcement. Motivated by findings that test driven error exposure and feedback based correction outperform repetitive practice, we propose Diagnostic-driven Progressive Evolution (DPE), a spiral loop where diagnosis steers data generation and reinforcement, and each iteration re-diagnoses the updated model to drive the next round of targeted improvement. DPE has two key components. First, multiple agents annotate and quality control massive unlabeled multimodal data, using tools such as web search and image editing to produce diverse, realistic samples. Second, DPE attributes failures to specific weaknesses, dynamically adjusts the data mixture, and guides agents to generate weakness focused data for targeted reinforcement. Experiments on Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B-Instruct show stable, continual gains across eleven benchmarks, indicating DPE as a scalable paradigm for continual LMM training under open task distributions. Our code, models, and data are publicly available at https://github.com/hongruijia/DPE.

While Large Language Model (LLM) agents have shown encouraging progress in financial reasoning, their effectiveness in realistic financial workflows is severely hindered by the lack of holistic benchmarks and the fragility of unstructured reasoning. To address this gap, we introduce ProFinR, the first Professional Finance Reasoning benchmark, covering four financial domain tasks and comprising 528 expert-designed problems. To solve these complex financial reasoning questions, we construct the Financial Tool Universe, a tool library containing 53 domain-specific tools organized into 13 categories. Building on the tool library, we introduce ProFinAgent, a structured agent framework based on Directed Acyclic Graph (DAG) and Case-Based Memory (CBM). Compared with strictly sequential workflows, ProFinAgent coordinates tool execution through a DAG planner for parallel efficiency and uses CBM to retrieve prior cases for more reliable decision-making. Experimental results demonstrate that ProFinAgent achieves a 49.81% performance gain over state-of-the-art baselines with a 47.1% reduction in inference latency. Our code is available at https://github.com/tsinghua-fib-lab/ProFinAgent.

Retrieval-Augmented Generation (RAG) systems are increasingly deployed to provide query-based access to large knowledge bases, thereby introducing concrete privacy risks whereby the underlying corpus may be partially or fully extracted through the deployed service. Existing extraction attacks typically rely on locally driven search strategies, in which newly extracted content is inferred or expanded based on previously recovered fragments. However, real-world knowledge bases are often multi-source and heterogeneous, with pronounced semantic discontinuities across domains. Such gaps can trap extraction methods that rely on local semantic continuity in local optima, severely limiting large-scale corpus reconstruction. In this paper, we introduce an extraction framework (GeoEx) designed to navigate and reconstruct heterogeneous RAG knowledge bases without any prior knowledge. The framework plans extraction directly in the embedding space of a proxy retrieval model to improve global coverage, and employs an embedding inversion module to convert latent vectors into executable queries. We further propose a composite geometric strategy that combines orthogonal query synthesis for cross-domain exploration with local embedding perturbations for dense extraction within discovered clusters. Experiments on mixed corpora spanning eight diverse domains and multiple retrievers and LLMs show that GeoEx significantly outperforms baselines in both extraction coverage and query efficiency.

Spatio-temporal forecasting is critical for real-world applications like traffic management, yet capturing complex interactions under high-noise conditions remains challenging. While current methods have shown improved accuracy using spatial physical priors, they often struggle with evolving temporal correlations and systematic errors. In this work, we propose a nested forecasting framework that couples future macro-level regional trends with micro-level historical observations, enabling top-down guidance from abstract future representations for fine-grained forecasting. Specifically, we construct semantically coherent regions via spectral clustering and design a progressive coarse-to-fine predictor to inject macro-dynamics into node-level forecasting. Extensive experiments on multiple real-world datasets demonstrate that our method consistently outperforms state-of-the-art baselines, validating the effectiveness of future macro-guided nested forecasting.

LLM-based agents solve complex tasks through iterative reasoning, tool use, and environment interaction, where each intermediate thought directly shapes subsequent actions. Small deviations in these thoughts can therefore propagate into unsafe behaviors, yet existing guardrails typically operate only on final outputs or require intrusive model modifications. We introduce Thought-Aligner, a lightweight plug-in safety model that performs causal correction on unsafe thoughts before action execution, without altering the underlying agent. The corrected thoughts are fed back into the agent, steering its decision process and tool use toward safer trajectories. Because it operates solely at the thought level, Thought-Aligner is model-agnostic and can be integrated into diverse agent frameworks. We train Thought-Aligner via two-stage contrastive learning on paired safe and unsafe thoughts generated across ten risk scenarios. Experiments on diverse agent-safety benchmarks and six LLMs show that Thought-Aligner increases behavioral safety from about 50% without protection to around 90% on average, exceeding state-of-the-art guardrails by roughly 23%, while also improving helpfulness by about 5%. The method incurs low per-step latency and minimal overhead, enabling scalable and practical deployment. We publicly release Thought-Aligner-7B at https://huggingface.co/WhitzardAgent/Thought-Aligner-7B.

Generating realistic adversarial examples for tabular data remains challenging due to heterogeneous feature types and asymmetric inter-feature dependencies. Existing approaches typically rely on handcrafted constraints or undirected similarity criteria to delimit the feasible attack region, which often fail to capture the structural dependency governing tabular generation. Consequently, standard attacks typically produce perturbations that are statistically likely yet semantically inconsistent and prone to optimization stagnation via gradient masking. To address this, we propose LCSA, a white-box framework that formulates adversarial generation as optimization over structurally admissible perturbations. LCSA leverages an ensemble of heterogeneous neural Structural Causal Models to infer dependencies and introduces a structure-aware ripple mechanism. Unlike attacks that perturb features in isolation, this mechanism propagates updates downstream, acting as a structural preconditioner that conditions gradient flow to mitigate masking effects. Extensive experiments demonstrate that LCSA outperforms state-of-the-art baselines in 45 of 50 evaluated configurations, yielding adversarial examples with superior structural consistency and transferability.

While Reinforcement Learning from Human Feedback (RLHF) is the standard paradigm for aligning large language models with human preferences, its effectiveness in pluralistic settings has been called into question. Notably, recent work by Gölz et al. (2025) demonstrated that the *distortion* — defined as the multiplicative gap between the average user utility of the RLHF policy and the optimal average utility — can scale exponentially with the Bradley-Terry temperature parameter $\beta$ when users have heterogeneous preferences. In this work, we present a fine-grained analysis of the distortion of RLHF with reward clipping and demonstrate that such exponential degradation is not a fundamental property of the algorithm but rather a consequence of distribution mismatch between the distribution generating preference data ($\mu$) and the KL reference policy ($\pi_{\mathrm{ref}})$. To this end, we establish tight upper and lower bounds on the distortion of RLHF across multiple regimes of the KL regularization strength. We show that in a representative regime, under the Bradley–Terry model, the distortion is $\tilde{\Theta}(\beta B + \beta)$, where $B$ is an upper bound on the log density ratio between $\mu$ and $\pi_{\mathrm{ref}}$. In particular, when there is no distribution mismatch (i.e., $\mu = \pi_{\mathrm{ref}}$), RLHF achieves the optimal distortion of $O(\beta)$ up to a constant. Our results suggest that, to reasonably maximize average utility with RLHF, it is preferable to use on-policy sampled preference data or to fine-tune before RLHF on data from a source close to $\mu$.