ICLR 2026 论文列表

Harnessing the power of LLMs requires a delicate dance between being helpful and harmless, leading to two critical challenges: vulnerability to adversarial attacks that elicit unsafe content, and a tendency for overrefusal on benign but sensitive prompts. Current approaches often navigate this dance with safeguard models that completely reject any content that contains unsafe portions. This approach cuts the music entirely—it may exacerbate overrefusals and fails to provide nuanced guidance for queries it refuses. To teach models a more coordinated choreography, we propose WaltzRL, a novel multi-agent reinforcement learning framework that formulates safety alignment as a collaborative, positive-sum game. WaltzRL jointly trains a conversation agent and a feedback agent, where the latter is incentivized to provide useful suggestions that improve the safety and helpfulness of the conversation agent's responses. At the core of WaltzRL is a Dynamic Improvement Reward (DIR) that evolves over time based on how well the conversation agent incorporates the feedback. At inference time, unsafe or overrefusing responses from the conversation agent are improved rather than discarded. The feedback agent is deployed together with the conversation agent and only engages adaptively when needed, preserving helpfulness and low latency on safe queries. Our experiments, conducted across five diverse datasets, demonstrate that WaltzRL significantly reduces both unsafe responses (e.g., from 39.0% to 4.6% on WildJailbreak) and overrefusals (from 45.3% to 9.9% on OR-Bench) compared to various baselines. By enabling the conversation and feedback agents to co-evolve and adaptively apply feedback, WaltzRL enhances LLM safety without degrading general capabilities, thereby advancing the Pareto front between helpfulness and harmlessness.

Attribute correlations in the training data will compromise the ability of a deep generative model (DGM) to synthesize images with under-represented attribute combinations ($\textit{i.e.,}$ minority samples). Existing approaches mitigate this by data re-sampling to remove attribute correlations seen by the DGM, using a classifier to provide $\textit{pseudo-supervision}$ for the generation of counterfactual samples, or incorporating inductive bias to explicitly decompose the generation into independent causal mechanisms. We present ProReGen, a $\textit{progressive residual generation}$ approach inspired by the classical Robinson's transformation to partial out from an image attribute $\mathbf{x}_2$ its component $m(\mathbf{x}_1)$ that is predictable by other image attributes $\mathbf{x}_1$, and the residual $\gamma = \mathbf{x}_2 - m(\mathbf{x}_1)$ that is not. This simplifies the original problem of learning a DGM $g(\mathbf{x}_1, \mathbf{x}_2)$ conditioned on correlated inputs, to learning $\tilde{g}(\mathbf{x}_1, \gamma)$ conditioned on orthogonal inputs. It further allows us to progressively learn $\tilde{g}$ by first shifting the burden to abundant majority samples to learn the generator $\tilde{g}(\mathbf{x}_1, \gamma = 0)$, and then expanding it with additional layers $g\_{\text{res}}$ to resolve its difference to $\tilde{g}(\mathbf{x}_1, \gamma)$ using residual attribute $\gamma$ on limited minority samples. On three benchmark datasets with curated varying strengths of attribute correlations, we demonstrate that ProReGen---with input orthogonalizaton and progressive residual learning---improved the correctness and quality of generations compared to existing strategies.

Solving Traveling Salesman Problem (TSP) is NP-hard yet fundamental for wide real-world applications. Classical exact methods face challenges in scaling, and heuristic methods often require domain-specific parameter calibration. While learning-based approaches have shown promise, they suffer from poor generalization and limited scalability due to fixed training data. This work proposes ViTSP, a novel framework that leverages pre-trained vision language models (VLMs) to visually guide the solution process for large-scale TSPs. The VLMs function to identify promising small-scale subproblems from a visualized TSP instance, which are then efficiently optimized using an off-the-shelf solver to improve the global solution. ViTSP bypasses the dedicated model training at the user end while maintaining effectiveness across diverse instances. Experiments on real-world TSP instances ranging from 1k to 88k nodes demonstrate that ViTSP consistently achieves solutions with average optimality gaps below 0.2\%, outperforming existing learning-based methods. Under the same runtime budget, it surpasses the best-performing heuristic solver, LKH-3, by reducing its gaps by 12\% to 100\%, particularly on very-large-scale instances with more than 10k nodes. Our framework offers a new perspective in hybridizing pre-trained generative models and operations research solvers in solving combinatorial optimization problems, with practical implications for integration into more complex logistics systems. The code is available at https://anonymous.4open.science/r/ViTSP_codes-6683.

This work presents MAC-Flow, a simple yet expressive framework for multi-agent coordination. We argue that requirements of effective coordination are twofold: *(i)* a rich representation of the diverse joint behaviors present in offline data and *(ii)* the ability to act efficiently in real time. However, prior approaches often sacrifice one for the other, *i.e.*, denoising diffusion-based solutions capture complex coordination but are computationally slow, while Gaussian policy-based solutions are fast but brittle in handling multi-agent interaction. MAC-Flow addresses this trade-off by first learning a flow-based representation of joint behaviors, and then distilling it into decentralized one-step policies that preserve coordination while enabling fast execution. Across four different benchmarks, including $12$ environments and $34$ datasets, MAC-Flow alleviates the trade-off between performance and computational cost, specifically achieving about $\boldsymbol{\times14.5}$ faster inference compared to diffusion-based MARL methods, while maintaining good performance. At the same time, its inference speed is similar to that of prior Gaussian policy-based offline MARL methods.

Visual prompted object detection enables interactive and flexible definition of target categories, thereby facilitating open-vocabulary detection. Since visual prompts are derived directly from image features, they often outperform text prompts in recognizing rare categories. Nevertheless, research on visual prompted detection has been largely overlooked, and it is typically treated as a byproduct of training text prompted detectors, which hinders its development. To fully unlock the potential of visual-prompted detection, we investigate the reasons why its performance is suboptimal and reveal that the underlying issue lies in the absence of global discriminability in visual prompts. Motivated by these observations, we propose DETR-ViP, a robust object detection framework that yields class-distinguishable visual prompts. On top of basic image-text contrastive learning, DETR-ViP incorporates global prompt integration and visual-textual prompt relation distillation to learn more discriminative prompt representations. In addition, DETR-ViP employs a selective fusion strategy that ensures stable and robust detection. Extensive experiments on COCO, LVIS, ODinW, and Roboflow100 demonstrate that DETR-ViP achieves substantially higher performance in visual prompt detection compared to other state-of-the-art counterparts. A series of ablation studies and analyses further validate the effectiveness of the proposed improvements and shed light on the underlying reasons for the enhanced detection capability of visual prompts.

The fields of explainable AI and mechanistic interpretability aim to uncover the internal structure of neural networks, with circuit discovery as a central tool for understanding model computations. Existing approaches, however, rely on manual inspection and remain limited to toy tasks. Automated interpretability offers scalability by analyzing isolated features and their activations, but it often misses interactions between features and depends strongly on external LLMs and dataset quality. Transcoders have recently made it possible to separate feature attributions into input-dependent and input-invariant components, providing a foundation for more systematic circuit analysis. Building on this, we propose WeightLens and CircuitLens, two complementary methods that go beyond activation-based analysis. WeightLens interprets features directly from their learned weights, removing the need for explainer models or datasets while matching or exceeding the performance of existing methods on context-independent features. CircuitLens captures how feature activations arise from interactions between components, revealing circuit-level dynamics that activation-only approaches cannot identify. Together, these methods increase interpretability robustness and enhance scalable mechanistic analysis of circuits while maintaining efficiency and quality.

Transformers excel through content-addressable retrieval and the ability to exploit contexts of, in principle, unbounded length. We recast associative memory at the level of probability measures, treating a context as a distribution over tokens and viewing attention as an integral operator on measures. Concretely, for mixture contexts $\nu = I^{-1} \sum_{i=1}^I \mu^{(i)}$ and a query $x_{\mathrm{q}}(i^\*)$, the task decomposes into (i) recall of the relevant component $\mu^{(i^\*)}$ and (ii) prediction from $(\mu_{i^\*},x_{\mathrm{q}})$. We study learned softmax attention (not a frozen kernel) trained by empirical risk minimization and show that a shallow measure-theoretic Transformer composed with an MLP learns the recall-and-predict map under a spectral assumption on the input densities. We further establish a matching minimax lower bound with the same rate exponent (up to multiplicative constants), proving sharpness of the convergence order. The framework offers a principled recipe for designing and analyzing Transformers that recall from arbitrarily long, distributional contexts with provable generalization guarantees.

Multimodal Large Language Models (MLLMs) have achieved impressive progress in vision–language alignment, yet they remain limited in visual–spatial reasoning. We first identify that this limitation arises from the attention mechanism: visual tokens are overshadowed by language tokens, preventing the model from consistently recognizing the same visual cues across frames. To address this challenge, we draw a novel connection between the self-expressiveness property in sparse subspace clustering and the attention mechanism in Transformers. Building on this insight, we propose VideoAnchor, a plug-and-play module that leverages subspace affinities to reinforce visual cues across frames without retraining, effectively anchoring attention to shared visual structures. Extensive experiments across benchmarks and backbone models show consistent performance gains — e.g., 3.2% and 4.6% improvements on VSI-Bench and Video-MME (spatial-related tasks) with InternVL2-8B and Qwen2.5VL-72B—while qualitative analyses demonstrate more coherent subspace partitions and stronger visual grounding.

AdamW has long been the dominant optimizer in language model pretraining, despite numerous claims that alternative optimizers offer 1.4 to 2$\times$ speedup. We posit that two methodological shortcomings have obscured fair comparisons and hindered practical adoption: (i) unequal hyperparameter tuning and (ii) limited or misleading evaluation setups. To address these two issues, we conduct a systematic study of ten deep learning optimizers across four model scales (0.1B–1.2B parameters) and data-to-model ratios (1--8$\times$ the Chinchilla optimum). We find that fair and informative comparisons require rigorous hyperparameter tuning and evaluations across a range of model scales and data-to-model ratios, performed at the end of training. First, optimal hyperparameters for one optimizer may be suboptimal for another, making blind hyperparameter transfer unfair. Second, the actual speedup of many proposed optimizers over well-tuned baselines is lower than claimed and decreases with model size to only 1.1$\times$ for 1.2B parameter models. Thirdly, comparing intermediate checkpoints before reaching the target training budgets can be misleading, as rankings between two optimizers can flip during training due to learning rate decay. Through our thorough investigation, we find that all the fastest optimizers such as Muon and Soap, use matrices as preconditioners --- multiplying gradients with matrices rather than entry-wise scalars. However, the speedup of matrix-based optimizers is inversely proportional to model scale, decreasing from 1.4$\times$ over AdamW for 0.1B parameter models to merely 1.1$\times$ for 1.2B parameter models.

Modern methods of generative modelling and unpaired data translation based on Schrodinger bridges and stochastic optimal control theory aim to transform an initial density to a target one in an optimal way. In the present paper, we assume that we only have access to i.i.d. samples from initial and final distributions. This makes our setup suitable for both generative modelling and unpaired data translation. Relying on the stochastic optimal control approach, we choose an Ornstein-Uhlenbeck process as the reference one and estimate the corresponding Schrodinger potential. Introducing a risk function as the Kullback-Leibler divergence between couplings, we derive tight bounds on generalization ability of an empirical risk minimizer in a class of Schrodinger potentials including Gaussian mixtures. Thanks to the mixing properties of the Ornstein-Uhlenbeck process, we almost achieve fast rates of convergence up to some logarithmic factors in favourable scenarios. We also illustrate performance of the suggested approach with numerical experiments.

Recent advances in large language models (LLMs) and agent system designs have empowered agents with unprecedented levels of capability. However, existing agent benchmarks are showing a trend of rapid ceiling-hitting by newly developed agents, making it difficult to meet the demands for evaluating agent abilities. To address this problem, we propose the Trajectory-based Reproducible Agent- benchmark Complexity Evolution (TRACE) framework. This framework takes an original task from an existing benchmark and encourages agents to freely explore and evolve it into a new task with higher difficulty while recording traceable agent trajectories. The framework proceeds in three stages: (1) evolutionary proposal mining, which provides task evolution proposals through preliminary exploration and divergent thinking; (2) problem formation and free exploration, where proposals are conceptualized into feasible problem candidates and the agents then explore them freely while recording their execution trajectories; and (3) multi-level validation, which ensures that the evolved tasks are accompanied by validatable and reproducible trajectories. Experiments on the GAIA benchmark demonstrate that the TRACE framework consistently enhances task complexity while improving the reliability of correctness through validatable execution trajectories. This work marks a paradigm shift from static, manually curated benchmarks to dynamic, self-evolving evaluation systems, providing a sustainable and challenging runway for agent development

Large Reasoning Models (LRMs) exhibit strong performance, yet often produce rationales that sound plausible but fail to reflect their true decision process, undermining reliability and trust. We introduce a formal framework for *reasoning faithfulness*, defined by two testable conditions: *stance consistency* (a coherent stance linking reasoning to answer) and *causal influence* (the stated reasoning causally drives the answer under output-level interventions), explicitly decoupled from accuracy. To operationalize this, we present **RFEval**, a benchmark of 7,186 instances across seven tasks that probes faithfulness via controlled, output-level counterfactual interventions. Evaluating twelve open-source LRMs, we find unfaithfulness in 49.7% of outputs, predominantly from post-intervention stance inconsistency. Failures are concentrated in brittle, convergent domains such as math and code, and correlate more with post-training regimes than with scale: within-family ablations indicate that adding current RL-style objectives on top of supervised fine-tuning can *reduce* reasoning faithfulness, even when accuracy is maintained. Crucially, *accuracy is neither a sufficient nor a reliable proxy for faithfulness*: once controlling for model and task, the accuracy–faithfulness link is weak and statistically insignificant. Our work establishes a rigorous methodology for auditing LRM reliability and shows that trustworthy AI requires optimizing not only for correct outcomes but also for the structural integrity of the reasoning process.

In control problems and basic scientific modeling, it is important to compare observations with dynamical simulations. For example, comparing two neural systems can shed light on the nature of emergent computations in the brain and deep neural networks. Recently, Ostrow et al. (2023) introduced Dynamical Similarity Analysis (DSA), a method to measure the similarity of two systems based on their recurrent dynamics rather than geometry or topology. However, DSA does not consider how inputs affect the dynamics, meaning that two similar systems, if driven differently, may be classified as different. Because real-world dynamical systems are rarely autonomous, it is important to account for the effects of input drive. To this end, we introduce a novel metric for comparing both intrinsic (recurrent) and input-driven dynamics, called InputDSA (iDSA). InputDSA extends the DSA framework by estimating and comparing both input and intrinsic dynamic operators using a variant of Dynamic Mode Decomposition with control (DMDc) based on subspace identification. We demonstrate that InputDSA can successfully compare partially observed, input-driven systems from noisy data. We show that when the true inputs are unknown, surrogate inputs can be substituted without a major deterioration in similarity estimates. We apply InputDSA on Recurrent Neural Networks (RNNs) trained with Deep Reinforcement Learning, identifying that high-performing networks are dynamically similar to one another, while low-performing networks are more diverse. Lastly, we apply InputDSA to neural data recorded from rats performing a cognitive task, demonstrating that it identifies a transition from input-driven evidence accumulation to intrinsically-driven decision-making. Our work demonstrates that InputDSA is a robust and efficient method for comparing intrinsic dynamics and the effect of external input on dynamical systems.

Data teams at frontier AI companies routinely train small proxy models to make critical decisions about pretraining data recipes for full-scale training. However, the community has a limited understanding of whether and when conclusions drawn from small-scale experiments reliably transfer to large-scale production training. In this work, we uncover a critical issue in the standard practice of training small proxy models on each data recipe with a single set of hyperparameters. We demonstrate that each dataset requires its own optimal training configuration, and that dataset rankings can completely reverse with even minor adjustments to proxy training hyperparameters. Furthermore, this creates a disconnect from the actual model development pipeline, where hyperparameter optimization is a standard step. Consequently, we propose that the objective of data selection should be to identify the dataset that yields the best performance after its own hyperparameter optimization. We introduce a simple yet effective patch to the current proxy-model-based method: training proxy models with sufficiently small learning rates produces dataset rankings that strongly correlate with those obtained when large-scale models are properly tuned for each dataset. Theoretically, we prove that, for random-feature models, this approach preserves the ordering of datasets according to their optimal achievable losses. Empirically, we validate this approach through comprehensive experiments across 23 data recipes covering four critical dimensions of data curation decisions faced in production settings, demonstrating dramatic improvements in proxy model reliability.

PDE surrogate models such as FNO and PINN struggle to predict solutions across inputs with diverse physical units and scales, limiting their out-of-distribution (OOD) generalization. We propose a $\pi$-invariant test-time projection that aligns test inputs with the training distribution by solving a log-space least squares problem that preserves Buckingham $\pi$-invariants. For PDEs with multidimensional spatial fields, we use geometric representative $\pi$-values to compute distances and project inputs, overcoming degeneracy and singular points that limit prior $\pi$-methods. To accelerate projection, we cluster the training set into K clusters, reducing the complexity from O(MN) to O(KN) for the M training and N test samples. Across wide input scale ranges, tests on 2D thermal conduction and linear elasticity achieve an average MAE reduction up to $\approx 91\\%$ with minimal overhead. This training-free, model-agnostic method is expected to apply to more diverse PDE-based simulations.

Contact forces introduce discontinuities into robot dynamics that severely limit the use of simulators for gradient-based optimization. Penalty-based simulators such as MuJoCo, soften contact resolution to enable gradient computation. However, realistically simulating hard contacts requires stiff solver settings, which leads to incorrect simulator gradients when using automatic differentiation. Contrarily, using non-stiff settings strongly increases the sim-to-real gap. We analyze penalty-based simulators to pinpoint why gradients degrade under hard contacts. Building on these insights, we propose DiffMJX, which couples adaptive time integration with penalty-based simulation to substantially improve gradient accuracy. A second challenge is that contact gradients vanish when bodies separate. To address this, we introduce contacts from distance (CFD) which combines penalty-based simulation with straight-through estimation. By applying CFD exclusively in the backward pass, we obtain informative pre-contact gradients while retaining physical realism.

Federated Learning enables collaborative fine-tuning of Large Language Models (LLMs) across decentralized Non-Independent and Identically Distributed (Non-IID) clients, but such models' massive parameter sizes lead to significant memory and communication challenges. This work introduces Meerkat, a sparse zeroth-order optimization (ZO) method designed for federated LLM fine-tuning. By limiting fine-tuning to a transferable, static, extremely sparse subset of parameters, Meerkat achieves remarkable communication efficiency, enabling cost-effective high-frequency synchronization. With theoretical analysis and experiments, we show that this high-frequency communication effectively mitigates Non-IID data challenges and leads to superior performance compared to full-parameter ZO. Furthermore, experiment results show that Meerkat outperforms existing sparsity baselines with better performance at the same communication frequency. To further handle Non-IID drift, Meerkat leverages traceable local updates and forms a virtual path for each client. This virtual path mechanism reveals the GradIP phenomenon: the inner products between LLM pre-training gradients maintained by server and client gradients estimated via ZO converges for extreme Non-IID clients but oscillates for IID ones. This distinct behavior provides a signal for identifying clients with extreme data heterogeneity. Using this signal, Meerkat-vp is proposed to analyze GradIP trajectories to identify extreme Non-IID clients and applies early stopping to enhance aggregated model quality. Experiments confirm that Meerkat and Meerkat-vp significantly improve the efficiency and effectiveness of ZO federated LLM fine-tuning.

Large language models (LLMs) with reasoning abilities have demonstrated growing promise for tackling complex scientific problems. Yet such tasks are inherently domain-specific, unbounded and open-ended, demanding exploration across vast and flexible solution spaces. Existing approaches, whether purely learning-based or reliant on carefully designed workflows, often suffer from limited exploration efficiency and poor generalization. To overcome these challenges, we present **HELIX**---a **H**ierarchical **E**volutionary reinforcement **L**earning framework with **I**n-context e**X**periences. HELIX introduces two key novelties: (i) a diverse yet high-quality pool of candidate solutions that broadens exploration through in-context learning, and (ii) reinforcement learning for iterative policy refinement that progressively elevates solution quality. This synergy enables the discovery of more advanced solutions. On the circle packing task, HELIX achieves a new state-of-the-art with a sum of radii of 2.635983 using only a 14B model. Across standard machine learning benchmarks, HELIX further surpasses GPT-4o with a carefully engineered pipeline, delivering an average F1 improvement of 5.95 points on the Adult and Bank Marketing datasets and a 40.5\% reduction in RMSE on Boston Housing.

Flow map models such as Consistency Models (CM) and Mean Flow (MF) enable few-step generation by learning the long jump of the ODE solution of diffusion models, yet training remains unstable, sensitive to hyperparameters, and costly. Initializing from a pre-trained diffusion model helps, but still requires converting infinitesimal steps into a long-jump map, leaving instability unresolved. We introduce *mid-training*, the first concept and practical method that inserts a lightweight intermediate stage between the (diffusion) pre-training and the final flow map training (i.e., post-training) for vision generation. Concretely, *Consistency Mid-Training* (CMT) is a compact and principled stage that trains a model to map points along a solver trajectory from a pre-trained model, starting from a prior sample, directly to the solver-generated clean sample. It yields a trajectory-consistent and stable initialization. This initializer outperforms random and diffusion-based baselines and enables fast, robust convergence without heuristics. Initializing post-training with CMT weights further simplifies flow map learning. Empirically, CMT achieves state-of-the-art two-step FIDs of 1.97 (CIFAR-10), 1.32 (ImageNet $64\times64$), and 1.84 (ImageNet $512\times512$), using up to $98$\% less training data and GPU time than CMs. On ImageNet $256\times256$, it attains 1-step FID 3.34 with $\sim50$\% less training than MF from scratch (FID 3.43). On MSCOCO T2I, CMT reaches the best FID with $\sim47$\% less training. This establishes CMT as a principled, efficient, and general framework for training flow map models. Code and models are available at https://github.com/sony/cmt.