ICLR 2026 Papers

While reinforcement learning from verifiable rewards (RLVR) typically has relied on a single binary verification signal, symbolic proof assistants in formal reasoning offer rich, fine-grained structured feedback. This gap between structured processes and unstructured rewards highlights the importance of feedback that is both dense and sound. In this work, we demonstrate that the Lean proof assistant itself can serve as a symbolic process oracle, supplying both outcome-level and fine-grained tactic-level verified feedback during training. Proof attempts are parsed into tactic sequences, and Lean's elaboration marks both locally sound steps and the earliest failing step, yielding dense, verifier-grounded credit signals rooted in type theory. We incorporate these structured rewards into a GRPO-style reinforcement learning objective with first-error propagation and first-token credit methods that balances outcome- and process-level advantages. Experiments with STP-Lean and DeepSeek-Prover-V1.5 show that tactic-level supervision outperforms outcome-only baselines in most settings, delivering improvements on benchmarks such as MiniF2F and ProofNet. Beyond empirical gains, our study highlights a broader perspective: symbolic proof assistants are not only verifiers at evaluation time, but can also act as process-level reward oracles during training. This opens a path toward reinforcement learning frameworks that combine the scalability of language models with the reliability of symbolic verification for formal reasoning.

Safety evaluation of multimodal foundation models often treats vision and language inputs separately, missing risks from joint interpretation where benign content becomes harmful in combination. Existing approaches also fail to distinguish clearly unsafe content from borderline cases, leading to problematic over-blocking or under-refusal of genuinely harmful content. We present Vision Language Safety Understanding (VLSU), a comprehensive framework to systematically evaluate multimodal safety through fine-grained severity classification and combinatorial analysis across 17 distinct safety patterns. Using a multi-stage pipeline with real-world images and human annotation, we construct a large-scale benchmark of 8,187 samples spanning 15 harm categories. Our evaluation of eleven state-of-the-art models reveals systematic joint understanding failures: while models achieve 90\%+ accuracy on clear unimodal safety signals, performance degrades substantially to 20-55\% when joint image-text reasoning is required to determine the safety label. Most critically, 34\% of errors in joint image-text safety classification occur despite correct classification of the individual modalities, further demonstrating absent compositional reasoning capabilities. Additionally, we find that models struggle to balance refusing unsafe content while still responding to borderline cases that deserve engagement. For example, we find that instruction framing can reduce the over-blocking rate on borderline content from 62.4\% to 10.4\% in Gemini-1.5, but only at the cost of under-refusing on unsafe content with refusal rate dropping from 90.8\% to 53.9\%. Overall, our framework exposes weaknesses in joint image-text understanding and alignment gaps in current models, and provides a critical test bed to enable the next milestones in research on robust vision–language safety.

Human mobility trajectory data supports a wide range of applications, including urban planning, intelligent transportation systems, and public safety monitoring. However, large-scale, high-quality mobility datasets are difficult to obtain due to privacy concerns. Raw trajectory data may reveal sensitive user information, such as home addresses, routines, or social relationships, making it crucial to develop privacy-preserving alternatives. Recent advances in deep generative modeling have enabled synthetic trajectory generation, but existing methods either lack formal privacy guarantees or suffer from reduced utility and scalability. Differential Privacy (DP) has emerged as a rigorous framework for data protection, and recent efforts such as PATE-GAN and PATEGAIL integrate DP with generative adversarial learning. While promising, these methods struggle to generalize across diverse trajectory patterns and often incur significant utility degradation. In this work, we propose a new framework that builds on PATEGAIL by introducing a sensitivity-aware noise injection module that dynamically adjusts privacy noise based on sample-level sensitivity. This design significantly improves trajectory fidelity, downstream task performance, and scalability under strong privacy guarantees. We evaluate our method on real-world mobility datasets and demonstrate its superiority over state-of-the-art baselines in terms of privacy-utility trade-off.

This paper introduces LLM2Fx-Tools, a multimodal tool-calling framework that generates executable sequences of audio effects (Fx-chain) for music post-production. LLM2Fx-Tools uses a large language model (LLM) to understand audio inputs, select audio effects types, determine their order, and estimate parameters, guided by chain-of-thought (CoT) planning. We also present LP-Fx, a new instruction-following dataset with structured CoT annotations and tool calls for audio effects modules. Experiments show that LLM2Fx-Tools can infer an Fx-chain and its parameters from pairs of unprocessed and processed audio, enabled by autoregressive sequence modeling, tool calling, and CoT reasoning. We further validate the system in a style transfer setting, where audio effects information is transferred from a reference source and applied to new content. Finally, LLM-as-a-judge evaluation demonstrates that our approach generates appropriate CoT reasoning and responses for music production queries. To our knowledge, this is the first work to apply LLM-based tool calling to audio effects modules, enabling interpretable and controllable music production.

Hallucinations are a common issue that undermine the reliability of large language models (LLMs). Recent studies have identified a specific subset of hallucinations, known as confabulations, which arise due to predictive uncertainty of LLMs. To detect confabulations, various methods for estimating predictive uncertainty in natural language generation (NLG) have been developed. These methods are typically evaluated by correlating uncertainty estimates with the correctness of generated text, with question-answering (QA) datasets serving as the standard benchmark. However, commonly used approximate correctness functions have substantial disagreement between each other and, consequently, in the ranking of the uncertainty estimation methods. This allows one to inflate the apparent performance of uncertainty estimation methods. We propose using several alternative risk indicators for risk correlation experiments that improve robustness of empirical assessment of UE algorithms for NLG. For QA tasks, we show that marginalizing over multiple LLM-as-a-judge variants leads to reducing the evaluation biases. Furthermore, we explore structured tasks as well as out of distribution and perturbation detection tasks which provide robust and controllable risk indicators. Finally, we propose to use an Elo rating of uncertainty estimation methods to give an objective summarization over extensive evaluation settings.

Diffusion-based video generation can create realistic videos, yet existing image and text-based conditioning fails to offer precise motion control. Prior methods for motion control typically rely on displacement-based conditioning and require model-specific fine-tuning, which is computationally expensive and restrictive. We introduce Time-to-Move (TTM), a training-free, plug-and-play framework for motion- and appearance-controlled video generation with image-to-video (I2V) diffusion models. Our key insight is to use crude reference animations, obtained through user-friendly manipulations such as cut-and-drag or depth-based reprojection, as direct motion guidance, analogous to using coarse layout input in image editing. To integrate these signals, we adapt SDEdit to the video domain while anchoring the appearance with image conditioning. We further propose dual-clock denoising, a region-dependent strategy that enforces strong alignment in motion-specified regions and grants flexibility elsewhere, balancing fidelity to user intent with natural dynamics. This lightweight modification of the sampling process incurs no additional training or runtime cost and is compatible with any backbone. Extensive experiments on object and camera motion benchmarks show that TTM matches or exceeds existing training-based baselines in realism and motion control. Beyond this, TTM introduces a unique capability: precise appearance control through pixel-level conditioning, exceeding the limits of text-only prompting.

Recent advances in Vision-Language Models (VLMs) have propelled embodied agents by enabling direct perception, reasoning, and planning task-oriented actions from visual inputs. However, such vision-driven embodied agents open a new attack surface: visual backdoor attacks, where the agent behaves normally until a visual trigger appears in the scene, then persistently executes an attacker-specified multi-step policy. We introduce BEAT, the first framework to inject such visual backdoors into VLM-based embodied agents using objects in the environments as triggers. Unlike textual triggers, object triggers exhibit wide variation across viewpoints and lighting, making them difficult to implant reliably. BEAT addresses this challenge by (1) constructing a training set that spans diverse scenes, tasks, and trigger placements to expose agents to trigger variability, and (2) introducing a two-stage training scheme that first applies supervised fine-tuning (SFT) and then our novel Contrastive Trigger Learning (CTL). CTL formulates trigger discrimination as preference learning between trigger-present and trigger-free inputs, explicitly sharpening the decision boundaries to ensure precise backdoor activation. Across various embodied agent benchmarks and VLMs, BEAT achieves attack success rates up to 80\%, while maintaining strong benign task performance, and generalizes reliably to out-of-distribution trigger placements. Notably, compared to naive SFT, CTL boosts backdoor activation accuracy up to 39\% under limited backdoor data. These findings expose a critical yet unexplored security risk in VLM-based embodied agents, underscoring the need for robust defenses before real-world deployment.

Chromosomes are the carriers of genetic information. Further understanding their 3D structure can help reveal gene-regulatory mechanisms and cellular functions. A standard pipeline for reconstructing the chromosome 3D structure first applies the single-cell Hi-C high-throughput screening method to measure pairwise interactions between DNA fragments at different resolutions; then it adopts computational methods to reconstruct the 3D structures from these contacts. These include traditional numerical methods and deep learning models, which struggle with limited model expressiveness and poor generalization across resolutions. To solve this issue, we propose InertialGenome, a novel transformer-based framework for robust and resolution-agnostic chromosome reconstruction. InertialGenome first adopts the inertial frame for the pose canonicalization. Then, based on such an invariant frame, it proposes a Transformer with geometry-aware positional encoding, leveraging Nyström estimation. To verify the effect of InertialGenome, we evaluate our model on two single-cell 3D reconstruction datasets with four resolutions, reaching superior performance over all four computational baselines. In addition to the structure metrics, we observe that InertialGenome outperforms when analyzing the function of reconstructed structures on two validation tasks. Finally, we leverage InertialGenome for cross-resolution transfer learning, yielding up to a 5\% improvement from low to high resolution.

Inspired by the effectiveness of genetic algorithms and the importance of synthesizability in molecular design, we present SynGA, a simple genetic algorithm that operates directly over synthesis routes. Our method features custom crossover and mutation operators that explicitly constrain it to synthesizable molecular space. By modifying the fitness function, we demonstrate the effectiveness of SynGA on a variety of design tasks, including synthesizable analog search and sample-efficient property optimization, for both 2D and 3D objectives. Furthermore, by coupling SynGA with a machine learning-based filter that focuses the building block set, we boost SynGA to state-of-the-art performance. For property optimization, this manifests as a model-based variant SynGBO, which employs SynGA and block filtering in the inner loop of Bayesian optimization. Since SynGA is lightweight and enforces synthesizability by construction, our hope is that SynGA can not only serve as a strong standalone baseline but also as a versatile module that can be incorporated into larger synthesis-aware workflows in the future.

Multi-agent learning faces a fundamental tension: leveraging distributed collaboration without sacrificing the personalization needed for diverse agents. This tension intensifies when aiming for full personalization while adapting to unknown heterogeneity levels—gaining collaborative speedup when agents are similar, without performance degradation when they are different. Embracing the challenge, we propose personalized collaborative learning (PCL), a novel framework for heterogeneous agents to collaboratively learn personalized solutions with seamless adaptivity. Through carefully designed bias correction and importance correction mechanisms, our method AffPCL robustly handles both environment and objective heterogeneity. We prove that AffPCL reduces sample complexity over independent learning by a factor of $\max\\{n^{-1}, \delta\\}$, where $n$ is the number of agents and $\delta\in[0,1]$ measures their heterogeneity. This *affinity-based* acceleration automatically interpolates between the linear speedup of federated learning in homogeneous settings and the baseline of independent learning, without requiring prior knowledge of the system. Our analysis further reveals that an agent may obtain linear speedup even by collaborating with arbitrarily dissimilar agents, unveiling new insights into personalization and collaboration in the high heterogeneity regime.

The pursuit of robot generalists—instructable agents capable of performing diverse tasks across diverse environments—demands rigorous and scalable evaluation. Yet real-world testing of robot policies remains fundamentally constrained: it is labor-intensive, slow, unsafe at scale, and difficult to reproduce. Existing simulation benchmarks are similarly limited, as they train and test policies within the same synthetic domains and cannot assess models trained primarily on real-world demonstrations, which is the dominant paradigm for today’s vision-language-action (VLA) models. As policies expand in scope and complexity, these barriers only intensify, since defining ``success" in robotics often hinges on nuanced human judgments of execution quality. In this paper, we introduce a new benchmarking framework that overcomes these challenges by shifting VLA evaluation into large-scale simulated environments augmented with online human feedback. Leveraging advances in vision-language models, 2D-to-3D generative modeling, and differentiable rendering, our approach automatically converts video demonstrations from widely used robot datasets into simulated counterparts. Within these digital twins, we assess VLA policies using both automated VLM-guided scoring and scalable human preference judgments collected from crowdworkers—transforming human involvement from tedious scene setup, resetting, and safety supervision into lightweight preference comparisons.

Recent advances in alignment techniques such as Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and Direct Preference Optimization (DPO) have improved the safety of large language models (LLMs). However, these LLMs remain vulnerable to jailbreak attacks that disguise harmful intent through indirect or deceptive phrasing. Using causal intervention, we empirically demonstrate that this vulnerability stems from shallow alignment mechanisms that lack deep reasoning, often rejecting harmful prompts without truly understanding why they are harmful. To mitigate this vulnerability, we propose enhancing alignment through reasoning-aware post-training. We construct and release a novel Chain-of-Thought (CoT) fine-tuning dataset that includes both utility-oriented and safety-critical prompts with step-by-step rationales. Fine-tuning on this dataset encourages models to produce principled refusals grounded in reasoning, outperforming standard SFT baselines. Furthermore, inspired by failure patterns in CoT fine-tuning, we introduce **Alignment-Weighted DPO**, which targets the most problematic parts of an output by assigning different preference weights to the reasoning and final-answer segments. This produces finer-grained, targeted updates than vanilla DPO and improves robustness to diverse jailbreak strategies. Extensive experiments across multiple safety and utility benchmarks show that our method consistently improves alignment robustness while maintaining overall model utility.

Urban development impacts over half of the global population, making human-centered understanding of its structural and perceptual changes essential for smart city planning. While Multimodal Large Language Models (MLLMs) have shown remarkable capabilities across various domains, existing benchmarks that explore their performance in urban environments remain limited, lacking systematic exploration of temporal evolution and subjective perception of urban environment that aligns with human perception. To address these limitations, we propose UrbanFeel, a comprehensive benchmark designed to evaluate the performance of MLLMs in urban development understanding and subjective environmental perception. UrbanFeel comprises 14.3K carefully constructed visual questions spanning three cognitively progressive dimensions: Static Scene Perception, Temporal Change Perception, and Subjective Environmental Perception. We collect multi-temporal single-view and panoramic street-view images from 11 representative cities worldwide, and generate high-quality question-answer pairs through a hybrid pipeline of spatial clustering, rule-based generation, model-assisted prompting, and manual annotation. Through extensive evaluation of 20 state-of-the-art MLLMs, we observe that Gemini-2.5 Pro achieves the best overall performance, with its accuracy approaching human expert levels and narrowing the average gap to just 1.5%. Most models perform well on tasks grounded in scene understanding. In particular, some models even surpass human annotators in pixel-level change detection. However, performance drops notably in tasks requiring temporal reasoning over urban development. Additionally, in the subjective perception dimension, several models reach human-level or even higher consistency in evaluating dimension such as beautiful and safety. Our results suggest that MLLMs are demonstrating rudimentary emotion understanding capabilities. Our UrbanFeel benchmark will be made publicly available.

Subject-driven generation faces a fundamental challenge: achieving high subject fidelity while maintaining semantic alignment with textual descriptions. While recent GRPO-based approaches have shown promise in aligning generative models with human preferences, they apply uniform optimization across all denoising timesteps, ignoring the temporal dynamics of how textual and visual conditions influence generation. We present Consis-GCPO, a causal reinforcement learning framework that reformulates multi-modal condition generation through discrete-time causal modeling. Our key insight is that different conditioning signals exert varying influence throughout the denoising process—text guides semantic structure in early steps while visual references anchor details in later stages. By introducing decoupled causal intervention trajectories, we quantify instantaneous causal effects at each timestep, transforming these measurements into temporally-weighted advantages for targeted optimization. This approach enables precise tracking of textual and visual contributions, ensuring accurate credit assignment for each conditioning modality. Extensive experiments demonstrate that Consis-GCPO significantly advances personalized generation, achieving superior subject consistency while preserving strong text-following capabilities, particularly excelling in complex multi-subject scenarios.

Many alignment methods, including reinforcement learning from human feedback (RLHF), rely on the Bradley-Terry reward assumption, which is not always sufficient to capture the full range and complexity of general human preferences. We explore RLHF under a general preference framework by modeling the alignment problem as a two-player zero-sum game in a game-theoretic framework, where the Nash equilibrium policy guarantees a 50\% win rate against any competing policy. However, previous self-play algorithms for finding the Nash policy either diverge or only converge to a Nash policy in a modified game, even in a simple synthetic setting, thereby failing to maintain the 50\% win rate guarantee against all other policies. We propose a meta-algorithm, **Co**nvergent **M**eta **Al**ignment Algorithm (COMAL), for language model alignment with general preferences, inspired by convergent algorithms in game theory. We provide theoretical analysis that our meta-algorithm converges to an exact Nash policy in the last iterate and demonstrate its effectiveness on a range of synthetic and preference optimization datasets. COMAL is simple and can be integrated with many existing methods designed for preference optimization with minimal changes, and empirically it consistently maintains above 60.2\% and 56.8\% win rates, when applied to Llama-3-8B-Instruct and Qwen2.5-7B, against all compared algorithms under controlled evaluations.

We introduce *Einstein Fields*, a neural representation designed to compress computationally intensive *four-dimensional* numerical relativity simulations into compact implicit neural network weights. By modeling the *metric*, the core tensor field of general relativity, Einstein Fields enable the derivation of physical quantities via automatic differentiation. Unlike conventional neural fields (e.g., signed distance, occupancy, or radiance fields), Einstein Fields fall into the class of *Neural Tensor Fields* with the key difference that, when encoding the spacetime geometry into neural field representations, dynamics emerge naturally as a byproduct. Our novel implicit approach demonstrates remarkable potential, including continuum modeling of four-dimensional spacetime, mesh-agnosticity, storage efficiency, derivative accuracy, and ease of use. It achieves up to a $\mathtt{4,000}$-fold reduction in storage memory compared to discrete representations while retaining a numerical accuracy of five to seven decimal places. Moreover, in single precision, differentiation of the Einstein Fields-parameterized metric tensor is up to five orders of magnitude more accurate compared to naive finite differencing methods. We demonstrate these properties on several canonical test beds of general relativity and numerical relativity simulation data, while also releasing an open-source $\mathtt{JAX}$-based library: https://github.com/AndreiB137/EinFields, taking the first steps to studying the potential of machine learning in numerical relativity.

While most autoregressive LLMs are constrained to one-by-one decoding, diffusion LLMs (dLLMs) have attracted growing interest for their potential to dramatically accelerate inference through parallel decoding. Despite this promise, the conditional independence assumption in dLLMs causes parallel decoding to ignore token dependencies, inevitably degrading generation quality when these dependencies are strong. However, existing works largely overlook these inherent challenges, and evaluations on standard benchmarks (e.g., math and coding) are not sufficient to capture the quality degradation caused by parallel decoding. To address this gap, we first provide an information-theoretic analysis of parallel decoding. We then conduct case studies on analytically tractable synthetic list operations from both data distribution and decoding strategy perspectives, offering quantitative insights that highlight the fundamental limitations of parallel decoding. Building on these insights, we propose **ParallelBench**, the first benchmark specifically designed for dLLMs, featuring realistic tasks that are trivial for humans and autoregressive LLMs yet exceptionally challenging for dLLMs under parallel decoding. Using ParallelBench, we systematically analyze both dLLMs and autoregressive LLMs, revealing that: (i) dLLMs under parallel decoding can suffer dramatic quality degradation in real-world scenarios, and (ii) current parallel decoding strategies struggle to adapt their degree of parallelism based on task difficulty, thus failing to achieve meaningful speedup without compromising quality. Our findings underscore the pressing need for innovative decoding methods that can overcome the current speed-quality trade-off. We are releasing our benchmark to help accelerate the development of truly efficient dLLMs.

Video generation models, as one form of world models, has emerged as one of the most exciting frontiers in AI, promising agents the ability to imagine the future by modeling the temporal evolution of complex scenes. In autonomous driving, this vision gives rise to driving world models—generative simulators that imagine ego and agent futures, enabling scalable simulation, safe testing of corner cases, and rich synthetic data generation. Yet, despite fast-growing research activity, the field lacks a rigorous benchmark to measure progress and guide priorities. Existing evaluations remain limited: generic video metrics overlook safety-critical imaging factors; trajectory plausibility is rarely quantified; temporal and agent-level consistency is neglected; and controllability with respect to ego conditioning is ignored. Moreover, current datasets fail to cover the diversity of conditions required for real-world deployment. To address these gaps, we present DrivingGen, the first comprehensive benchmark for generative driving world models. DrivingGen combines a diverse evaluation dataset—curated from both driving datasets and internet-scale video sources, spanning varied weather, time of day, geographic regions, and complex maneuvers—with a suite of new metrics that jointly assess visual realism, trajectory plausibility, temporal coherence, and controllability. Benchmarking 14 state-of-the-art models reveals clear trade-offs: general models look better but break physics, while driving-specific ones capture motion realistically but lag in visual quality. DrivingGen offers a unified evaluation framework to foster reliable, controllable, and deployable driving world models, enabling scalable simulation, planning, and data-driven decision-making.

Diffusion models have recently emerged as a powerful class of generative models and have achieved state-of-the-art performance in various image synthesis tasks. However, training diffusion models generally requires large amounts of data and suffer from overfitting when the dataset size is limited. To address these limitations, we propose a novel method called WILD-Diffusion, which is inspired by Wasserstein Distributionally Robust Optimization (WDRO), an important and elegant mathematical formulation from robust optimization area. Specifically, WILD-Diffusion utilizes WDRO to iteratively generate new training samples within a Wasserstein distance based uncertainty set centered at the limited data data distribution. This carefully designed method can progressively augment the training set throughout the training process and effectively overcome the obstacles caused by the limited data issue. Moreover, we establish the convergence guarantee for our algorithm even though the mixture of diffusion process and WDRO brings significant challenges to our analysis in theory. Finally, we conduct a set of experiments to verify the effectiveness of our proposed method. With WILD-Diffusion, we can achieve more than a $10$% reduction in FID using only $20$% of the training data across different datasets. Moreover, our method can attain state-of-the-art FID with as few as $100$ images, both in pretrained and non-pretrained settings.