ICLR 2026 Papers

Predicting individualized potential outcomes in sequential decision-making is central for optimizing therapeutic decisions in personalized medicine (e.g., which dosing sequence to give to a cancer patient). However, predicting potential out- comes over long horizons is notoriously difficult. Existing methods that break the curse of the horizon typically lack strong theoretical guarantees such as orthogonality and quasi-oracle efficiency. In this paper, we revisit the problem of predicting individualized potential outcomes in sequential decision-making (i.e., estimating Q-functions in Markov decision processes with observational data) through a causal inference lens. In particular, we develop a comprehensive theoretical foundation for meta-learners in this setting with a focus on beneficial theoretical properties. As a result, we yield a novel meta-learner called DRQ-learner and establish that it is: (1) doubly robust (i.e., valid inference under model misspecification), (2) Neyman-orthogonal (i.e., insensitive to first-order estimation errors in the nuisance functions), and (3) achieves quasi-oracle efficiency (i.e., behaves asymptotically as if the ground-truth nuisance functions were known). Our DRQ-learner is applicable to settings with both discrete and continuous state spaces. Further, our DRQ-learner is flexible and can be used together with arbitrary machine learning models (e.g., neural networks). We validate our theoretical results through numerical experiments, thereby showing that our meta-learner outperforms state-of-the-art baselines.

Physics-Informed Neural Networks (PINNs) have emerged as a powerful tool for solving PDEs, yet training is difficult due to a multi-objective loss that couples PDE residuals, initial/boundary conditions, and auxiliary physics terms. Existing remedies often yield infeasible scaling factors or conflicting update directions, resulting in degraded performance. In this paper, we show that training PINNs requires jointly considering feasible scaling factors and a non-conflict direction. Through a geometric analysis of per-loss gradients, we define the $\textit{harmonized cone}$ as the intersection of their primal and dual cones, which characterizes directions that are simultaneously feasible and non-conflicting. Building on this, we propose $HARMONIC$ (HARMONIzed Cone gradient descent), a training procedure that computes updates within the harmonized cone by leveraging the Double Description method to aggregate extreme rays. Theoretically, we establish convergence guarantees in nonconvex settings and prove the existence of a nontrivial harmonized cone. Across standard PDE benchmarks, $HARMONIC$ generally outperforms state-of-the-art methods while ensuring feasible and non-conflict updates.

Rotary Position Embeddings (RoPE) are widely adopted in LLMs, and it is commonly believed that larger base frequencies $\theta$ yield better long-context performance. In this paper, we show that a high-norm RoPE dimension, referred to as the “frequency band,” consistently emerges across multiple models, and we focus on this band to reveal the trade-offs inherent in RoPE. We find that replacing the RoPE dimensions below the frequency band with NoPE during inference has little effect on performance, indicating that these lower-frequency dimensions are only weakly utilized. We further find that the location of the frequency band depends on the RoPE base $\theta$ and the training sequence length. Moreover, the band forms early during pre-training and persists even after context extension via position interpolation. Notably, we show that aligning $\theta$ with the training length shifts the band toward lower frequencies and improves extrapolation, whereas increasing $\theta$ enhances interpolation but reduces extrapolation, revealing a clear trade-off between interpolation and extrapolation. We believe this work is a step toward a sharper understanding of positional embeddings in LLMs, with falsifiable diagnostics and practical guidance for choosing $\theta$ that support scaling to longer contexts.

Unified multimodal Large Language Models (MLLMs) hold great promise for seamlessly integrating understanding and generation. However, monolithic autoregressive architectures, despite their elegance and conversational fluency, suffer from a fundamental semantic–structural conflict: optimizing for low-level reconstructability in generation leads to catastrophic forgetting of high-level semantic understanding. We present ORION, a unified framework that resolves this conflict through Decoupling and Alignment. A non-linear vision head decouples structural pressures from shared representations, while a novel Representation Consistency Loss explicitly aligns semantics during generation. Together with a curated progressive training recipe and high-quality multimodal data, our method enables balanced optimization of both capabilities. Built purely on a monolithic autoregressive backbone without task-specific separate parameters, ORION achieves performance on par with or exceeding recent state-of-the-art unified models that rely on more complex designs. These results validate monolithic autoregression as a simple, effective, and competitive path toward truly integrated multimodal intelligence.

Post-training of large language models involves a fundamental trade-off between supervised fine-tuning (SFT), which efficiently mimics demonstrations but tends to memorize, and reinforcement learning (RL), which achieves better generaliza- tion at higher computational cost. Dynamic Fine-Tuning (DFT) recently emerged as a promising middle ground, reweighting SFT objectives with token probabili- ties and achieving improvements in certain reasoning domains, though it exhibits instability in other tasks. We provide a analysis of DFT through the reward- weighted regression (RWR) framework, revealing that it corresponds to a spe- cific auxiliary distribution choice that yields provably tighter RL bounds than standard SFT. However, our analysis also uncovers a critical limitation: this con- struction lacks distributional anchoring, leading to progressive drift that under- mines training stability. To address this, we propose Anchored Supervised Fine- Tuning (ASFT), which augments DFT’s reweighting with lightweight KL regu- larization to preserve tightness while ensuring stability. Empirically, ASFT con- sistently outperforms both SFT and DFT across mathematical reasoning, medical knowledge grounding, and code generation, achieving substantial improvements with minimal computational overhead. Our RWR framework provides a system- atic lens for understanding post-training methods and demonstrates that principled theoretical analysis leads to both stronger guarantees and practical gains.

In the past year, distillation has seen a renewed prominence in large language model (LLM) pretraining, exemplified by the Llama-3.2 and Gemma model families. While distillation has historically been shown to improve statistical modeling, its effects on new paradigms key to modern LLMs—such as test-time scaling and in-context learning—remain underexplored. In this work, we make three main contributions. First, we show that pretraining with distillation yields models that exhibit remarkably better test-time scaling. Second, we observe that this benefit comes with a trade-off: distillation impairs in-context learning capabilities, particularly the one modeled via induction heads. Third, to demystify these findings, we study distilled pretraining in a sandbox of a bigram model, which helps us isolate the common principal factor behind our observations. Finally, using these insights, we shed light on various design choices for pretraining that should help practitioners going forward.

Control-flow hijacking attacks manipulate orchestration mechanisms in multi-agent systems into performing unsafe actions that compromise the system and exfiltrate sensitive information. Recently proposed defenses, such as LlamaFirewall, rely on alignment checks of inter-agent communications to ensure that all agent invocations are "related to" and "likely to further" the original objective. We start by demonstrating control-flow hijacking attacks that evade these defenses even if alignment checks are performed by advanced LLMs. We argue that the safety and functionality objectives of multi-agent systems fundamentally conflict with each other. This conflict is exacerbated by the brittle definitions of "alignment" and the checkers' incomplete visibility into the execution context. We then propose, implement, and evaluate ControlValve, a new defense based on the principles of control-flow integrity and least privilege. ControlValve (1) generates permitted control-flow graphs for multi-agent systems, and (2) enforces that all executions comply with these graphs, along with contextual rules (generated in a zero-shot manner) for each agent invocation.

We introduce M3-Agent, a novel multimodal agent framework equipped with long-term memory. Like humans, M3-Agent can process real-time visual and auditory inputs to build and update episodic and semantic memories, gradually accumulating world knowledge. Its memory is organized in an entity-centric, multimodal manner, enabling deeper and more consistent understanding of the environment. Given an instruction, M3-Agent autonomously performs multi-turn reasoning and retrieves relevant memories to complete tasks. To evaluate memory effectiveness and memory-based reasoning in multimodal agents, we develop M3-Bench, a long-video question answering benchmark comprising 100 newly recorded robot-perspective videos (M3-Bench-robot) and 920 diverse web-sourced videos (M3-Bench-web). We annotate QA pairs designed to test capabilities essential for agent applications, such as person understanding, general knowledge extraction, and cross-modal reasoning. Experimental results show that M3-Agent, trained via reinforcement learning, outperforms the strongest baseline, a prompting agent using Gemini-1.5-pro and GPT-4o, achieving 6.7%, 7.7%, and 5.3% higher accuracy on M3-Bench-robot, M3-Bench-web and VideoMME-long, respectively. Our work advances multimodal agents toward more human-like long-term memory and provides insights for their practical design. Models, datasets and code are available at https://github.com/ByteDance-Seed/m3-agent.

Emotion recognition in speech presents a complex multimodal challenge, requiring comprehension of both linguistic content and vocal expressivity, particularly prosodic features such as fundamental frequency, intensity, and temporal dynamics. Although large language models (LLMs) have shown promise in reasoning over textual transcriptions for emotion recognition, they typically neglect fine-grained prosodic information, limiting their effectiveness and interpretability. In this work, we propose VowelPrompt, a linguistically grounded framework that augments LLM-based emotion recognition with interpretable, fine-grained vowel-level prosodic cues. Drawing on phonetic evidence that vowels serve as primary carriers of affective prosody, VowelPrompt extracts pitch-, energy-, and duration-based descriptors from time-aligned vowel segments, and converts these features into natural language descriptions for better interpretability. Such a design enables LLMs to jointly reason over lexical semantics and fine-grained prosodic variation. Moreover, we adopt a two-stage adaptation procedure comprising supervised fine-tuning (SFT) followed by Reinforcement Learning with Verifiable Reward (RLVR), implemented via Group Relative Policy Optimization (GRPO), to enhance reasoning capability, enforce structured output adherence, and improve generalization across domains and speaker variations. Extensive evaluations across diverse benchmark datasets demonstrate that VowelPrompt consistently outperforms state-of-the-art emotion recognition methods under zero-shot, fine-tuned, cross-domain, and cross-linguistic conditions, while enabling the generation of interpretable explanations that are jointly grounded in contextual semantics and fine-grained prosodic structure.

Models are susceptible to adversarially out-of-distribution (OOD) data despite large training-compute investments into their robustification. Zaremba et al. (2025) make progress on this problem at test time, showing LLM reasoning improves satisfaction of model specifications designed to thwart attacks, resulting in a correlation between reasoning effort and robustness to jailbreaks. However, this benefit of test compute fades when attackers are given access to gradients or multimodal inputs. We address this gap, clarifying that inference-compute offers benefits even in such cases. Our approach argues that compositional generalization, through which OOD data is understandable via its in-distribution (ID) components, enables adherence to defensive specifications on adversarially OOD inputs. Namely, we posit the Robustness from Inference Compute Hypothesis (RICH): inference-compute defenses profit as the model's training data better reflects the attacked data’s components. We empirically support this hypothesis across vision language model and attack types, finding robustness gains from test-time compute if specification following on OOD data is unlocked by compositional generalization. For example, InternVL 3.5 gpt-oss 20B gains little robustness when its test compute is scaled, but such scaling adds significant robustness if we first robustify its vision encoder. This correlation of inference-compute's robustness benefit with base model robustness is the rich-get-richer dynamic of the RICH: attacked data components are more ID for robustified models, aiding compositional generalization to OOD data. Thus, we advise layering train-time and test-time defenses to obtain their synergistic benefit.

Large language model (LLM) simulations of human behavior have the potential to revolutionize the social and behavioral sciences, if and only if they faithfully reflect real human behaviors. Current evaluations are fragmented, based on bespoke tasks and metrics, creating a patchwork of incomparable results. To address this, we introduce SimBench, the first large-scale, standardized benchmark for a robust, reproducible science of LLM simulation. By unifying 20 diverse datasets covering tasks from moral decision-making to economic choice across a large global participant pool, SimBench provides the necessary foundation to ask fundamental questions about when, how, and why LLM simulations succeed or fail. We show that, while even the best LLMs today have limited simulation ability (score: 40.80/100), performance scales log-linearly with model size. Simulation performance is not improved by increased inference-time compute. We discover an alignment-simulation trade-off: instruction-tuning improves performance on low-entropy (consensus) questions but degrades it on high-entropy (diverse) ones. Models particularly struggle when simulating specific demographic groups. Finally, we demonstrate that simulation ability correlates most strongly with deep, knowledge-intensive reasoning (MMLU-Pro, r=0.939). By making progress measurable, we aim to accelerate the development of more faithful LLM simulators.

Diffusion policies for visuomotor robot manipulation tasks achieve remarkable dexterity and robustness while only training on a small number of task demonstrations. However, the reason for this performance remains a mystery. In this paper, we offer a surprising hypothesis: diffusion policies essentially memorize an action lookup table---\emph{and this is beneficial}. We posit that, at runtime, diffusion policies find the closest training image to the test image in a latent space, and recall the associated training action (i.e. action chunk), offering reactivity without the need for action generalization. This is effective in the sparse data regime, where there is not enough data density for the model to learn action generalization. We support this claim with systematic empirical evidence, showing that even when conditioned on highly out of distribution (OOD) images, Diffusion Policy still outputs an action chunk from the training data. We evaluate and compare three representative policy families on the same data set: Diffusion Policy, Action Chunking with Transformers (ACT), and GR00T, a pre-trained generalist Vision-Language-Action (VLA) model. We show that Diffusion Policy gives strong action memorization giving surprising robustness in OOD regimes, ACT shows action interpolation with poor robustness in OOD regimes, and GR00T (benefiting from substantial pre-training) shows both action interpolation and OOD robustness. As a simple alternative to Diffusion Policy, we introduce the Action Lookup Table (ALT) policy, showing that an explicit lookup table policy can perform comparably in this low data regime. Despite its simplicity, ALT attains Diffusion Policy–level performance while also providing faster inference and explicit OOD detection via latent-distance thresholds. These results reframe diffusion policies for robot manipulation as reactive memory retrieval under data sparsity, and provide practical tools for interpreting, evaluating, and monitoring such policies. More information can be found at: \url{https://stanfordmsl.github.io/alt/}.

Recent work has framed decision-making as a sequence modeling problem using generative models such as diffusion models. Although promising, these approaches often overlook latent factors that exhibit evolving dynamics, elements that are fundamental to environment transitions, reward structures, and high-level agent behavior. Explicitly modeling these hidden processes is essential for both precise dynamics modeling and effective decision-making. In this paper, we propose a unified framework that explicitly incorporates latent dynamic inference into generative decision-making from minimal yet sufficient observations. We theoretically show that under mild conditions, the latent process can be identified from small temporal blocks of observations. Building on this insight, we introduce Ada-Diffuser, a causal diffusion model that learns the temporal structure of observed interactions and the underlying latent dynamics simultaneously, and furthermore, leverages them for planning and control. With a proper modular design, Ada-Diffuser supports both planning and policy learning tasks, enabling adaptation to latent variations in dynamics, rewards, and even recovering hidden action variables from action-free demonstrations. Extensive experiments on locomotion and robotic manipulation benchmarks demonstrate the model’s effectiveness in accurate latent inference, long-horizon planning, and adaptive policy learning.

Diffusion language models have recently emerged as a competitive alternative to autoregressive language models. Beyond next-token generation, they are more efficient and flexible by enabling parallel and any-order token generation. However, despite empirical successes, their computational power and fundamental limitations remain poorly understood. In this paper, we formally study whether non-autoregressive generation in Masked Diffusion Models (MDM) enables solving problems beyond the reach of Auto-Regressive Models (ARM). Our results show that MDM with sufficiently large context length is computationally universal with decoding steps matching the optimal parallel time complexity in PRAM. However, when controlling for other factors, MDM's flexibility to generate in any-order does not expand what ARM can already solve. To address this, we propose a new form of generation called any-process generation, which extends MDM with capabilities to remask, insert and delete tokens, allowing self-correction, length-variable editing, and adaptive parallelism. Theoretically and empirically, we demonstrate these capabilities enable scalability to significantly harder reasoning problems that are otherwise intractable for ARM and vanilla MDM. Additionally, they prove essential for generation tasks where objects naturally evolve through non-sequential processes, crucial for extending current LLMs beyond natural language to domains such as coding and science.

Reinforcement learning algorithms are typically designed for discrete-time dynamics, even though the underlying real-world control systems are often continuous in time. In this paper, we study the problem of continuous-time reinforcement learning, where the unknown system dynamics are represented using nonlinear ordinary differential equations (ODEs). We leverage probabilistic models, such as Gaussian processes and Bayesian neural networks, to learn an uncertainty-aware model of the underlying ODE. Our algorithm, COMBRL, greedily maximizes a weighted sum of the extrinsic reward and model epistemic uncertainty. This yields a scalable and sample-efficient approach to continuous-time model-based RL. We show that COMBRL achieves sublinear regret in the reward-driven setting, and in the unsupervised RL setting (i.e., without extrinsic rewards), we provide a sample complexity bound. In our experiments, we evaluate COMBRL in both standard and unsupervised RL settings and demonstrate that it scales better, is more sample-efficient than prior methods, and outperforms baselines across several deep RL tasks.

Precipitation forecasting relies on heterogeneous data sets. Weather radar is accurate, but coverage is geographically limited and costly to maintain. Weather stations provide accurate but sparse point measurements, while satellites offer dense, high-resolution coverage without direct rainfall retrieval. To overcome these limitations, we propose Query-Conditioned Gaussian Splatting (QCGS), the first framework to fuse automatic weather station (AWS) observations with satellite imagery for generating radar-like rainfall fields. Unlike conventional 2D Gaussian splatting, which renders the entire image plane, QCGS selectively renders only queried rainfall regions, avoiding unnecessary computation in non-precipitating areas while preserving sharp precipitation structures. The framework combines a radar point proposal network that identifies rainfall-support locations with an implicit neural representation (INR) network that predicts Gaussian parameters for each point. QCGS enables efficient, resolution-flexible rainfall field generation in real time. Through extensive evaluation with benchmark precipitation products, QCGS demonstrates over 50\% improvement in RMSE compared to conventional gridded rainfall products, and consistently maintains high performance across multiple spatiotemporal scales.

We revisit FedExProx -- a recently proposed distributed optimization method designed to enhance convergence properties of parallel proximal algorithms via extrapolation. In the process, we uncover a surprising flaw: its known theoretical guarantees on quadratic optimization tasks are no better than those offered by the vanilla Gradient Descent (GD) method. Motivated by this observation, we develop a novel analysis framework, establishing a tighter linear convergence rate for non-strongly convex quadratic problems. By incorporating both computation and communication costs, we demonstrate that FedExProx can indeed provably outperform GD, in stark contrast to the original analysis. Furthermore, we consider partial participation scenarios and analyze two adaptive extrapolation strategies -- based on gradient diversity and Polyak stepsizes -- again significantly outperforming previous results. Moving beyond quadratics, we extend the applicability of our analysis to general functions satisfying the Polyak-Łojasiewicz condition, outperforming the previous strongly convex analysis while operating under weaker assumptions. Backed by empirical results, our findings point to a new and stronger potential of FedExProx, paving the way for further exploration of the benefits of extrapolation in federated learning.

The reliance on implicit point matching via attention has become a core bottleneck in drag-based editing, resulting in a fundamental compromise on weakened inversion strength and costly test-time optimization (TTO). This compromise severely limits the generative capabilities, suppressing high-fidelity inpainting and text-guided creation. In this paper, we introduce LazyDrag, the first drag-based image editing method for Multi-Modal Diffusion Transformers, which directly eliminates the reliance on implicit point matching. In concrete terms, our method generates an explicit correspondence map from user drag inputs as a reliable reference to boost the attention control. This reliable reference opens the potential for a stable full-strength inversion process, which is the first in the drag-based editing task. It obviates the necessity for TTO and unlocks the generative capability of models. Therefore, LazyDrag naturally unifies precise geometric control with text guidance, enabling complex edits that were previously out of reach: opening the mouth of a dog and inpainting its interior, generating new objects like a ``tennis ball'', or for ambiguous drags, making context-aware changes like moving hands into pockets. Moreover, LazyDrag supports multi-round edits with simultaneous move and scale operations. Evaluated on DragBench, our method outperforms baselines in drag accuracy and perceptual quality, as validated by mean distances, VIEScore and user studies. LazyDrag not only sets new state-of-the-art performance, but also paves a new way to editing paradigms. Here is the project website.

Large language and vision-language models have inspired end-to-end vision-language-action (VLA) systems in robotics, yet existing robot datasets remain costly, embodiment-specific, and insufficient, limiting robustness and generalization. Recent approaches address this by adopting a plan-then-execute paradigm, where high-level plans are generated before translating into low-level actions, but their success depends on fine-grained intermediate supervision that current datasets lack. To fill this gap, we present the RoboInter Manipulation Suite, a unified resource for data, benchmarking, and modeling of intermediate representations. It includes RoboInter-Tool, a lightweight GUI for semi-automatic per-frame annotation of embodied videos, and RoboInter-Data, a human-verified dataset with over 200k episodes across 571 diverse scenes, offering dense per-frame alignment across more than nine intermediate categories and surpassing prior work in both scale and quality. Building on this foundation, RoboInter-VQA introduces 8 spatial and 20 temporal embodied QA categories to benchmark and enhance the embodied capabilities of current large vision-language models, while RoboInter-VLA provides a flexible plan-then-execute framework with modular and end-to-end variants that link planning to execution. Together, these contributions establish RoboInter Manipulation Suite as a foundation for advancing generalizable and robust robotic learning through fine-grained intermediate supervision.