ICLR 2026 Papers

Multi-Agent System (MAS) powered by Visual Language Models (VLMs) enables challenging tasks but suffers from a novel failure term, multi-agent visual hallucination snowballing, where hallucinations are seeded in a single agent and amplified by following ones due to the over-reliance on textual flow to relay visual information. Through turn-, layer-, and token-wise attention analyses, we provide detailed insights into the essence of hallucination snowballing regarding the reduction of visual attention allocation. It leads us to identify a subset of vision tokens with a unimodal attention peak in middle layers that best preserve visual evidence but gradually diminish in deeper agent turns, resulting in the visual hallucination snowballing in MAS. Thus, we propose ViF, a lightweight, plug-and-play mitigation paradigm that relays inter-agent messages with Visual Flow powered by the selected visual relay tokens and applies attention reallocation to amplify this pattern. The experiment results demonstrate that our method markedly reduces hallucination snowballing, consistently improving the performance across eight benchmarks based on four common MAS structures and ten base models. The implementation source code will be made publicly available.

This paper introduces Diffusion via Autoregressive (D-AR) models, a new paradigm recasting the pixel diffusion process as a vanilla autoregressive procedure in the standard next-token-prediction fashion. We start by designing the tokenizer that converts an image into the sequence of discrete tokens, where tokens in different positions can be decoded into different diffusion denoising steps in the pixel space. Thanks to the diffusion property, these tokens naturally follow a coarse-to-fine order, which directly lends itself to autoregressive modeling. Then, we apply standard next-token prediction to these tokens, without modifying any underlying designs (either causal masks or training/inference strategies), and such sequential autoregressive token generation directly mirrors the diffusion procedure in image space. That is, once the autoregressive model generates an increment of tokens, we can directly decode these tokens into the corresponding diffusion denoising step on pixels in a streaming manner. Our pipeline naturally reveals several intriguing properties, for example, it supports consistent previews when generating only a subset of tokens and enables zero-shot layout-controlled synthesis. On the standard ImageNet benchmark, our method achieves 2.09 and 2.00 FID using a 775M and 1.4B Llama backbone with 256 discrete tokens. We hope our work can inspire future research on unified autoregressive architectures of visual synthesis, especially with large language models.

Recursive transformers reuse parameters and iterate over hidden states multiple times, decoupling compute depth from parameter depth. However, under matched compute, recursive models with fewer parameters often lag behind non-recursive counterparts. By probing hidden states, we trace this performance gap to two primary bottlenecks: __undifferentiated computation__, where the core is forced to adopt a similar computational pattern at every iteration, and __information overload__, where long-lived and transient information must coexist in a single hidden state. To address the issues, we introduce a **Me**mory-as-**S**tate-**H**ighways **(MeSH)** scheme, which externalizes state management into an explicit memory buffer and employs lightweight routers to dynamically diversify computation across iterations. Probing visualizations confirm that MeSH successfully resolves the pathologies by inducing functional specialization across iterations. On the Pythia suite (160M–6.9B), MeSH-enhanced recursive transformers consistently improve over recursive baselines and outperforms its larger non-recursive counterpart at the 1.4B scale, improving average downstream accuracy by +1.06\% with 33\% fewer non-embedding parameters. Our analysis establishes MeSH as a scalable and principled architecture for building stronger recursive models.

Intelligent embodied agents not only need to accomplish preset tasks, but also learn to align with individual human needs and preferences. Extracting reward signals from human language preferences allows an embodied agent to adapt through reinforcement learning. However, human language preferences are unconstrained, diverse, and dynamic, making constructing learnable reward from them a major challenge. We present ROSETTA, a framework that uses foundation models to ground and disambiguate unconstrained natural language preference, construct multi-stage reward functions, and implement them with code generation. Unlike prior works requiring extensive offline training to get general reward models or fine-grained correction on a single task, ROSETTA allows agents to adapt online to preference that evolves and is diverse in language and content. We test ROSETTA on both short-horizon and long-horizon manipulation tasks and conduct extensive human evaluation, finding that ROSETTA outperforms SOTA baselines and achieves 87% average success rate and 86% human satisfaction across 116 preferences.

Flow and diffusion models are typically pre-trained on limited available data (e.g., molecular samples), covering only a fraction of the valid design space (e.g., the full molecular space). As a consequence, they tend to generate samples from only a narrow portion of the feasible domain. This is a fundamental limitation for scientific discovery applications, where one typically aims to sample valid designs beyond the available data distribution. To this end, we address the challenge of leveraging access to a verifier (e.g., an atomic bonds checker), to adapt a pre-trained flow model so that its induced density expands beyond regions of high data availability, while preserving samples validity. We introduce formal notions of strong and weak verifiers and propose algorithmic frameworks for global and local flow expansion via probability-space optimization. Then, we present Flow Expander (FE), a scalable mirror descent scheme that provably tackles both problems by verifier-constrained entropy maximization over the flow process noised state space. Next, we provide a thorough theoretical analysis of the proposed method, and state convergence guarantees under both idealized and general assumptions. Ultimately, we empirically evaluate our method on both illustrative, yet visually interpretable settings, and on a molecular design task showcasing the ability of FE to expand a pre-trained flow model increasing conformer diversity while preserving validity.

Language models can transfer hidden biases during distillation. For example, a teacher that "likes owls" can make its student "like owls" too, even when the training data consists only of lists of numbers. This surprising phenomenon is called *subliminal learning*. Subliminal learning can be expected under soft distillation, where the student is trained on the teacher's full next-token distribution. But the fact that this also occurs under hard distillation—where the student only sees sampled tokens—raises a deeper question: *when and how does subliminal learning actually occur?* We answer this question through controlled experiments and mechanistic analysis. Our results show that subliminal learning does not need (global) token entanglement or logit leakage. Instead, it comes down to a small set of *divergence tokens*—rare cases where teachers with different biases would predict different tokens. Masking out these tokens mostly removes the hidden bias transfer. Mechanistically, divergence tokens reveal that early layers are critical. Surprisingly, finetuning even a single such early layer is sufficient for subliminal learning. Finally, we find that subliminal learning is fragile. Even small changes, like paraphrasing prompts, are usually sufficient to suppress it.

Reinforcement learning finetuning (RFT) has significantly advanced the reasoning capabilities of large language models (LLMs) by enabling long chains of thought, multi-turn self-correction, and effective tool use. While recent works attempt to extend RFT to vision-language models (VLMs), these efforts largely produce text-only reasoning conditioned on original image inputs, falling short of incorporating visual reasoning in the response. In contrast, test-time methods like Visual Sketchpad incorporate visual steps but lack training mechanisms. We introduce VTool-R1, the first RFT framework that trains VLMs to generate multimodal chains of thought by interleaving text and intermediate visual reasoning steps. VTool-R1 integrates Python-based visual editing tools into the RFT process, enabling VLMs to learn when and how to generate visual reasoning steps that benefit final reasoning. Trained with outcome-based rewards, our approach elicits strategic visual tool use for multi-modal reasoning without relying on process-based supervision. Extensive experiments on structured visual reasoning over charts and tables show that VTool-R1 enhances reasoning performance by teaching VLMs to "think with images" and generate multimodal chain of thoughts with tools. To support future research in multi-turn multi-modal reasoning, we will release our full training and evaluation code as open source.

A persistent paradox in Continual Learning is that neural networks often retain linearly separable representations of past tasks even when their output predictions fail. We formalize this distinction as the gap between deep (feature-space) and shallow (classifier-level) forgetting. We demonstrate that experience replay affects these two levels asymmetrically: while even minimal buffers anchor feature geometry and prevent deep forgetting, mitigating shallow forgetting requires substantially larger buffers. To explain this, we extend the Neural Collapse framework to sequential training. We theoretically model deep forgetting as a geometric drift toward out-of-distribution subspaces, proving that replay guarantees asymptotic separability. In contrast, we show that shallow forgetting stems from an under-determined classifier optimization: the strong collapse of buffer data leads to rank-deficient covariances and inflated means, blinding the classifier to the true population boundaries. Our work unifies continual learning with OOD detection and challenges the reliance on large buffers, suggesting that explicitly correcting the statistical artifacts of Neural Collapse could unlock robust performance with minimal replay.

Multi-Agent System (MAS) and Reinforcement Learning (RL) are both widely adopted to improve large language model (LLM) agentic performance. MAS strengthens task-specialized performance via role-based orchestration; RL leverages environment rewards to train stronger policies, such as Group Relative Policy Optimization (GRPO)-style optimization. Yet applying on-policy RL training to MAS is underexplored. While promising, it poses several challenges. On the algorithm side, Standard GRPO grouping assumptions fail in MAS because prompts differ by role and turn. On the system side, the training system needs to support MAS-workflow-based rollouts and on-policy updates for both single and multiple policy models. To address these issues, we introduce AT-GRPO, consisting of (i) an Agent- and Turn-wise grouped RL algorithm tailored for MAS and (ii) a system to support both single-policy and multi-policy training. Across game, plan, coding, and math tasks, AT-GRPO demonstrates substantial performance gains across diverse domains. Especially on long-horizon planning tasks, AT-GRPO boosts accuracy from a 14.0–47.0% single-agent RL baseline to 96.0–99.5%. Furthermore, it improves reasoning performance, with an average gain of 3.87–7.62% on coding and 9.0-17.93% on math. The code are available at https://github.com/pettingllms-ai/PettingLLMs.

Diffusion Large Language Models (DLLMs) are emerging as a powerful alternative to the dominant Autoregressive Large Language Models, offering efficient parallel generation and capable global context modeling. However, the practical application of DLLMs is hindered by a critical architectural constraint: the need for a statically predefined generation length. This static length allocation leads to a problematic trade-off: insufficient lengths cripple performance on complex tasks, while excessive lengths incur significant computational overhead and sometimes result in performance degradation. While the inference framework is rigid, we observe that the model itself possesses internal signals that correlate with the optimal response length for a given task. To bridge this gap, we leverage these latent signals and introduce DAEDAL, a novel training-free denoising strategy that enables Dynamic Adaptive Length Expansion for Diffusion Large Language Models. DAEDAL operates in two phases: 1) Before the denoising process, DAEDAL starts from a short initial length and iteratively expands it to a coarse task-appropriate length, guided by a sequence completion metric. 2) During the denoising process, DAEDAL dynamically intervenes by pinpointing and expanding insufficient generation regions through mask token insertion, ensuring the final output is fully developed. Extensive experiments on DLLMs demonstrate that DAEDAL achieves performance comparable, and in some cases superior, to meticulously tuned fixed-length baselines, while simultaneously enhancing computational efficiency by achieving a higher effective token ratio. By resolving the static length constraint, DAEDAL unlocks new potential for DLLMs, bridging a critical gap with their Autoregressive counterparts and paving the way for more efficient and capable generation.

While deep learning has achieved impressive performance in time series forecasting, it becomes increasingly crucial to understand its decision-making process for building trust in high-stakes scenarios. Existing interpretable models often provide only local and partial explanations, lacking the capability to reveal how heterogeneous and interacting input variables jointly shape the overall temporal patterns in the forecast curve. We propose ProtoTS, a novel interpretable forecasting framework that achieves both high accuracy and transparent decision-making through modeling prototypical temporal patterns. ProtoTS computes instance-prototype similarity based on a denoised representation that preserves abundant heterogeneous information. The prototypes are organized hierarchically to capture global temporal patterns with coarse prototypes while capturing finer-grained local variations with detailed prototypes, enabling expert steering and multi-level interpretability. Experiments on multiple realistic benchmarks, including a newly released LOF dataset, show that ProtoTS not only exceeds existing methods in forecast accuracy but also delivers expert-steerable interpretations for better model understanding and decision support. The source code is available at https://anonymous.4open.science/r/ProtoTS-D791/.

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for \textit{dishonesty} as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are crafted to be subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool \emph{all} output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a \emph{honeypot} against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using them as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

Scaling unsupervised skill discovery algorithms to high-DoF agents remains challenging. As dimensionality increases, the exploration space grows exponentially, while the manifold of meaningful skills remains limited. Therefore, semantic meaningfulness becomes essential to effectively guide exploration in high-dimensional spaces. In this work, we present **Reference-Grounded Skill Discovery (RGSD)**, a novel algorithm that grounds skill discovery in a semantically meaningful latent space using reference data. RGSD first performs contrastive pretraining to embed motions on a unit hypersphere, clustering each reference trajectory into a distinct direction. This grounding enables skill discovery to simultaneously involve both imitation of reference behaviors and the discovery of semantically related diverse behaviors. On a simulated SMPL humanoid with $359$-D observations and $69$-D actions, RGSD successfully imitates skills such as walking, running, punching, and sidestepping, while also discover variations of these behaviors. In downstream locomotion tasks, RGSD leverages the discovered skills to faithfully satisfy user-specified style commands and outperforms imitation-learning baselines, which often fail to maintain the commanded style. Overall, our results suggest that lightweight reference-grounding offers a practical path to discovering semantically rich and structured skills in high-DoF systems.

In this work, we investigate the synergy between supervised fine-tuning (SFT) and reinforcement learning (RL) in developing strong reasoning models. We begin by curating the SFT training data through two scaling strategies: increasing the number of collected prompts and the number of generated responses per prompt. Both approaches yield notable improvements in reasoning performance, with scaling the number of prompts resulting in more substantial gains. We then explore the following questions regarding the synergy between SFT and RL: (i) Does a stronger SFT model consistently lead to better final performance after large-scale RL training? (ii) How can we determine an appropriate sampling temperature during RL training to effectively balance exploration and exploitation for a given SFT initialization? Our findings suggest that (i) holds true, provided effective RL training is conducted, particularly when the sampling temperature is carefully chosen to maintain the temperature-adjusted entropy around 0.3, a setting that strikes a good balance between exploration and exploitation. Notably, the performance gap between initial SFT models narrows significantly throughout the RL process. Built on a strong SFT foundation and SFT–RL synergy, our AceReason-Nemotron-1.1 7B model significantly outperforms AceReason-Nemotron-1.0 and achieves new state-of-the-art performance among Qwen2.5-7B-based reasoning models on challenging math and code benchmarks, thereby demonstrating the effectiveness of our post-training recipe.

The Schrödinger Bridge provides a principled framework for modeling stochastic processes between distributions; however, existing methods are limited by energy-conservation assumptions, which constrains the bridge's shape preventing it from model varying-energy phenomena. To overcome this, we introduce the non-conservative generalized Schrödinger bridge (NCGSB), a novel, energy-varying reformulation based on contact Hamiltonian mechanics. By allowing energy to change over time, the NCGSB provides a broader class of real-world stochastic processes, capturing richer and more faithful intermediate dynamics. By parameterizing the Wasserstein manifold, we lift the bridge problem to a tractable geodesic computation in a finite-dimensional space. Unlike computationally expensive iterative solutions, our contact Wasserstein geodesic (CWG) is naturally implemented via a ResNet architecture and relies on a non-iterative solver with near-linear complexity. Furthermore, CWG supports guided generation by modulating a task-specific distance metric. We validate our framework on tasks including manifold navigation, molecular dynamics predictions, and image generation, demonstrating its practical benefits and versatility.

Scene Graph Generation (SGG) is pivotal for structured visual understanding, yet it remains hindered by a fundamental limitation: the reliance on fixed, frozen semantic representations from pre-trained language models. These semantic priors, while beneficial in other domains, are inherently misaligned with the dynamic, context-sensitive nature of visual relationships, leading to biased and suboptimal performance. In this paper, we transcend the traditional one-stage v.s. two-stage architectural debate and identify this representational bottleneck as the core issue. We introduce Adaptive Prompt Tuning (APT), a universal paradigm that converts frozen semantic features into dynamic, context-aware representations through lightweight, learnable prompts. APT acts as a plug-in module that can be seamlessly integrated into existing SGG frameworks. Extensive experiments demonstrate that APT achieves +2.7 improvement in mR@100 on PredCls, +3.6 gain in F@100 and up to +6.0 gain in mR@50 in open-vocabulary novel splits. Notably, it achieves this with less than 0.5M additonal parameters (<1.5\% overhead) and reduced 7.8\%-25\% training time, establishing a new state-of-the-art while offering a unified, efficient, and scalable solution for future SGG research. The source code of APT is available at <https://anonymous.4open.science/r/APT-1D24>.

Graph generation plays an important role in various domains such as molecular design, protein prediction, and drug discovery. However, generating graph-structured data poses challenges due to the complex dependencies inherent in graphs, spanning from intricate local substructures to broad global topologies. Although recent advances in graph-generative models have made notable progress, most existing methods still leverage the node-level generative paradigms and struggle with graphs that exhibit pronounced sparsity and complicated multiscale relationships. To address these challenges, we propose a unified latent diffusion model that jointly learns local and global topological information, enabling effective and efficient graph generation. Besides, our approach introduces a dual conditioning mechanism designed to promote dynamic interaction between local and global information, equipping the generative model with global and local awareness to better capture the coupled dependencies within graphs. Our method can largely promote the joint modeling of global and local information and substantially improve the quality of the generated graphs. Extensive experiments consistently demonstrate the effectiveness of our proposed method.

Realistic and interactive traffic simulation is essential for training and evaluating autonomous driving systems. However, most existing data-driven simulation methods rely on static initialization or log-replay data, limiting their ability to model dynamic, long-horizon scenarios with evolving agent populations. We propose InfGen, a scenario generation framework that outputs agent states and trajectories in an autoregressive manner. InfGen represents the entire scene as a sequence of tokens—including traffic light signals, agent states, and motion vectors—and uses a transformer model to simulate traffic over time. This design enables InfGen to continuously insert new agents into traffic, supporting infinite scene generation. Experiments demonstrate that InfGen produces realistic, diverse, and adaptive traffic behaviors. Furthermore, reinforcement learning policies trained in InfGen-generated scenarios achieve superior robustness and generalization, validating its utility as a high-fidelity simulation environment for autonomous driving. Code and models will be made publicly available.

VLM-based mobile agents are increasingly popular due to their capabilities to interact with smartphone GUIs and XML-structured texts and to complete daily tasks. However, existing online benchmarks fail to obtain stable critical reward signals under dynamic environmental changes, and neglect the influence of noise components and interactive instructions. Offline benchmarks evaluate the agents through single-path trajectories, which stand in contrast to the inherently multi-solution characteristics of GUI tasks. To address these limitations, we introduce SMAN-Bench, a benchmark designed to evaluate agents under Single-path, Multi-path, Ambiguous, and Noisy task settings. We employ a slot-based instruction generation method to match templates with GUI trajectories from an existing, graph-structured, unlabeled mobile corpus. SMAN-Bench includes a common task split, with offline multi-path evaluation to assess the agent’s ability to obtain step rewards during task execution. It contains a noisy split based on pop-ups and ad apps, and a contaminated split named AITZ-Noise to simulate a realistic noisy environment. Furthermore, an ambiguous instruction split with preset Q&A interactions is released to evaluate the agent’s proactive interaction capabilities. Our evaluation covers mobile agent frameworks like AppAgent-v1, Mobile-Agent-v2, and Mobile-Agent-E, and includes both open-source and closed-source mobile fundamental models, as well as several multimodal reasoning models.