ICLR 2026 Papers

While Multimodal Large Language Models (MLLMs) excel at holistic understanding, they struggle with the dense world, i.e., complex scenes requiring fine-grained analysis of intricate details and object inter-relationships. Region-level MLLMs have been a promising step. However, previous attempts are generally optimized to understand given regions in isolation, neglecting crucial global contexts. To address this, we introduce Grasp Any Region (GAR) for comprehensive region-level visual understanding. Empowered by an effective RoI-aligned feature replay technique, GAR supports (1) precise perception by leveraging necessary global contexts, and (2) modeling interactions between multiple prompts. Together, it then naturally achieves (3) advanced compositional reasoning to answer specific free-form questions about any region, shifting the paradigm from passive description to active dialogue. Moreover, we construct GARBench, which not only provides a more accurate evaluation of single-region comprehension, but also, more importantly, measures interactions and complex reasoning across multiple regions. Empirically, GAR-1B not only maintains the state-of-the-art captioning capabilities, e.g., outperforming DAM-3B +4.5 on DLC-Bench, but also excels at modeling relationships between multiple prompts with advanced comprehension capabilities, even surpassing InternVL3-78B on GARBench-VQA. More importantly, our zero-shot GAR-8B even outperforms in-domain VideoRefer-7B on VideoRefer-BenchQ, indicating its strong comprehension capabilities can be easily transferred to videos. Code and data will be released to the community.

Recurrent large language models (Recurrent LLMs) offer linear computational complexity as efficient alternatives to quadratic self-attention-based LLMs (Self-Attention LLMs). However, Recurrent LLMs underperform on long-context tasks due to limited fixed-size memory. Previous research focused on architectural innovations to enhance memory capacity, but failed to match Self-Attention LLM performance. We argue this limitation stems from processing entire contexts at once being ill-suited for Recurrent LLMs. We propose Smooth Reading, a co-design of recurrent architecture and inference method. It introduces a end-to-end multi-round inference method that processes context incrementally and iteratively summarizes information, reducing memory demands. Methodologically, we reveal architecture-inference interactions play an important role for performance, efficiency and scalability, shedding light on future Recurrent LLM design. Besides, our method substantially bridges the performance gap between Recurrent and Self-Attention LLMs on long-context tasks while preserving efficiency advantages. Smooth Reading boosts SWA-3B-4k from 5.68% lower to 3.61% higher performance than Self-Attention LLMs on LongBench, while maintaining 2.5× faster training and 2× faster inference at 64k context.

Recent advances in few-step generative models (typically $1$-$8$ steps), such as consistency models, have yielded impressive performance. However, their broader adoption is hindered by significant challenges, including substantial computational overhead, the reliance on complex multi-component loss functions, and intricate multi-stage training strategies that lack end-to-end simplicity. These limitations impede their scalability and stability, especially when applied to large-scale models. To address these issues, we introduce **$N$-th order Recursive Consistent velocity field estimation for Generative Modeling (RCGM)**, a novel framework that unifies many existing approaches. Within this framework, we reveal that conventional one-step methods, such as consistency and MeanFlow models, are special cases of 1st-order RCGM. This insight enables a natural extension to higher-order scenarios ($N \geq 2$), which exhibit markedly improved training stability and achieve state-of-the-art (SOTA) performance. For instance, on ImageNet $256\times256$, RCGM enables a $675\text{M}$ parameter diffusion transformer to achieve a $1.48$ FID score in just $2$ sampling steps. Crucially, RCGM facilitates the stable full-parameter training of a large-scale ($20\textrm{B}$) unified multi-modal model, attaining a $0.86$ GenEval score in $2$ steps. In contrast, conventional 1st-order approaches, such as consistency and MeanFlow models, typically suffer from training instability, model collapse, or memory constraints under comparable settings. *Code will be publicly available.*

Reward hacking, where a reasoning model exploits loopholes in a reward function to achieve high rewards without solving the intended task, poses a significant threat. This behavior may be explicit, i.e. verbalized in the model's chain-of-thought (CoT), or implicit, where the CoT appears benign thus bypasses CoT monitors. To detect implicit reward hacking, we propose TRACE (Truncated Reasoning AUC Evaluation). Our key observation is that hacking occurs when exploiting the loophole is easier than solving the actual task. This means that the model is using less `effort' than required to achieve high reward. TRACE quantifies effort by measuring how early a model's reasoning becomes sufficient to pass a verifier. We progressively truncate a model's CoT at various lengths and measure the verifier-passing rate at each cutoff. A hacking model, which takes a reasoning shortcut, will achieve a high passing rate with only a small fraction of its CoT, yielding a large area under the accuracy-vs-length curve. TRACE achieves over 65% gains over our strongest 72B CoT monitoring baseline in math, and over 30% gains over a 32B monitoring baseline in code. We further show that TRACE can discover unknown loopholes in the training environment. Overall, TRACE offers a scalable unsupervised approach for oversight where current monitoring methods prove ineffective.

Large language models (LLMs) are emerging as powerful tools for scientific law discovery, a foundational challenge in AI-driven science. However, existing benchmarks for this task suffer from a fundamental methodological trilemma, forcing a trade-off between scientific relevance, scalability, and resistance to memorization. Furthermore, they oversimplify discovery as static function fitting, failing to capture the authentic scientific process of uncovering embedded laws through the interactive exploration of complex model systems. To address these critical gaps, we introduce **NewtonBench**, a benchmark comprising 324 scientific law discovery tasks across 12 physics domains. Our design mitigates the evaluation trilemma by using counterfactual law shifts - systematic alterations of canonical laws - to generate a vast suite of problems that are scalable, scientifically relevant, and memorization-resistant. Moreover, we elevate the evaluation from static function fitting to interactive model discovery, requiring agents to experimentally probe simulated complex systems to uncover hidden principles. Our extensive evaluation of 11 state-of-the-art LLMs reveals a clear but fragile capability for discovery in frontier models: this ability degrades precipitously with increasing system complexity and exhibits extreme sensitivity to observational noise. Notably, we uncover a paradoxical effect of tool assistance: providing a code interpreter can hinder more capable models by inducing a premature shift from exploration to exploitation, causing them to satisfice on suboptimal solutions. These results demonstrate that robust, generalizable discovery in complex, interactive environments remains the core challenge for the future of automated science. By providing a scalable, robust, and scientifically authentic testbed, NewtonBench offers a crucial tool for measuring true progress and guiding the development of next-generation AI agents capable of genuine scientific discovery.

Neural synchrony is hypothesized to help the brain organize visual scenes into structured multi-object representations. In machine learning, synchrony-based models analogously learn object-centric representations by storing binding in the phase of complex-valued features. Rotating Features (RF) instantiate this idea with vector-valued activations, encoding object presence in magnitudes and affiliation in orientations. We propose Orthogonal Rotating Features (OrthoRF), which enforces orthogonality in RF’s orientation space via an inner-product loss and architectural modifications. This yields sharper phase alignment and more reliable grouping. In evaluations of unsupervised object discovery, including settings with overlapping objects, noise, and out-of-distribution tests, OrthoRF matches or outperforms current models while producing more interpretable representations, and it eliminates the post-hoc clustering required by many synchrony-based approaches. Unlike current models, OrthoRF also recovers occluded object parts, indicating stronger grouping under occlusion. Overall, orthogonality emerges as a simple, effective inductive bias for synchrony-based object-centric learning.

Generating aesthetic posters is more challenging than simple design images: it requires not only precise text rendering but also the seamless integration of abstract artistic content, striking layouts, and overall stylistic harmony. To address this, we propose PosterCraft, a unified framework that abandons prior modular pipelines and rigid, predefined layouts, allowing the model to freely explore coherent, visually compelling compositions. PosterCraft employs a carefully designed, cascaded workflow to optimize the generation of high-aesthetic posters: (i) large-scale text-rendering optimization on our newly introduced Text-Render-2M dataset; (ii) region-aware supervised finetuning on HQ-Poster-100K; (iii) aesthetic-text reinforcement learning via best-of-n preference optimization; and (iv) joint vision–language feedback refinement. Each stage is supported by a fully automated data-construction pipeline tailored to its specific needs, enabling robust training without complex architectural modifications. Evaluated on multiple experiments, PosterCraft significantly outperforms open-source baselines in rendering accuracy, layout coherence, and overall visual appeal—approaching the quality of SOTA commercial systems.

Modern language models (LMs) still rely on fixed, pre-defined subword tokenizations. Once a tokenizer is trained, the LM can only operate at this fixed level of granularity, which often leads to brittle and counterintuitive behaviors even in otherwise strong reasoning models. We introduce \textbf{ByteFlow Net}, a new architecture that removes tokenizers entirely and instead enables models to learn their own segmentation of raw byte streams into semantically meaningful units. Our approach is grounded in information theory: ByteFlow Net performs compression-driven segmentation based on coding rate of latent representation, allowing the model to dynamically evaluate the information cost of grouping bytes and decide chunk boundaries during processing. Unlike prior self-tokenizing methods that depend on brittle heuristics with human-designed inductive biases, ByteFlow Net adapts its internal representation granularity to the input itself. Experiments demonstrate that this compression-based chunking strategy yields substantial performance gains, with ByteFlow Net outperforming both BPE-based Transformers and previous byte-level architectures. These results suggest that end-to-end, tokenizer-free modeling is not only feasible but also more effective, opening a path toward more adaptive, robust, and information-grounded language models.

The choice of prior is central to solving ill-posed imaging inverse problems, making it essential to select one consistent with the measurements $y$ to avoid severe bias. In Bayesian inverse problems, this could be achieved by evaluating the model evidence $p(y \mid M)$ under different models $M$ that specify the prior and then selecting the one with the highest value. Diffusion models are the state-of-the-art approach to solving inverse problems with a data-driven prior; however, directly computing the model evidence with respect to a diffusion prior is intractable. Furthermore, most existing model evidence estimators require either many pointwise evaluations of the unnormalized prior density or an accurate clean prior score. We propose DiME, an estimator of the model evidence under a diffusion prior by integrating over the time-marginals of posterior sampling methods. Our method leverages the large amount of intermediate samples that are naturally obtained during the reverse diffusion sampling process to obtain an accurate estimation of the model evidence using only a handful of posterior samples (e.g., 20). We demonstrate how to implement our estimator in tandem with recent diffusion posterior sampling methods. Empirically, our estimator matches the model evidence when it can be computed analytically, and it is able to both select the correct diffusion model prior and diagnose prior misfit under different highly ill-conditioned, non-linear inverse problems, including a real-world black hole imaging problem.

Large language models (LLMs) suffer from hallucination and context forgetting. Prior studies suggest that attention drift is a primary cause of these problems, where LLMs' focus shifts towards newly generated tokens and away from the initial input context. To address this, we make use of a related, intrinsic characteristic of LLMs: attention sink – the tendency to consistently allocate high attention to the very first token (i.e., ⟨BOS⟩) of a sequence. Concretely, we propose an advanced context anchoring method, SINKTRACK, which treats ⟨BOS⟩ as an information anchor and injects key contextual features (such as those derived from the input image or instruction) into its representation. As such, LLM remains anchored to the initial input context throughout the entire generation process. SINKTRACK is training-free, plug-and-play, and introduces negligible inference overhead. Experiments demonstrate that SINKTRACK mitigates hallucination and context forgetting across both textual (e.g., +18.9% on QuAC with Llama3.1-8B-Instruct) and multi-modal (e.g., +23.0% on M3CoT with Qwen2.5-VL-7B-Instruct) tasks. Its consistent gains across different architectures and scales underscore the robustness and generalizability. We also analyze its underlying working mechanism from the perspective of information delivery. Our source code is available at anonymous GitHub.

The ability to robustly identify causal relationships is essential for autonomous decision-making and adaptation to novel scenarios. However, accurately inferring causal structure requires integrating both world knowledge and abstract logical reasoning. In this work, we investigate the interaction between these two capabilities through the representative task of causal reasoning over narratives. Through controlled synthetic, semi-synthetic and real-world experiments, we find that state-of-the-art large language models (LLMs) often rely on superficial heuristics—for example, inferring causality from event order or recalling memorized world knowledge without attending to context. Furthermore, we show that simple reformulations of the task can elicit more robust reasoning behavior. Our evaluation spans a range of causal structures, from linear chains to complex graphs involving colliders and forks. These findings uncover systematic patterns in how LLMs perform causal reasoning and lay the groundwork for developing methods that better align LLM behavior with principled causal inference.

Prevalent retrieval-based tool-use pipelines struggle with a dual semantic challenge: their retrievers often employ encoders that fail to capture complex semantics, while the Large Language Model (LLM) itself lacks intrinsic tool knowledge from its natural language pretraining. Generative methods offer a powerful alternative by unifying selection and execution, tasking the LLM to directly learn and generate tool identifiers. However, the common practice of mapping each tool to a unique new token introduces substantial limitations: it creates a scalability and generalization crisis, as the vocabulary size explodes and each tool is assigned a semantically isolated token. This approach also creates a semantic bottleneck that hinders the learning of collaborative tool relationships, as the model must infer them from sparse co-occurrences of monolithic tool IDs within a vast library. To address these limitations, we propose ToolWeaver, a novel generative tool learning framework that encodes tools into hierarchical sequences. This approach makes vocabulary expansion logarithmic to the number of tools. Crucially, it enables the model to learn collaborative patterns from the dense co-occurrence of shared codes, rather than the sparse co-occurrence of monolithic tool IDs. We generate these structured codes through a novel tokenization process designed to weave together a tool's intrinsic semantics with its extrinsic co-usage patterns. These structured codes are then integrated into the LLM through a generative alignment stage, where the model is fine-tuned to produce the hierarchical code sequences. Evaluation results with nearly 47,000 tools show that ToolWeaver significantly outperforms state-of-the-art methods, establishing a more scalable, generalizable, and semantically-aware foundation for advanced tool-augmented agents.

Current deep learning models for image segmentation often lack reliable uncertainty quantification, particularly at the image-specific level. While Conformal Risk Control (CRC) offers marginal statistical guarantees, achieving image-conditional coverage, which ensures prediction sets reliably capture ground truth for individual images, remains a significant challenge. This paper introduces a novel approach to address this gap by learning image-adaptive thresholds for conformal image segmentation. We first propose AT (Adaptive Thresholding), which frames threshold prediction as a supervised regression task. Building upon the insights from AT, we then introduce COAT (Conditional Optimization for Adaptive Thresholding), an innovative end-to-end differentiable framework. COAT directly optimizes image-conditional coverage by using a soft approximation of the True Positive Rate (TPR) as its loss function, enabling direct gradient-based learning of optimal image-specific thresholds. This novel differentiable miscoverage loss is key to enhancing conditional coverage. Our methods provide a robust pathway towards more trustworthy and interpretable uncertainty estimates in image segmentation, offering improved conditional guarantees crucial for safety-critical applications. The code is available at https://github.com/bjbbbb/Conditional-Optimization-for-Adaptive-Thresholding.

Diffusion models have demonstrated strong potential in language modeling, offering various advantages over traditional autoregressive approaches. Their ability to generate and revise entire responses in parallel enables faster generation and built-in self-correction mechanisms. Most modern diffusion-based language models employ masked diffusion, where decoding involves iteratively processing masked tokens based on a binary decision: either retaining the mask or replacing it with the predicted token. However, this binary choice discards valuable predictive information when the mask is retained. To address this limitation, we introduce soft-masking (SM), a novel method that dynamically blends the embedding of the mask token with the embeddings of the top-k predicted tokens from the previous decoding step, for each retained mask. This provides the model with a more informative prior, preserving context from earlier computations and allowing partial information about masked tokens to propagate beyond a single step. We propose a training methodology that efficiently adapts masked diffusion language models to incorporate SM. We demonstrate that training a 169M parameter model from scratch with SM yields superior perplexity and MAUVE scores compared to binary masking baselines. Similarly, a pretrained model can be enhanced with SM through continued pretraining. Finally, we finetune two state-of-the-art diffusion models, Dream-7B and Dream-Coder-7B, with SM. SM consistently improves performance across multiple coding benchmarks, particularly in high-throughput settings.

Large Language Model (LLM) pretraining, finetuning, and evaluation rely on input-space reconstruction and generative capabilities. Yet, it has been observed in vision that embedding-space training objectives, e.g., with Joint Embedding Predictive Architectures (JEPAs), are far superior to their input-space counterpart. That mismatch in how training is achieved between language and vision opens up a natural question: {\em can language training methods learn a few tricks from the vision ones?} The lack of JEPA-style LLM is a testimony of the challenge in designing such objectives for language. In this work, we propose a first step in that direction where we develop LLM-JEPA, a JEPA based solution for LLMs applicable both to finetuning and pretraining. Thus far, LLM-JEPA is able to outperform the standard LLM training objectives by a significant margin across models, all while being robust to overfiting. Those findings are observed across numerous datasets (NL-RX, GSM8K, Spider, RottenTomatoes) and various models from the Llama3, OpenELM, Gemma2 and Olmo families. Code: \url{https://github.com/galilai-group/llm-jepa}.

Time series forecasting plays a significant role in finance, energy, meteorology, and IoT applications. Recent studies have leveraged the generalization capabilities of large language models (LLMs) to adapt to time series forecasting, achieving promising performance. However, existing studies focus on token-level modal alignment, instead of bridging the intrinsic modality gap between linguistic knowledge structures and time series data patterns, greatly limiting the semantic representation. To address this issue, we propose a novel Semantic-Enhanced LLM (SE-LLM) that explores the inherent periodicity and anomalous characteristics of time series to embed into the semantic space to enhance the token embedding. This process enhances the interpretability of tokens for LLMs, thereby activating the potential of LLMs for temporal sequence analysis. Moreover, existing Transformer-based LLMs excel at capturing long-range dependencies but are weak at modeling short-term anomalies in time-series data. Hence, we propose a plugin module embedded within self-attention that models long-term and short-term dependencies to effectively adapt LLMs to time-series analysis. Our approach freezes the LLM and reduces the sequence dimensionality of tokens, greatly reducing computational consumption. Experiments demonstrate the superiority performance of our SE-LLM against the state-of-the-art (SOTA) methods.

Significant recent work has studied the ability of gradient descent to recover a hidden planted direction $\theta^\star \in S^{d-1}$ in different high-dimensional settings, including tensor PCA and single-index models. The key quantity that governs the ability of gradient descent to traverse these landscapes is the information exponent $k^\star$ (Ben Arous et al., (2021)), which corresponds to the order of the saddle at initialization in the population landscape. Ben Arous et al., (2021) showed that $n \gtrsim d^{\max(1, k^\star-1)}$ samples were necessary and sufficient for online SGD to recover $\theta^\star$, and Ben Arous et al., (2020) proved a similar lower bound for Langevin dynamics. More recently, Damian et al., (2023) showed it was possible to circumvent these lower bounds by running gradient descent on a smoothed landscape, and that this algorithm succeeds with $n \gtrsim d^{\max(1, k^\star/2)}$ samples, which is optimal in the worst case. This raises the question of whether it is possible to achieve the same rate without explicit smoothing. In this paper, we show that Langevin dynamics can succeed with $n \gtrsim d^{ k^\star/2 }$ samples if one considers the average iterate, rather than the last iterate. The key idea is that the combination of noise-injection and iterate averaging is able to emulate the effect of landscape smoothing. We apply this result to both the tensor PCA and single-index model settings. Finally, we conjecture that minibatch SGD can also achieve the same rate without adding any additional noise.

Can language models (LMs) self-refine their own responses? This question is increasingly relevant as more than 10% of real-world user interactions involve refinement requests (see Appendix G). Yet prior studies have largely tested LMs on verifiable tasks such as competition math or symbolic reasoning with simplified scaffolds, whereas users often pose open-ended queries and provide varying degrees of feedback about what went wrong. The recent advent of reasoning models that exhibit self-reflection patterns in their chain-of-thought further motivates this question. To address it, we introduce RefineBench, a benchmark of 1,002 challenging problems across 11 domains paired with a checklist-based evaluation framework. We evaluate two refinement modes: (1) guided refinement, where an LM is provided natural language feedback, and (2) self-refinement, where LMs attempt to improve without guidance. In the self-refinement setting, even frontier LMs such as Gemini 2.5 Pro and GPT-4.1 achieve modest baseline scores of 31.1 and 23.4, respectively, and most models fail to consistently improve across iterations (e.g., Gemini-2.5-Pro gains only +1.8%, while DeepSeek-R1 declines by –0.2%). By contrast, in guided refinement, both proprietary LMs and large open-weight LMs (>70B) can leverage targeted feedback to refine responses to near-perfect levels within five turns. These findings suggest that frontier LMs require breakthroughs to self-refine effectively when their initial responses are incorrect, and that RefineBench provides a valuable testbed for tracking progress.

Large language models (LLMs) aligned for safety often suffer from over-refusal—the tendency to reject seemingly toxic or benign prompts by misclassifying them as toxic. This behavior undermines models' helpfulness and restricts usability in sensitive or nuanced contexts. While prior work has proposed mitigation strategies such as data augmentation and activation steering, these approaches often face a trade-off: reducing over-refusal typically degrades the model’s ability to reject genuinely harmful content. We argue that this issue arises from the ambiguous influence of toxic and seemingly toxic prompts on the model’s learning dynamics. To address it, we introduce a preceding alignment stage, DCR: $\textbf{D}$iscernment via $\textbf{C}$ontrastive $\textbf{R}$efinement. Both theoretically and empirically, we demonstrate that contrastive refinement improves an LLM’s capacity to distinguish truly toxic prompts from superficially toxic ones. Evaluation across diverse benchmarks shows that our method effectively reduces over-refusal while preserving the safety benefits of alignment. Importantly, it achieves this with minimal degradation of general capabilities, offering a more principled and robust direction for safety alignment.