ICLR 2026 Papers

Incorporating metadata in Large Language Models (LLMs) pretraining has recently emerged as a promising approach to accelerate training. However prior work highlighted only one useful signal—URLs, leaving open the question of whether other forms of metadata could yield greater benefits. In this study, we investigate a wider range of metadata types and find other types of metadata, such as fine-grained indicators of document quality that can also accelerate pretraining when prepended. We identify a common feature among effective metadata: they encode information at a finer granularity. We further introduce metadata appending as a means of improving training efficiency, where predicting an appropriate metadata as auxiliary task can help speed up pretraining. In addition, learnable meta-tokens trained with masked loss can recover part of the speedup by inducing quality-aware latent structure. Using probing, we analyze latent representations to understand how metadata shapes learning. Together, these results yield practical guidelines for integrating metadata to improve both the efficiency and effectiveness of LLM pretraining.

Transformer LLMs have been shown to exhibit strong reasoning ability that scales with inference-time compute, most prominently through token-space “thinking” (i.e., chains of thought). A growing line of work pushes this extra computation into the model’s latent space (adjacent to standard decoding) which we term Auxiliary Latent-Space Computation (ALSC). Existing ALSC methods largely fall into three buckets: (i) token-mediated latent or special-token rollouts, (ii) residual/activation steering, and (iii) memory compression via cache pruning, merging, or summarization. An underexplored alternative is memory consolidation and reconsolidation, two processes in the brain that are responsible for stabilising newly formed memory traces, and, upon recall, transiently rendering established traces plastic such they can integrate new contextual information before restabilising. In a Transformer LLM, this can be seen as analogous to performing in-place global rewrites of incoming KV segments, and rewrites of past segments conditioned on newly observed tokens. In this work, we give a theoretical justification as to why memory (re)consolidation via KV cache rewrites is beneficial for improved reasoning. We do this through the lens of Information Bottleneck (IB) theory, which posits that model generalisation emerges from an optimal balance between input information compression and retention of predictive information in latent representations. We prove using IB theory that Vanilla decoder-only Transformers are inherently constrained in their ability to form task-optimal sequence representations. We then introduce the Bottlenecked Transformer, which augments a decoder-only backbone LLM with a lightweight Cache Processor, an auxiliary Transformer that performs periodic, non-causal, in-place KV rewrites at newline-delimited reasoning step boundaries. The processor consolidates recently written KV entries and reconsolidates a small, top-$k$ attention-selected set of prior entries, conditioned on recent context. We evaluate our Bottlenecked Transformer architecture on seven mathematical reasoning benchmarks, with four backbone LLMs. Our model sees consistent performance gains over vanilla Transformers and pause-token augmented Transformer baselines, with gains of up to +6.6pp for selected tasks and backbones.

The LLM-as-a-judge paradigm is widely used in both evaluating free-text model responses and reward modeling for model alignment and finetuning. Recently, finetuning judges with judge-specific data has emerged as an often preferred choice over directly prompting frontier models as judges, as the former achieves better performance with smaller model sizes while being more robust to common biases. However, the standard evaluation ignores several practical concerns of finetuned judges regarding their real world deployment. In this paper, we identify and formalize three aspects that affect the *shelf life* of these judges: *future-proofing* and *backward-compatibility* – how well judges finetuned on responses by today's generator models perform on responses by future models or past models, as well as *question generalization* – how well judges generalize to unseen questions at test time. We study these three aspects under a unified framework with varying train and test distributions in two reasoning datasets, three SFT- and DPO-based fine-tuning algorithms, and three different backbone models. Experiments suggest that future-proofing is challenging for most models, while backward-compatibility is relatively easy, with DPO-trained models consistently *improving* performance. We further find that continual learning provides a more balanced adaptation to shifts between older and newer response distributions than training solely on stronger or weaker responses. Moreover, all models observe certain degrees of performance degradation when moving from questions seen during training to unseen ones, showing that current judges do not fully generalize to unseen questions. These findings provide insights into practical considerations for developing and deploying judge models in the face of ever-changing generators.

Modern conversational agents such as ChatGPT and Alexa+ have become indispensable in everyday life. To handle diverse business requirements and enable agentic capabilities, these LLM-based systems often rely on predefined policies, which specify instructions such as model metadata, response styles, and tool-using rules. These policies, typically implemented as in-context prompts, are becoming increasingly complex and lengthy, posing challenges for models in faithfully following them. Moreover, they impose a large fixed computational cost regardless of the input query. As multimodal conversational agents emerge, complex policies that govern multimodal tasks and even involve visual instructions are becoming increasingly necessary, yet they have been rarely studied in previous work. In particular, prior work on prompt compression has focused solely on reducing the length of task templates and demonstrations, which require limited reasoning compared to policies. Meanwhile, related work on policy alignment has been limited to internalizing text-only safety instructions. To bridge this gap, we introduce Multimodal Policy Internalization (MPI), a new task that aims to internalize reasoning-intensive multimodal policies into the parameters of a large multimodal model, enabling stronger policy-following behavior without requiring the policy to be included in-context during inference. MPI presents unique challenges from both data and algorithmic perspectives. We construct two new datasets that cover complex decision-making and tool-using tasks across both synthetic and real-world visual inputs. We investigate diverse internalization strategies and propose a novel three-stage training framework, TriMPI, which enables stronger guidance from the original policy during internalization. Specifically, we first introduce a continual pretraining stage before supervised finetuning, which directly injects policy knowledge into the model. We then propose PolicyRollout, a simple yet effective extension to GRPO-style RL algorithms, which enables more grounded exploration by augmenting the rollout space with policy-aware responses. We show significant improvements of TriMPI over strong baselines in end-to-end performance, generalization capability, and robustness to catastrophic forgetting. As the first work on multimodal policy internalization, we aim to build a strong foundation for future research by providing datasets, training recipes, and comprehensive evaluations.

Generalist visual captioning goes beyond a simple appearance description task, but requires integrating a series of visual cues into a caption and handling various visual domains. In this task, current open-source models present a large performance gap with commercial ones, which limits various applications such as data synthesis. To bridge the gap, this paper proposes CapFlow, a novel multi-agent collaboration workflow. CapFlow demonstrates for the first time that, by capitalizing on open-source models, it is possible to achieve caption quality on par with GPT-4.1 in various domains with an 89.5\% reduction in costs. By leveraging CapFlow as the data synthesizer, we produce high-quality visual captions from image and video domains at scale, and obtain a generalist visual captioner via fine-tuning, namely MetaCaptioner. Through extensive experiments, we show that MetaCaptioner not only achieves comparable captioning capabilities with commercial models but also reaches top-tier multimodal performance in the open-source community. We hope CapFlow and MetaCaptioner can benefit future multimodal research by providing a strong and cost-effective visual captioning solution. Our source code and models will be publicly released.

Time series anomaly detection plays a critical role in many dynamic systems. However, previous approaches have primarily relied on unimodal numerical data, overlooking the importance of complementary information from other modalities. In this paper, we propose a novel multimodal time series anomaly detection model (MindTS) that focuses on addressing two key challenges: (1) how to achieve semantically consistent alignment across heterogeneous multimodal data, and (2) how to filter out redundant modality information to enhance cross-modal interaction effectively. To address the first challenge, we propose Fine-grained Time-text Semantic Alignment. It integrates exogenous and endogenous text information through cross-view text fusion and a multimodal alignment mechanism, achieving semantically consistent alignment between time and text modalities. For the second challenge, we introduce Content Condenser Reconstruction, which filters redundant information within the aligned text modality and performs cross-modal reconstruction to enable interaction. Extensive experiments on six real-world multimodal datasets demonstrate that the proposed MindTS achieves competitive or superior results compared to existing methods.

When reading, we often have specific information that interests us in a text. For example, you might be reading this paper because you are curious about LLMs for eye movements in reading, the experimental design, or perhaps you wonder ``This sounds like science fiction. Does it actually work?''. More broadly, in daily life, people approach texts with any number of text-specific goals that guide their reading behavior. In this work, we ask, for the first time, whether open-ended reading goals can be automatically decoded solely from eye movements in reading. To address this question, we introduce goal decoding tasks and evaluation frameworks using large-scale eye tracking for reading data in English with hundreds of text-specific information seeking tasks. We develop and compare several discriminative and generative multimodal text and eye movements LLMs for these tasks. Our experiments show considerable success on the task of selecting the correct goal among several options, and even progress towards free-form textual reconstruction of the precise goal formulation. These results open the door for further scientific investigation of goal driven reading, as well as the development of educational and assistive technologies that will rely on real-time decoding of reader goals from their eye movements.

We introduce Genie Envisioner (GE), a unified world foundation platform for robotic manipulation that jointly learns visual representations and action policies within a single video-generative framework. At its core, GE-Base is a large-scale instruction-conditioned video diffusion model that captures the spatial, temporal, and semantic dynamics of real-world robotic interactions in a structured latent space. Building on this foundation, GE-Act employs a lightweight flow-matching decoder to map latent representations into executable action trajectories, enabling precise and generalizable policy inference across diverse embodiments with minimal supervision. Trained on over 1 million manipulation episodes, GE supports both short- and long-horizon tasks, and generalizes across embodiments. All code, models, and benchmarks will be released publicly.

Recent strides in multimodal large language models (MLLMs) have demonstrated significant progress in many reasoning tasks, but they still fail in Abstract Visual Reasoning (AVR) tasks. Our experimental findings indicate that the core bottleneck lies not only in the reasoning capabilities of MLLMs but more critically in their absence of fine-grained perception. To address this issue, we present VisuRiddles, a dedicated resource for AVR research. It consists of (i) a benchmark, collected from real-world data, for the systematic evaluation of MLLMs' AVR capabilities, and (ii) a synthesizer, which automatically generates AVR instances enriched with perceptual descriptions and reasoning chains, enabling supervised training and deeper investigation. Building on VisuRiddles, we propose a two-stage training paradigm that progressively enhances perceptual ability and strengthens reasoning, producing the Perception-Augmented Visual Reasoner (PAVR). Experiments demonstrate that PAVR unifies perception and reasoning, substantially outperforming both open-source and commercial MLLMs, thereby underscoring fine-grained perception as the primary bottleneck in AVR.

We introduce a **variational reasoning** framework for language models that treats thinking traces as latent variables and optimizes them through variational inference. Starting from the evidence lower bound (ELBO), we extend it to a multi-trace objective for tighter bounds and propose a forward-KL formulation that stabilizes the training of the variational posterior. We further show that rejection sampling finetuning and binary-reward RL, including GRPO, can be interpreted as local forward-KL objectives, where *an implicit weighting by model accuracy* naturally arises from the derivation and reveals a previously unnoticed bias toward easier questions. We empirically validate our method on the Qwen 2.5 and Qwen 3 model families across a wide range of reasoning tasks. Overall, our work provides a principled probabilistic perspective that unifies variational inference with RL-style methods and yields stable objectives for improving the reasoning ability of language models.

Bidirectional language models (LMs) consistently show stronger context understanding than unidirectional models, yet the theoretical reason remains unclear. We present a simple information bottleneck (IB) perspective: bidirectional representations preserve more mutual information (MI) about both the input and the target, yielding richer features for downstream tasks. We adopt a layer–wise view and hypothesize that, at comparable capacity, bidirectional layers retain more useful signal than unidirectional ones. To test this claim empirically, we present Flow Neural Information Bottleneck (FlowNIB), a lightweight, post-hoc framework capable of estimating comparable mutual information values for individual layers in LMs, quantifying how much mutual information each layer carries for a dataset. FlowNIB takes three inputs—(i) the original LM’s inputs/dataset, (ii) ground–truth labels, and (iii) layer activations—simultaneously estimates the mutual information for both the input–layer and layer–label pairs. Empirically, bidirectional LM layers exhibit higher mutual information than similar—and even larger—unidirectional LMs. As a result, bidirectional LMs outperform unidirectional LMs across extensive experiments on NLU benchmarks (e.g., GLUE), commonsense reasoning, and regression tasks, demonstrating superior context understanding.

Agent systems powered by large language models (LLMs) have demonstrated impressive performance on repository-level code-generation tasks. However, for tasks such as website codebase generation, which depend heavily on visual effects and user-interaction feedback, current code agents rely only on simple code execution for feedback and verification. This approach fails to capture the actual quality of the generated code. In this paper, we propose WebGen-Agent, a novel website-generation agent that leverages comprehensive and multi-level visual feedback to iteratively generate and refine the website codebase. Detailed and expressive text descriptions and suggestions regarding the screenshots and GUI-agent testing of the websites are generated by a visual language model (VLM), together with scores that quantify their quality. The screenshot and GUI-agent scores are further integrated with a backtracking and select-best mechanism, enhancing the performance of the agent. Utilizing the accurate visual scores inherent in the WebGen-Agent workflow, we further introduce Step-GRPO with Screenshot and GUI-Agent Feedback to improve the ability of LLMs to act as the agent-engine model. By using the screenshot and GUI-agent scores at each step as the reward in Step-GRPO, we provide a dense and reliable process supervision signal, which effectively improves the model's website-generation ability. On the WebGen-Bench dataset, WebGen-Agent increases the accuracy of Claude 3.5 Sonnet from 26.4\% to 51.9\% and its appearance score from 3.0 to 3.9, outperforming the previous state-of-the-art agent system. Additionally, our training approach increases the accuracy of Qwen2.5-Coder-7B-Instruct from 38.9\% to 45.4\% and raises the appearance score from 3.4 to 3.7.

Although reinforcement learning with verifiable rewards (RLVR) shows promise in improving the reasoning ability of large language models (LLMs), the scaling up dilemma remains due to the reliance on human-annotated labels especially for complex tasks. Recent self-rewarding methods provide a label-free alternative that exhibits the eliciting potential of LLM reasoning, but they often suffer from the non-negligible training collapse issue, as the single-view supervision signal easily forms the self-consistent illusion, yielding the reward hacking. Inspired by the success of self-supervised learning, we propose \textit{Co-rewarding}, a novel self-supervised RL framework that improves training stability by seeking complementary supervision from another views. Specifically, we instantiate Co-rewarding in two ways: (1) \textit{Co-rewarding-I} is a data-side instantiation that derives reward signals from contrastive agreement across semantically analogous questions; and (2) \textit{Co-rewarding-II} is a model-side instantiation that maintains a slowly-updated reference teacher with pseudo labels to realize self-distillation. Intuitively, such instantiations introduce different levels of discrepancy to increase the difficulty of training collapse on trivial reasoning solutions. We also explore their orthogonally combined version to further boost the performance. Empirically, Co-rewarding exhibits stable training across various setups, and outperforms other self-rewarding baselines by $+3.31\%$ improvements on average on multiple mathematical reasoning benchmarks, especially by $+7.49\%$ on Llama-3.2-3B-Instruct. Notably, Co-rewarding reaches or even surpasses ground-truth (GT) labeled reward in several cases of RLVR, such as achieving a Pass@$1$ of $94.01\%$ on GSM8K with Qwen3-8B-Base.

The growing use of large language models (LLMs) for sensitive applications has highlighted the need for effective watermarking techniques to ensure the provenance and accountability of AI-generated text. However, most existing watermarking methods require access to the decoding process, limiting their applicability in real-world settings. One illustrative example is the use of LLMs by dishonest reviewers in the context of academic peer review, where conference organizers have no access to the model used but still need to detect AI-generated reviews. Motivated by this gap, we introduce In-Context Watermarking (ICW), which embeds watermarks into generated text solely through prompt engineering, leveraging LLMs' in-context learning and instruction-following abilities. We investigate four ICW strategies at different levels of granularity, each paired with a tailored detection method. We further examine the Indirect Prompt Injection (IPI) setting as a specific case study, in which watermarking is covertly triggered by modifying input documents such as academic manuscripts. Our experiments validate the feasibility of ICW as a model-agnostic, practical watermarking approach. Moreover, our findings suggest that as LLMs become more capable, ICW offers a promising direction for scalable and accessible content attribution.

Recent vision-language models have strong perceptual ability but their implicit reasoning is hard to explain and easily generates hallucinations on complex queries. Compositional methods improve interpretability, but most rely on a single agent or hand-crafted pipeline and cannot decide when to collaborate across complementary agents or compete among overlapping ones. We introduce MATA (Multi-Agent hierarchical Trainable Automaton), a multi-agent system presented as hierarchical finite-state automaton for visual reasoning whose top-level transitions are chosen by a trainable hyper agent. Each agent corresponds to a state in the hyper automaton, and runs a small rule-based sub-automaton for reliable micro-control. All agents read and write a shared memory, yielding transparent execution history. To supervise the hyper agent’s transition policy, we build transition-trajectory trees and transform to memory-to-next-state pairs, forming the MATA-SFT-90K dataset for supervised finetuning (SFT). The finetuned LLM as the transition policy understands the query and the capacity of agents, and it can efficiently choose the optimal agent to solve the task. Across multiple visual reasoning benchmarks, MATA achieves the state-of-the-art results compared with monolithic and compositional baselines. The code and dataset are available at https://github.com/ControlNet/MATA.

The recent advancement of Artificial Intelligence Generated Content (AIGC) has led to significant strides in modeling human interaction, particularly in the context of multi-modal dialogue. While current methods impressively generates realistic dialogue in speech and vision modalities, challenges remain in multi-modal conditional dialogue generation. This paper focuses on the natural alignment between speech, vision, and text, aiming at expressive dialogue generation through multi-modal conditional control. Since existing datasets lack the richness and diversity in dialogue expressiveness, we introduce a novel multi-modal dialogue annotation pipeline to exploit meaningful dialogues from movies and TV series with fine-grained annotations across multi-modalities. The resultant dataset, MM-Dia, provides over 360 hours and 54,700 dialogues, facilitating the Multimodal Dialogue Generation task through explicit control over style-controllable dialogue speech synthesis. While the proposed benchmark, MM-Dia-Bench, containing 309 dialogues that are highly expressive with visible dual/single speaker scenes, supporting the evaluation of implicit cross-modal control through downstream multi-modal dialogue generation tasks to assess the audio-visual style consistency across modalities. Our experiments demonstrate the effectiveness of our data in enhancing style controllability and reveal limitations in current frameworks' ability to replicate human interaction expressiveness, providing new insights and challenges for multi-modal conditional dialogue generation. Code, demo and data will be released at: https://mmdiaiclr26.github.io/mmdiaiclr26/.

Object hallucination in Multimodal Large Language Models (MLLMs) is a persistent failure mode that causes the model to perceive objects absent in the image. This weakness of MLLMs is currently studied using static benchmarks with fixed visual scenarios, which preempts the possibility of uncovering model-specific or unanticipated hallucination vulnerabilities. We introduce GHOST (Generating Hallucinations via Optimizing Stealth Tokens), a method designed to stress-test MLLMs by actively generating images that induce hallucination. GHOST is fully automatic and requires no human supervision or prior knowledge. It operates by optimizing in the image embedding space to mislead the model while keeping the target object absent, and then guiding a diffusion model conditioned on the embedding to generate natural-looking images. The resulting images remain visually natural and close to the original input, yet introduce subtle misleading cues that cause the model to hallucinate. We evaluate our method across a range of models, including reasoning models like GLM-4.1V-Thinking, and achieve a hallucination success rate exceeding 28%, compared to around 1% in prior data-driven discovery methods. We confirm that the generated images are both high-quality and object-free through quantitative metrics and human evaluation. Also, GHOST uncovers transferable vulnerabilities: images optimized for Qwen2.5-VL induce hallucinations in GPT-4o at a 66.5% rate. Finally, we show that fine-tuning on our images mitigates hallucination, positioning GHOST as both a diagnostic and corrective tool for building more reliable multimodal systems.

We introduce a novel framework for measuring Overton pluralism in LLMs—the extent to which diverse viewpoints are represented in model outputs. We (i) formalize Overton pluralism as a set-coverage metric (OVERTONSCORE), (ii) conduct a large-scale US-representative human study (N=1209; 60 questions; 8 LLMs), and (iii) develop an automated benchmark that closely reproduces human judgments. On average, models achieve OVERTONSCOREs of 0.35 – 0.41, with Deepseek V3 performing best; yet all models remain far below the theoretical maximum of 1.0, revealing substantial headroom for improvement. Because repeated large-scale human studies are costly and slow, scalable evaluation tools are essential for model development. Hence, we propose an automated benchmark that achieves high rank correlation with human judgments ($\rho=0.88$), providing a practical proxy while not replacing human assessment. By turning pluralistic alignment from a normative aim into a measurable benchmark, our work establishes a foundation for systematic progress toward more pluralistic LLMs.

Vision Transformers are widely adopted as the backbone of vision foundation models, but they are known to produce high-norm artifacts that degrade representation quality. When knowledge distillation transfers these features to students, high-norm artifacts dominate the objective, so students overfit to artifacts and underweight informative signals, diminishing the gains from larger models. Prior work attempted to remove artifacts but encountered an inherent trade-off between artifact suppression and preserving informative signals from teachers. To address this, we introduce Singular Nullspace-Guided Energy Reallocation (SiNGER), a novel distillation framework that suppresses artifacts while preserving informative signals. The key idea is principled teacher feature refinement: during refinement, we leverage the nullspace-guided perturbation to preserve information while suppressing artifacts. Then, the refined teacher's features are distilled to a student. We implement this perturbation efficiently with a LoRA-based adapter that requires minimal structural modification. Extensive experiments show that \oursname consistently improves student models, achieving state-of-the-art performance in multiple downstream tasks and producing clearer and more interpretable representations.