ICLR 2026 Papers

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.

Alignment of large language models (LLMs) has predominantly relied on pairwise preference optimization, where annotators select the better of two responses to a prompt. While simple, this approach overlooks the opportunity to learn from richer forms of human feedback, such as multiwise comparisons and top-$k$ rankings. We introduce $\textit{Ranked Choice Preference Optimization}$ (RCPO), a unified framework that bridges preference optimization with (ranked) choice modeling via maximum likelihood estimation. RCPO supports both utility based and rank based models, subsumes several pairwise methods (such as DPO and SimPO) as special cases, and provide principled training objectives for richer feedback formats. We instantiate this framework with two representative models (Multinomial Logit and Mallows-RMJ). Experiments on Llama-3-8B-Instruct, Gemma-2-9B-it, and Mistral-7B-Instruct across in-distribution and out-of-distribution settings show that RCPO consistently outperforms competitive baselines. RCPO shows that directly leveraging ranked preference data, combined with the right choice models, yields more effective alignment. It offers an extensible foundation for incorporating (ranked) choice modeling into LLM training.

Large language models (LLMs) have recently achieved impressive progress in knowledge-intensive and reasoning tasks. However, their tendency to produce fabricated or factually inconsistent content remains a fundamental challenge to their practical deployment. To address this issue, we propose Negative-Direction Aware Decoding (NDAD), a novel decoding method that identifies and exploits hallucination signals as repulsive directions in the model’s representation space, thereby improving factual adherence without retraining. Specifically, NDAD elicits hallucination-leaning signals by selectively masking critical attention heads, which exposes unstable hypotheses that the model would otherwise amplify during generation. To regulate the influence of these signals, NDAD employs two complementary weights: a global alignment weight measuring how well the induced signal aligns with the layer’s native activations (thus quantifying its referential utility) and a local weight estimating whether low-probability tokens in the masked distribution are likely to evolve toward the final output. Based on the weights, we derive a latent hallucination distribution that serves as the negative direction. A lightweight gradient-descent step then subtracts mass from hallucination-prone regions of the output distribution, adjusting the final logits while preserving the model’s high-confidence predictions. Extensive experiments across multiple LLMs and diverse benchmark datasets demonstrate that NDAD consistently enhances factual reliability without requiring additional training or external knowledge.

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.

Unsupervised domain adaptive hashing transfers knowledge from labeled source domains to unlabeled target domains, addressing domain shift challenges in real-world retrieval tasks. Existing methods face two critical limitations: target domain noise severely misleads model training, and indiscriminate domain alignment strategies treat all target samples equally, potentially distorting essential feature structures. We propose an uncertainty-aware adaptive hashing approach that addresses these challenges through a hierarchical conformal calibration framework. At the semantic level, we employ conformal inference to generate confidence prediction sets, replacing single pseudo-labels with set-based predictions whose sizes directly quantify sample reliability for weighted pseudo-label learning and domain alignment. This enables the model to focus on reliable samples while suppressing noise. At the representation level, we predict the stability of individual hash bits, where bit-level confidence guides a robust weighted quantization loss and enables dynamic weighted Hamming distance during retrieval, fundamentally enhancing hash code quality and retrieval robustness. Through this hierarchical calibration mechanism, our method achieves more adaptive and robust cross-domain knowledge transfer. Extensive experiments on multiple benchmark datasets demonstrate significant improvements over existing approaches, validating the effectiveness and superiority of our method.

Recent advances in large multi-modal generative models have demonstrated impressive capabilities in multi-modal generation, including image and video generation. These models are typically built upon multi-step frameworks like diffusion and flow matching, which inherently limits their inference efficiency (requiring 40-100 Number of Function Evaluations (NFEs)). While various few-step methods aim to accelerate the inference, existing solutions have clear limitations. Prominent distillation-based methods, such as progressive and consistency distillation, either require an iterative distillation procedure or show significant degradation at very few steps (< 4-NFE). Meanwhile, integrating adversarial training into distillation (e.g., DMD/DMD2 and SANA-Sprint) to enhance performance introduces training instability, added complexity, and high GPU memory overhead due to the auxiliary trained models. To this end, we propose TwinFlow, a simple yet effective framework for training 1-step generative models that bypasses the need for distillation from pre-trained models and avoids standard adversarial training, making it ideal for building large-scale, efficient models. On text-to-image tasks, our method achieves a GenEval score of 0.83 in 1-NFE, outperforming strong baselines like SANA-Sprint (a GAN loss-based framework) and RCGM (a consistency-based framework). **Notably, we demonstrate the scalability of TwinFlow by transforming Qwen-Image-20B---the current largest open-source multi-modal generative model---into an efficient few-step generator**. With just 1-NFE, our approach matches the performance of the original 100-NFE model on both the GenEval and DPG-Bench benchmarks, reducing computational cost by $100\times$ with minor quality degradation. Our code and models will be made publicly available.

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/.

Learned optimizers (LOs) have the potential to significantly reduce the wall-clock training time of neural networks. However, they can struggle to optimize unseen tasks (*meta-generalize*), especially when training networks wider than those seen during meta-training. To address this, we derive the Maximal Update Parametrization ($\mu$P) for two state-of-the-art learned optimizer architectures and propose a simple meta-training recipe for $\mu$-parameterized LOs ($\mu$LOs). Our empirical evaluation demonstrates that LOs meta-trained with our recipe substantially improve meta-generalization to wider unseen tasks when compared to LOs trained under standard parametrization (SP) using the same compute budget. We also empirically observe that $\mu$LOs exhibit unexpectedly improved meta-generalization to deeper networks ($5\times$ meta-training) and surprising generalization to much longer training horizons ($25\times$ meta-training) when compared to SP LOs.

Personalized large language models (LLMs) rely on memory retrieval to incorporate user-specific histories, preferences, and contexts. Existing approaches either overload the LLM by feeding all the user's past memory into the prompt, which is costly and unscalable, or simplify retrieval into a one-shot similarity search, which captures only surface matches. Cognitive science, however, shows that human memory operates through a dual process: Familiarity, offering fast but coarse recognition, and Recollection, enabling deliberate, chain-like reconstruction for deeply recovering episodic content. Current systems lack both the ability to perform recollection retrieval and mechanisms to adaptively switch between the dual retrieval paths, leading to either insufficient recall or the inclusion of noise. To address this, we propose RF-Mem (Recollection–Familiarity Memory Retrieval), a familiarity uncertainty-guided dual-path memory retriever. RF-Mem measures the familiarity signal through the mean score and entropy. High familiarity leads to the direct top-$K$ Familiarity retrieval path, while low familiarity activates the Recollection path. In the Recollection path, the system clusters candidate memories and applies $\alpha$-mix with the query to iteratively expand evidence in embedding space, simulating deliberate contextual reconstruction. This design embeds human-like dual-process recognition into the retriever, avoiding full-context overhead and enabling scalable, adaptive personalization. Experiments across three benchmarks and corpus scales demonstrate that RF-Mem consistently outperforms both one-shot retrieval and full-context reasoning under fixed budget and latency constraints. Our code can be found in the Supplementary Materials.

The {\em stop gradient} and {\em exponential moving average} iterative procedures are commonly used in non-contrastive approaches to self-supervised learning to avoid representation collapse, with excellent performance in downstream applications in practice. This presentation investigates these procedures from the dual viewpoints of optimization and dynamical systems. We show that, in general, although they {\em do not} optimize the original objective, or {\em any} other smooth function, they {\em do} avoid collapse. Following [Tian et al. 2021], but without any of the extra assumptions used in their proofs, we then show using a dynamical system perspective that, in the linear case, minimizing the original objective function without the use of a stop gradient or exponential moving average {\em always} leads to collapse. Conversely, we characterize explicitly the equilibria of the dynamical systems associated with these two procedures in this linear setting as algebraic varieties in their parameter space, and show that they are, in general, {\em asymptotically stable}. Our theoretical findings are illustrated by empirical experiments with real and synthetic data.

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.

Unlike the foundation model scaling laws seen in natural language processing and computer vision, biological foundation models scale relatively poorly. For example, the ESM-2 family of protein language models plateaus at 650M-3B parameters on ProteinGym benchmarks. We address this limitation by introducing Reverse Distillation, a principled framework that decomposes large protein language model representations into orthogonal subspaces guided by smaller models of the same family. We hypothesize that this decomposition matches the natural hierarchy of protein properties, where broad features like secondary structure are robustly captured by compact, smaller models while the residual capacity of larger models specializes in protein-family specific functions. Our method is theoretically grounded and enables monotonic scaling---larger reverse-distilled models consistently outperform their smaller counterparts, overcoming the scaling plateau. Moreover, on ProteinGym benchmarks, reverse-distilled ESM-2 variants broadly outperform their respective baseline models at the same embedding dimensionality. Our approach offers a generalizable framework for disentangling hierarchical feature spaces in foundation model embeddings, with potential applications across biology and other domains where scaling challenges persist.

Category-Agnostic Pose Estimation (CAPE) localizes keypoints across diverse object categories with a single model, using one or few annotated support images. Recent works have shown that using a pose-graph (i.e., treating keypoints as nodes in a graph rather than isolated points) helps handle occlusions and break symmetry. However, these methods assume a given pose-graph with equal-weight edges, leading to suboptimal results. We introduce EdgeCape, a novel framework that overcomes these limitations by predicting the graph's edge weights in order to optimize localization. To further leverage structural (i.e., graph) priors, we propose integrating Markov Attention Bias, which modulates the self-attention interaction between nodes based on the number of hops between them. We show that this improves the model’s ability to capture global spatial dependencies. Evaluated on the MP-100 benchmark, which includes 100 categories and over 20K images, EdgeCape achieves state-of-the-art results in the 1-shot and 5-shot settings, significantly improving keypoint localization accuracy.

Neural network certification methods heavily rely on convex relaxations to provide robustness guarantees. However, these relaxations are often imprecise: even the most accurate single-neuron relaxation is incomplete for general ReLU networks, a limitation known as the *single-neuron convex barrier*. While multi-neuron relaxations have been heuristically applied to address this issue, two central questions arise: (i) whether they overcome the convex barrier, and if not, (ii) whether they offer theoretical capabilities beyond those of single-neuron relaxations. In this work, we present the first rigorous analysis of the expressiveness of multi-neuron relaxations. Perhaps surprisingly, we show that they are inherently incomplete, even when allocated sufficient resources to capture finitely many neurons and layers optimally. This result extends the single-neuron barrier to a *universal convex barrier* for neural network certification. On the positive side, we show that completeness can be achieved by either (i) augmenting the network with a polynomial number of carefully designed ReLU neurons or (ii) partitioning the input domain into convex sub-polytopes, thereby distinguishing multi-neuron relaxations from single-neuron ones which are unable to realize the former and have worse partition complexity for the latter. Our findings establish a foundation for multi-neuron relaxations and point to new directions for certified robustness, including training methods tailored to multi-neuron relaxations and verification methods with multi-neuron relaxations as the main subroutine.

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.

Noncontact exfiltration of electronic screen content poses a security challenge, with side-channel incursions as the principal vector. We introduce an optical projection side-channel paradigm that confronts two core instabilities: (i) the near-singular Jacobian spectrum of projection mapping breaches Hadamard stability, rendering inversion hypersensitive to perturbations; (ii) irreversible compression in light transport obliterates global semantic cues, magnifying reconstruction ambiguity. Exploiting passive speckle patterns formed by diffuse reflection, our Irradiance Robust Radiometric Inversion Network (IR$^4$Net) fuses a Physically Regularized Irradiance Approximation (PRIrr‑Approximation), which embeds the radiative transfer equation in a learnable optimizer, with a contour-to-detail cross-scale reconstruction mechanism that arrests noise propagation. Moreover, an Irreversibility Constrained Semantic Reprojection (ICSR) module reinstates lost global structure through context-driven semantic mapping. Evaluated across four scene categories, IR$^4$Net achieves fidelity beyond competing neural approaches while retaining resilience to illumination perturbations.

Chain-of-Thought (CoT) reasoning has advanced the capabilities and transparency of language models (LMs); however, reasoning chains can contain inaccurate statements that reduce performance and trustworthiness. To address this, we propose to augment each reasoning step in a CoT with a latent veracity (or correctness) variable. To efficiently explore this expanded space, we introduce Veracity Search (VS), a discrete search algorithm over veracity assignments. It performs otherwise intractable inference in the posterior distribution over latent veracity values by leveraging the LM's joint likelihood over veracity and the final answer as a proxy reward. This efficient inference-time verification method facilitates supervised fine-tuning of an Amortized Veracity Inference (AVI) machine by providing pseudo-labels for veracity. AVI generalizes VS, enabling accurate zero-shot veracity inference in novel contexts. Empirical results demonstrate that VS reliably identifies errors in logical (ProntoQA), mathematical (GSM8K), and commonsense (CommonsenseQA) reasoning benchmarks, with AVI achieving comparable zero-shot accuracy. Finally, we demonstrate the utility of latent veracity inference for providing feedback during self-correction and self-improvement.

Facility Location is a fundamental problem in clustering and unsupervised learning. Recently, significant attention has been given to studying this problem in the classical online setting enhanced with machine learning advice. While (almost) tight bounds exist for the fractional version of the problem, the integral version remains less understood, with only weaker results available. In this paper, we address this gap by presenting the first online rounding algorithms for the facility location problem, and by showing their applications to online facility location with machine learning advice.