ICLR 2026 Papers

We introduce k-NLPmeans and k-LLMmeans, text-clustering variants of k-means that periodically replace numeric centroids with textual summaries. The key idea—summary-as-centroid—retains k-means assignments in embedding space while producing human-readable, auditable cluster prototypes. The method is LLM-optional: k-NLPmeans uses lightweight, deterministic summarizers, enabling offline, low-cost, and stable operation; k-LLMmeans is a drop-in upgrade that uses an LLM for summaries under a fixed per-iteration budget whose cost does not grow with dataset size. We also present a mini-batch extension for real-time clustering of streaming text. Across diverse datasets, embedding models, and summarization strategies, our approach consistently outperforms classical baselines and approaches the accuracy of recent LLM-based clustering without extensive LLM calls. Finally, we provide a case study on sequential text streams and release a StackExchange-derived benchmark for evaluating streaming text clustering.

Self-supervised learning (SSL) has emerged as a powerful pretraining strategy to learn transferable representations from unlabeled data. Yet, it remains unclear how long SSL models should be pretrained for such representations to emerge. Contrary to the prevailing heuristic that longer pretraining translates to better downstream performance, we identify a transferability trade-off: across diverse SSL settings, intermediate checkpoints often yield stronger out-of-domain (OOD) generalization, whereas additional pretraining primarily benefits in-domain (ID) accuracy. From this observation, we hypothesize that SSL progresses through learning phases that can be characterized through the lens of critical periods (CP). Prior work on CP has shown that supervised learning models exhibit early phases of high plasticity, followed by a consolidation phase where adaptability declines but task-specific performance keeps increasing. Since traditional CP analysis depends on supervised labels, for SSL we rethink CP in two ways. First, we inject deficits to perturb the pretraining data and measure the quality of learned representations via downstream tasks. Second, to estimate network plasticity during pretraining we compute the Fisher Information matrix on pretext objectives, quantifying the sensitivity of model parameters to the supervisory signal defined by the pretext tasks. We conduct several experiments to demonstrate that SSL models do exhibit their own CP, with CP closure marking a sweet spot where representations are neither underdeveloped nor overfitted to the pretext task. Leveraging these insights, we propose CP-guided checkpoint selection as a mechanism for identifying intermediate checkpoints during SSL that improve OOD transferability. Finally, to balance the transferability trade-off, we propose CP-guided self-distillation, which selectively distills layer representations from the sweet spot (CP closure) checkpoint into their overspecialized counterparts in the final pretrained model.

The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models human social learning as joint probabilistic inference over structured, executable world models given sensorimotor and linguistic data. We make this possible by turning a pretrained language model into a probabilistic model of how humans share advice conditioned on their beliefs, allowing our agents both to generate advice for others and to interpret linguistic input as evidence during Bayesian inference. Using behavioral experiments and simulations across 10 video games, we show how linguistic guidance can shape exploration and accelerate learning by reducing risky interactions and speeding up key discoveries in both humans and models. We further explore how knowledge can accumulate across generations through iterated learning experiments and demonstrate successful knowledge transfer between humans and models—revealing how structured, language-compatible representations might facilitate human-machine collaborative learning.

Vision-language-action (VLA) models aim to understand natural language instructions and visual observations and execute corresponding actions as an embodied agent. Recent advancements have integrated future images into the understanding-action loop, enabling foresight-driven policies that reduce abstract action prediction to a more tractable inverse kinematics problem. However, existing models either rely on external experts for modality unification or treat image generation and action prediction as separate processes, limiting the benefits of direct synergy between these tasks. In this work, we propose Unified Diffusion VLAs, which tightly couple understanding, generation, and action in a mutually reinforcing manner. Our method optimizes the generation of actions and images jointly through a synchronous denoising diffusion process, where action tokens progressively attend to future image tokens. This iterative refinement enables actions to evolve from initialization with sufficient visual guidance, ensuring precise action execution. We introduce a hybrid attention mechanism and the Joint Discrete Denoising Diffusion Process (JD3P), which integrates multiple modalities into a unified trajectory. We also propose a two-stage training pipeline and several inference-time techniques that optimize performance and efficiency. Our approach achieves state-of-the-art performance on benchmarks such as CALVIN, LIBERO, and SimplerEnv, and we demonstrate its effectiveness through ablation studies and real-world evaluations.

Time series prediction underpins a broad range of downstream tasks across many scientific domains. Recent advances and increasing adoption of black-box machine learning models for time series prediction highlight the critical need for reliable uncertainty quantification. While conformal prediction has gained attention as a reliable uncertainty quantification method, conformal prediction for time series faces two key challenges: (1) adaptively leveraging correlations in features and non-conformity scores to overcome the exchangeability assumption, and (2) constructing prediction sets for multi-dimensional outcomes. To address these challenges jointly, we propose a novel conformal prediction method for time series using flow with classifier-free guidance. We provide coverage guarantees by establishing exact non-asymptotic marginal coverage and a finite-sample bound on conditional coverage for the proposed method. Evaluations on real-world time series datasets demonstrate that our method constructs significantly smaller prediction sets than existing conformal prediction methods while maintaining target coverage.

Reliable cognitive load (CL) prediction in real-world settings is fundamentally constrained by the poor and variable quality of physiological signals. In safety-critical tasks such as driving, degraded signal quality can severely compromise prediction accuracy, limiting the deployment of existing models outside controlled lab conditions. To address this challenge, we propose CogMoE, a signal quality–guided Mixture-of-Experts (MoE) framework that dynamically adapts to heterogeneous and noisy inputs. CogMoE flexibly integrates physiological modalities, including EEG, ECG, EDA, and gaze, through quality-aware gating, enabling context-sensitive fusion. The framework operates in two stages: (1) quality-aware multi-modal synchronization and recovery to mitigate artifacts, temporal misalignment, and missing data, and (2) signal-quality-specific expert modeling via a cross-modal MoE transformer that regulates information flow based on signal reliability. To further improve stability, we introduce CORTEX Loss, which balances reconstruction fidelity and expert utilization under noise. Experiments on CL-Drive and ADABase show that CogMoE outperforms strong baselines, delivering consistent improvements across diverse signal qualities.

Achieving streaming, fine-grained control over the outputs of autoregressive video diffusion models remains challenging, making it difficult to ensure that they consistently align with user expectations. To bridge this gap, we propose \textbf{stReaming drag-oriEnted interactiVe vidEo manipuLation (REVEL)}, a new task that enables users to modify generated videos \emph{anytime} on \emph{anything} via fine-grained, interactive drag. Beyond DragVideo and SG-I2V, REVEL unifies drag-style video manipulation as editing and animating video frames with both supporting user-specified translation, deformation, and rotation effects, making drag operations versatile. In resolving REVEL, we observe: \emph{i}) drag-induced perturbations accumulate in latent space, causing severe latent distribution drift that halts the drag process; \emph{ii}) streaming drag is easily disturbed by context frames, thereby yielding visually unnatural outcomes. We thus propose a training-free approach, \textbf{DragStream}, comprising: \emph{i}) an adaptive distribution self-rectification strategy that leverages neighboring frames' statistics to effectively constrain the drift of latent embeddings; \emph{ii}) a spatial-frequency selective optimization mechanism, allowing the model to fully exploit contextual information while mitigating its interference via selectively propagating visual cues along generation. Our method can be seamlessly integrated into existing autoregressive video diffusion models, and extensive experiments firmly demonstrate the effectiveness of our DragStream.

Scaling Transformers to ultra-long contexts is bottlenecked by the $O(n^2 d)$ cost of self-attention. Existing methods reduce this cost along the sequence axis through local windows, kernel approximations, or token-level sparsity, but these approaches consistently degrade accuracy. In this paper, we instead explore an orthogonal axis: \emph{feature sparsity}. We propose \textbf{Sparse Feature Attention (SFA)}, where queries and keys are represented as $k$-sparse codes that preserve high-dimensional expressivity while reducing the cost of attention from $\Theta(n^2 d)$ to $\Theta(n^2 k^2/d)$. To make this efficient at scale, we introduce \textbf{FlashSFA}, an IO-aware kernel that extends FlashAttention to operate directly on sparse overlaps without materializing dense score matrices. Across GPT-2 and Qwen3 pretraining, SFA matches dense baselines while improving speed by up to $2.5\times$ and reducing FLOPs and KV-cache by nearly 50\%. On synthetic and downstream benchmarks, SFA preserves retrieval accuracy and robustness at long contexts, outperforming short-embedding baselines that collapse feature diversity. These results establish feature-level sparsity as a complementary and underexplored axis for efficient attention, enabling Transformers to scale to orders-of-magnitude longer contexts with minimal quality loss.

We present the first comprehensive evaluation of AI agents against human cybersecurity professionals in a live enterprise environment. We evaluate ten cybersecurity professionals alongside six existing AI agents and ARTEMIS, our new agent scaffold, on a large university network consisting of $\sim$8,000 hosts across 12 subnets. ARTEMIS is a multi-agent framework featuring dynamic prompt generation, arbitrary sub-agents, and automatic vulnerability triaging. In our comparative study, ARTEMIS placed second overall, discovering 9 valid vulnerabilities with an 82\% valid submission rate and outperforming 9 of 10 human participants. While existing scaffolds such as Codex and CyAgent underperformed relative to most human participants, ARTEMIS demonstrated technical sophistication and submission quality comparable to the strongest participants. AI agents offer advantages in systematic enumeration, parallel exploitation, and cost---certain ARTEMIS variants cost $18/hour versus $60/hour for professional penetration testers. We also identify key capability gaps: AI agents exhibit higher false-positive rates and struggle with GUI-based tasks.

Large Language Models (LLMs) often struggle with challenging, multi-step reasoning problems due to a fundamental learning gap -- Reinforcement Learning with Verifiable Rewards (RLVR) suffers from sparse rewards when correct solutions are rarely sampled, while Supervised Fine-Tuning (SFT) tends to overfit to long demonstrations through rigid token mimicry. To bridge this gap, we introduce Supervised Reinforcement Learning (SRL), a framework that reformulates problem-solving as a sequence of logical actions. SRL trains the model to learn from each action, where the model first generates an internal reasoning monologue and then commits to an action. The model receives dense rewards based on the similarity between its actions and the expert’s at each step, providing a richer signal than RLVR. More importantly, by only rewarding the action, SRL allows the model flexibility in its self-generated thought process, promoting stronger reasoning abilities than SFT. On challenging mathematical reasoning benchmarks, SRL significantly outperforms both methods. Furthermore, a curriculum that cold-start with SRL before refining with RLVR achieves the strongest results. SRL also generalizes effectively to agentic software engineering tasks, establishing it as a robust framework for various reasoning tasks.

Humans ponder before articulating complex sentence elements, enabling deeper cognitive processing through focused effort. In this work, we introduce this pondering process into language models by repeatedly invoking the forward process within a single token generation step. During pondering, instead of generating an actual token sampled from the prediction distribution, the model ponders by yielding a weighted sum of all token embeddings according to the predicted token distribution. The generated embedding is then fed back as input for another forward pass. We show that the model can learn to ponder in this way through self-supervised learning, without any human annotations. Experiments across three widely used open-source architectures—GPT-2, Pythia, and LLaMA—and extensive downstream task evaluations demonstrate the effectiveness and generality of our method. On 9 downstream benchmarks, our pondering-enhanced Pythia models significantly outperform the official Pythia models. Notably, our PonderPythia models demonstrate remarkable effectiveness: PonderPythia-2.8B surpasses Pythia-6.9B and rivals Pythia-12B, while our PonderPythia-1B matches TinyLlama-1.1B, a model trained on 10 times more data.

Flow-matching policies hold great promise for reinforcement learning (RL) by capturing complex, multi-modal action distributions. However, their practical application is often hindered by prohibitive inference latency and ineffective online exploration. Although recent works have employed one-step distillation for fast inference, the structure of the initial noise distribution remains an overlooked factor that presents significant untapped potential. This overlooked factor, along with the challenge of controlling policy stochasticity, constitutes two critical areas for advancing distilled flow-matching policies. To overcome these limitations, we propose GoldenStart (GS-flow), a policy distillation method with Q-guided priors and explicit entropy control. Instead of initializing generation from uninformed noise, we introduce a Q-guided prior modeled by a conditional VAE. This state-conditioned prior repositions the starting points of the one-step generation process into high-Q regions, effectively providing a ``golden start'' that shortcuts the policy to promising actions. Furthermore, for effective online exploration, we enable our distilled actor to output a stochastic distribution instead of a deterministic point. This is governed by entropy regularization, allowing the policy to shift from pure exploitation to principled exploration. Our integrated framework demonstrates that by designing the generative startpoint and explicitly controlling policy entropy, it is possible to achieve efficient and exploratory policies, bridging the generative models and the practical actor-critic methods. We conduct extensive experiments on offline and online continuous control benchmarks, where our method significantly outperforms prior state-of-the-art approaches.

Vision-Language Models (VLMs), such as CLIP, have shown strong zero-shot generalization but remain highly vulnerable to adversarial perturbations, posing serious risks in real-world applications. Test-time defenses for VLMs have recently emerged as a promising and efficient approach to defend against adversarial attacks without requiring costly large-scale retraining. In this work, we uncover a surprising phenomenon: under diverse input transformations, adversarial images in CLIP’s feature space consistently shift along a dominant direction, in contrast to the dispersed patterns of clean images. We hypothesize that this dominant shift, termed the Defense Direction, opposes the adversarial shift, pointing features back toward their correct class centers. Building on this insight, we propose Directional Bias-guided Defense (DBD), a test-time framework that estimates the Defense Direction and employs a DB-score–based two-stream reconstruction strategy to recover robust representations. Experiments on 15 datasets demonstrate that DBD not only achieves SOTA adversarial robustness while preserving clean accuracy, but also reveals the counterintuitive result that adversarial accuracy can even surpass clean accuracy. This demonstrates that adversarial perturbations inherently encode directional priors about the true decision boundary.

While search-augmented large language models (LLMs) exhibit impressive capabilities, their reliability in complex multi-hop reasoning remains limited. This limitation arises from three fundamental challenges: decomposition errors, where tasks are incorrectly broken down; retrieval missing, where key evidence fails to be retrieved; and reasoning errors, where flawed logic propagates through the reasoning chain. A single failure in any of these stages can derail the final answer. We propose Erasable Reinforcement Learning (ERL), a novel framework that transforms fragile reasoning into a robust process. ERL explicitly identifies faulty steps, erases them, and regenerates reasoning in place, preventing defective logic from propagating through the reasoning chain. This targeted correction mechanism turns brittle reasoning into a more resilient process. Models trained with ERL, termed ESearch, achieve substantial improvements on HotpotQA, MuSiQue, 2Wiki, and Bamboogle, with the 3B model achieving +8.48% EM and +11.56% F1, and the 7B model achieving +5.38% EM and +7.22% F1 over previous state-of-the-art(SOTA) results. These findings suggest that erasable reinforcement learning provides a powerful paradigm shift for robust multi-step reasoning in LLMs.

Predicting molecule-protein interactions (MPIs) is a fundamental task in computational biology, with crucial applications in drug discovery and molecular function annotation. However, existing MPI models face two major challenges. First, the scarcity of labeled molecule-protein pairs significantly limits model performance, as available datasets capture only a small fraction of biological relevant interactions. Second, most methods rely solely on molecular and protein features, ignoring broader biological context—such as genes, metabolic pathways, and functional annotations—that could provide essential complementary information. To address these limitations, our framework first aggregates diverse biological datasets, including molecular, protein, genes and pathway-level interactions, and then develop an optimal transport-based approach to generate high-quality pseudo-labels for unlabeled molecule-protein pairs, leveraging the underlying distribution of known interactions to guide label assignment. By treating pseudo-labeling as a mechanism for bridging disparate biological modalities, our approach enables the effective use of heterogeneous data to enhance MPI prediction. We evaluate our framework on multiple MPI datasets including virtual screening tasks and protein retrieval tasks, demonstrating substantial improvements over state-of-the-art methods in prediction accuracies and zero shot ability across unseen interactions. Beyond MPI prediction, our approach provides a new paradigm for leveraging diverse biological data sources to tackle problems traditionally constrained by single or bi-modal learning, paving the way for future advances in computational biology and drug discovery.

Unsupervised pre-training can equip reinforcement learning agents with prior knowledge and accelerate learning in downstream tasks. A promising direction, grounded in human development, investigates agents that learn by setting and pursuing their own goals. The core challenge lies in how to effectively generate, select, and learn from such goals. Our focus is on broad distributions of downstream tasks where solving every task zero-shot is infeasible. Such settings naturally arise when the target tasks lie outside of the pre-training distribution or when their identities are unknown to the agent. In this work, we (i) optimize for efficient multi-episode exploration and adaptation within a meta-learning framework, and (ii) guide the training curriculum with evolving estimates of the agent’s post-adaptation performance. We present ULEE, an unsupervised meta-learning method that combines an in-context learner with an adversarial goal-generation strategy that maintains training at the frontier of the agent’s capabilities. On XLand-MiniGrid benchmarks, ULEE pre-training yields improved exploration and adaptation abilities that generalize to novel objectives, environment dynamics, and map structures. The resulting policy attains improved zero-shot and few-shot performance, and provides a strong initialization for longer fine-tuning processes. It outperforms learning from scratch, DIAYN pre-training, and alternative curricula.

We address the 3D reconstruction of human faces from a single RGB image. To this end, we propose Pixel3DMM, a set of highly-generalized vision transformers which predict per-pixel geometric cues in order to constrain the optimization of a 3D morphable face model (3DMM). We exploit the latent features of the DINO foundation model, and introduce a tailored surface normal and uv-coordinate prediction head. We train our model by registering three high-quality 3D face datasets against the FLAME mesh topology, which results in a total of over 1,000 identities and 976K images. For 3D face reconstruction, we propose a FLAME fitting opitmization that solves for the 3DMM parameters from the uv-coordinate and normal estimates. To evaluate our method, we introduce a new benchmark for single-image face reconstruction, which features high diversity facial expressions, viewing angles, and ethnicities. Crucially, our benchmark is the first to evaluate both posed and neutral facial geometry. Ultimately, our method outperforms the state-of-the-art (SoTA) by over 15\% in terms of geometric accuracy for posed facial expressions.

Offline reinforcement learning (RL) is a powerful approach for data-driven decision-making and control. Compared to model-free methods, offline model-based reinforcement learning (MBRL) explicitly learns world models from a static dataset and uses them as surrogate simulators, improving the data efficiency and enabling the learned policy to potentially generalize beyond the dataset support. However, there could be various MDPs that behave identically on the offline dataset and dealing with the uncertainty about the true MDP can be challenging. In this paper, we propose modeling offline MBRL as a Bayes Adaptive Markov Decision Process (BAMDP), which is a principled framework for addressing model uncertainty. We further propose a novel Bayes Adaptive Monte-Carlo planning algorithm capable of solving BAMDPs in continuous state and action spaces with stochastic transitions. This planning process is based on Monte Carlo Tree Search and can be integrated into offline MBRL as a policy improvement operator in policy iteration. Our "RL + Search" framework follows in the footsteps of superhuman AIs like AlphaZero, improving on current offline MBRL methods by incorporating more computation input. The proposed algorithm significantly outperforms state-of-the-art offline RL methods on twelve D4RL MuJoCo tasks and three challenging, stochastic tokamak control tasks.

Conformal prediction provides prediction sets with coverage guarantees. The informativeness of conformal prediction depends on its efficiency, typically quantified by the expected size of the prediction set. Prior work on the efficiency of conformalized regression commonly treats the miscoverage level $\alpha$ as a fixed constant. In this work, we establish non-asymptotic bounds on the deviation of the prediction set length from the oracle interval length for conformalized quantile and median regression trained via SGD, under mild assumptions on the data distribution. Our bounds of order $\mathcal{O}(1/\sqrt{n} + 1/(\alpha^2 n) + 1/\sqrt{m} + \exp(-\alpha^2 m))$ capture the joint dependence of efficiency on the proper training set size $n$, the calibration set size $m$, and the miscoverage level $\alpha$. The results identify phase transitions in convergence rates across different regimes of $\alpha$, offering guidance for allocating data to control excess prediction set length. Empirical results are consistent with our theoretical findings.