ICLR 2026 论文列表

Large language models deployed in sensitive applications increasingly require the ability to unlearn specific knowledge, such as user requests, copyrighted materi- als, or outdated information, without retraining from scratch to ensure regulatory compliance, user privacy, and safety. This task, known as machine unlearning, aims to remove the influence of targeted data (forgetting) while maintaining per- formance on the remaining data (retention). A common approach is to formu- late this as a multi-objective problem and reduce it to a single-objective prob- lem via scalarization, where forgetting and retention losses are combined using a weighted sum. However, this often results in unstable training dynamics and degraded model utility due to conflicting gradient directions. To address these challenges, we propose OFMU, a penalty-based bi-level optimization framework that explicitly prioritizes forgetting while preserving retention through a hierar- chical structure. Our method enforces forgetting via an inner maximization step that incorporates a similarity-aware penalty to decorrelate the gradients of the for- get and retention objectives, and restores utility through an outer minimization step. To ensure scalability, we develop a two-loop algorithm with provable conver- gence guarantees under both convex and non-convex regimes. We further provide a rigorous theoretical analysis of convergence rates and show that our approach achieves better trade-offs between forgetting efficacy and model utility compared to prior methods. Extensive experiments across vision and language benchmarks demonstrate that OFMU consistently outperforms existing unlearning methods in both forgetting efficacy and retained utility.

Accurate recall from large-scale memories remains a core challenge for memory-augmented AI assistants performing question answering (QA), especially in similarity-dense scenarios where existing methods mainly rely on semantic distance to the query for retrieval. Inspired by how humans link information associatively, we propose AssoMem, a novel framework constructing an associative memory graph that anchors dialogue utterances to automatically extracted clues. This structure provides a rich organizational view of the conversational context and facilitates importance-aware ranking. Further, AssoMem integrates multi-dimensional retrieval signals—relevance, importance, and temporal alignment—using an adaptive mutual information (MI)-driven fusion strategy. Extensive experiments across three benchmarks and a newly introduced dataset, MeetingQA, demonstrate that AssoMem consistently outperforms state-of-the-art baselines, verifying its superiority in context-aware memory recall.

Neural scaling laws are widely used for performance projection and resource planning, yet their sensitivity to data quality interventions remains poorly understood. We present an empirical study of how interventions—deduplication, heuristic filtering, and LLM-guided rewriting—reshape scaling behavior in large language model training. Using QualityPajama, a suite of 23 systematically filtered and synthetic datasets, we train over 2,000 models (100M–8B parameters, 100M–200B tokens) to measure how data quality affects scaling-law parameters and compute-optimal design decisions. Our results show that data interventions reshape scaling dynamics in non-trivial ways not captured by current theory, simultaneously moving exponents, coefficients, and constants in conflicting directions that exert opposing forces on loss. For example, an intervention may improve constants but hurt the exponents. Strategies that appear optimal at small scale can reverse at larger scale, and compute-optimal token–parameter ratios can vary by orders of magnitude depending on the intervention. These findings demonstrate that data curation and scaling strategy are deeply intertwined, and that evaluating interventions only at fixed scales can lead to misleading conclusions. We recommend evaluating interventions through their full scaling trajectories using scaling law projections.

Detecting missing persons in forest environments remains a challenge, as dense canopy cover often conceals individuals from detection in top-down or oblique aerial imagery typically captured by Unmanned Aerial Vehicles (UAVs). While UAVs are effective for covering large, inaccessible areas, their aerial perspectives often miss critical visual cues beneath the forest canopy. This limitation underscores the need for under-canopy perspectives better suited for detecting missing persons in such environments. To address this gap, we introduce ForestPersons, a novel large-scale dataset specifically designed for under-canopy person detection. ForestPersons contains 96,482 images and 204,078 annotations collected under diverse environmental and temporal conditions. Each annotation includes a bounding box, pose, and visibility label for occlusion-aware analysis. ForestPersons provides ground-level and low-altitude perspectives that closely reflect the visual conditions encountered by Micro Aerial Vehicles (MAVs) during forest Search and Rescue (SAR) missions. Our baseline evaluations reveal that standard object detection models, trained on prior large-scale object detection datasets or SAR-oriented datasets, show limited performance on ForestPersons. This indicates that prior benchmarks are not well aligned with the challenges of missing person detection under the forest canopy. We offer this benchmark to support advanced person detection capabilities in real-world SAR scenarios. The dataset is publicly available at https://huggingface.co/datasets/anonreviewer2026/under-canopy-benchmark-anon.

Inference-time computation, analogous to human System 2 Thinking, has recently become popular for improving model performance. However, most existing approaches suffer from several limitations: they are modality-specific (e.g., working only in text), problem-specific (e.g., verifiable domains like math and coding), or require additional supervision/training on top of unsupervised pretraining (e.g., verifiers or verifiable rewards). In this paper, we ask the question “Is it possible to generalize these System 2 Thinking approaches, and develop models that learn to think solely from unsupervised learning?” We find the answer is yes, by learning to explicitly verify the compatibility between inputs and candidate-predictions, and then re-framing prediction problems as optimization with respect to this verifier. Specifically, we train Energy-Based Transformers (EBTs)---a new class of Energy-Based Models (EBMs)---to assign an energy value to every input and candidate-prediction, enabling predictions through energy minimization until convergence. To support this approach, we introduce several key techniques for stable and parallelizable training, which enable the emergence of strong System 2 Thinking capabilities and scalable EBMs. Across discrete and continuous modalities, we find EBTs outperform the Transformer++ approach, scaling up to 35% faster during pretraining, and improving inference-time performance by up to 29%. EBTs also surpass Diffusion Transformers on image denoising while requiring 99% fewer forward passes. Moreover, System 2 Thinking with EBTs yields larger performance gains on data that is farther out-of-distribution, and EBTs achieve better results than existing models on most downstream tasks despite achieving the same or worse pretraining performance, enabling EBTs to generalize better than existing approaches. Consequently, EBTs are a flexible and exciting new approach for scaling both the learning and thinking capabilities of models.

Accurately estimating the Q-function is a central challenge in offline reinforcement learning. However, existing approaches often rely on a single global Q-function, which struggles to capture the compositional nature of tasks involving diverse subtasks. We propose In-context Compositional Q-Learning ($\texttt{ICQL}$), the first offline RL framework that formulates Q-learning as a contextual inference problem, using linear Transformers to adaptively infer local Q-functions from retrieved transitions without explicit subtask labels. Theoretically, we show that under two assumptions—linear approximability of the local Q-function and accurate weight inference from retrieved context—$\texttt{ICQL}$ achieves bounded Q-function approximation error, and supports near-optimal policy extraction. Empirically, $\texttt{ICQL}$ substantially improves performance in offline settings: improving performance in Kitchen tasks by up to 29. 46\%, and in Gym and Adroit tasks by up to 6\%. These results highlight the underexplored potential of in-context learning for robust and compositional value estimation, positioning $\texttt{ICQL}$ as a principled and effective framework for offline RL.

The tendency of users to anthropomorphise large language models (LLMs) is of growing societal interest. Here, we present AnthroBench: a novel empirical method and tool for evaluating anthropomorphic LLM behaviours in realistic settings. Our work introduces three key advances; first, we develop a multi-turn evaluation of 14 distinct anthropomorphic behaviours, moving beyond single-turn assessment. Second, we present a scalable, automated approach by leveraging simulations of user interactions, enabling efficient and reproducible assessment. Third, we conduct an interactive, large-scale human subject study (N=1101) to empirically validate that the model behaviours we measure predict real users’ anthropomorphic perceptions. We find that all evaluated LLMs exhibit similar behaviours, primarily characterised by relationship-building (e.g., empathy and validation) and first-person pronoun use. Crucially, we observe that the majority of these anthropomorphic behaviors only first occur after multiple turns, underscoring the necessity of multi-turn evaluations for understanding complex social phenomena in human-AI interaction. Our work provides a robust empirical foundation for investigating how design choices influence anthropomorphic model behaviours and for progressing the ethical debate on the desirability of these behaviours.

A recent breakthrough in nonconvex optimization is the online-to-nonconvex conversion framework of Cutkosky et al. (2023), which reformulates the task of finding an $\varepsilon$-first-order stationary point as an online learning problem. When both the gradient and the Hessian are Lipschitz continuous, instantiating this framework with two different online learners achieves a complexity of $ \mathcal{O}(\varepsilon^{-1.75}\log(1/\varepsilon)) $ in the deterministic case and a complexity of $ \mathcal{O}(\varepsilon^{-3.5}) $ in the stochastic case. However, this approach suffers from several limitations: (i) the deterministic method relies on a complex double-loop scheme that solves a fixed-point equation to construct hint vectors for an optimistic online learner, introducing an extra logarithmic factor; (ii) the stochastic method assumes a bounded second-order moment of the stochastic gradient, which is stronger than standard variance bounds; and (iii) different online learning algorithms are used in the two settings. In this paper, we address these issues by introducing an online optimistic gradient method based on a novel **doubly optimistic hint function**. Specifically, we use the gradient at an extrapolated point as the hint, motivated by two optimistic assumptions: that the difference between the hint and the target gradient remains near constant, and that consecutive update directions change slowly due to smoothness. Our method eliminates the need for a double loop and removes the logarithmic factor. Furthermore, by simply replacing full gradients with stochastic gradients and under the standard assumption that their variance is bounded by $\sigma^2$, we obtain a unified algorithm with complexity $\mathcal{O}(\varepsilon^{-1.75} + \sigma^2 \varepsilon^{-3.5})$, smoothly interpolating between the best-known deterministic rate and the optimal stochastic rate.

Personalized alignment is crucial for enabling Large Language Models (LLMs) to engage effectively in user-centric interactions. However, current methods face a dual challenge: they fail to infer users' deep implicit preferences (including unstated goals, semantic context and risk tolerances), and they lack the defensive reasoning required to navigate real-world ambiguity. This cognitive gap leads to responses that are superficial, brittle and short-sighted. To address this, we propose Critique-Driven Reasoning Alignment (CDRA), which reframes alignment from a scalar reward-matching task into a structured reasoning process. First, to bridge the preference inference gap, we introduce the DeepPref benchmark. This dataset, comprising 3000 preference-query pairs across 20 topics, is curated by simulating a multi-faceted cognitive council that produces critique-annotated reasoning chains to deconstruct query semantics and reveal latent risks. Second, to instill defensive reasoning, we introduce the Personalized Generative Process Reward Model (Pers-GenPRM), which frames reward modeling as a personalized reasoning task. It generates a critique chain to evaluate a response's alignment with user preferences before outputting a final score based on this rationale. Ultimately, this interpretable, structured reward signal guides policy model through Critique-Driven Policy Alignment, a process-level online reinforcement learning algorithm integrating both numerical and natural language feedback. Experiments demonstrate that CDRA excels at discovering and aligning with users' true preferences while executing robust reasoning. Our code and dataset are available at https://anonymous.4open.science/r/Deep-pref-9DE9.

Visual generation models have made remarkable progress in creating realistic images from text prompts, yet struggle with complex prompts that specify multiple objects with precise spatial relationships and attributes. Effective handling of such prompts requires explicit reasoning about the semantic content and spatial layout. We present GoT-R1, a framework that applies reinforcement learning to enhance semantic-spatial reasoning in autoregressive visual generation models. Leveraging the natural affinity between autoregressive architectures and sequential reasoning, our approach builds upon the Generation Chain-of-Thought framework to enable models to autonomously discover effective reasoning strategies beyond predefined templates. To achieve this, we propose a dual-stage multi-dimensional reward framework that leverages MLLMs to evaluate both the reasoning process and final output, enabling effective supervision across the entire generation pipeline. The reward system assesses semantic alignment, spatial accuracy, and visual quality in a unified approach. Experimental results demonstrate significant improvements on T2I-CompBench and GenEval benchmark, particularly in compositional tasks involving precise spatial relationships and attribute binding. GoT-R1 advances the state-of-the-art in autoregressive image generation by successfully transferring sophisticated reasoning capabilities from language models to the visual generation domain. Code is available at https://github.com/gogoduan/GoT-R1.

Graph outlier detection is a critical task for identifying rare, deviant patterns in graph-structured data. However, prevalent methods based on graph convolution are fundamentally challenged by the ''Homophily Trap'': the aggregation of features from neighboring nodes inadvertently contaminates the representations of normal nodes near anomalies, blurring their distinctions. To overcome this limitation, we propose a Clustering-guided Edge Reweighting framework for Graph Outlier Detection (CER-GOD), which jointly optimizes a self-discriminative masking spoiler with an adaptive clustering-based outlier detector. The masking spoiler learns to selectively weaken the influence of heterogeneous neighbors, preserving the discriminative power of node embeddings. This process is guided by the clustering detector, which generates pseudo-labels in an unsupervised manner, thereby eliminating the need for predefined anomaly thresholds. To ensure robust optimization and prevent class collapse—a failure mode exacerbated by the homophily trap—we introduce a diversity loss that stabilizes the clustering process. Our end-to-end framework demonstrates superior performance on multiple benchmark datasets, establishing a new state-of-the-art by effectively dismantling the homophily trap.

Learned Sparse Retrieval (LSR) combines the efficiency of bi-encoders with the transparency of lexical matching, but existing approaches struggle to scale beyond English. We introduce MILCO, an LSR architecture that maps queries and documents from different languages into a shared English lexical space via a multilingual connector. MILCO is trained with a specialized two-stage regime that combines Sparse Alignment Pretraining with contrastive training to provide representation transparency and effectiveness while mitigating semantic collapse. Motivated by the observation that uncommon entities are often lost when projected into English, we propose a new LexEcho head, which enhances robustness by augmenting the English lexical representation with a source-language view obtained through a special [ECHO] token. MILCO achieves state-of-the-art multilingual and cross-lingual LSR performance, outperforming leading dense, sparse, and multi-vector baselines such as BGE-M3 and Qwen3-Embed on standard multilingual benchmarks, while supporting dynamic efficiency through post-hoc pruning. Notably, when using mass-based pruning to reduce document representations to only 30 active dimensions on average, MILCO 560M outperforms the similarly-sized Qwen3-Embed 0.6B with 1024 dimensions, , while achieving 3$\times$ lower retrieval latency and 10$\times$ smaller index size.

Adapters have become a widely adopted strategy for efficient fine-tuning of foundation models, particularly in resource-constrained settings. However, their performance under extreme data scarcity—common in medical imaging due to high annotation costs, privacy regulations, and fragmented datasets—remains underexplored. In this work, we present the first comprehensive study of adapter-based fine-tuning for vision foundation models in low-data medical imaging scenarios. We find that, contrary to their promise, conventional Adapters can degrade performance under severe data constraints, performing even worse than simple linear probing when trained on less than 1\% of the corresponding training data. Through systematic analysis, we identify a sharp reduction in Effective Receptive Field (ERF) as a key factor behind this degradation. Motivated by these findings, we propose the Dual-Kernel Adapter (DKA), a lightweight module that expands spatial context via large-kernel convolutions while preserving local detail with small-kernel counterparts. Extensive experiments across diverse classification and segmentation benchmarks show that DKA significantly outperforms existing Adapter methods, establishing new leading results in both data-constrained and data-rich regimes.

Sampling efficiency is a key bottleneck in reinforcement learning with verifiable rewards. Existing group-based policy optimization methods, such as GRPO, allocate a fixed number of rollouts for all training prompts. This uniform allocation implicitly treats all prompts as equally informative, and could lead to inefficient computational budget usage and impede training progress. We introduce VIP, a Variance-Informed Predictive allocation strategy that allocates a given rollout budget to the prompts in the incumbent batch to minimize the expected gradient variance of the policy update. At each iteration, VIP uses a lightweight Gaussian process model to predict per-prompt success probabilities based on recent rollouts. These probability predictions are translated into variance estimates, which are then fed into a convex optimization problem to determine the optimal rollout allocations under a hard compute budget constraint. Empirical results show that VIP consistently improves sampling efficiency and achieves higher performance than uniform or heuristic allocation strategies in multiple benchmarks.

Simulated environments are a key piece in the success of Reinforcement Learning (RL), allowing practitioners and researchers to train decision making agents without running expensive experiments on real hardware. Simulators remain a security blind spot, however, enabling adversarial developers to alter the dynamics of their released simulators for malicious purposes. Therefore, in this work we highlight a novel threat, demonstrating how simulator dynamics can be exploited to stealthily implant action-level backdoors into RL agents. The backdoor then allows an adversary to reliably activate targeted actions in an agent upon observing a predefined "trigger", leading to potentially dangerous consequences. Traditional backdoor attacks are limited in their strong threat models, assuming the adversary has near full control over an agent's training pipeline, enabling them to both alter and observe agent's rewards. As these assumptions are infeasible to implement within a simulator, we propose a new attack "Daze" which is able to reliably and stealthily implant backdoors into RL agents trained for real world tasks without altering or even observing their rewards. We provide formal proof of Daze's effectiveness in guaranteeing attack success across general RL tasks along with extensive empirical evaluations on both discrete and continuous action space domains. We additionally provide the first example of RL backdoor attacks transferring to real, robotic hardware. These developments motivate further research into securing all components of the RL training pipeline to prevent malicious attacks.

Self-supervised pre-training has emerged as a critical paradigm for learning transferable representations from unlabeled medical volumetric data. Masked autoencoder based methods have garnered significant attention, yet their application to volumetric medical image faces fundamental limitations from the discrete voxel-level reconstruction objective, which neglects comprehensive anatomical structure continuity. To address this challenge, We propose MedGMAE, a novel framework that replaces traditional voxel reconstruction with 3D Gaussian primitives reconstruction as new perspectives on representation learning. Our approach learns to predict complete sets of 3D Gaussian parameters as semantic abstractions to represent the entire 3D volume, from sparse visible image patches. MedGMAE demonstrates dual utility across medical imaging applications. For representation learning, sparse Gaussian prediction produces superior encoder representations that outperform traditional MAE baselines on downstream segmentation, classification, and registration tasks. For volumetric reconstruction, the Gaussian decoder leverages pretrained anatomical priors to accelerate 3D CT volume reconstruction convergence. Extensive experiments across multiple medical imaging datasets demonstrate that our approach achieves superior performance, establishing a new paradigm for medical image pre-training. The code will be available in https://github.com/windrise/MedGMAE.

Large Vision-Language Models (LVLMs) have demonstrated remarkable multimodal capabilities, but they inherit the tendency to hallucinate from their underlying language models. While visual contrastive decoding has been proposed to mitigate this issue, existing methods often apply generic visual augmentations that disregard the specific context provided by the text query, limiting their effectiveness. This study introduces a novel training-free decoding strategy that addresses these limitations, featuring two key contributions. First, a self-augmentation prompting strategy that leverages the intrinsic knowledge of the model to dynamically align semantics between the query and the visual augmentation. Second, an adaptive thresholding algorithm that adaptively adjusts next token candidate size based on the output sparsity, utilizing full information from the logit distribution. Extensive experiments across four LVLMs and seven benchmarks demonstrate that the proposed decoding significantly enhances factual consistency compared to state-of-the-art decoding methods. This work highlights the importance of integrating query-dependent augmentation and entropy-aware decoding for improving effective generation of LVLMs. The source code will be released upon acceptance.

Prior work suggests that language models, while trained on next token prediction, show implicit planning behavior: they may select the next token in preparation to a predicted future token, such as a likely rhyming word, as supported by a prior qualitative study of Claude 3.5 Haiku using a cross-layer transcoder. We propose much simpler techniques for assessing implicit planning in language models. With case studies on rhyme poetry generation and question answering, we demonstrate that our methodology easily scales to many models. Across models, we find that the generated rhyme (e.g. "-ight") or answer to a question ("whale") can be manipulated by steering at the end of the preceding line with a vector, affecting the generation of intermediate tokens leading up to the rhyme or answer word. We show that implicit planning is a universal mechanism, present in smaller models than previously thought, starting from 1B parameters. Our methodology offers a widely applicable direct way to study implicit planning abilities of LLMs. More broadly, understanding planning abilities of language models can inform decisions in AI safety and control.

Efficiently editing knowledge stored in Large Language Models (LLMs) enables model updates without large-scale training. One promising solution is Locate-and-Edit (L\&E), allowing simultaneous modifications of a massive number of factual knowledge. However, such editing may compromise the general abilities of LLMs and even result in forgetting edited facts when scaling up to thousands of edits. In this paper, we model existing linear L\&E methods as querying a Key-Value (KV) database. From this perspective, we then propose NeuralDB, an editing framework that explicitly represents the edited facts as a neural KV database equipped with a non-linear gated retrieval module. With simple modification over L\&E methods, our framework not only significantly extends the capacity of knowledge editing but also eliminates the associated side effects. Comprehensive experiments involving the editing of 10,000 facts were conducted on the ZsRE and CounterFact datasets, including GPT2-XL, GPT-J (6B) and Llama-3 (8B). The results demonstrate that NeuralDB excels in all metrics of editing success while maintaining original performance evaluated by six representative text understanding and generation tasks. Further experiments indicate that NeuralDB maintains its effectiveness even when scaled to 100,000 facts (\textbf{50}$\mathbf{\times}$ more than in prior work).