ICLR 2026 Papers

Large vision-language models (VLMs) achieve strong performance in Visual Question Answering but still rely heavily on supervised fine-tuning (SFT) with massive labeled datasets, which is costly due to human annotations. Crucially, real-world datasets often exhibit *human uncertainty* (**HU**) — variation in human confidence across annotations, but standard SFT simply optimizes toward the most frequent label, disregarding HU distributions. This leaves two open questions: *How does HU affect SFT*, and *how can HU be effectively leveraged in training?* In this work, we first conduct a systematic evaluation of VLMs across varying HU levels. We have two key findings: (i) surprisingly, high-HU samples contribute little, or even degrade, model performance, and (ii) naively training on the full dataset yields under-calibrated models that fail to capture HU distributions. Motivated by these findings, we introduce **HaDola**, a **h**uman uncertainty-**a**ware **d**ata selection and aut**o**matic **la**beling framework. HaDola operates in four stages: **discriminate**, **self-annotate**, **error trigger**, and **training**, to iteratively identify harmful samples, prioritize informative ones, and bootstrap from a small seed set (5% of data). Our approach substantially reduces reliance on costly HU annotations and makes VLMs more accurate and better calibrated. Extensive experiments on VQAv2 and VizWiz datasets demonstrate that HaDola consistently matches or outperforms state-of-the-art baselines, with less training data. Our work highlights the importance of explicitly modeling HU in SFT, suggesting better utilization of HU is more effective than merely scaling up dataset size.

The Mixture of Experts (MoE) is an effective architecture for scaling large language models by leveraging sparse expert activation to balance performance and efficiency. However, under expert parallelism, MoE suffers from inference inefficiencies due to imbalanced token-to-expert assignment, where underloaded experts complete computations early but must wait for overloaded experts, leading to global delays. We define this phenomenon as the \textbf{\textit{Straggler Effect}}, as the most burdened experts dictate the overall inference latency. To address this, we first propose \textit{\textbf{Capacity-Aware Token Drop}}, which enforces expert capacity limits by discarding excess tokens from overloaded experts, effectively reducing load imbalance with minimal performance impact (e.g., $30\%$ speedup with only $0.9\%$ degradation on OLMoE). Next, given the presence of low-load experts remaining well below the capacity threshold, we introduce \textit{\textbf{Capacity-Aware Expanded Drop}}, which allows tokens to include additional local experts in their candidate set before enforcing strict local capacity constraints, thereby improving load balance and enhancing the utilization of underused experts. Extensive experiments on both language and multimodal MoE models demonstrate the effectiveness of our approach, yielding substantial gains in expert utilization, model performance, and inference efficiency, e.g., applying Expanded Drop to Mixtral-8$\times$7B-Instruct yields a {0.2\%} average performance improvement and a {1.85$\times$} inference speedup. The code is released at: https://github.com/CASE-Lab-UMD/Capacity-Aware-MoE.

State Space Models (SSMs), developed to tackle long sequence modeling tasks efficiently, offer both parallelizable training and fast inference. At their core are recurrent dynamical systems that maintain a hidden state, with update costs scaling with the state dimension. A key design challenge is striking the right balance between maximizing expressivity and limiting this computational burden. Control theory, and more specifically Hankel singular value analysis, provides a potent framework for the measure of energy for each state, as well as the balanced truncation of the original system down to a smaller representation with performance guarantees. Leveraging the eigenvalue stability properties of Hankel matrices, we apply this lens to SSMs during training, where only dimensions of high influence are identified and preserved. Our approach applies to Linear Time-Invariant SSMs such as Linear Recurrent Units, but is also extendable to selective models. Experiments show that in-training reduction significantly accelerates optimization while preserving expressivity, with compressed models retaining task-critical structure lost by models trained directly at smaller dimension. In other words, SSMs that begin large and shrink during training achieve computational efficiency while maintaining higher performance.

Visual Question Answering (VQA) enables targeted and context-dependent analysis of medical images, such as chest X-rays (CXRs). However, existing VQA datasets for CXRs are typically constrained by simplistic and brief answer formats, lacking localization annotations (e.g., bounding boxes) and structured tags (e.g., region or radiological finding/disease tags). To address these limitations, we introduce MIMIC-Ext-CXR-QBA (abbr. CXR-QBA), a large-scale CXR VQA dataset derived from MIMIC-CXR, comprising 42 million QA-pairs with multi-granular, multi-part answers, detailed bounding boxes, and structured tags. We automatically generated our VQA dataset from scene graphs (also made available), which we constructed using LLM-based information extraction from radiology reports. After automatic quality assessment, we identified 31M pre-training and 7.5M fine-tuning grade QA-pairs, providing the largest and most sophisticated VQA dataset for CXRs to date. Tools for using our dataset and the construction pipeline are available at https://github.com/philip-mueller/mimic-ext-cxr-qba/ .

Direct Preference Optimization (DPO) has emerged as a powerful framework for aligning large language models (LLMs) with human preferences via pairwise comparisons. However, its performance is highly sensitive to the quality of training samples: when the reference policy is poorly aligned with human preferences, ambiguous pairs can dominate the gradient signal and degrade generalization. To address this, we propose RAPPO($\textbf{R}$eliable $\textbf{A}$lignment for $\textbf{P}$reference $\textbf{P}$olicy $\textbf{O}$ptimization), a simple sample-aware modification of the DPO loss that mitigates reference-policy misalignment by filtering out the hardest, most ambiguous samples. We theoretically show that RAPPO yields improved generalization guarantees. RAPPO is lightweight and requires only a few lines of code to be integrated into any existing DPO-type algorithm. Surprisingly, With this simple modification, our simulations across a broad suite of alignment tasks and benchmarks show consistent gains over DPO and recent state-of-the-art baselines. On the PKU-SafeRLHF benchmark, RAPPO attains helpfulness $0.693$ ($+34.8\%$ over DPO) and harmlessness $0.357$ ($-21.0\%$ vs DPO).

As artificial intelligence (AI) assistants become more widely adopted in safety-critical domains, it becomes important to develop safeguards against potential failures or adversarial attacks. A key prerequisite to developing these safeguards is understanding the ability of these AI assistants to mislead human teammates. We investigate this attack problem within the context of an intellective strategy game where a team of three humans and one AI assistant collaborate to answer a series of trivia questions. Unbeknownst to the humans, the AI assistant is adversarial. Leveraging techniques from Model-Based Reinforcement Learning (MBRL), the AI assistant learns a model of the humans' trust evolution and uses that model to manipulate the group decision-making process to harm the team. We evaluate two models -- one inspired by literature and the other data-driven -- and find that both can effectively harm the human team. Moreover, we find that in this setting while our data-driven model is the most capable of accurately predicting how human agents appraise their teammates given limited information on prior interactions, the model based on principles of cognitive psychology does not lag too far behind. Finally, we compare the performance of state-of-the-art LLM models to human agents on our influence allocation task to evaluate whether the LLMs allocate influence similarly to humans or if they are more robust to our attack. These results enhance our understanding of decision-making dynamics in small human-AI teams and lay the foundation for defense strategies.

Hand gesture classification using high-quality structured data such as videos, images, and hand skeletons is a well-explored problem in computer vision. Alternatively, leveraging low-power, cost-effective bio-signals, e.g. surface electromyography (sEMG), allows for continuous gesture prediction on wearable devices. In this work, we aim to enhance EMG representation quality by aligning it with embeddings obtained from structured, high-quality modalities that provide richer semantic guidance, ultimately enabling zero-shot gesture generalization. Specifically, we propose EMBridge, a cross-modal representation learning framework that bridges the modality gap between EMG and pose. EMBridge learns high-quality EMG representations by introducing a Querying Transformer (Q-Former), a masked pose reconstruction loss, and a community-aware soft contrastive learning objective that aligns the relative geometry of the embedding spaces. We evaluate EMBridge on both in-distribution and unseen gesture classification tasks and demonstrate consistent performance gains over all baselines. To the best of our knowledge, EMBridge is the first cross-modal representation learning framework to achieve zero-shot gesture classification from wearable EMG signals, showing potential toward real-world gesture recognition on wearable devices.

For RL algorithms, appropriate entropy control is crucial to their effectiveness. To control the policy entropy, a commonly used method is entropy regularization, which is adopted in various popular RL algorithms including PPO, SAC and A3C. Although entropy regularization proves effective in robotic and games RL conventionally, studies found that it gives weak to no gains in LLM-RL training. In this work, we study the issues of entropy bonus in LLM-RL setting. Specifically, we first argue that the conventional entropy regularization suffers from the LLM's extremely large response space and the sparsity of the optimal outputs. As a remedy, we propose AEnt, an entropy control method that utilizes a new clamped entropy bonus with an automatically adjusted coefficient. The clamped entropy is evaluated with the re-normalized policy defined on certain smaller token space, which encourages exploration within a more compact response set. In addition, the algorithm automatically adjusts entropy coefficient according to the clamped entropy value, effectively controlling the entropy-induced bias while leveraging the entropy's benefits. AEnt is tested in math-reasoning tasks under different base models and datasets, and it is observed that AEnt outperforms the baselines consistently across multiple benchmarks.

Speech brain–computer interfaces (BCIs) aim to restore communication for people with paralysis by translating neural activity into text. Most systems use cascaded frameworks that decode phonemes before assembling sentences with an n-gram language model (LM), preventing joint optimization of all stages simultaneously. Here, we introduce an end-to-end BraIn-to-Text (BIT) framework that translates neural activity into coherent sentences using a single differentiable neural network. Central to our approach is a cross-task, cross-species pretrained neural encoder, whose representations transfer to both attempted and imagined speech. In a cascaded setting with an n-gram LM, the pretrained encoder establishes a new state-of-the-art (SOTA) on the Brain-to-Text ’24 and ’25 benchmarks. Integrated end-to-end with audio large language models (LLMs) and trained with contrastive learning for cross-modal alignment, BIT reduces the word error rate (WER) of the prior end-to-end method from 24.69% to 10.22%. Notably, we find that small-scale audio-LLMs markedly improve end-to-end decoding. Beyond record-setting performance, BIT aligns attempted and imagined speech embeddings to enable cross-task generalization. Altogether, our approach advances the integration of large, diverse neural datasets, paving the way for an end-to-end decoding framework that supports seamless, differentiable optimization.

Large Language Models (LLMs) have recently alleviated limitations in outdated internal knowledge and computational inaccuracies by invoking external tools such as search engines and code generation. While reinforcement learning (RL) has substantially enhanced tool usage in LLMs, most existing agentic RL approaches rely solely on outcome-only reward signals, which assign credit at a coarse granularity and often induce gradient conflict (e.g., correct tool calls may be penalized due to incorrect final answers). To address this, we propose the *Tool-call Reward Model* (TRM), a specialized process reward model meticulously designed to evaluate and reward each tool invocation. Since previous PRM research has predominantly focused on traditional reasoning tasks such as step-wise mathematical reasoning, the introduction of TRM brings two unique challenges: (1) limited understanding of how to construct effective TRMs, including data requirements and model size; and (2) difficulties integrating TRM with classical RL algorithms such as PPO and GRPO, where naive adaptation may lead to reward hacking (minimizing tool calls to avoid penalties). To tackle these challenges, we establish a systematic TRM construction workflow and propose refined credit assignment and turn-level advantage estimation for effective integration with PPO and GRPO. Experiments show that a 3B TRM trained on 10K samples achieves robust performance. On search-based QA and Python code-based math tasks, integrating TRM consistently outperforms outcome-only reward RL methods across models of different sizes.

Conditioning is crucial for stable training of full-head 3D-aware GANs. Without any conditioning signal, the model suffers from severe mode collapse, making it impractical to training (\cref{fig:intro}(a,b)). However, a series of previous full-head 3D-aware GANs conventionally choose the view angle as the conditioning input, which leads to a bias in the learned 3D full-head space along the conditional view direction. This is evident in the significant differences in generation quality and diversity between the conditional view and non-conditional views of the generated 3D heads, resulting in global incoherence across different head regions (\cref{fig:intro}(d-i)). In this work, we propose to use \textit{view-invariant semantic feature} as the conditioning input, thereby decoupling the generative capability of 3D heads from the viewing direction. To construct a view-invariant semantic condition for each training image, we create a novel synthesized head image dataset. We leverage FLUX.1 Kontext to extend existing high-quality frontal face datasets to a wide range of view angles. The image clip feature extracted from the frontal view is then used as a shared semantic condition across all views in the extended images, ensuring semantic alignment while eliminating directional bias. This also allows supervision from different views of the same subject to be consolidated under a shared semantic condition, which accelerates training (\cref{fig:intro}(c)) and enhances the global coherence of the generated 3D heads (\cref{fig:teaser}). Moreover, as GANs often experience slower improvements in diversity once the generator learns a few modes that successfully fool the discriminator, our semantic conditioning encourages the generator to follow the true semantic distribution, thereby promoting continuous learning and diverse generation. Extensive experiments on full-head synthesis and single-view GAN inversion demonstrate that our method achieves significantly higher fidelity, diversity, and generalizability.

Do large language models (LLMs) represent concepts abstractly, i.e., independent of input format? We revisit Function Vectors (FVs), compact representations of in-context learning (ICL) tasks that causally drive task performance. Across multiple LLMs, we show that FVs are not fully invariant: FVs are nearly orthogonal when extracted from different input formats (e.g., open-ended vs. multiple-choice), even if both target the same concept. We identify Concept Vectors (CVs), which carry more stable concept representations. Like FVs, CVs are composed of attention head outputs; however, unlike FVs, the constituent heads are selected using Representational Similarity Analysis (RSA) based on whether they encode concepts consistently across input formats. While these heads emerge in similar layers to FV-related heads, the two sets are largely distinct, suggesting different underlying mechanisms. Steering experiments reveal that FVs excel in-distribution, when extraction and application formats match (e.g., both open-ended in English), while CVs generalize better out-of-distribution across both question types (open-ended vs. multiple-choice) and languages. Our results show that LLMs do contain abstract concept representations, but these differ from those that drive ICL performance.

Vision-language models (VLMs) have demonstrated impressive generalization across multimodal tasks, yet most evaluation benchmarks remain Western-centric, leaving open questions about their performance in culturally diverse and multilingual settings. To address this gap, we introduce IndicVisionBench, the first large-scale benchmark centered on the Indian subcontinent. Covering English and 10 Indian languages, our benchmark spans 3 multimodal tasks, including Optical Character Recognition (OCR), Multimodal Machine Translation (MMT), and Visual Question Answering (VQA), covering 6 kinds of question types. Our final benchmark consists of a total of ~5K images and 37K+ QA pairs across 13 culturally grounded topics. In addition, we release a paired parallel corpus of annotations across 10 Indic languages, creating a unique resource for analyzing cultural and linguistic biases in VLMs. We evaluate a broad spectrum of 8 models, from proprietary closed-source systems to open-weights medium and large-scale models. Our experiments reveal substantial performance gaps, underscoring the limitations of current VLMs in culturally diverse contexts. By centering cultural diversity and multilinguality, IndicVisionBench establishes a reproducible evaluation framework that paves the way for more inclusive multimodal research.

Counterfactual reasoning, a hallmark of intelligence, consists of three steps: inferring latent variables from observations (abduction), constructing alternative situations (interventions), and predicting the outcomes of the alternatives (prediction). This skill is essential for advancing LLMs' causal understanding and expanding their applications in high-stakes domains such as scientific research and healthcare. However, existing efforts in assessing LLM's counterfactual reasoning capabilities tend to skip the abduction step, effectively reducing to interventional reasoning and leading to over-estimated LLM performance. To address this, we introduce executable counterfactuals, a novel framework that operationalizes causal reasoning through code and math problems. Our framework explicitly requires all three steps of counterfactual reasoning and enables scalable synthetic data creation with varying difficulty, creating a new frontier for evaluating and improving LLM's reasoning. Our results reveal substantial drop in accuracy (25-40%) from interventional to counterfactual reasoning for state-of-the-art models such as o4-mini and Claude-4-Sonnet. To address this gap, we construct a training set comprising counterfactual code problems having if-condition and test on out-of-distribution code structures (e.g., having while-loop); we also test whether a model trained on code would generalize to counterfactual math word problems. While Supervised Finetuning (SFT) on stronger models' reasoning traces improves in-distribution performance of Qwen models, it leads to a decrease in accuracy on out-of-distribution tasks such as counterfactual math problems. In contrast, reinforcement learning (RL) induces the core cognitive behaviors and generalizes to new distributions, yielding substantial accuracy gains over the base model on both code (improvement of 1.5x-2x) and counterfactual math problems. Analysis of the reasoning traces further reinforces these findings and highlights the promise of RL with scalable data generation for improving LLMs' counterfactual reasoning.

Scientific discovery is increasingly constrained by costly experiments and limited budgets, making efficient optimization essential for AI for science. Bayesian Optimization (BO), while widely adopted for balancing exploration and exploitation, suffers from slow cold-start performance and poor scalability in high-dimensional settings, limiting its effectiveness in real-world scientific applications. To address these challenges, we propose LLM-Guided Bayesian Optimization (LGBO), the first LLM preference-guided BO framework that continuously integrates the semantic reasoning of large language models (LLMs) into the optimization loop. Unlike prior works that use LLMs only for warm-start initialization or candidate generation, LGBO introduces a region-lifted preference mechanism that embeds LLM-driven preferences into every iteration, shifting the surrogate mean in a stable and controllable way. Theoretically, we prove that LGBO is not perform significantly worse than standard BO in the worst case, while achieving significantly faster convergence when preferences align with the objective. Empirically, LGBO achieves consistent improvements across diverse dry benchmarks in physics, chemistry, biology, and materials science. Most notably, in a new wet-lab optimization of Fe–Cr battery electrolytes, LGBO reaches \textbf{90\% of the best observed value within 6 iterations}, whereas standard BO and existing LLM-augmented baselines require more than 10 iterations. Together, the results suggest that LGBO offers a promising direction for integrating LLMs into scientific optimization workflows.

This paper studies federated multi-armed bandit (MAB) problems where multiple agents working together to solve a common MAB problem through a communication network. We focus on the heterogeneous setting in which no single agent can identify the global best arm using only local biased observations. In this setting, different agents may select the same arm at the same time step but receive varying rewards. We propose a novel algorithm called \textsc{FedFTRL} for this problem, which is the first work to achieve near-optimal regret guarantees in both stochastic and adversarial environments. Notably, in the adversarial regime, our algorithm achieves $O(\sqrt{T})$ regret which is a significant improvement over the state-of-the-art regret of $O(T^{\frac{2}{3}})$ \citep{yi2023doubly}. We also provide numerical evaluations comparing our algorithm with baseline methods, demonstrating the effectiveness of our approach on both synthetic and real-world datasets.

Recent advancements in DeepFakes attribution technologies have significantly enhanced forensic capabilities, enabling the extraction of traces left by generative models (GMs) in images, making DeepFakes traceable back to their source GMs. Meanwhile, several attacks have attempted to evade attribution models (AMs) for exploring their limitations, calling for more robust AMs. However, existing attacks fail to eliminate GMs' traces, thus can be mitigated by defensive measures. In this paper, we identify that untraceable DeepFakes can be achieved through a multiplicative attack, which can fundamentally eliminate GMs' traces, thereby evading AMs even enhanced with defensive measures. We design a universal and black-box attack method that trains an adversarial model solely using real data, applicable for various GMs and agnostic to AMs. Experimental results demonstrate the outstanding attack capability and universal applicability of our method, achieving an average attack success rate (ASR) of 97.08\% against 6 advanced AMs across 12 GMs. Even in the presence of defensive mechanisms, our method maintains an ASR exceeding 72.39\%. Our work underscores the potential challenges posed by multiplicative attacks and highlights the need for more robust AMs. The code is available at \textbf{\textit{https://anonymous.4open.science/r/TEST-F4B1}}.

Periodicity is a fundamental characteristic of time series data and has long played a central role in forecasting. Recent deep learning methods strengthen the exploitation of periodicity by treating patches as basic tokens, thereby improving predictive effectiveness. However, their efficiency remains a bottleneck due to large parameter counts and heavy computational costs. This paper provides, for the first time, a clear explanation of why patch-level processing is inherently inefficient, supported by strong evidence from real-world data. To address these limitations, we introduce a phase perspective for modeling periodicity and present an efficient yet effective solution, PhaseFormer. PhaseFormer features phase-wise prediction through compact phase embeddings and efficient cross-phase interaction enabled by a lightweight routing mechanism. Extensive experiments demonstrate that PhaseFormer achieves state-of-the-art performance on the evaluated benchmarks with around 1k parameters, consistently across benchmark datasets. Notably, it excels on large-scale and complex datasets, where models with comparable efficiency often struggle. This work marks a significant step toward truly efficient and effective time series forecasting. Code is available at this repository: https://anonymous.4open.science/r/ICLR26-PhaseFormer-17678.

Spiking Neural Networks (SNNs), with their temporal processing capabilities and biologically plausible dynamics, offer a natural platform for unsupervised representation learning. However, current unsupervised SNNs predominantly employ shallow architectures or localized plasticity rules, limiting their ability to model long-range temporal dependencies and maintain temporal feature consistency. This results in semantically unstable representations, thereby impeding the development of deep unsupervised SNNs for large-scale temporal video data. We propose PredNext, which explicitly models temporal relationships through cross-view future Step Prediction and Clip Prediction. This plug-and-play module seamlessly integrates with diverse self-supervised objectives. We firstly establish standard benchmarks for SNN self-supervised learning on UCF101, HMDB51, and MiniKinetics, which are substantially larger than conventional DVS datasets. PredNext delivers significant performance improvements across different tasks and self-supervised methods. PredNext achieves performance comparable to ImageNet-pretrained supervised weights through unsupervised training solely on UCF101. Additional experiments demonstrate that PredNext, distinct from forced consistency constraints, substantially improves temporal feature consistency while enhancing network generalization capabilities. This work provides a effective foundation for unsupervised deep SNNs on large-scale temporal video data.