ICLR 2026 Papers

Reinforcement learning (RL) algorithms have been used recently to align diffusion models with downstream objectives such as aesthetic quality and text-image consistency by fine-tuning them to maximize a single reward function under a fixed KL regularization. However, this approach is inherently restrictive in practice, where alignment must balance multiple, often conflicting objectives. Moreover, user preferences vary across prompts, individuals, and deployment contexts, with varying tolerances for deviation from a pre-trained base model. We address the problem of inference-time multi-preference alignment: given a set of basis reward functions and a reference KL regularization strength, can we design a fine-tuning procedure so that, at inference time, it can generate images aligned with any user-specified linear combination of rewards and regularization, without requiring additional fine-tuning? We propose Diffusion Blend, a novel approach to solve inference-time multi-preference alignment by blending backward diffusion processes associated with fine-tuned models, and we instantiate this approach with three algorithms: DB-MPA for multi-reward alignment, DB-KLA for KL regularization control, and DB-MPA-LS for approximating DB-MPA without additional inference cost. Extensive experiments show that Diffusion Blend algorithms consistently outperform relevant baselines and closely match or exceed the performance of individually fine-tuned models, enabling efficient, user-driven alignment at inference-time.

Shapley values have emerged as a central game-theoretic tool in explainable AI (XAI). However, computing Shapley values exactly requires $2^d$ game evaluations for a model with $d$ features. Lundberg and Lee's KernelSHAP algorithm has emerged as a leading method for avoiding this exponential cost. KernelSHAP approximates Shapley values by approximating the game as a linear function, which is fit using a small number of game evaluations for random feature subsets. In this work, we extend KernelSHAP by approximating the game via higher degree polynomials, which capture non-linear interactions between features. Our resulting PolySHAP method yields empirically better Shapley value estimates for various benchmark datasets, and we prove that these estimates are consistent. Moreover, we connect our approach to paired sampling (antithetic sampling), a ubiquitous modification to KernelSHAP that improves empirical accuracy. We prove that paired sampling outputs exactly the same Shapley value approximations as second-order PolySHAP, without ever fitting a degree 2 polynomial. To the best of our knowledge, this finding provides the first strong theoretical justification for the excellent practical performance of the paired sampling heuristic.

Prompt engineering has become central to eliciting the capabilities of large language models (LLMs). At its core lies prompt selection - efficiently identifying the most effective prompts. However, most prior investigations overlook a key challenge: the inherently multi-faceted nature of prompt performance, which cannot be captured by a single metric. To fill this gap, we study the multi-objective prompt selection problem under two practical settings: Pareto prompt set recovery and best feasible prompt identification. Casting the problem into the pure-exploration bandits framework, we adapt provably efficient algorithms from multi-objective bandits and further introduce a novel design for best feasible arm identification in structured bandits, with theoretical guarantees on the identification error in the linear case. Extensive experiments across multiple LLMs show that the bandit-based approaches yield significant improvements over baselines, establishing a principled and efficient framework for multi-objective prompt optimization.

While machine-generated texts (MGTs) offer great convenience, they also pose risks such as disinformation and phishing, highlighting the need for reliable detection. Metric-based methods, which extract statistically distinguishable features of MGTs, are often more practical than complex model-based methods that are prone to overfitting. Given their diverse designs, we first place representative metric-based methods within a unified framework, enabling a clear assessment of their advantages and limitations. Our analysis identifies a core challenge across these methods: the token-level detection score is easily biased by the inherent randomness of the MGTs generation process. To address this, we theoretically and empirically reveal two relationships of context detection scores that may aid calibration: Neighbor Similarity and Initial Instability. We then propose a Markov-informed score calibration strategy that models these relationships using Markov random fields, and implements it as a lightweight component via a mean-field approximation, allowing our method to be seamlessly integrated into existing detectors. Extensive experiments in various real-world scenarios, such as cross-LLM and paraphrasing attacks, demonstrate significant gains over baselines with negligible computational overhead. The code is available at \url{https://anonymous.4open.science/r/MRF-Enhance}.

Recent findings raise concerns about whether the evaluation of Multiple-Choice Question Answering (MCQA) accurately reflects the comprehension abilities of large language models. This paper explores the concept of \textit{choice sensitivity}, which refers to the tendency for model decisions to be more influenced by the answer options than by a genuine understanding of the question. We introduce a new scoring method called **Normalized Probability Shift by the Question (NPSQ)**, designed to isolate the impact of the question itself and provide a more reliable assessment of comprehension. Through experiments involving various input formats, including cloze, symbols, and hybrid formats, we find that traditional scoring methods — such as those based on log-likelihood or its length-normalized variant — are vulnerable to superficial characteristics of the answer choices. In contrast, NPSQ remains stable even when modifications are made to the answer options.

Generative models, like flows and diffusions, have recently emerged as popular and efficacious policy parameterizations in robotics. There has been much speculation as to the factors underlying their successes, ranging from capturing multimodal action distributions to expressing more complex behaviors. In this work, we perform a comprehensive evaluation of popular generative control policies (GCPs) on common behavior cloning (BC) benchmarks. We find that GCPs do not owe their success to their ability to capture multimodality or to express more complex observation-to-action mappings. Instead, we find that their advantage stems from iterative computation, provided that intermediate steps are supervised during training and this supervision is paired with a suitable level of stochasticity. As a validation of our findings, we show that a minimal iterative policy (MIP), a lightweight two-step regression-based policy, essentially matches the performance of flow GCPs. Our results suggest that the distribution-fitting component of GCPs is less salient than commonly believed and point toward new design spaces focusing solely on control performance. Videos and supplementary materials are available at https://anonymous.4open.science/w/mip-anonymous/.

Recent advances in large language model (LLM) reasoning have shown that reasoning ability can emerge through reinforcement learning (RL). However, despite these successes, RL in its current form remains insufficient to induce capabilities that exceed the limitations of the base model, as it is primarily optimized based on existing knowledge of the model. To address this limitation, we employ supervised fine-tuning (SFT) to learn what RL cannot, which enables the incorporation of new knowledge and reasoning patterns by leveraging high-quality demonstration data. We analyze the training dynamics of RL and SFT for LLM reasoning and find that RL excels at improving performance on questions within the model's original capabilities, while SFT is more effective at enabling progress on questions beyond the current scope of the model. Motivated by the complementary strengths of RL and SFT, we introduce \textbf{ReLIFT} (\textbf{Re}inforcement \textbf{L}earning \textbf{I}nterleaved with Online \textbf{F}ine-\textbf{T}uning), a novel training strategy. ReLIFT employs RL for general training, but interleaves it with targeted SFT on challenging questions for which high-quality solutions are collected online. By alternating between RL and SFT, ReLIFT addresses model weaknesses as they emerge. Empirically, ReLIFT outperforms previous RLVR methods by an average of +6.7 points across a suite of six benchmarks (five math reasoning and one out-of-distribution). More importantly, ReLIFT surpasses baselines such as individual RL, individual SFT, and various hybrid approaches while reducing the required training time. These results provide compelling evidence that ReLIFT is a powerful and resource-efficient paradigm for developing capable reasoning models. The code is available at \href{https://anonymous.4open.science/r/Learning-What-Reinforcement-Learning-Can-t-6AFF/}{here}.

Few-Shot Class Incremental Learning (FSCIL) is crucial for adapting to the complex open-world environments. Contemporary prospective learning-based space construction methods struggle to balance old and new knowledge, as prototype bias and rigid structures limit the expressive capacity of the embedding space. Different from these strategies, we rethink the optimization dilemma from the perspective of feature-structure dual consistency, and propose a Consistency-driven Calibration and Matching (ConCM) framework that systematically mitigates the knowledge conflict inherent in FSCIL. Specifically, inspired by hippocampal associative memory, we design a memory-aware prototype calibration that extracts generalized semantic attributes from base classes and reintegrates them into novel classes to enhance the conceptual center consistency of features. Further, to consolidate memory associations, we propose dynamic structure matching, which adaptively aligns the calibrated features to a session-specific optimal manifold space, ensuring cross-session structure consistency. This process requires no class-number priors and is theoretically guaranteed to achieve geometric optimality and maximum matching. On large-scale FSCIL benchmarks including mini-ImageNet, CIFAR100 and CUB200, ConCM achieves state-of-the-art performance, with harmonic accuracy gains of up to 3.41% in incremental sessions. Code is available at: https://anonymous.4open.science/r/ConCM-7385

Personalization has become a pivotal field of study in contemporary intelligent systems. While large language models (LLMs) excel at general knowledge tasks, they often struggle with personalization, i.e., adapting their outputs to individual user expectations. Existing approaches that steer LLM behavior to meet users’ implicit preferences and behavior patterns, primarily relying on tune-free methods (e.g., RAG, PAG) or parameter fine-tuning methods (e.g., LoRA), face challenges in effectively balancing effectiveness and efficiency. Moreover, the mechanisms underlying personalized preferences remain underexplored. To address these challenges, we first uncover key patterns of user-specific information embedded in the representation space. Specifically, we find that (1) personalized information lies within a low-rank subspace represented by vectors, and (2) these vectors demonstrate both a collective shift shared across users and a personalized shift unique to each individual user. Building on these insights, we introduce PerFit, a novel two-stage solution that directly fine-tunes interventions in the hidden representation space by addressing both collective and user-specific shifts, thereby achieving precise steering of LLM with minimal parameter overhead. Experimental results demonstrate that \perfit delivers strong performance across six datasets while \cutting the number of parameters by an average of 92.3% compared to the state-of-the-art method.

Sharpness-Aware Minimization (SAM) has proven effective in enhancing deep neural network training by simultaneously minimizing the training loss and the sharpness of the loss landscape, thereby guiding models toward flatter minima that are empirically linked to improved generalization. From another perspective, generalization can be seen as a model’s ability to remain stable under distributional variability. In particular, effective learning requires that updates derived from different subsets or resamplings of the same data distribution remain consistent. In this work, we investigate the connection between the flatness induced by SAM and the alignment of gradients across random subsets of the data distribution, and propose \textit{Align-SAM} as a novel strategy to further enhance model generalization. Align-SAM extends the core principles of SAM by promoting optimization toward flatter minima on a primary subset (the training set), while simultaneously enforcing low loss on an auxiliary subset drawn from the same distribution. This dual-objective approach leads to solutions that are not only resilient to local perturbations but also robust against distributional shifts in each training iteration. Empirical evaluations demonstrate that Align-SAM consistently improves generalization across diverse datasets and challenging settings, including scenarios with noisy labels and limited data availability.

Parameter-Efficient Fine-Tuning (PEFT) methods have become crucial for rapidly adapting large language models (LLMs) to downstream tasks. Prefix-Tuning, an early and effective PEFT technique, demonstrated the ability to achieve performance comparable to full fine-tuning with significantly reduced computational and memory overhead. However, despite its earlier success, its effectiveness in training modern state-of-the-art LLMs has been very limited. In this work, we demonstrate empirically that Prefix-Tuning underperforms on LLMs because of an inherent tradeoff between the contribution of input prompt and parameterized prefix within the attention head. This motivates us to introduce PrefixMemory-Tuning, an architecture that generalizes the principles of Prefix-Tuning while addressing its shortcomings by shifting the prefix module out of the attention head itself and improving its expressiveness. Our experiments show that, across diverse benchmarks, PrefixMemory-Tuning consistently outperforms existing Prefix-Tuning methods. Notably, it achieves competitive performance with modern PEFTs on several general benchmarks, highlighting a potential extension of Prefix-Tuning approaches to become state-of-the-art. Our findings suggest that by overcoming its inherent limitations, Prefix-Tuning can remain a competitive and relevant research direction in the landscape of parameter-efficient LLM adaptation.

In this work, we present a simple yet theoretically motivated improvement to Supervised Fine-Tuning (SFT) for the Large Language Model (LLM), addressing its limited generalization compared to reinforcement learning (RL). Through mathematical analysis, we reveal that standard SFT gradients implicitly encode a problematic reward structure that may severely restrict the generalization capabilities of model compared to RL. To rectify this, we propose Dynamic Fine-Tuning (DFT), stabilizing gradient updates for each token by dynamically rescaling the objective function with the probability of this token. With just a single-line change, the method outperforms standard SFT on multiple difficult benchmarks and base models, from math reasoning to code generation and multi-modal tasks, demonstrating improved generalization. Additionally, DFT achieves competitive results in offline RL settings, and further boosts the effectiveness of subsequent RL training, providing an effective yet streamlined alternative. The experiments further demonstrate that DFT not only strengthens SFT performance but also consistently improves the effectiveness of subsequent RL training. By bridging theoretical insights with practical solutions, this work advances the state of SFT. The source code will be publicly released.

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.