ICLR 2026 论文列表

In post-training for reasoning Large Language Models (LLMs), the current state of practice trains LLMs in two independent stages: Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR, shortened as "RL" below). In this work, we challenge whether high SFT scores translate to improved performance after RL. We provide extensive counter-examples where this is not true. We find high SFT scores can be biased toward simpler or more homogeneous data and are not reliably predictive of subsequent RL gains or scaled-up post-training effectiveness. In some cases, RL training on models with improved SFT performance could lead to substantially worse outcome compared to RL on the base model without SFT. We study alternative metrics and identify generalization loss on held-out reasoning examples and Pass@large k performance to provide strong proxies for the RL outcome. We trained hundreds of models up to 12B-parameter with SFT and RLVR via GRPO and ran extensive evaluations on 7 math benchmarks with up to 256 repetitions, spending $>$1M GPU hours. Experiments include models from Llama3, Mistral-Nemo, Qwen3 and multiple state-of-the-art SFT/RL datasets. Compared to directly predicting from pre-RL performance, prediction based on generalization loss and Pass@large k achieves substantial higher precision, improving $R^2$ coefficient and Spearman's rank correlation coefficient by up to 0.5 (2x). This provides strong utility for broad use cases. For example, in most experiments, we find SFT training on unique examples for a one epoch underperforms training on half examples for two epochs, either after SFT or SFT-then-RL; With the same SFT budget, training only on short examples may lead to better SFT performance, though, it often leads to worse outcome after RL compared to training on examples with varying lengths. This work develops an enhanced evaluation tool that will be open-sourced.

Many methods have been developed to predict static protein structures, however understanding the dynamics of protein structure is essential for elucidating biological function. While molecular dynamics (MD) simulations remain the in silico gold standard, its high computational cost limits scalability. We present DynaProt, a lightweight, SE(3)-invariant framework that predicts rich descriptors of protein dynamics directly from static structures. By casting the problem through the lens of multivariate Gaussians, DynaProt estimates dynamics at two complementary scales: (1) per-residue marginal anisotropy as covariance matrices capturing local flexibility, and (2) joint scalar covariances encoding pairwise dynamic coupling across residues. From these dynamics outputs, DynaProt achieves high accuracy in predicting residue-level flexibility (RMSF) and, remarkably, enables reasonable reconstruction of the full covariance matrix for fast ensemble generation. Notably, it does so using orders of magnitude fewer parameters than prior methods. Our results highlight the potential of direct protein dynamics prediction as a scalable alternative to existing methods.

Anticipating future actions in videos is challenging, as the observed frames provide only evidence of past activities, requiring the inference of latent intentions to predict upcoming actions. Existing transformer-based approaches, which rely on dot-product attention over pixel representations, often lack the high-level semantics necessary to model video sequences for effective action anticipation. As a result, these methods tend to overfit to explicit visual cues present in the past frames, limiting their ability to capture underlying intentions and degrading generalization to unseen samples. To address this, we propose Action-Guided Attention (AGA), an attention mechanism that explicitly leverages predicted action sequences as queries and keys to guide sequence modeling. Our approach fosters the attention module to emphasize relevant moments from the past based on the upcoming activity and combine this information with the current frame embedding via a dedicated gating function. The design of AGA enables post-training analysis of the knowledge discovered from the training set. Experiments on the widely adopted EPIC-Kitchens-100 benchmark demonstrate that AGA generalizes well from validation to unseen test sets. Post-training analysis can further examine the action dependencies captured by the model and the counterfactual evidence it has internalized, offering transparent and interpretable insights into its anticipative predictions.

The development of state-of-the-art large language models is commonly understood as a two-stage process involving pre-training and post-training. We point out the need for an additional intermediate stage called reinforcement mid-training with potential for strong performance gains. In this paper, we formally define the problem and identify three key challenges: (1) inefficient training due to excessive reasoning steps, (2) disregard of the imbalanced token entropy distribution, and (3) underutilization of token information. To address these challenges, we propose RMT, a framework for efficient, adaptive, and unified reinforcement mid-training with various innovative components. In particular, we first introduce a dynamic token budget mechanism that constrains unnecessary reasoning steps and mitigates model overthinking. Next, we design a curriculum-based adaptive sampling method that fosters a progressive learning trajectory from easy to hard tokens. Finally, we present a dual training strategy that combines reinforcement learning with next-token prediction, ensuring targeted learning on key tokens and full exploitation of all token information. Extensive experiments demonstrate the superiority of RMT over state-of-the-art methods, achieving up to +64.91% performance improvement with only 21% of the reasoning length in language modeling. We also show that checkpoints obtained after reinforcement mid-training can benefit the subsequent post-training, yielding up to +18.76% improvement in the mathematical domain.

The success of teams in robotics, nature, and society often depends on the division of labor among diverse specialists; however, a principled explanation for when such diversity surpasses a homogeneous team is still missing. Focusing on multi-agent task allocation problems, we study this question from the perspective of reward design: what kinds of objectives are best suited for heterogeneous teams? We first consider an instantaneous, non-spatial setting where the global reward is built by two generalized aggregation operators: an inner operator that maps the N agents’ effort allocations on individual tasks to a task score, and an outer operator that merges the M task scores into the global team reward. We prove that the curvature of these operators determines whether heterogeneity can increase reward, and that for broad reward families this collapses to a simple convexity test. Next, we ask what incentivizes heterogeneity to emerge when embodied, time-extended agents must learn an effort allocation policy. To study heterogeneity in such settings, we use multi-agent reinforcement learning (MARL) as our computational paradigm, and introduce Heterogeneity Gain Parameter Search (HetGPS), a gradient-based algorithm that optimizes the parameter space of underspecified MARL environments to find scenarios where heterogeneity is advantageous. Across different environments, we show that HetGPS rediscovers the reward regimes predicted by our theory to maximize the advantage of heterogeneity, both validating HetGPS and connecting our theoretical insights to reward design in MARL. Together, these results help us understand when behavioral diversity delivers a measurable benefit.

Time series forecasting has attracted significant attention in the field of AI. Previous works have revealed that the Channel-Independent (CI) strategy improves forecasting performance by modeling each channel individually, but it often suffers from poor generalization and overlooks meaningful inter-channel interactions. Conversely, Channel-Dependent (CD) strategies aggregate all channels, which may introduce irrelevant information and lead to oversmoothing. Despite recent progress, few existing methods offer the flexibility to adaptively balance CI and CD strategies in response to varying channel dependencies. To address this, we propose a generic plugin xCPD, that can adaptively model the channel-patch dependencies from the perspective of graph spectral decomposition. Specifically, xCPD first projects multivariate signals into the frequency domain using a shared graph Fourier basis, and groups patches into low-, mid-, and high-frequency bands based on their spectral energy responses. xCPD then applies a channel-adaptive routing mechanism that dynamically adjusts the degree of inter-channel interaction for each patch, enabling selective activation of frequency-specific experts. This facilitates fine-grained, input-aware modeling of smooth trends, local fluctuations, and abrupt transitions. xCPD can be seamlessly integrated on top of existing CI and CD forecasting models, consistently enhancing both accuracy and generalization across benchmarks.

Systematic reviews (SR), in which experts summarize and analyze evidence across individual studies to provide insights on a specialized topic, are a cornerstone for evidence-based clinical decision-making, research, and policy. Given the exponential growth of scientific articles, there is growing interest in using large language models (LLMs) to automate SR generation. However, the ability of LLMs to critically assess evidence and reason across multiple documents to provide recommendations at the same proficiency as domain experts remains poorly characterized. We therefore ask: Can LLMs match the conclusions of systematic reviews written by clinical experts when given access to the same studies? To explore this question, we present MedEvidence, a benchmark pairing findings from 100 SRs with the studies they are based on. We benchmark 24 LLMs on MedEvidence, including reasoning, non-reasoning, medical specialist, and models across varying sizes (from 7B-700B). Through our systematic evaluation, we find that reasoning does not necessarily improve performance, larger models do not consistently yield greater gains, and knowledge-based fine-tuning degrades accuracy on MedEvidence. Instead, most models exhibit similar behavior: performance tends to degrade as token length increases, their responses show overconfidence, and, contrary to human experts, all models show a lack of scientific skepticism toward low-quality findings. These results suggest that more work is still required before LLMs can reliably match the observations from expert-conducted SRs, even though these systems are already deployed and being used by clinicians. We release our codebase and benchmark to the broader research community to further investigate LLM-based SR systems.

Evaluation of Text to Speech (TTS) systems is challenging and resource-intensive. Subjective metrics such as Mean Opinion Score (MOS) are not easily comparable between works. Objective metrics are frequently used, but rarely validated against subjective ones. Both kinds of metrics are challenged by recent TTS systems capable of producing synthetic speech indistinguishable from real speech. In this work, we introduce Text to Speech Distribution Score 2 (TTSDS2), a more robust and improved version of TTSDS. Across a range of domains and languages, it is the only one out of 16 compared metrics to correlate with a Spearman correlation above 0.50 for every domain and subjective score evaluated. We also release a range of resources for evaluating synthetic speech close to real speech: A dataset with over 11,000 subjective opinion score ratings; a pipeline for recreating a multilingual test dataset to avoid data leakage; and a benchmark for TTS in 14 languages.

While recent flow-based image editing models demonstrate general-purpose capabilities across diverse tasks, they often struggle to specialize in challenging scenarios---particularly those involving large-scale shape transformations. When performing such structural edits, these methods either fail to achieve the intended shape change or inadvertently alter non-target regions, resulting in degraded background quality. We propose $\textbf{Follow-Your-Shape}$, a training- and mask-free framework that supports precise and controllable editing of object shapes while strictly preserving non-target content. Motivated by the divergence between inversion and editing trajectories, we compute a $\textbf{Trajectory Divergence Map (TDM)}$ by comparing token-wise velocity differences between the inversion and denoising paths. The TDM enables precise localization of editable regions and guides a $\textbf{Scheduled KV Injection}$ mechanism that ensures stable and faithful editing. To facilitate a rigorous evaluation, we introduce $\textit{\textbf{ReShapeBench}}$, a new benchmark comprising 120 new images and enriched prompt pairs specifically curated for shape-aware editing. Experiments demonstrate that our method achieves superior editability and visual fidelity, particularly in tasks requiring large-scale shape replacement.

Contrastive learning, commonly applied in large-scale multimodal models, often relies on data from diverse and often unreliable sources, which can include misaligned or mislabeled text-image pairs. This frequently leads to robustness issues and hallucinations, ultimately causing performance degradation. Data valuation is an efficient way to detect and trace these misalignments. Nevertheless, existing methods are computationally expensive for large-scale models. Although computationally efficient, classical influence functions are inadequate for contrastive learning models, as they were initially designed for pointwise loss. Furthermore, contrastive learning involves minimizing the distance between positive sample modalities while maximizing the distance between negative sample modalities. This necessitates evaluating the influence of samples from both perspectives. To tackle these challenges, we introduce the Extended Influence Function for Contrastive Loss (ECIF), an influence function crafted for contrastive loss. ECIF considers both positive and negative samples and provides a closed-form approximation of contrastive learning models, eliminating the need for retraining. Building upon ECIF, we develop a series of algorithms for data evaluation, misalignment detection, and misprediction trace-back tasks. Experimental results demonstrate that our ECIF advances the transparency and interpretability of CLIP-style embedding models by offering a more accurate assessment of data impact and model alignment compared to traditional baseline methods.

Return-conditioned supervised learning (RCSL) algorithms have demonstrated strong generative capabilities in offline reinforcement learning (RL) by learning action distributions based on both the state and the return. However, many existing approaches treat RL as a conditional sequence modeling task, which can lead to an overreliance on suboptimal past experiences, impairing decision-making and reducing the effectiveness of trajectory synthesis. To address these limitations, we propose a novel approach, the Q-Augmented Dual-Feature Fusion Decision Transformer (QDFFDT), which adaptively combines both global sequence features and local immediate features through a learnable fusion mechanism. This model improves generalization across different tasks without the need for extensive hyperparameter tuning. Experimental results on the D4RL benchmark show that QDFFDT outperforms current methods, establishing new state-of-the-art performance and demonstrating the power of adaptive feature fusion.

In clinical applications, the utility of segmentation models is often based on the accuracy of derived downstream metrics such as organ size, rather than by the pixel-level accuracy of the segmentation masks themselves. Thus, uncertainty quantification for such metrics is crucial for decision-making. Conformal prediction (CP) is a popular framework to derive such principled uncertainty guarantees, but applying CP naively to the final scalar metric is inefficient because it treats the complex, non-linear segmentation-to-metric pipeline as a black box. We introduce COMPASS, a practical framework that generates efficient, metric-based CP intervals for image segmentation models by leveraging the inductive biases of their underlying deep neural networks. COMPASS performs calibration directly in the model's representation space by perturbing intermediate features along low-dimensional subspaces maximally sensitive to the target metric. We prove that COMPASS achieves valid marginal coverage under the assumption of exchangeability. Empirically, we demonstrate that COMPASS produces significantly tighter intervals than traditional CP baselines on four medical image segmentation tasks for area estimation of skin lesions and anatomical structures. Furthermore, we show that leveraging learned internal features to estimate importance weights allows COMPASS to also recover target coverage under covariate shifts. COMPASS paves the way for practical, metric-based uncertainty quantification for medical image segmentation.

Recent advances in multi-modal large reasoning models (MLRMs) have shown significant ability to interpret complex visual content. While these models possess impressive reasoning capabilities, they also introduce novel and underexplored privacy risks. In this paper, we identify a novel category of privacy leakage in MLRMs: Adversaries can infer sensitive geolocation information, such as users' home addresses or neighborhoods, from user-generated images, including selfies captured in private settings. To formalize and evaluate these risks, we propose a three-level privacy risk framework that categorizes image based on contextual sensitivity and potential for geolocation inference. We further introduce DoxBench, a curated dataset of 500 real-world images reflecting diverse privacy scenarios divided into 6 categories. Our evaluation across 13 advanced MLRMs and MLLMs demonstrates that most of these models outperform non-expert humans in geolocation inference and can effectively leak location-related private information. This significantly lowers the barrier for adversaries to obtain users' sensitive geolocation information. We further analyze and identify two primary factors contributing to this vulnerability: (1) MLRMs exhibit strong geolocation reasoning capabilities by leveraging visual clues in combination with their internal world knowledge; and (2) MLRMs frequently rely on privacy-related visual clues for inference without any built-in mechanisms to suppress or avoid such usage. To better understand and demonstrate real-world attack feasibility, we propose GeoMiner, a collaborative attack framework that decomposes the prediction process into two stages consisting of clue extraction and reasoning to improve geolocation performance. Our findings highlight the urgent need to reassess inference-time privacy risks in MLRMs to better protect users' sensitive information.

The human brain exhibits remarkable efficiency in processing sequential information, a capability deeply rooted in the temporal selectivity and stochastic competition of neuronal activation. Current continual learning in spiking neural networks (SNNs) faces a critical challenge: balancing task-specific selectivity with adaptive resource allocation and enhancing the robustness with perturbations to mitigate catastrophic forgetting. Considering the intrinsic temporal dynamics of spiking neurons instead of traditional K-winner-take-all (K-WTA) based on firing rate, we explore how to leave networks robust to temporal perturbations in SNNs on lifelong learning tasks. In this paper, we propose Randomized Temporal K-winner-take-all (RTK-WTA) SNNs for lifelong learning, a biologically grounded approach that integrates trace-dependent neuronal activation with probabilistic top-k selection. By dynamically prioritizing neurons based on their spatiotemporal relevance, RTK-WTA SNNs emulate the brain’s ability to modulate neural resources in spatial and temporal dimensions while introducing controlled randomness to prevent overlapping task representations. The proposed RTK-WTA SNNs enhance inter-class margins and robustness through expanded feature space utilization theoretically. The experimental results show that RTK-WTA surpasses deterministic K-WTA by 3.07–5.0\% accuracy on splitMNIST and splitCIFAR100 with elastic weight consolidation. Controlled stochasticity balances temporal coherence and adaptability, offering a scalable framework for lifelong learning in neuromorphic systems.

Neural Radiance Fields (NeRFs) have emerged as a groundbreaking paradigm for representing 3D objects and scenes by encoding shape and appearance information into the weights of a neural network. Recent studies have demonstrated that these weights can be used as input for frameworks designed to address deep learning tasks; however, such frameworks require NeRFs to adhere to a specific, predefined architecture. In this paper, we introduce the first framework capable of processing NeRFs with diverse architectures and performing inference on architectures unseen at training time. We achieve this by training a Graph Meta-Network within an unsupervised representation learning framework, and show that a contrastive objective is conducive to obtaining an architecture-agnostic latent space. In experiments conducted across 13 NeRF architectures belonging to three families (MLPs, tri-planes, and, for the first time, hash tables), our approach demonstrates robust performance in classification, retrieval, and language tasks involving multiple architectures, even unseen at training time, while also matching or exceeding the results of existing frameworks limited to single architectures. Our code and data are available at [this https URL](https://cvlab-unibo.github.io/gmnerf).

Current AI safety frameworks, which often treat harmfulness as binary, lack the flexibility to handle borderline cases where humans meaningfully disagree. To build more pluralistic systems, it is essential to move beyond consensus and instead understand where and why disagreements arise. We introduce PluriHarms, a benchmark designed to systematically study human harm judgments across two key dimensions—the harm axis (benign to harmful) and the agreement axis (agreement to disagreement). Our scalable framework generates prompts that capture diverse AI harms and human values while targeting cases with high disagreement rates, validated by human data. The benchmark includes 150 prompts with 15,000 ratings from 100 human annotators, enriched with demographic and psychological traits and prompt-level features of harmful actions, effects, and values. Our analyses show that prompts that relate to imminent risks and tangible harms amplify perceived harmfulness, while annotator traits (e.g., toxicity experience, education) and their interactions with prompt content explain systematic disagreement. We benchmark AI safety models and alignment methods on PluriHarms, finding that while personalization significantly improves prediction of human harm judgments, considerable room remains for future progress. By explicitly targeting value diversity and disagreement, our work provides a principled benchmark for moving beyond "one-size-fits-all" safety toward pluralistically safe AI.

Generative models such as diffusion and flow-matching offer expressive policies for offline reinforcement learning (RL) by capturing rich, multimodal action distributions, but their iterative sampling introduces high inference costs and training instability due to gradient propagation across sampling steps. We propose the \textit{Single-Step Completion Policy} (SSCP), a generative policy trained with an augmented flow-matching objective to predict direct completion vectors from intermediate flow samples, enabling accurate, one-shot action generation. In an off-policy actor-critic framework, SSCP combines the expressiveness of generative models with the training and inference efficiency of unimodal policies, without requiring long backpropagation chains. Our method scales effectively to offline, offline-to-online, and online RL settings, offering substantial gains in speed and adaptability over diffusion-based baselines. We further extend SSCP to goal-conditioned RL (GCRL), enabling flat policies to exploit subgoal structures without explicit hierarchical inference. SSCP achieves strong results across standard offline RL and GCRL benchmarks, positioning it as a versatile, expressive, and efficient framework for deep RL and sequential decision-making.

Bayesian optimization in large unstructured discrete spaces is often hindered by the computational cost of maximizing acquisition functions due to the absence of gradients. We propose a scalable alternative based on Thompson sampling that eliminates the need for acquisition function maximization by directly parameterizing the probability that a candidate yields the maximum reward. Our approach, *Thompson Sampling via Fine-Tuning* (ToSFiT) leverages the prior knowledge embedded in prompt-conditioned large language models, and incrementally adapts them toward the posterior. Theoretically, we derive a novel regret bound for a variational formulation of Thompson Sampling that matches the strong guarantees of its standard counterpart. Our analysis reveals the critical role of careful adaptation to the posterior probability of maximality—a principle that underpins our ToSFiT algorithm. Empirically, we validate our method on three diverse tasks: FAQ response refinement, thermally stable protein search, and quantum circuit design. Within a collection of methods covering Bayesian optimization, reinforcement learning, and evolutionary search, ToSFiT exhibits both state-of-the-art sample efficiency and computational efficiency.

Large language models are becoming general knowledge engines for diverse applications. However, their computations are deeply entangled after training, resisting modularization which complicates interpretability, auditing, and long-term maintenance. We introduce Jet Expansions, a framework for expanding computational graphs using jet operators that generalize truncated Taylor series. Our method systematically decomposes language models into explicit input-to-output computational paths and complementary remainders. This functional decomposition provides a principled, knife-like operator for cutting through entanglement in LLMs, enabling scalable model inspection. We demonstrate how Jet Expansions ground and subsume the popular interpretability technique Logit Lens, reveal a (super-)exponential path structure with respect to recursive residual depth, and support several interpretability applications, including sketching a transformer language model with $n$-gram statistics extracted from its computations and indexing model toxicity levels *without* curated benchmarks.