ICLR 2026 Papers

The InfoNCE objective, originally introduced for contrastive representation learning, has become a popular choice for mutual information (MI) estimation, despite its indirect connection to MI. In this paper, we demonstrate why InfoNCE should not be regarded as a valid MI estimator, and we introduce a simple modification, which we refer to as *InfoNCE-anchor*, for accurate MI estimation. Our modification introduces an auxiliary \emph{anchor} class, enabling consistent density ratio estimation and yielding a plug-in MI estimator with significantly reduced bias. Beyond this, we generalize our framework using proper scoring rules, which recover InfoNCE-anchor as a special case when the log score is employed. This formulation unifies a broad spectrum of contrastive objectives, including NCE, InfoNCE, and $f$-divergence variants, under a single principled framework. Empirically, we find that InfoNCE-anchor with the log score achieves the most accurate MI estimates; however, in self-supervised representation learning experiments, we find that the anchor does not improve the downstream task performance. These findings corroborate that contrastive representation learning benefits not from accurate MI estimation per se, but from the learning of structured density ratios.

Recent studies have explored leveraging the world knowledge and cognitive capabilities of Vision-Language Models (VLMs) to address the long-tail problem in end-to-end autonomous driving. However, existing methods typically formulate trajectory planning as a language modeling task, where physical actions are output in the language space, potentially leading to issues such as format-violating outputs, infeasible actions, and slow inference speeds. In this paper, we propose ReCogDrive, a novel **Re**inforced **Cog**nitive framework for end-to-end autonomous **Driv**ing, unifying driving understanding and planning by integrating an autoregressive model with a diffusion planner. First, to instill human driving cognition into the VLM, we introduce a hierarchical data pipeline that mimics the sequential cognitive process of human drivers through three stages: generation, refinement, and quality control. Building on this cognitive foundation, we then address the language-action mismatch by injecting the VLM's learned driving priors into a diffusion planner to efficiently generate continuous and stable trajectories. Furthermore, to enhance driving safety and reduce collisions, we introduce a Diffusion Group Relative Policy Optimization (DiffGRPO) stage, reinforcing the planner for enhanced safety and comfort. Extensive experiments on the NAVSIM and Bench2Drive benchmarks demonstrate that ReCogDrive achieves state-of-the-art performance. Additionally, qualitative results across diverse driving scenarios and DriveBench highlight the model's scene comprehension. All code, model weights, and datasets will be made publicly available to facilitate subsequent research.

Fine-tuning large language models (LLMs) on a mixture of diverse datasets poses challenges due to data imbalance and heterogeneity. Existing methods often address these issues across datasets (globally) but overlook the imbalance and heterogeneity within individual datasets (locally), which limits their effectiveness. We introduce Hierarchical Balancing Optimization (HBO), a novel method that enables LLMs to autonomously adjust data allocation during fine-tuning both across datasets (globally) and within each individual dataset (locally). HBO employs a bilevel optimization strategy with two types of actors: a Global Actor, which balances data sampling across different subsets of the training mixture, and several Local Actors, which optimizes data usage within each subset based on difficulty levels. These actors are guided by reward functions derived from the LLM's training state, which measure learning progress and relative performance improvement. We evaluate HBO on three LLM backbones across nine diverse tasks in multilingual and multitask setups. Results show that HBO consistently outperforms existing baselines, achieving significant accuracy gains. Our in-depth analysis further demonstrates that both the global actor and local actors of HBO effectively adjust data usage during fine-tuning. HBO provides a comprehensive solution to the challenges of data imbalance and heterogeneity in LLM fine-tuning, enabling more effective training across diverse datasets.

Fully polynomial neural networks—models whose computations comprise only additions and multiplications—are attractive for privacy-preserving inference under homomorphic encryption (HE). Yet most prior systems obtain such models by *post-hoc* replacement of nonlinearities with high-degree or cascaded polynomials, which inflates HE cost and makes training numerically fragile and hard to scale. We introduce **ULD-Net**, a pretraining methodology that enables *ultra-low-degree* (multiplicative depth ≤ 3 for each operator) fully polynomial networks to be trained from scratch at ImageNet and transformer scale while maintaining high accuracy. The key is a polynomial-only normalization, **PolyNorm**, coupled with a principled choice of normalization axis that keeps activations in a well-conditioned range across deep stacks of polynomial layers. Together with a special set of polynomial-aware operator replacements, such as polynomial activation functions and linear attention, ULD-Net delivers stable optimization without resorting to high-degree approximations. Experimental results demonstrate that ULD-Net outperforms several state-of-the-art open-source fully and partially polynomial approaches across both CNNs and ViTs on diverse datasets, in terms of both accuracy and HE inference latency. Specifically, ULD-Net achieves +0.39% accuracy and a 2.76× speedup compared to the best fully polynomial baseline; up to +3.33% accuracy and a 3.17× speedup compared to the best partial polynomial baseline. Applying ULD-Net to ViT-Small and ViT-Base yields 76.7% and 75.2% top-1 accuracy on ImageNet, demonstrating the first fully polynomial models scaled to the ViT/ImageNet level.

We investigate the role of learning rate scheduling in the large-scale pre-training of large language models, focusing on its influence on downstream performance after supervised fine-tuning (SFT). Decay-based learning rate schedulers are widely used to minimize pre-training loss. However, despite their widespread use, how these schedulers affect performance after SFT remains underexplored. In this paper, we examine Warmup-Stable-Only (WSO), which maintains a constant learning rate after warmup without any decay. Through experiments with 1B and 8B parameter models, we show that WSO consistently outperforms decay-based schedulers in terms of performance after SFT, even though decay-based schedulers may exhibit better performance after pre-training. The result also holds across different regimes with mid-training and over-training. Loss landscape analysis further reveals that decay-based schedulers lead models into sharper minima, whereas WSO preserves flatter minima that support adaptability. These findings indicate that applying LR decay to improve pre-training metrics may compromise downstream adaptability. Our work also provides practical guidance for training and model release strategies, highlighting that pre-training models with WSO enhances their adaptability for downstream tasks.

For multi-stage recommenders in industry, a user request would first trigger a simple and efficient retriever module that selects and ranks a list of relevant items, then the recommender calls a slower but more sophisticated reranking model that refines the item list exposure to the user. To consistently optimize the two-stage retrieval reranking framework, most efforts have focused on learning reranker-aware retrievers. In contrast, there has been limited work on how to achieve a retriever-aware reranker. In this work, we provide evidence that the retriever scores from the previous stage are informative signals that have been underexplored. Specifically, we first empirically show that the reranking task under the two-stage framework is naturally a noise reduction problem on the retriever scores, and theoretically show the limitations of naive utilization techniques of the retriever scores. Following this notion, we derive an adversarial framework DNR that associates the denoising reranker with a carefully designed noise generation module. The resulting DNR solution extends the conventional score error minimization loss with three augmented objectives, including: 1) a denoising objective that aims to denoise the noisy retriever scores to align with the user feedback; 2) an adversarial retriever score generation objective that improves the exploration in the retriever score space; and 3) a distribution regularization term that aims to align the distribution of generated noisy retriever scores with the real ones. We conduct extensive experiments on three public datasets and an industrial recommender system, together with analytical support, to validate the effectiveness of the proposed DNR.

Self-supervised contrastive learning (CL) has achieved remarkable empirical success, often producing representations that rival supervised pre-training on downstream tasks. Recent theory explains this by showing that the CL loss closely approximates a supervised surrogate, Negatives-Only Supervised Contrastive Learning (NSCL), as the number of classes grows. Yet this loss-level similarity leaves an open question: {\em Do CL and NSCL also remain aligned at the representation level throughout training, not just in their objectives?} We address this by analyzing the representation alignment of CL and NSCL models trained under shared randomness (same initialization, batches, and augmentations). First, we show that their induced representations remain similar: specifically, we prove that the similarity matrices of CL and NSCL stay close under realistic conditions. Our bounds provide high-probability guarantees on alignment metrics such as centered kernel alignment (CKA) and representational similarity analysis (RSA), and they clarify how alignment improves with more classes, higher temperatures, and its dependence on batch size. In contrast, we demonstrate that parameter-space coupling is inherently unstable: divergence between CL and NSCL weights can grow exponentially with training time. Finally, we validate these predictions empirically, showing that CL–NSCL alignment strengthens with scale and temperature, and that NSCL tracks CL more closely than other supervised objectives. This positions NSCL as a principled bridge between self-supervised and supervised learning.

The *Learnware* paradigm aims to help users solve machine learning tasks by leveraging existing well-trained models rather than starting from scratch. A learnware comprises a submitted model paired with a *specification* sketching its capabilities. For an open platform with continuously uploaded models, these specifications are essential to enabling users to identify helpful models, eliminating the requirement for prohibitively costly per-model evaluations. In previous research, specifications based on privacy-preserving reduced sets succeed in enabling learnware identification through distribution matching, but suffer from high sample complexity for learnwares from high-dimensional, unstructured data like images or text. In this paper, we formalize **Pa**rameter **Ve**ctor (PAVE) specification for learnware identification, which utilizes the changes in pre-trained model parameters to inherently encode the model capability and task requirements, offering an effective solution for these learnwares. Theoretically, from the neural tangent kernel perspective, we establish a tight connection between PAVE and prior specifications, providing a theoretical explanation for their shared underlying principles. We further approximate the parameter vector in a low-rank space and analyze the approximation error bound, highly reducing the computational and storage overhead. Extensive empirical studies demonstrate that PAVE specification excels at identifying CV and NLP learnwares for reuse on given user tasks, and succeeds in identifying helpful learnwares from open learnware repository with corrupted model quality for the first time. Reusing identified learnware to solve user tasks can even outperform user-fine-tuned pre-trained models in data-limited scenarios.

Learning from Label Proportions~(LLP) aims to use bags of instances associated with the proportions of each label within the bag to learn an instance-level classifier. Proportion matching is a widely used strategy that aligns the average model outputs of all instances in a bag with the label proportions in order to induce the classifier. However, simply fitting the label proportions does not encourage discriminative instance-level predictions and may cause over-smoothing problems, resulting in poor classification performance. In this paper, we propose a novel LLP approach that can mitigate the over-smoothing problems with theoretical guarantees. Rather than fitting the label proportions directly, we treat them as targets for an auxiliary proportional value classification task to induce the target classifier. Our approach only requires the incorporation of an aggregation function after the classification layer. We also introduce an efficient computational approach with a divide-and-conquer strategy. Extensive experiments on various benchmark datasets and under different bag-generation strategies demonstrate that our approach achieves superior performance compared with state-of-the-art LLP methods. The code is publicly available at https://github.com/TianhaoMa5/ICLR2026_LLP-PVC.

Pretrained vision-language models (VLMs), such as CLIP, have shown promise in federated learning (FL) by bringing strong multimodal representations to edge devices. However, continual adaptation remains a core challenge in practical federated settings, where task distributions evolve over time and data remain non-IID across clients. In this emerging area, recent works adopt parameter-efficient fine-tuning (PEFT) as a lightweight way to reduce communication overhead, yet they fail to preserve satisfactory performance under continual learning conditions. Meanwhile, traditional federated continual learning (FCL) methods lack the capacity to maintain cross-modal alignment crucial to VLM performance. We introduce Fed-Duet, a novel Dual Expert-orchestrated framework for efficient federated continual learning in vision-language models. Fed-Duet features a dual-expert adaptation mechanism, combining server-coordinated semantic prompts with client-personalized modular adapters. These pathways are dynamically fused via a cross-attention mechanism, enabling effective knowledge transfer while preserving multimodal alignment and mitigating forgetting. We evaluate Fed-Duet across multiple challenging continual learning tasks in federated vision-language settings and demonstrate that it achieves superior performance and stability compared to existing approaches. Our work highlights the importance of coordinated expert composition in enabling scalable and robust multimodal continual learning. The code is available at https://anonymous.4open.science/r/FedDuet-0426/.

Large language models (LLMs) achieve strong performance across diverse tasks but suffer from high inference latency due to their autoregressive generation. Speculative Decoding (SPD) mitigates this issue by verifying candidate tokens from a smaller draft model in parallel, yet its strict exact-match verification discards many semantically valid continuations. We propose Training-Free Loosely Speculative Decoding (FLy), a novel method that loosens the rigid verification criterion by leveraging the target model’s own corrective behavior to judge whether a draft–target mismatch remains semantically valid. FLy introduces a two-tier mechanism: an entropy-level gate that identifies whether the current token allows multiple plausible alternatives or is nearly deterministic, and a token-level deferred window that distinguishes genuine errors from differently worded yet semantically correct variants. To further reduce latency, we design a multi-level acceleration strategy that accelerates not only the target model but also the drafter itself. Owing to its training-free design, FLy composes seamlessly with arbitrary draft–target pairs and generalizes across models and domains without hyperparameter re-tuning. Experiments show that FLy preserves $\geq$99\% of the target model’s accuracy while achieving an average 2.81$\times$ speedup on Llama-3.1-70B-Instruct and 5.07$\times$ speedup on the 405B variant. Notably, on out-of-domain datasets, our method remains highly effective and outperforms the training-based method EAGLE-3 by 1.62$\times$.

Large language models (LLMs) often experience language confusion, which is the unintended mixing of languages during text generation. Current solutions to this problem either necessitate model retraining or cannot differentiate between harmful confusion and acceptable code-switching. This paper introduces the \textbf{Language Confusion Gate} (LCG), a lightweight, plug-in solution that filters tokens during decoding without altering the base LLM. The LCG is trained using norm-adjusted self-distillation to predict appropriate language families and apply masking only when needed. Our method is based on the findings that language confusion is infrequent, correct-language tokens are usually among the top predictions, and output token embedding norms are larger for high-resource languages, which biases sampling. When evaluated across various models, including Qwen3, GPT-OSS, Gemma3, Llama3.1, LCG decreases language confusion significantly—often by an order of magnitude—without negatively impacting task performance.

Large language model (LLM)-based multi-agent systems have shown strong capabilities in tasks such as code generation and collaborative reasoning. However, the effectiveness and robustness of these systems critically depend on their communication topology, which is often fixed or statically learned, ignoring real-world dynamics such as model upgrades, API (or tool) changes, or knowledge source variability. To address this limitation, we propose CARD (Conditional Agentic Graph Designer), a conditional graph-generation framework that instantiates AMACP, a protocol for adaptive multi-agent communication. CARD explicitly incorporates dynamic environmental signals into graph construction, enabling topology adaptation at both training and runtime. Through a conditional variational graph encoder and environment-aware optimization, CARD produces communication structures that are both effective and resilient to shifts in model capability or resource availability. Empirical results on HumanEval, MATH, and MMLU demonstrate that CARD consistently outperforms static and prompt-based baselines, achieving higher accuracy and robustness across diverse conditions. The source code is available at: \url{https://anonymous.4open.science/r/agentgraph-FF9A}.

Large reasoning models such as OpenAI o1 and DeepSeek-R1 have demonstrated remarkable performance in complex reasoning tasks. A critical component of their training is the incorporation of reference-based reward systems within reinforcement learning (RL), where model outputs are evaluated against ground truth references. However, existing reward benchmarks focus on preference comparisons between responses rather than evaluating verification against ground truth references, leaving a critical gap in our ability to evaluate verification systems used in reasoning model training. In this paper, we introduce VerifyBench and its challenging variant VerifyBench-Hard, two benchmarks specifically designed to assess reference-based reward systems. These benchmarks are constructed through meticulous data collection and curation, followed by careful human annotation to ensure high quality. Our comprehensive evaluation reveals that while larger model-based verifiers show promise on standard cases, all current systems demonstrate substantial room for improvement on challenging instances. Through systematic analysis of performance patterns across reasoning tasks and error categories, we provide insights for advancing reference-based reward systems. These benchmarks establish a standardized framework for improving verification accuracy, ultimately enhancing reasoning capabilities in models trained via RL.

Distinguishing a model's lack of knowledge (epistemic uncertainty) from inherent task randomness (aleatoric uncertainty) is crucial for reliable AI. However, standard evaluation metrics of confidence scores target different aspects. AUC and accuracy capture predictive signal, proper scoring rules capture overall uncertainty, and calibration metrics isolate part of the epistemic uncertainty but ignore heterogeneity of the errors within bins, known as grouping loss. We close this evaluation gap by introducing asymptotically consistent and sample-efficient lower-bound estimators for the grouping loss and excess risk, i.e. suboptimality of a prediction. Our estimators complement existing calibration metrics to provide a more complete, fine-grained assessment of epistemic uncertainty. Applied to LLM question-answering with inherent aleatoric noise, our estimator reveals substantial grouping loss which decreases with model scale but is amplified by instruction tuning. The local nature of our estimators provides actionable insights: they automatically identify subgroups with systematic over- or under-confidence for interpretable audits. We also demonstrate that it reveals better the need of post-training. Finally, we leverage our estimator to design efficient LLM cascades that defer to stronger models, achieving higher accuracy at a lower cost than competing approaches.

Multi-track music generation has garnered significant research interest due to its precise mixing and remixing capabilities. However, existing models often overlook essential attributes such as rhythmic stability and synchronization, leading to a focus on differences between tracks rather than their inherent properties. In this paper, we introduce SyncTrack, a synchronous multi-track waveform music generation model designed to capture the unique characteristics of multi-track music. SyncTrack features a novel architecture that includes track-shared modules to establish a common rhythm across all tracks and track-specific modules to accommodate diverse timbres and pitch ranges. Each track-shared module employs two cross-track attention mechanisms to synchronize rhythmic information, while each track-specific module utilizes learnable instrument priors to better represent timbre and other unique features. Additionally, we enhance the evaluation of multi-track music quality by introducing rhythmic consistency through three novel metrics: Inner-track Rhythmic Stability (IRS), Cross-track Beat Synchronization (CBS), and Cross-track Beat Dispersion (CBD). Experiments demonstrate that SyncTrack significantly improves the multi-track music quality by enhancing rhythmic consistency. Code is available at https://anonymous.4open.science/r/SyncTrack/.

A central challenge in image-based Model-Based Reinforcement Learning (MBRL) is to learn representations that distill essential information from irrelevant visual details. While promising, reconstruction-based methods often waste capacity on large task-irrelevant regions. Decoder-free methods instead learn robust representations by leveraging Data Augmentation (DA), but reliance on such external regularizers limits versatility. We propose R2-Dreamer, a decoder-free MBRL framework with a self-supervised objective that serves as an internal regularizer, preventing representation collapse without resorting to DA. The core of our method is a \emph{redundancy-reduction} objective inspired by Barlow Twins, which can be easily integrated into existing frameworks. On DeepMind Control Suite and Meta-World, R2-Dreamer is competitive with strong baselines such as DreamerV3 and TD-MPC2 while training 1.59$\times$ faster than DreamerV3, and yields substantial gains on DMC-Subtle with tiny task-relevant objects. These results suggest that an effective internal regularizer can enable versatile, high-performance decoder-free MBRL. Code is available at https://github.com/NM512/r2dreamer.

Current speech language models require their core component, the speech codec, to discretize continuous speech signals into tokens that not only capture high-level cues for autoregressive modeling but also preserve sufficient acoustic details for perceptual quality. To address this need, we propose Gogo, a group-wise granularity-ordered codec that quantizes each group of frames into tokens arranged from coarse to fine, where coarse tokens encode high-level abstractions and fine tokens progressively recover low-level details. Building on the granularity-ordering property of Gogo, we introduce GogoSpeech, a two-stage speech language model that performs speech generation by first constructing a coarse speech backbone at an extremely low token rate and then enriching the backbone with fine-grained acoustic details. Considering the inherently non-uniform information distribution in speech signals, we further design a Group Relative Policy Optimization (GRPO)-trained token allocator that adaptively allocates token budgets to groups based on group-wise complexity. Experimental results demonstrate that Gogo delivers state-of-the-art reconstruction performance across most metrics at a token rate of 47. Moreover, evaluations on zero-shot text-to-speech tasks show that GogoSpeech enables efficient generation by adaptively reducing the average token rate, and attains state-of-the-art results in long-form speech generation.

Modern AI algorithms increasingly adopt fused kernels for performance, but implementing them remains complex due to the lack of fine-grained control in existing compilers like Triton. We introduce TileLang, a controllable programming system for fused neural kernels. TileLang provides explicit tile-level primitives for memory placement, data movement, and parallel scheduling. To guide developers in hardware-aware programming, the TileLang introduces two key techniques: tile inference which models tile programs as fused graphs and automatically deduces tile configuration from partial annotations; and tile recommendation that suggests efficient tile configurations based on hardware profiles and heuristics. TileLang makes it easy to express a wide range of fused attention kernels in under 80 lines of Python code, reducing code size by up to 90% compared to manual implementations. Evaluations show that TileLang achieves up to 5x speedup over Triton on NVIDIA H100 and up to 6 on AMD GPUs, demonstrating its ability to bridge programmability and performance.