ICLR 2026 论文列表

LLMs are known to store vast amounts of knowledge in their parametric memory. However, learning and recalling facts from this memory is known to be unreliable, depending largely on the prevalence of particular facts in the training data and other factors which are poorly understood. Practitioners are lacking tools which will allow them to ensure that the models learn a given body of knowledge reliably and consistently. To this end, we propose Active Reading: a framework where we train models to study a given set of material with self-generated learning strategies. First, we demonstrate models trained with Active Reading on expert domains absorb significantly more knowledge than vanilla finetuning and other data augmentations. We train expert 8B models that achieve 66% on a Wikipedia-grounded subset of SimpleQA (+313% relative over vanilla finetuning) and 26% on FinanceBench (+160% relative over vanilla finetuning) by applying Active Reading to the source documents for each benchmark. Finally, we show that Active Reading can be utilized at pre-training scale to build more factual models. As a demonstration of this, we release WikiExpert-8B, a Wikipedia-expert model trained on 1 trillion generated tokens, which outcompetes models with hundreds of billions of parameters on factual QA.

Krotov and Hopfield (2021) proposed a biologically plausible two-layer associative memory network with memory storage capacity exponential in the number of visible neurons. However, the capacity was only linear in the number of hidden neurons. This limitation arose from the choice of nonlinearity between the visible and hidden units, which enforced winner-takes-all dynamics in the hidden layer, thereby restricting each hidden unit to encode only a single memory. We overcome this limitation by introducing a novel associative memory network with a threshold nonlinearity that enables distributed representations. In contrast to winner-takes-all dynamics, where each hidden neuron is tied to an entire memory, our network allows hidden neurons to encode basic components shared across many memories. Consequently, complex patterns are represented through combinations of hidden neurons. These representations reduce redundancy and allow many correlated memories to be stored compositionally. Thus, we achieve much higher capacity: exponential in the number of hidden units, provided the number of visible units is sufficiently larger than the number of hidden neurons. Exponential capacity arises because all binary states of the hidden units can become stable memory patterns with an appropriately chosen threshold. Moreover, the distributed hidden representation, which has much lower dimensionality than the visible layer, preserves class-discriminative structure, supporting efficient nonlinear decoding. These results establish a new regime for associative memory, enabling high-capacity, robust, and scalable architectures consistent with biological constraints.

Large language models (LLMs) have revolutionized natural language processing. Understanding their internal mechanisms is crucial for developing more interpretable and optimized architectures. Mechanistic interpretability has led to the development of various methods for assessing layer relevance, with cosine similarity being a widely used tool in the field. On this work, we demonstrate that cosine similarity is a poor proxy for the actual performance degradation caused by layer removal. Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. On the other hand, empirical evidence from a range of LLMs confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate, leading to misleading interpretations of a transformer's internal mechanisms. We propose a more robust metric for assessing layer relevance: the actual drop in model accuracy resulting from the removal of a layer. Even though it is a computationally costly metric, this approach offers a more accurate picture of layer importance, allowing for more informed pruning strategies and lightweight models. Our findings have significant implications for the development of interpretable LLMs and highlight the need to move beyond cosine similarity in assessing layer relevance.

Recent diffusion-based Multimodal Large Language Models (dMLLMs) suffer from high inference latency and therefore rely on caching techniques to accelerate decoding. However, the application of cache mechanisms often introduces undesirable repetitive text generation, a phenomenon we term the Repeat Curse. To better investigate underlying mechanism behind this issue, we analyze repetition generation through the lens of information flow. Our work reveals three key findings: (1) context tokens aggregate semantic information as anchors and guide the final predictions; (2) as information propagates across layers, the entropy of context tokens converges in deeper layers, reflecting the model’s growing prediction certainty; (3) Repetition is typically linked to disruptions in the information flow of context tokens and to the inability of their entropy to converge in deeper layers. Based on these insights, we present CoTA, a plug-and-play method for mitigating repetition. CoTA enhances the attention of context tokens to preserve intrinsic information flow patterns, while introducing a penalty term to the confidence score during decoding to avoid outputs driven by uncertain context tokens. With extensive experiments, CoTA demonstrates significant effectiveness in alleviating repetition and achieves consistent performance improvements on general tasks. Code will be made available.

The attention mechanism in Transformers is an important primitive for accurate and scalable sequence modeling. Its quadratic-compute and linear-memory complexity however remain significant bottlenecks. Linear attention and state-space models enable linear-time, constant-memory sequence modeling and can moreover be trained efficiently through matmul-rich parallelization across sequence length. However, at their core these models are still RNNs, and thus their use of a fixed-size hidden state to model the context is a fundamental limitation. This paper develops log-linear attention, an attention mechanism that balances linear attention's efficiency and the expressiveness of softmax attention. Log-linear attention replaces the fixed-size hidden state with a logarithmically growing set of hidden states. We show that with a particular growth function, log-linear attention admits a similarly matmul-rich parallel form whose compute cost is log-linear in sequence length. Log-linear attention is a general framework and can be applied on top of existing linear attention variants. As case studies, we instantiate log-linear variants of two recent architectures---Mamba-2 and Gated DeltaNet---and find they perform well compared to their linear-time variants.

Conformal prediction can be used to construct prediction sets that cover the true outcome with a desired probability, but can sometimes lead to large prediction sets that are costly in practice. The most useful outcome is a singleton prediction---an unambiguous decision---yet existing efficiency-oriented methods primarily optimize average set size. Motivated by this, we propose a new non-conformity score that is motivated by minimizing the probability of producing non-singleton sets while maintaining coverage. Starting from a non-convex constrained optimization problem as a motivation, we provide a convex-geometric reformulation and associated algorithm for computing the non-conformity score and associated split conformal prediction sets in $O(K)$ time for $K$-class problems. Using this score in split conformal prediction, we introduce Singleton-Optimized Conformal Prediction (SOCOP). We evaluate our method in experiments on image classification and LLM multiple-choice answering, comparing with standard non-conformity scores such as the (negative) label probability estimates and their cumulative distribution function; both of which are motivated by aiming to optimize average length. The results show that SOCOP increases singleton frequency (sometimes by over 20\%) compared to the above scores, with minimal impact on average set size.

Tool-Integrated Reasoning (TIR) enables large language models (LLMs) to enhance their internal reasoning ability by integrating external tools. However, models with TIR often exhibit suboptimal behaviors, including insufficient tool calls, excessive tool calls, and overthinking after receiving tool call results. How to empower LLMs to perform TIR efficiently and accurately, while stabilizing the reasoning process, remains an open challenge. In this paper, we first analyze the impact of tool calls on model reasoning from the perspective of information entropy. We find that when tool call results are provided, the information entropy of subsequent reasoning content will show a clear trend of change, and the overall information entropy of the reasoning chain will vary depending on the number of tool calls. Based on these observations, we propose Tool-Light, a framework designed to encourage LLMs to perform TIR efficiently and accurately. Our framework consists of dataset construction and multi-stage fine-tuning. For dataset construction, we use the trained model for continuous self-evolved sampling, integrating two methods: vanilla sampling and entropy-guided sampling. At the same time, during the sampling process, we design strict criteria for selecting positive-negative pairs. For the training process, we introduce a two-stage method, which includes a Supervised Fine-Tuning (SFT), and Self-Evolved Direct Preference Optimization (DPO). Test results on 10 datasets reveal the effectiveness of Tool-Light, significantly improving the efficiency and accuracy of the model in completing TIR tasks.

Recent attempts to transfer features from 2D Vision–Language Models (VLMs) to 3D semantic segmentation expose a persistent trade-off. Directly projecting 2D features into 3D yields noisy and fragmented predictions, whereas enforcing geometric coherence necessitates costly training pipelines and large-scale, annotated 3D data. We argue that this limitation stems from the dominant \textit{segmentation-and-matching} paradigm, which fails to reconcile 2D semantics with 3D geometric structure. The geometric cues are not eliminated during the 2D-to-3D transfer but remain latent within the noisy and view-aggregated features. To exploit this property, we propose \textbf{GeoPurify} that applies a small Student Affinity Network to purify 2D VLM-generated 3D point features using geometric priors distilled from a 3D self-supervised teacher model. During inference, we devise a Geometry-Guided Pooling module to further denoise the point cloud and ensure the semantic and structural consistency. Benefiting from latent geometric information and the learned affinity network, GeoPurify effectively mitigates the trade-off and achieves superior data efficiency. Extensive experiments on major 3D benchmarks demonstrate that GeoPurify achieves or surpasses state-of-the-art performance while utilizing only \textbf{$\sim$1.5\%} of the training data.

Classifier-Free Guidance (CFG) is a widely used technique for conditional generation and improving sample quality in continuous diffusion models, and its extensions to discrete diffusion has recently started to be investigated. In order to improve the algorithms in a principled way, this paper starts by analyzing the exact effect of CFG in the context of a low-dimensional masked diffusion model, with a special emphasis on the guidance schedule. Our analysis shows that high guidance early in sampling (when inputs are heavily masked) harms generation quality, while late-stage guidance has a larger effect. These findings provide a theoretical explanation for empirical observations in recent studies on guidance schedules. The analysis also reveals an imperfection of the current CFG implementations. These implementations can unintentionally cause imbalanced transitions, such as unmasking too rapidly during the early stages of generation, which degrades the quality of the resulting samples. To address this, we draw insight from the analysis and propose a novel classifier-free guidance mechanism. Intuitively, our method smoothens the transport between the data distribution and the initial (masked) distribution, which results in improved sample quality. Remarkably, our method is achievable via a simple one-line code change. Experiments on conditional image and text generation empirically confirm the efficacy of our method.

Meta-learning offers a principled framework leveraging *task-invariant* priors from related tasks, with which *task-specific* models can be fine-tuned on downstream tasks, even with limited data records. Gradient-based meta-learning (GBML) relies on gradient descent (GD) to adapt the prior to a new task. Albeit effective, these methods incur high computational overhead that scales linearly with the number of GD steps. To enhance efficiency and scalability, existing methods approximate the gradient of prior parameters (meta-gradient) via truncated backpropagation, yet suffer large approximation errors. Targeting accurate approximation, this work puts forth binomial GBML (BinomGBML), which relies on a truncated binomial expansion for meta-gradient estimation. This novel expansion endows more information in the meta-gradient estimation via efficient parallel computation. As a running paradigm applied to model-agnostic meta-learning (MAML), the resultant BinomMAML provably enjoys error bounds that not only improve upon existing approaches, but also decay super-exponentially under mild conditions. Numerical tests corroborate the theoretical analysis and showcase boosted performance with slightly increased computational overhead.

Recent advances in scientific machine learning (SciML) have enabled neural operators (NOs) to serve as powerful surrogates for modeling the dynamic evolution of physical systems governed by partial differential equations (PDEs). While existing approaches focus primarily on learning simulations from the target PDE, they often overlook more fundamental physical principles underlying these equations. Inspired by how numerical solvers are compatible with simulations of different settings of PDEs, we propose a multiphysics training framework that jointly learns from both the original PDEs and their simplified basic forms. Our framework enhances data efficiency, reduces predictive errors, and improves out-of-distribution (OOD) generalization, particularly in scenarios involving shifts of physical parameters and synthetic-to-real transfer. Our method is architecture-agnostic and demonstrates consistent improvements in normalized root mean square error (nRMSE) across a wide range of 1D/2D/3D PDE problems. Through extensive experiments, we show that explicit incorporation of fundamental physics knowledge significantly strengthens the generalization ability of neural operators. We will release models and codes at https://sites.google.com/view/sciml-fundemental-pde.

Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in complex problem-solving through Chain-of-Thought (CoT) reasoning. However, the multi-step nature of CoT introduces new safety challenges that extend beyond conventional language model alignment. We identify a failure mode in current safety CoT tuning methods: the snowball effect, where minor reasoning deviations progressively amplify throughout the thought process, leading to either harmful compliance or excessive refusal. This effect stems from models being trained to imitate perfect reasoning scripts without learning to self-correct. To address this limitation, we propose AdvChain, an alignment paradigm that teaches models dynamic self-correction through adversarial CoT tuning. Our method involves constructing a dataset containing Temptation-Correction and Hesitation-Correction samples, where models learn to recover from harmful reasoning drifts and unnecessary cautions. Extensive experiments show that AdvChain significantly enhances robustness against jailbreak attacks and CoT hijacking while substantially reducing over-refusal on benign prompts, achieving a superior safety-utility balance without compromising reasoning capabilities. Our work establishes a new direction for building more robust and reliable reasoning models.

While Large Language Models (LLMs) have shown remarkable advancements in reasoning and tool use, they often fail to generate optimal, grounded solutions under complex constraints. Real-world travel planning exemplifies these challenges, evaluating agents’ abilities to handle constraints that are explicit, implicit, and even evolving based on interactions with dynamic environments and user needs. In this paper, we present ATLAS, a general multi-agent framework designed to effectively handle such complex nature of constraints awareness in real-world travel planning tasks. ATLAS introduces a principled approach to address the fundamental challenges of constraint-aware planning through dedicated mechanisms for dynamic constraint management, iterative plan critique, and adaptive interleaved search. ATLAS demonstrates state-of-the-art performance on the TravelPlanner benchmark, improving the final pass rate from 23.3% to 44.4% over its best alternative. More importantly, our work is the first to demonstrate quantitative effectiveness on real-world travel planning tasks with live information search and multi-turn feedback. In this realistic setting, ATLAS showcases its superior overall planning performance, achieving an 84% final pass rate which significantly outperforms baselines including ReAct (59%) and a monolithic agent (27%).

Instruction tuning improves the LLMs performance but depends on high-quality training data. Recently, LLMs have been used to synthesize data, enhancing training with seeds like question-answer (QA) pairs. However, this synthesis often results in instruction examples similar to the seeds, lacking diversity and biasing real applications. Thus, we propose to extract instruction tuning data from web corpus with much rich knowledge. The most straightforward strategy is to quickly retrieve domain specific documents from the corpus and then extract all QA pairs of these documents for tuning LLMs, which has two main limitations. (1) Extracting all QA pairs using LLMs is prohibitively expensive; and (2) These extracted pairs are not all beneficial for the downstream applications, and incorporating all of them for tuning may even hurt the model performance. To overcome the limitations, we introduce $\texttt{EQUAL}$, an $\textbf{E}$ffective and scalable data extraction framework that iteratively interleaves document selection and extract high-$\textbf{QUAL}$ity QA pairs to optimize instruction tuning. $\texttt{EQUAL}$ first clusters the document set based on the embeddings generated by contrastive learning. Then it leverages the multi-armed bandit based strategy to quickly identify document clusters where can extract high-quality QA pairs for training. This iterative framework significantly reduces computational costs while improving model performance much. Experiments on AutoMathText, KnowledgePile and StackOverflow across 13 downstream tasks demonstrate that $\texttt{EQUAL}$ reduces computational costs by 5–10$\times$ while improving accuracy by 2.5\% on LLaMA-3.1-8B, Qwen2.5-7B and Mistral-7B. Code and data is available at https://anonymous.4open.science/r/EQUAL-DD20.

Evaluating language models fairly is becoming harder as static benchmarks available on the internet risk contamination by training data. This makes it unclear whether models are truly reasoning or just recalling answers. In this paper, we introduce $\textbf{BeyondBench}$, an evaluation framework that avoids this problem by using $\textbf{algorithmic problem generation}$. Unlike traditional benchmarks that risk contamination from internet-scale training data, $\textbf{BeyondBench}$ creates mathematically grounded problems on the fly, ensuring each test remains fresh and uncontaminated. Our framework covers $\textbf{44 algorithmic tasks}$ with a total of $\textbf{117 variations}$, grouped into three difficulty levels: the $\textit{Easy Suite}$ (29 tasks) for basic arithmetic and statistics, the $\textit{Medium Suite}$ (5 tasks, 49 variations) for sequence patterns and reasoning, and the $\textit{Hard Suite}$ (10 tasks, 68 variations) tackling NP-complete and constraint satisfaction problems. Each task generates problems from a combinatorial space larger than $10^{15}$ unique instances, with solutions verified deterministically by mathematical proofs. We evaluated $\textbf{101 language models}$, including 85 open-source and 16 closed-source models, spanning sizes from 0.5B to 141B parameters and multiple quantization schemes. All evaluations use three-fold cross-validation to ensure statistical robustness. Our results show consistent reasoning deficiencies across model families, with performance degrading sharply as problem complexity increases from polynomial to exponential. In our Hard Suite evaluations, models such as Gemini-2.5-pro, Llama-3.3-70B, and Qwen2.5-72B achieved average accuracies of 56.38%, 26.91%, and 33.60%, respectively. Moreover, we observe that performance drops drastically without tool usage, with GPT-5, GPT-5-mini, and GPT-5-nano showing a $\textbf{decline}$ of 16.81% 28.05%, and 47.59% accuracy on the hard suite. The contamination resistance of $\textbf{BeyondBench}$ rests on three guarantees: (i) the problem space is vastly larger than any static dataset, (ii) every instance has a unique, verifiable solution, and (iii) isomorphic transformations generate semantically equivalent but syntactically new problems. $\textbf{BeyondBench}$ redefines reasoning evaluation through genuine algorithmic problem-solving capability, ensuring a fair and meaningful evaluation.

Gradient orthogonalization is a simple strategy that shows great utility in speeding up gradient descent. The Muon optimizer (Keller et al., 2024b) combines gradient orthogonalization with first-order momentum and achieves significant improvement in data efficiency over Adam/AdamW for language model training. However, when using model parallelism, gradient orthogonalization introduces additional overhead compared to coordinate-wise optimizers (such as AdamW) due to additional gather and scatter operations on gradient matrix shards from different devices. This additional communication can amount to a throughput hit of 5\%-10\% compared to Adam/AdamW. To remedy this, we propose Muon with Block-Periodic Orthogonalization (MuonBP), which applies orthogonalization independently to matrix shards on each device and periodically performs full orthogonalization to maintain training stability at scale. We show how to adjust the learning rate from the baseline to MuonBP and give convergence guarantees for this algorithm. Crucially, our theory dictates that we use two stepsizes: one for the blockwise orthogonalization steps, and one for the full orthogonalization steps. Our method is simple, requires minimal hyperparameter adjustments, and achieves competitive iteration complexity compared with the baselines Muon and Dion while providing per-iteration throughput comparable to coordinate-wise methods such as AdamW. When training an 8B model with eight-way tensor parallelism and ZeRO optimizer state sharding, MuonBP achieves 8\% throughput increase compared to Muon with no degradation in performance.

Calibration has emerged as a foundational goal in trustworthy machine learning, in part because of its strong decision theoretic semantics. Independent of the underlying distribution, and independent of the decision maker's utility function, calibration promises that amongst all policies mapping predictions to actions, the uniformly best policy is the one that trusts the predictions and acts as if they were correct. But this is true only of fully calibrated forecasts, which are tractable to guarantee only for very low dimensional prediction problems. For higher dimensional prediction problems (e.g. when outcomes are multiclass), weaker forms of calibration have been studied that lack these decision theoretic properties. In this paper we study how a conservative decision maker should map predictions endowed with these weaker (partial) calibration guarantees to actions, in a way that is robust in a minimax sense: i.e. to maximize their expected utility in the worst case over distributions consistent with the calibration guarantees. We characterize their minimax optimal decision rule via a duality argument, and show that surprisingly, trusting the predictions and acting accordingly is recovered in this minimax sense by decision calibration (and any strictly stronger notion of calibration), a substantially weaker and more tractable condition than full calibration. For calibration guarantees that fall short of decision calibration, the minimax optimal decision rule is still efficiently computable, and we provide an empirical evaluation of a natural one that applies to any regression model solved to optimize squared error.

The 3D occupancy prediction task has witnessed remarkable progress in recent years, playing a crucial role in vision-based autonomous driving systems. While traditional methods are limited to fixed semantic categories, recent approaches have moved towards predicting text-aligned features to enable open-vocabulary text queries in real-world scenes. However, there exists a trade-off in text-aligned scene modeling: sparse Gaussian representation struggles to capture small objects in the scene, while dense representation incurs significant computational overhead. To address these limitations, we present **PG-Occ**, an innovative **P**rogressive **G**aussian Transformer Framework that enables open-vocabulary 3D occupancy prediction. Our framework employs progressive online densification, a feed-forward strategy that gradually enhances the 3D Gaussian representation to capture fine-grained scene details. By iteratively enhancing the representation, the framework achieves increasingly precise and detailed scene understanding. Another key contribution is the introduction of an anisotropy-aware sampling strategy with spatio-temporal fusion, which adaptively assigns receptive fields to Gaussians at different scales and stages, enabling more effective feature aggregation and richer scene information capture. Through extensive evaluations, we demonstrate that **PG-Occ** achieves *state-of-the-art* performance with a relative **14.3\% mIoU improvement** over the previous best performing method. The source code and models will be made publicly available upon publication.

Graph Neural Networks (GNNs) have made significant advances in representation learning on various types of graph-structured data. However, GNNs struggle to simultaneously model heterophily and homophily, a challenge that is amplified under self-supervised learning (SSL) where no labels are available to guide the training process. This paper presents H$^3$GNNs, an end-to-end graph SSL framework designed to harmonize heterophily and homophily through two complementary innovative perspectives: (i) Representation Harmonization via Joint Structural Node Encoding. Nodes are embedded into a unified latent space that retains both node specificity and graph structural awareness for harmonizing heterophily and homophily. Node specificity is learned via linear and non-linear node feature projections. Graph structural awareness is learned via a proposed Weighted Graph Convolutional Network (WGCN). A self-attention module enables the model learning-to-adapt to varying levels of patterns. (ii) Objective Harmonization via Predictive Architecture with Node-Difficulty–Aware Masking. A teacher network processes the full graph. A student network receives a partially masked graph. The student is trained end-to-end, while the teacher is an exponential moving average of the student. The proxy task is to train the student to predict the teacher’s embeddings for all nodes (masked and unmasked). To keep the objective informative across the graph, two masking strategies that guide selection toward currently hard nodes while retaining exploration are proposed. Theoretical underpinnings of H$^3$GNNs are also analyzed in detail. Comprehensive evaluations on benchmarks demonstrate that H$^3$GNNs achieves state-of-the-art performance on heterophilic graphs (e.g., +7.1% on Texas, +9.6% on Roman-Empire over the prior art) while matching SOTA on homophilic graphs, and delivering strong computational efficiency.