ICLR 2026 Papers

A common approach to mechanistic interpretability is to causally manipulate model representations via targeted interventions in order to understand what those representations encode. Here we ask whether such interventions create out-of-distribution (divergent) representations, and whether this raises concerns about how faithful their resulting explanations are to the target model in its natural state. First, we demonstrate theoretically and empirically that common causal intervention techniques often do shift internal representations away from the natural distribution of the target model. Then, we provide a theoretical analysis of two cases of such divergences: "harmless" divergences that occur in the behavioral null-space of the layer(s) of interest, and "pernicious" divergences that activate hidden network pathways and cause dormant behavioral changes. Finally, in an effort to mitigate the pernicious cases, we apply and modify the Counterfactual Latent (CL) loss from Grant (2025) allowing representations from causal interventions to remain closer to the natural distribution, reducing the likelihood of harmful divergences while preserving the interpretive power of the interventions. Together, these results highlight a path towards more reliable interpretability methods.

This work proposes a compositional AND–OR template for art and design that encodes the part–relation–geometry organization of images in a structured and interpretable form. Within a maximum-entropy log-linear model, we define a unified consistency score as log-likelihood gain against a reference distribution and decompose it into term-level evidence, enabling an evidence-to-prescription mapping for actionable composition guidance. Learning is performed by a penalized EM-style block-pursuit with sparsity and local mutual exclusivity: object templates are learned first and reused as scene terminals to induce scene templates. A semi-supervised structural expansion, which is triggered by matching gain and structural-consistency thresholds, bootstraps new branches from unlabeled, high-quality images. Evaluations on a curated compositional dataset and AVA/AADB themes show strong agreement with expert paradigms, interpretable parse trees, and competitive performance with deep baselines while exhibiting higher alignment with human ratings. The learned templates also act as lightweight structural conditions to steer AIGC generation and layout design. Overall, the framework delivers a transferable structural prior with favorable data/parameter efficiency and a unified pathway for explainable visual assessment and generation.

Deep Q-learning jointly learns representations and values within monolithic networks, promising beneficial co-adaptation between features and value estimates. Although this architecture has attained substantial success, the coupling between representation and value learning creates instability as representations must constantly adapt to non-stationary value targets, while value estimates depend on these shifting representations. This is compounded by high variance in bootstrapped targets, which causes bias in value estimation in off-policy methods. We introduce Stackelberg Coupled Representation and Reinforcement Learning (SCORER), a framework for value-based RL that views representation and Q-learning as two strategic agents in a hierarchical game. SCORER models the Q-function as the leader, which commits to its strategy by updating less frequently, while the perception network (encoder) acts as the follower, adapting more frequently to learn representations that minimize Bellman error variance given the leader's committed strategy. Through this division of labor, the Q-function minimizes MSBE while perception minimizes its variance, thereby reducing bias accordingly, with asymmetric updates allowing stable co-adaptation, unlike simultaneous parameter updates in monolithic solutions. Our proposed SCORER framework leads to a bi-level optimization problem whose solution is approximated by a two-timescale algorithm that creates an asymmetric learning dynamic between the two players. Extensive experiments on DQN and its variants demonstrate that gains stem from algorithmic insight rather than model complexity.

We study offline change point localization and inference in dynamic multilayer random dot product graphs (D-MRDPGs), where at each time point, a multilayer network is observed with shared node latent positions and time-varying, layer-specific connectivity patterns. We propose a novel two-stage algorithm that combines seeded binary segmentation with low-rank tensor estimation, and establish its consistency in estimating both the number and locations of change points. Furthermore, we derive the limiting distributions of the refined estimators under both vanishing and non-vanishing jump regimes. To the best of our knowledge, this is the first result of its kind in the context of dynamic network data. We also develop a fully data-driven procedure for constructing confidence intervals. Extensive numerical experiments demonstrate the superior performance and practical utility of our methods compared to existing alternatives.

Translating the internal representations and computations of models into concepts that humans can understand is a key goal of interpretability. While recent dictionary learning methods such as Sparse Autoencoders (SAEs) provide a promising route to discover human-interpretable features, they often only recover token-specific, noisy, or highly local concepts. We argue that this limitation stems from neglecting the temporal structure of language, where semantic content typically evolves smoothly over sequences. Building on this insight, we introduce Temporal Sparse Autoencoders (T-SAEs), which incorporate a novel contrastive loss encouraging consistent activations of high-level features over adjacent tokens. This simple yet powerful modification enables SAEs to disentangle semantic from syntactic features in a self-supervised manner. Across multiple datasets and models, T-SAEs recover smoother, more coherent semantic concepts without sacrificing reconstruction quality. Strikingly, they exhibit clear semantic structure despite being trained without explicit semantic signal, offering a new pathway for unsupervised interpretability in language models.

This paper presents *the first study of grokking in practical LLM pretraining*. Specifically, we investigate when an LLM memorizes the training data, when its generalization on downstream tasks starts to improve, and what happens if there is a lag between the two. Unlike existing works studying when a small model generalizes to limited and specified tasks during thousands epochs' training on algorithmic data, we focus on a practical setting for LLMs, i.e., near single-pass pretraining of next-token prediction on a cross-domain, large-scale corpus, and generalization on diverse benchmark tasks covering math/commonsense reasoning, code generation, and domain-specific retrieval. Our study, *for the first time, verifies that grokking still emerges in pretraining mixture-of-experts (MoE) LLMs*, though different local data groups may enter their grokking stages asynchronously due to the heterogeneity of their distributions and attributions to others. To find a mechanistic interpretation of this local grokking, we investigate the dynamics of training data's pathways (i.e., expert choices across layers in MoE). Our primary discovery is that *the pathways evolve from random, non-smooth across layers, instance-specific to more structured and transferable across samples*, despite the converged pretraining loss. This depicts a transition from memorization to generalization. Two novel metrics are developed to quantify these patterns: one computes the pathway similarity between samples, while the other measures the consistency of aggregated experts between subsequent layers for each sample. These training data based metrics induce near zero cost but can faithfully track and monitor the generalization of LLMs on downstream tasks, reducing reliance on costly instruction tuning and benchmark evaluations. We also ground our findings in a theoretical analysis of one-layer MoE, showing that more structured pathways improve the generalization bound.

Inspired by behavioral science, we propose Behavior Learning (BL), a novel general-purpose machine learning framework that learns interpretable and identifiable optimization structures from data, ranging from single optimization problems to hierarchical compositions. It unifies predictive performance, intrinsic interpretability, and identifiability, with broad applicability to scientific domains involving optimization. BL parameterizes a compositional utility function built from intrinsically interpretable modular blocks, which induces a data distribution for prediction and generation. Each block represents and can be written in symbolic form as a utility maximization problem (UMP), a foundational paradigm in behavioral science and a universal framework of optimization. BL supports architectures ranging from a single UMP to hierarchical compositions, the latter modeling hierarchical optimization structures that offer both expressiveness and structural transparency. Its smooth and monotone variant (IBL) guarantees identifiability under mild conditions. Theoretically, we establish the universal approximation property of both BL and IBL, and analyze the M-estimation properties of IBL. Empirically, BL demonstrates strong predictive performance, intrinsic interpretability and scalability to high-dimensional data. Code: https://github.com/MoonYLiang/Behavior-Learning; installable via pip install blnetwork.

Learning to Explain is a forward-looking paradigm recently proposed in the field of explainable AI, which envisions training explainers capable of producing high-quality explanations for target models efficiently. Although existing studies have made attempts through self-supervised optimization or learning from prior explanation methods, the Learning to Explain paradigm still faces three critical challenges: 1) self-supervised objectives often rely on assumptions about the target model or task, restricting their generalizability; 2) methods driven by prior explanations struggle to guarantee the quality of the supervisory signals; and 3) depending exclusively on either approach leads to poor convergence or limited explanation quality. To address these challenges, we propose a faithfulness-guided amortized explainer that 1) theoretically derives a self-supervised objective free from assumptions about the target model or task, 2) practically generates high-quality supervisory signals by deduplicating and filtering prior explanations, and 3) jointly optimizes both objectives via a dynamic weighting strategy, enabling the amortized explainer to produce more faithful explanations for complex, high-dimensional models. We re-formalize multiple well-validated faithfulness evaluation metrics within a unified notation system and theoretically prove that an explanation mapping can simultaneously achieve optimality across all these metrics. We aggregate prior explanation methods to generate high-quality supervised signals through deduplicating and faithfulness-based filtering. Our amortized explainer leverages dynamic weighting to guide optimization, initially emphasizing pattern consistency with the supervised signals for rapid convergence, and subsequently refining explanation quality by approximating the most faithful explanation mapping. Extensive experiments across various target models and image, text, and tabular tasks demonstrate that the proposed explainer consistently outperforms all prior explanation methods across all faithfulness metrics, highlighting its effectiveness and its potential to offer a systematic solution to the fundamental challenges of the Learning to Explain paradigm.

Explaining predictions of black-box time-series models remains a challenging problem due to the dynamically evolving patterns within individual sequences and their complex temporal dependencies. Unfortunately, existing explanation methods largely focus on point-wise explanations, which fail to capture broader temporal context, while methods that attempt to highlight interpretable temporal patterns (e.g., achieved by incorporating a regularizer or fixed-length patches) often lack principled definitions of meaningful segments. This limitation frequently leads to fragmented and confusing explanations for end users. As such, the notion of segment-wise explanations has remained underexplored, with little consensus on what constitutes an *interpretable* segment or how such segments should be identified. To bridge this gap, we define segment-wise explanation for black-box time-series models as the task of selecting contiguous subsequences that maximize their joint mutual information with the target prediction. Building on this formulation, we propose TimeSeg, a novel information-theoretic framework that employs reinforcement learning to sequentially identify predictive temporal segments at a per-instance level. By doing so, TimeSeg produces segment-wise explanations that capture holistic temporal patterns rather than fragmented points, providing class-predictive patterns in a human-interpretable manner. Extensive experiments on both synthetic and real‑world datasets demonstrate that TimeSeg produces more coherent and human-understandable explanations, while achieving performance that matches or surpasses existing methods on downstream tasks using the identified segments.

We partially reverse-engineer a convolutional recurrent neural network (RNN) trained with model-free reinforcement learning to play the box-pushing game Sokoban. We find that the RNN stores future moves (plans) as activations in particular channels of the hidden state, which we call *path channels*. A high activation in a particular location means that, when a box is in that location, it will get pushed in the channel's assigned direction. We examine the convolutional kernels between path channels and find that they encode the change in position resulting from each possible action, thus representing part of a learned *transition model*. The RNN constructs plans by starting at the boxes and goals. These kernels, *extend* activations in path channels forwards from boxes and backwards from the goal. Negative values are placed in channels at obstacles. This causes the extension kernels to propagate the negative value in reverse, thus pruning the last few steps and letting an alternative plan emerge; a form of backtracking. Our work shows that, a precise understanding of the plan representation allows us to directly understand the bidirectional planning-like algorithm learned by model-free training in more familiar terms.

In this work, we investigate the risks of collective financial fraud in large-scale multi-agent systems, driven by large language model (LLM) agents. We examine whether agents can collaborate in fraudulent activities, amplify the risks of such behaviors, and identify factors critical to fraud success. To facilitate this research, we introduce MultiAgentFraudBench, a large-scale benchmark for simulating financial fraud scenarios based on realistic online behaviors. The benchmark includes 21 typical online fraud scenarios, covering the full fraud lifecycle across both public and private domains. We explore the dynamics of fraud operations by analyzing interaction depth, hype-building effects, and collaboration failures. Finally, we propose two potential mitigation strategies: the development of monitor agents to block malicious agents and fostering group resilience through information sharing. Our findings highlight the real-world dangers of multi-agent fraud and suggest measures for reducing associated risks.

Mixture-of-Experts (MoE) architectures have become the key to scaling modern LLMs, yet little is understood about how their sparse routing dynamics respond to multilingual data. In this work, we analyze expert routing patterns using parallel multilingual datasets and present highly interpretable layer-wise phenomena. We find that MoE models route tokens in language-specific ways in the early and late decoder layers but exhibit significant cross-lingual routing alignment in middle layers, mirroring parameter-sharing trends observed in dense LLMs. In particular, we reveal a clear, strong correlation between a model's performance in a given language and how similarly its tokens are routed to English in these layers. Extending beyond correlation, we explore inference-time interventions that induce higher cross-lingual routing alignment. We introduce a method that steers the router by promoting middle-layer task experts frequently activated in English, and it successfully increases multilingual performance. These 1-2% gains are remarkably consistent across two evaluation tasks, three models, and 15+ languages, especially given that these simple interventions override routers of extensively trained, state-of-the-art LLMs. In comparison, interventions outside of the middle layers or targeting multilingual-specialized experts only yield performance degradation. Altogether, we present numerous findings that explain how MoEs process non-English text and demonstrate that generalization is limited by the model’s ability to leverage language-universal experts in all languages.

Using intelligent systems to perceive psychological and social behaviors, that is, the underlying affective, cognitive, and pathological states that are manifested through observable behaviors and social interactions, remains a challenge due to their complex, multifaceted, and personalized nature. Existing work tackling these dimensions through specialized datasets and single-task systems often miss opportunities for scalability, cross-task transfer, and broader generalization. To address this gap, we curate Human Behavior Atlas, a unified benchmark of diverse behavioral tasks designed to support the development of foundation models for understanding psychological and social behaviors. Human Behavior Atlas comprises over 100,000 samples spanning text, audio, and visual modalities, covering tasks on *affective states*, *cognitive states*, *pathologies*, and *social processes*. Our unification efforts can reduce redundancy and cost, enable training to scale efficiently across tasks, and enhance generalization of behavioral features across domains. On Human Behavior Atlas, we train three models: Omnisapiens-7B SFT, Omnisapiens-7B BAM, and Omnisapiens-7B RL. We show that training on Human Behavior Atlas enables models to consistently outperform existing multimodal LLMs across diverse behavioral tasks. Pretraining on Human Behavior Atlas also improves transfer to novel behavioral datasets; with the targeted use of behavioral descriptors yielding meaningful performance gains. The benchmark, models, and codes can be found at: https://github.com/MIT-MI/human_behavior_atlas.

The Linear Representation Hypothesis asserts that the embeddings learned by neural networks can be understood as linear combinations of features corresponding to high-level concepts. Based on this ansatz, sparse autoencoders (SAEs) have recently become a popular method for decomposing embeddings into a sparse combination of linear directions, which have been shown empirically to often correspond to human-interpretable semantics. However, recent attempts to apply SAEs to multimodal embedding spaces (such as the popular CLIP embeddings for image/text data) have found that SAEs often learn ``split dictionaries,'' where most of the learned sparse features are essentially unimodal, active only for data of a single modality. In this work, we study how to effectively adapt SAEs for the setting of multimodal embeddings while ensuring multimodal alignment. We first argue that the existence of a split dictionary decomposition on an aligned embedding space implies the existence of a non-split dictionary with improved modality alignment. Then, we propose a new SAE-based approach to multimodal embedding decomposition using cross-modal random masking and group-sparse regularization. We apply our method to popular embeddings for image/text (CLIP) and audio/text (CLAP) data and show that, compared to standard SAEs, our approach learns a more multimodal dictionary while reducing the number of dead neurons and improving feature semanticity. We finally demonstrate how this improvement in alignment of concepts between modalities can enable improvements in the interpretability and control of cross-modal tasks.

Transferability estimation metrics are used to find a high-performing pre-trained model for a given target task without fine-tuning models and without access to the source dataset. Despite the growing interest in developing such metrics, the benchmarks used to measure their progress have gone largely unexamined. In this work, we empirically show the shortcomings of widely used benchmark setups to evaluate transferability estimation metrics. We argue that the benchmarks on which these metrics are evaluated are fundamentally flawed. We empirically demonstrate that their unrealistic model spaces and static performance hierarchies artificially inflate the perceived performance of existing metrics, to the point where simple, dataset-agnostic heuristics can outperform sophisticated methods. Our analysis reveals a critical disconnect between current evaluation protocols and the complexities of real-world model selection. To address this, we provide concrete recommendations for constructing more robust and realistic benchmarks to guide future research in a more meaningful direction.

The growing adoption of spectrum-aware matrix-valued optimizers such as Muon and Shampoo in deep learning motivates a systematic study of their generalization properties and, in particular, when they might outperform competitive algorithms. We approach this question by introducing appropriate simplifying abstractions as follows: First, we use imbalanced data as a testbed. Second, we study the canonical form of such optimizers, which is Spectral Gradient Descent (SpecGD)—each update step is $\mathbf{U}\mathbf{V}^T$ where $\mathbf{U}\mathbf{\Sigma}\mathbf{ V}^T$ is the truncated SVD of the gradient. Third, within this framework we identify a canonical setting for which we precisely quantify when SpecGD outperforms vanilla Euclidean GD. For a Gaussian mixture data model and both linear and bilinear models, we show that unlike GD, which prioritizes learning dominant principal components of the data first, SpecGD learns all principal components of the data at equal rates. We demonstrate how this translates to a growing gap in balanced accuracy favoring SpecGD early in training and further show that the gap remains consistent even when the GD counterpart uses adaptive step-sizes via normalization. By extending the analysis to deep linear models, we show that depth amplifies these effects. We empirically verify our theoretical findings on a variety of imbalanced datasets. Our experiments compare practical variants of spectral methods, like Muon and Shampoo, against their Euclidean counterparts and Adam. The results validate our findings that these spectral optimizers achieve superior generalization by promoting a more balanced learning of the data's underlying components.

Reasoning models improve their problem-solving ability through inference-time scaling, allocating more compute via longer token budgets. Identifying which reasoning traces are likely to succeed remains a key opportunity: reliably predicting productive paths can substantially reduce wasted computation and improve overall efficiency. We introduce Latent-Trajectory signals that characterize the temporal evolution of a model's internal representations during the generation of intermediate reasoning tokens. By analyzing both the extent and temporal course of latent representational change, as well as its alignment with the final state, we show that these signals are strong predictors of solution accuracy, outperforming conventional output-based confidence measures. We use latent-trajectory signals to guide answer selection across multiple sampled generations, demonstrating that they make test-time scaling more effective and efficient, reducing token usage by up to 70% while preserving and even improving accuracy by 2.6% on average in comparison with majority voting. Finally, we show that these signals often emerge early in the reasoning trace, which enables early selection and allocation of compute to the most promising candidates during generation. Our findings contribute not only practical strategies for inference-time efficiency, but also a deeper interpretability perspective on how reasoning processes are represented and differentiated in latent space.

N:M sparsity has emerged as a hardware-friendly pruning strategy, notably supported by NVIDIA’s Sparse Tensor Cores. While efficient, its fixed sparsity ratio restricts flexibility, making it difficult to adapt pruning granularity to varying weight importance across layers and architectures. To overcome this limitation, we propose FlexHiNM, a hybrid framework that adaptively partitions each layer into three regions: dense, vector-pruned, and N:M sparse, enabling finer-grained control while preserving hardware compatibility. To better preserve salient weights, we extend this to FlexHiNM-GP, which incorporates Gyro-Permutation, an iterative channel-rearrangement algorithm. Through successive sampling, clustering, and assignment, Gyro-Permutation aligns high-importance weights with structured sparsity patterns and mitigates suboptimal configurations in multi-level pruning. During gradual pruning, FlexHiNM-GP further employs a differentiable masking mechanism based on the Hard Concrete distribution, enabling gradient-based mask learning and preventing over-aggressive early pruning. Experiments on vision and language benchmarks demonstrate that FlexHiNM-GP consistently surpasses strong structured baselines and approaches the performance of unstructured pruning, validating the effectiveness of combining hybrid sparsity with learned masks and permutation strategies.

Recent work has demonstrated that complex visual stimuli can be decoded from human brain activity using deep generative models, helping brain science researchers interpret how the brain represents real-world scenes. However, most current approaches leverage mapping brain signals into intermediate image or text feature spaces before guiding the generative process, masking the effect of contributions from different brain areas on the final reconstruction output. In this work, we propose \textit{NeuroAdapter}, a visual decoding framework that directly conditions a latent diffusion model on brain representations, bypassing the need for intermediate feature spaces. Our method demonstrates competitive visual reconstruction quality on public fMRI datasets compared to prior work, while providing greater transparency into how brain signals shape the generation process. To this end, we contribute an Image–Brain BI-directional interpretability framework (\textit{IBBI}) which investigates cross-attention mechanisms across diffusion denoising steps to reveal how different cortical areas influence the unfolding generative trajectory. Our results highlight the potential of end-to-end brain-to-image decoding and establish a path toward interpreting diffusion models through the lens of visual neuroscience.