NeurIPS 2026 Papers

Data attribution, which traces a model’s prediction back to specific training data, is an important tool for interpreting sophisticated AI models. The widely used TRAK algorithm addresses this challenge by first approximating the underlying model with a kernel machine and then leveraging techniques developed for approximating the leave-one-out (ALO) risk. Despite its strong empirical performance, the theoretical conditions under which TRAK approximations are accurate, as well as the regimes in which they break down, remain largely unexplored. In this paper, we provide a theoretical analysis of the TRAK algorithm, characterizing its performance and quantifying the errors introduced by the approximations on which the method relies. We show that although the approximations can incur significant errors, TRAK preserves the separation between highly influential and weakly influential data points. We corroborate our theoretical results through extensive simulations and empirical studies.

We present ImmuVis, a family of efficient foundation models for imaging mass cytometry (IMC), a high-throughput multiplex imaging technology that handles molecular marker measurements as image channels and enables large-scale spatial tissue profiling. Unlike natural images, multiplex imaging lacks a fixed channel space, as real-world marker sets vary across studies, violating a core assumption of standard vision backbones. To address this, ImmuVis introduces marker-adaptive hyperconvolutions that generate convolutional kernels from learned marker embeddings, enabling a single model to operate on arbitrary measured marker subsets without retraining. We pretrain ImmuVis on the largest IMC dataset to date, IMC17M (28 cohorts, 24,405 images, 265 markers, over 17M patches), using self-supervised masked reconstruction. ImmuVis outperforms state-of-the-art baselines and ablations in virtual staining and downstream classification tasks at substantially lower compute cost than transformer-based alternatives, and is the sole model that provides calibrated uncertainty via a heteroscedastic likelihood objective. These results position ImmuVis as a practical framework for real-world IMC modeling.

Reasoning about the 3D world from multi-view images remains a fundamental challenge for Multimodal Large Language Models (MLLMs). While modern MLLMs handle single-image inputs effectively, they struggle to integrate evidence across viewpoints into a coherent 3D understanding. A growing body of work attempts to close this gap by injecting 3D awareness into MLLMs, either by boosting fine-grained pixel-level cross-view correspondence or by fusing features from 3D geometry foundation models, yet a substantial gap to human reasoning persists. In this work, we revisit human spatial reasoning, which suggests that rather than relying on fine-grained geometry cues, humans roughly identify common objects across views, infer the relative geometry between viewpoints, and assemble a coarse 3D layout of the scene. Inspired by this process, we introduce Imagine3D-LLM, an MLLM that learns to assemble a similar compact 3D representation of the scene and conditions its answer on this representation. Concretely, we append a small set of learnable summary tokens after the image tokens, decode them into a compact 3D Gaussian Splatting representation supervised by a photometric reconstruction loss, and train jointly with the standard next-token prediction objective. Notably, although only the summary tokens receive direct reconstruction supervision, this objective also induces stronger cross-frame correspondence within the LLM's underlying image features, suggesting that learning to reconstruct propagates 3D-aware signals throughout the model. As a result, Imagine3D-LLM consistently outperforms prior approaches across multiple spatial reasoning and 3D understanding benchmarks, suggesting that imagining the scene can be more effective than being told its pixel-wise geometry. Our code and weights will be publicly released.

Modern AI agents can plan, reflect, reason, and act over long-horizon digital tasks. Even so, they cannot plan to steer a social interaction in real time. Doing so requires anticipating how latent factors like trust and resistance will evolve under the agent's actions, fast enough to plan within a conversational turn. Generative world models approach this by narrating possible futures, but autoregressive text generation is both too slow for real-time planning and fundamentally lossy. Representation-predictive methods can be superior, but are underexplored for interaction. We build LID-Bench, the first controlled testbed with oracle latent states for interaction dynamics, enabling systematic comparison of generative and representation-predictive world models against known dynamics. A bottleneck decomposition on three generative models spanning 117M--350M parameters reveals that all learn interaction dynamics internally, with text-rendering signal losses of 89--100\% observed by horizon $k{=}5$. Our representation-predictive model (Social-JEPA, 125M params, 500K trainable) bypasses text entirely, forecasting latent dynamics $2.8$--$3.9\times$ more accurately while performing rollouts $1{,}059$--$2{,}314\times$ faster---making real-time planning within conversational turn-taking latencies feasible.

Incorporating visual semantic representations as an intermediate step before image generation can reduce the modeling difficulty between text and images, thereby improving generation quality. Recent works such as X-Omni and BLIP3o-Next have explored this direction, but they typically use a two-stage external pipeline: a separate autoregressive model first generates semantic tokens, which are then fed as conditioning to an independent diffusion decoder. Since the decoder cannot jointly access the original input and the semantic plan, this design introduces an information bottleneck that limits detail preservation in downstream tasks such as editing. Internal architectures such as Transfusion, BAGEL, and Show-o2 avoid this bottleneck by enabling cross-modal interaction within a single model, but they still face the difficult text-to-pixel modeling gap without intermediate semantic guidance. We propose Visual Prompt Engineering (VPE), which can be seamlessly integrated into such internal frameworks. Specifically, the model first autoregressively generates visual semantic tokens ($e.g.$, SigLIP 2) as "visual prompts" that capture the semantic layout, then generates the full image tokens conditioned on this plan. We validate VPE across class-conditional generation, text-to-image generation, and image editing, covering various token types and model architectures. Results show that VPE can accelerate convergence, raise quality ceilings, and through internal integration, achieve substantially better editing preservation (PSNR: $26.76$ vs. $19.92$) than external alternatives of the same parameter scale, while maintaining competitive editing responsiveness. The code is available in supplementary material.

ImageNet class-conditional generation is a standard first test for new image generation methods. Progress on ImageNet is usually measured with one number: Fréchet Inception Distance (FID), computed by comparing 50K class-balanced generated samples with the ImageNet training set in the Inception-v3 feature space. The FID number was useful when the generation quality was low and the gap between methods was large. However, many recent methods already reach remarkable and close FID numbers, and we argue that the field should stop treating FID as decisive evidence of generation quality. To support this view, we re-evaluate 47 recent ImageNet class-conditional generators while keeping the generated samples fixed and changing only the evaluation protocol. We test multiple variants of FID, including different real reference sets, sample sizes, and feature extractors such as Inception-v3, CLIP, DINOv2, and VGG16. The same generated samples can receive different scores, resulting in varying rankings. Our observation means that a single FID protocol has a risk of being over-optimized and over-interpreted. We do not argue that FID should be removed. Instead, we recomend: keep the conventional ImageNet FID for continuity, but also report the results of some FID varaints to avoid overfitting to one evaluation recipe. If a method reaches comparable and stable ImageNet FID, authors should not be expected to keep chasing the best FID, and the paper should be evaluated based on their technical contributions. After passing the FID test, future work such as text-to-image training and human preference study, is important for the futher development of a generation method.

Illusory pattern perception is a well-documented human cognitive tendency to infer meaningful relationships in data that is actually random. Such a tendency, often described as “connecting the dots” where none exist, can result in systematic reasoning errors. This paper investigates whether Large Language Models (LLMs) exhibit such perceptual tendencies, which can lead to systematic errors in downstream applications. To our knowledge, this work presents the first systematic study of illusory pattern perception in LLMs, adapting classic psychological paradigms to three tasks with direct empirical comparison to human behaviors. We find that LLMs frequently exhibit stronger illusory pattern perception than humans. In particular, models tend to over-associate frequent positive attributes with majority groups or large organizations, and show increased tendencies to construct causal narratives from ambiguous events. To uncover the mechanism behind these behaviors, we develop a feature interpretability framework based on Sparse Autoencoders (SAEs) to analyze internal representations. Our results reveal that holistic frequency perception and analytic cognitive orientation are linked to the emergence of illusory perceptions. These findings highlight a previously underexplored cognitive-like illusion that may affect the reliability of LLM reasoning. Code available at [Anonymous Github](https://anonymous.4open.science/r/illusory-8E27/).

Causal discovery methods commonly assume that all data is independently and identically distributed (i.i.d.) and that there are no unmeasured variables affecting the system. In practice, these assumptions are often violated, leading to inaccurate inference. In this paper, we study how to identify hidden confounding and selection biases from causal mechanism shifts. In particular, we show that structural biases lead to dependent mechanism shifts. That is, by considering for which variables the mechanisms change given data from different environments, we can tell which variables are unbiased, which are subject to hidden confounding, and which are undergoing selection bias. We formalize this into an empirically testable criterion based on mutual information, and show under which conditions it identifies structural biases. To tell which nodes are subject to what kind of bias, we introduce the StruBI algorithm. Experiments on synthetic and real-world data show that StruBI works well in practice, accurately recovering affected variable sets and types of biases, outperforming the state-of-the-art by a wide margin.

Modern LLM APIs often reveal only top-$K$ logit scores and censor the remaining vocabulary. We study the per-position distribution-recovery limits of this access model. For censoring threshold $\tau$, the compatible teacher distributions form an identified set whose total-variation diameter is exactly $U_K=(V{-}K)\exp(\tau)/(Z_A+(V{-}K)\exp(\tau)),$ where $Z_A$ is the observed partition function. For KL recovery, we give a computable binary-endpoint lower bound and an asymptotically matching small-ambiguity upper bound, with an extension to reference-aware attackers. Experiments on a Qwen3 math-reasoning teacher reveal a layered extraction hierarchy: on-task top-$K$ distillation recovers 12% of private capability, full-logit distillation recovers 56% despite 99% KL closure, and generation-based extraction recovers 96%. Top-$K$ censoring therefore limits per-position distribution recovery but does not by itself prevent capability extraction, separating fidelity from transfer in prompt-only logit distillation.

The joint expectation of functions of potential outcomes $(Y_1,Y_0)$ is a fundamental quantity in causal inference, particularly when investigating the variation or heterogeneity of causal effects. This paper provides novel assumptions for their identification: comonotonicity and countercomonotonicity. These assumptions offer a unified framework for their identification across discrete, continuous, and mixed outcome variables. In the absence of these assumptions, we derive sharp bounds for two broad classes of functions, yielding new results for bounding moments of individual causal effects, which are central to measuring effect heterogeneity. We present corresponding estimation methods and illustrate them in simulation studies and real-world datasets.

Reconstructing the gene regulatory dynamics underlying tissue development from unpaired, cross-sectional spatial transcriptomic snapshots poses a significant challenge in single-cell biology. While recent optimal transport and flow-based methods have advanced trajectory interpolation, their unconstrained latent dynamics lack identifiability, hindering reliable discovery of the underlying regulatory mechanisms. To address this, we propose single-cell Identifiable Feedback-controlled Latent Flow (scIFLF), a generative framework that introduces a physics-inspired structural prior that decomposes the latent flow into a macroscopic developmental drift and a restorative feedback toward functional attractors. Crucially, the gene-regulatory Jacobian is shown to be uniquely identifiable, invariant to the latent ambiguity, enabling principled discovery of gene regulations directly from the learned flow. Implemented with a multimodal Neural ODE, scIFLF aligns latent trajectories using entropy-regularized optimal transport. Benchmark experiments demonstrate that scIFLF outperforms state-of-the-art methods in trajectory interpolation and spatial coherence, while successfully recovering key driver genes of tissue development and organogenesis, bridging the gap between deep generative flexibility and biological interpretability.

The Linear Representation Hypothesis (LRH) posits that Transformer activations decompose into sparse linear combinations of interpretable directions, with Sparse Autoencoders (SAEs) serving as the predominant method for recovering such decompositions. We propose that idempotency -- the requirement that re-encoding an SAE's output reproduces the same latent representation -- is a desirable property for any faithful dictionary learning method, and demonstrate empirically that most SAEs deviate very far from this ideal. We identify two underlying failure modes: exploding representation norms and unstable active feature sets. Further analysis reveals that even individual SAE latents cannot be reliably reconstructed under re-encoding, and that SAE outputs exhibit significant sensitivity to input scaling. We show that idempotency can be improved via an auxiliary loss term and demonstrate its potential to improve SAE interpretability.

Feature caching accelerates Diffusion Transformers by reusing intermediate computations across denoising timesteps, yet existing methods share a fundamental blind spot: they treat the standard reference denoising trajectory as the optimisation target, approximating its per-step states and thus inheriting its output as an implicit quality ceiling. We argue that this target is arbitrary --- what truly matters is the ideal result, not the intermediate path taken to reach it. We propose IdealCache, a training-free framework that abandons step-wise imitation of the reference trajectory. Instead, it anchors caching on a per-input ensemble of ideal trajectories constructed from candidate target results, aggregates their geometry into a single graph, and discovers optimal non-uniform schedules via a constrained shortest path on this graph. By decoupling from the reference process, IdealCache breaks the performance bottleneck of trajectory-approximation; our schedules can attain near-lossless fidelity in generation and even surpass on task-specific metrics (e.g., maintaining higher semantic consistency in image editing). Without retraining, IdealCache achieves $6.35\times$ speedup on Qwen-Image, $5.84\times$ on Qwen-Image-Edit, and $4.32\times$ on HunyuanVideo, with competitive or superior quality across tasks.

Image--text contrastive pretraining has become a dominant paradigm for visual representation learning, yet existing methods often yield representations that remain partially organized by modality rather than by semantics. We propose \textbf{ITO}, a framework addressing this limitation through two complementary mechanisms with distinct roles. \emph{Multimodal multiple alignment} enriches supervision by constructing diverse cross-modal correspondences from multi-view image augmentations, providing the primary source of discriminative gain. A lightweight \emph{training-time multimodal fusion} module then acts as a geometric regularizer, encouraging the encoders to produce features that are compatible under fusion and thereby reducing modality-induced separation. Crucially, the fusion module is discarded at inference, preserving the efficiency of standard dual-encoder architectures while incurring training-time overhead only. Extensive experiments across pretraining scales from millions to billions of image--text pairs show that ITO consistently outperforms strong baselines on classification, retrieval, and multimodal benchmarks. Our analysis further reveals that beyond accuracy gains, training-time fusion plays a stabilizing role in optimization, mitigating the late-stage overfitting commonly observed in aggressive contrastive learning.

Self-play fine-tuning enables large language models to improve beyond supervised fine-tuning without additional human annotations by contrasting annotated responses with self-generated ones. Many existing methods rely on a fixed divergence regime. SPIN is closely related to a KL-based regime, SPACE to a Jensen-Shannon-style objective via noise contrastive estimation, and SPIF to $\chi^2$-regularized self-play. Since these divergences exhibit different strengths depending on the distributional gap between model and target, no single choice appears to provide favorable learning dynamics across training stages. We propose IRIS (Interpolative R\'enyi Iterative Self-play), a R\'enyi-based self-play fine-tuning framework with a continuously adjustable objective. IRIS decomposes into two independent tilted risk terms over annotated and synthetic data, with exponential importance weights controlled by the order parameter $\alpha$. We show that several self-play objectives can be interpreted as limiting or representative regimes at particular values of $\alpha$, providing a unified theoretical perspective on these methods. An adaptive order schedule further adjusts $\alpha$ to the distributional gap, shifting from sharper importance weighting early in training to smoother refinement near convergence. Theoretically, we establish the fixed-point property of IRIS and analyze how $\alpha$ controls gradient concentration. Experiments on Zephyr-7B and Qwen2.5-3B across ten benchmarks show that IRIS improves upon baselines, reaching 44.57\% average score with gains across iterations. In our setting, IRIS with only 26$k$ annotated samples surpasses standard supervised fine-tuning trained on the full 200$k$ dataset.

Field trials use plot-level crop-state measurements to interpret how crops grow under real field conditions, but repeated measurements of canopy greenness, canopy amount, canopy closure, and growth stage require field visits, sensing campaigns, and post-processing. Modern wheat trials also collect weather records, trial metadata, field-camera images, UAV and satellite observations, proximal sensors, canopy products, and phenological surveys. Turning these records into a machine-learning benchmark is nontrivial because observations are sparse, asynchronous, unevenly collected, missing for structured reasons, and tied to measurement pipelines. We introduce INVITA-WheatFieldState, a benchmark derived from the INVITA wheat field-trial archive for estimating the crop state of a plot on a target date from observations available for that plot and date. Each plot-date example specifies a plot, date, and crop-state target, and is paired with available observations after temporal and source-provenance filtering. The derived dataset contains over 220k examples across NDVI, LAI, FCover, and Zadoks growth stage, covering canopy greenness, canopy amount, canopy closure, and phenological timing. LAI and FCover are released with provenance labels that distinguish product and proxy targets rather than pooled manual ground truth. The benchmark suite provides a plot-disjoint split matched to the plot-level experimental unit, validators, prediction schemas, regression representations, and prediction-level diagnostics. Results show that source, timing, metadata, and observation-availability structure are strong signals. Observation-set representations improve some all-example targets, while sensor-sequence, field-camera, and fusion representations remain target-dependent and coverage-bound. INVITA-WheatFieldState offers a concrete test case for ML methods that must learn from incomplete, asynchronous, and provenance-sensitive measurements used to study crop development in real fields.

Self-evolution offers a scalable path to stronger reasoning: a pretrained language model improves itself with only minimal external supervision. Yet existing methods either depend on extensively curated or teacher-generated training data, or, when the generator runs unsupervised, reward it by a difficulty heuristic that need not improve the solver. We introduce **INFUSER**, an iterative co-training framework with two co-evolving roles: a *Generator* that drafts questions and reference golden answers from a pool of unstructured, automatically collected documents, and a *Solver* that improves by training on them. The solver is trained with standard correctness rewards against the generator-provided answers, while the generator is rewarded by an *optimizer-aware influence score* that measures whether each proposed question would actually improve the solver on the target distribution. Because this continuous, noisy influence score is poorly served by standard GRPO, we propose **DuGRPO**, a dual-normalized variant of GRPO, for generator training. Together, these turn the document pool into an *adaptive curriculum* that favors questions useful to the current solver, not just hard ones. On Qwen3-8B-Base, INFUSER outperforms strong self-evolution baselines with over 20% relative improvement on Olympiad and SuperGPQA benchmarks, and an 8B INFUSER co-evolving generator outperforms a frozen 32B thinking generator on math and coding. Ablations confirm each design choice is necessary. Code is available at

In this paper, we introduce INEUS, a meshfree iterative neural solver for partial integro-differential equations (PIDEs). The method replaces the explicit evaluation of nonlocal jump integrals with single-jump sampling and reformulates PIDE solving as a sequence of recursive regression problems. Like Physics-Informed Neural Networks (PINNs), INEUS learns global solutions over the entire space-time domain, yet it offers a more efficient treatment of nonlocal terms and avoids the computationally expensive differentiation of full PIDE residuals. These features make INEUS particularly well suited for high-dimensional PDEs and PIDEs. Supported by a contraction-based convergence proof for linear PIDEs, our numerical experiments show that INEUS delivers accurate and scalable solutions for various high-dimensional linear and nonlinear examples.

Modeling Irregular Multivariate Time Series (IMTS) poses significant challenges due to asynchronous sampling and data sparsity. While existing methods focus on handling temporal irregularity after encoding, the tokenization stage itself remains underexplored. We propose IMTS-Tokenizer, a time aware tokenization framework built on the Irregular-time-aware Module (IRAM), which employs learnable temporal anchors and Gaussian-kernel attention to aggregate observations while preserving temporal information. To improve robustness under sparse and asynchronous sampling, the framework further incorporates Relative Time Feature (RTF) for pre-aggregation temporal encoding, Variable-Specific Anchor Initialization (VSA) for data-aligned anchor placement, anchor regularization for stable training, and multi-scale modeling. Experiments on four benchmarks demonstrate state-of-the-art (SOTA) performance, with the best results on seven out of eight metrics and statistically significant aggregate gains.