论文检索

Deep Learning · Generative Models and Autoencoders

Boyi Zeng, Shixiang Song, Siyuan Huang, Yixuan Wang, He Li, Ziwei He, Xinbing Wang, Zhiyu li, Zhouhan Lin

Humans ponder before articulating complex sentence elements, enabling deeper cognitive processing through focused effort. In this work, we introduce this pondering process into language models by repeatedly invoking the forward process within a single token generation step. During pondering, instead of generating an actual token sampled from the prediction distribution, the model ponders by yielding a weighted sum of all token embeddings according to the predicted token distribution. The generated embedding is then fed back as input for another forward pass. We show that the model can learn to ponder in this way through self-supervised learning, without any human annotations. Experiments across three widely used open-source architectures—GPT-2, Pythia, and LLaMA—and extensive downstream task evaluations demonstrate the effectiveness and generality of our method. On 9 downstream benchmarks, our pondering-enhanced Pythia models significantly outperform the official Pythia models. Notably, our PonderPythia models demonstrate remarkable effectiveness: PonderPythia-2.8B surpasses Pythia-6.9B and rivals Pythia-12B, while our PonderPythia-1B matches TinyLlama-1.1B, a model trained on 10 times more data.

Applications · Neuroscience, Cognitive Science

Liuyin Yang, Qiang Sun, Ang Li, Marc Van Hulle

Foundation models have recently emerged as a promising approach for learning generalizable EEG representations for brain–computer interfaces (BCIs). Yet, their true advantages over traditional methods—particularly classical non-neural approaches—remain unclear. In this work, we present a comprehensive benchmark of state-of-the-art EEG foundation models, evaluated across diverse datasets, decoding tasks, and six evaluation protocols, with rigorous statistical testing. We introduce spatiotemporal EEGFormer (ST-EEGFormer), a simple yet effective Vision Transformer (ViT)-based baseline, pre-trained solely with masked autoencoding (MAE) on over 8M EEG segments. Our results show that while fine-tuned foundation models perform well in data-rich, population-level settings, they often fail to significantly outperform compact neural networks or even classical non-neural decoders in data-scarce scenarios. Furthermore, linear probing remains consistently weak, and performance varies greatly across downstream tasks, with no clear scaling law observed among neural network decoders. These findings expose a substantial gap between pre-training and downstream fine-tuning, often diminishing the benefits of complex pre-training tasks. We further identify hidden architectural factors that affect performance and emphasize the need for transparent, statistically rigorous evaluation. Overall, this study calls for community-wide efforts to construct large-scale EEG datasets and for fair, reproducible benchmarks to advance EEG foundation models.

General Machine Learning · Representation Learning

Jihoon Tack, Jack Lanchantin, Jane Dwivedi-Yu, Andrew Cohen, Ilia Kulikov, Janice Lan, Shibo Hao, Yuandong Tian, Jason E Weston, Xian Li

Next token prediction has been the standard training objective used in large language model pretraining. Representations are learned as a result of optimizing for token-level perplexity. We propose Continuous Concept Mixing (CoCoMix), a novel pretraining framework that combines discrete next token prediction with continuous concepts. Specifically, CoCoMix predicts ``continuous concepts'' learned from a pretrained sparse autoencoder and mixes them into the model's hidden state by interleaving with token hidden representations. Through experiments on multiple benchmarks, including language modeling and downstream reasoning tasks, we show that CoCoMix is more sample efficient and consistently outperforms standard next token prediction and knowledge distillation. We find that combining both concept learning and interleaving in an end-to-end framework is critical to performance gains. Furthermore, CoCoMix enhances interpretability and steerability by allowing direct inspection and modification of the predicted concept, offering a transparent way to guide the model’s internal reasoning process.

Deep Learning · Generative Models and Autoencoders

Yinghui He, Abhishek Panigrahi, Yong Lin, Sanjeev Arora

Language models often show little to no improvement (i.e., “saturation”) when trained via vanilla supervised fine-tuning (SFT) on data similar to what they saw in their training set (e.g., MATH). We introduce a new fine-tuning strategy, STAT, to train such a student model by using the metacognition ability of a stronger large language model (LLM) as the teacher. The teacher uses the task dataset to create a list of skills needed for the task, and then labels each data point with its required skills (Didolkar et al., 2024). By monitoring the student’s answers, the teacher creates a Missing-Skill-Profile for the student, tracking how often they failed to apply each skill in their responses. We use this idea to build a modified training set in one of two ways. In STAT-Sel, the teacher uses an existing set of training examples but adaptively reweights them according to the Missing-Skill-Profile. In STAT-Syn, the teacher synthesizes additional examples involving missing skills. Across extensive experiments on Llama and Qwen models, our methods yield improvements of up to 7.5% on MATH, whereas SFT provides only limited gains. Furthermore, STAT enhances performance on out-of-distribution benchmarks (e.g., AIME24/25, AMC23, etc.) by an average of 4.6%. Crucially, we find that STAT is complementary to RL via GRPO (Shao et al., 2024): after the model is improved using STAT to address skill gaps, GRPO continues to add further gains. We conclude that skill-targeted adaptive training should broadly improve current training pipelines.

Deep Learning · Generative Models and Autoencoders

Yiming Huang, Tolga Birdal

Graph generation is a critical yet challenging task, as empirical analyses require a deep understanding of complex, non-Euclidean structures. Diffusion models have recently made significant advances in graph generation, but these models are typically adapted from image generation frameworks and overlook inherent higher-order topology, limiting their ability to capture graph topology. In this work, we propose Higher-order Guided Diffusion (HOG-Diff), a principled framework that progressively generates plausible graphs with inherent topological structures. HOG-Diff follows a coarse-to-fine generation curriculum, guided by higher-order topology and implemented via diffusion bridges. We further prove that our model admits stronger theoretical guarantees than classical diffusion frameworks. Extensive experiments across eight graph generation benchmarks, spanning diverse domains and including large-scale settings, demonstrate the scalability of our method and its superior performance on both pairwise and higher-order topological metrics. Our project page is available [here](https://circle-group.github.io/research/hog-diff/).

Deep Learning · Generative Models and Autoencoders

Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Michael Bronstein, Anru Zhang, Joey Bose, Alexander Tong

Diffusion language models have emerged as a powerful alternative to autoregressive models, enabling fast inference through more flexible and parallel generation paths. This flexibility of sampling is unlocked by new engineered sampling strategies, or *planners*, that select more favorable generation paths by iteratively planning---versus uniformly at random---where to denoise along the sequence. However, by modifying the reverse paths via planning, planners create an irrevocable mismatch between the uniformly random denoising paths during training and planning-based inference. In this paper, we systematically investigate the mismatch of discrete diffusion training and inference under planning and theoretically prove that the standard discrete diffusion training evidence lower bound (ELBO) does not accurately describe a denoiser that uses a non-uniform planner. To address this gap, we derive a new planned evidence lower bound (P-ELBO) that incorporates planner-based reverse dynamics directly into the training objective. Using the P-ELBO, we introduce *Planner Aware Path Learning* (PAPL), a novel training scheme that aligns training and inference under a planned denoiser. PAPL is implemented as a simple yet effective modification to the standard masked discrete diffusion loss, making it widely applicable and easy to adopt. Empirically, we show PAPL delivers consistent gains across domains, including a 40\% relative improvement in protein sequences, improved text generation with up to a $4\times$ relative MAUVE gain, and 23\% relative improvement in code generation HumanEval pass@10.

Deep Learning · Generative Models and Autoencoders

Qifan Li, Jiale Zou, Jinhua Zhang, Wei Long, Xingyu Zhou, Shuhang Gu

Vector-quantized based models have recently demonstrated strong potential for visual prior modeling. However, existing VQ-based methods simply encode visual features with nearest codebook items and train index predictor with code-level supervision. Due to the richness of visual signal, VQ encoding often leads to large quantization error. Furthermore, training predictor with code-level supervision can not take the final reconstruction errors into consideration, result in sub-optimal prior modeling accuracy. In this paper we address the above two issues and propose a Texture Vector-Quantization and a Reconstruction Aware Prediction strategy. The texture vector-quantization strategy leverages the task character of super-resolution and only introduce codebook to model the prior of missing textures. While the reconstruction aware prediction strategy makes use of the straight-through estimator to directly train index predictor with image-level supervision. Our proposed generative SR model TVQ&RAP is able to deliver photo-realistic SR results with small computational cost.

Deep Learning · Generative Models and Autoencoders

Fangyu Ding, Ding Ding, Sijin Chen, Kaibo Wang, Peng Xu, Zijin Feng, Haoli Bai, Kai Han, Youliang Yan, Binhang Yuan 等

While Masked Diffusion Language Models (MDLMs) relying on token masking and unmasking have shown promise in language modeling, their computational efficiency and generation flexibility remain constrained by the masking paradigm. In this paper, we propose Deletion-Insertion Diffusion language models (DID) that rigorously formulate token deletion and insertion as discrete diffusion processes, replacing the masking and unmasking processes in current MDLMs. DID improves training and inference efficiency by eliminating two major sources of computational overhead in MDLMs: the computations on non-informative 1) $\texttt{\}$ tokens inherent to its paradigm, and 2) $\texttt{\}$ tokens introduced in variable-length settings. Furthermore, DID offers greater flexibility by: 1) natively supporting variable-length sequences without requiring fixed-length padding, and 2) an intrinsic self-correction mechanism during generation due to insertion that dynamically adjusts token positions. To train DID, we design a score-based approach that assigns scores to token insertion operations and derive appropriate training objectives. The objectives involve subsequence counting problems, which we efficiently solve via a parallelized dynamic programming algorithm. Our experiments across fixed and variable-length settings demonstrate the advantage of DID over baselines of MDLMs and existing insertion-based LMs, in terms of modeling performance, sampling quality, and training/inference speed, without any hyperparameter tuning.

Deep Learning · Generative Models and Autoencoders

Johannes von Oswald, Nino Scherrer, Seijin Kobayashi, Luca Versari, Songlin Yang, Maximilian Schlegel, Kaitlin Maile, Yanick Schimpf, Oliver Sieberling, Alexander Meulemans 等

Sequence modeling is currently dominated by causal transformer architectures that use softmax self-attention. Although widely adopted, transformers require scaling memory and compute linearly during inference. A recent stream of work linearized the softmax operation, resulting in powerful recurrent neural network (RNN) models with constant memory and compute costs such as DeltaNet, Mamba or xLSTM. These models can be unified by noting that their recurrent layer dynamics can all be derived from an in-context regression objective, approximately optimized through an online learning rule. Here, we join this line of work and introduce a numerically stable, chunkwise parallelizable version of the recently proposed Mesa layer (von Oswald et al., 2024), which could only run sequentially in time and was therefore not scalable. This layer again stems from an in-context loss, but which is now minimized to optimality at every time point using a fast conjugate gradient solver. Through an extensive suite of experiments study up to the billion-parameter scale, we show that optimal test-time training enables reaching lower language modeling perplexity and higher downstream benchmark performance than previous RNNs, especially on tasks requiring long context understanding. This performance gain comes at the cost of additional flops spent during inference time. Our results are therefore intriguingly related to recent trends of increasing test-time compute to improve performance -- here by spending compute to solve sequential optimization problems within the neural network itself.

Deep Learning · Generative Models and Autoencoders

Ming Gui, Johannes Schusterbauer, Timy Phan, Felix Krause, Joshua Susskind, MIGUEL ANGEL BAUTISTA, Björn Ommer

We introduce Representation Tokenizer (RepTok), a generative modeling framework that represents an image using a single continuous latent token obtained from self-supervised vision transformers. Building on a pre-trained SSL encoder, we fine-tune only the semantic token embedding and pair it with a generative decoder trained end-to-end using a standard flow matching objective. This adaptation enriches the token with low-level, reconstruction-relevant details, enabling faithful image reconstruction. To preserve the favorable geometry of the original SSL space, we add a cosine-similarity loss that regularizes the adapted token, ensuring it remains smooth and suitable for generation.Our single-token formulation resolves the spatial redundancies of the 2D latent space, simplifies architectures, and significantly reduces training costs. Despite its simplicity and efficiency, RepTok achieves competitive results on class-conditional ImageNet generation and extends naturally to text-to-image synthesis, reaching competitive zero-shot performance on MS-COCO under extremely limited training budgets. Our findings highlight the potential of fine-tuned SSL representations as compact and effective latent spaces for efficient generative modeling. We will release our model to facilitate further research.

Deep Learning · Generative Models and Autoencoders

Todd Nief, David Reber, Sean Richardson, Ari Holtzman

When an LLM learns a new fact during finetuning (e.g., new movie releases, newly elected pope, etc.), where does this information go? Are entities enriched with relation information immediately, or do models recall information just-in-time before a prediction? Or, are "all of the above" true, with LLMs implementing multiple redundant heuristics? Existing localization approaches (e.g., activation patching) are ill-suited for this analysis because they usually replace parts of the residual stream, thus overriding previous information. To fill this interpretability gap, we propose dynamic weight grafting, an analysis technique that selectively grafts subsets of weights from a finetuned model onto a pretrained model. Using this technique, we show two separate pathways for retrieving finetuned relation information: 1) "enriching" the residual stream with relation information while processing the tokens that correspond to an entity (e.g., "Zendaya" in "Zendaya co-starred with Timothée Chalamet" and 2) "recalling" this information at the final token position before generating a target fact. In some cases, models need information from both of these pathways to correctly generate finetuned facts while, in other cases, either the "enrichment" or "recall" pathway alone is sufficient. We localize the "recall" pathway to model components---finding that "recall" occurs via both task-specific attention mechanisms and an entity-specific extraction step in the feedforward networks of the final layers before prediction. By targeting model components and parameters, as opposed to just activations, we are able to understand the mechanisms by which finetuned knowledge is retrieved during generation.

Deep Learning · Generative Models and Autoencoders

Reza Shirkavand, Xiaokai Wei, Chen Wang, Zheng Hui, Heng Huang, Michelle Gong

While collaborative filtering delivers predictive accuracy and efficiency, and Large Language Models (LLMs) enable expressive and generalizable reasoning, modern recommendation systems must bring these strengths together. Growing user expectations, such as natural-language queries and transparent explanations, further highlight the need for a unified approach. However, doing so is nontrivial. Collaborative signals are often token-efficient but semantically opaque, while LLMs are semantically rich but struggle to model implicit user preferences when trained only on textual inputs. This paper introduces Item-ID + Natural-language Mixture-of-Experts Language Model (IDIOMoE), which treats item interaction histories as a native dialect within the language space, enabling collaborative signals to be understood in the same way as natural language. By splitting the Feed Forward Network of each block of a pretrained LLM into a separate text expert and an item expert with token-type gating, our method avoids destructive interference between text and catalog modalities. IDIOMoE demonstrates strong recommendation performance across both public and proprietary datasets, while preserving the text understanding of the pretrained model.

Deep Learning · Generative Models and Autoencoders

Daisuke Oba, Danushka Bollegala, Masahiro Kaneko, Naoaki Okazaki

Masked Diffusion Language Models generate sequences via iterative sampling that progressively unmasks tokens. However, they still recompute the attention and feed-forward blocks for every token position at every step---even when many unmasked tokens are essentially fixed, resulting in substantial waste in compute. We propose **SureLock**: when the posterior at an unmasked position has stabilized across steps (our *sure* condition), we *lock* that position---thereafter skipping its query projection and feed-forward sublayers---while caching its attention keys and values so other positions can continue to attend to it. This reduces the dominant per-iteration computational cost from $O(N^2d)$ to $O(MNd)$ where $N$ is the sequence length, $M$ is the number of unlocked token positions, and $d$ is the model dimension. In practice, $M$ decreases as the iteration progresses, yielding substantial savings. On LLaDA-8B, SureLock reduces algorithmic FLOPs by 30--50\% relative to the same sampler without locking, while maintaining comparable generation quality. We also provide a theoretical analysis to justify the design rationale of SureLock: monitoring only the local KL at the lock step suffices to bound the deviation in final token probabilities. Our project page is available at https://daioba.github.io/surelock.

Deep Learning · Generative Models and Autoencoders

Mohammed Nowaz Rabbani Chowdhury, Kaoutar El Maghraoui, Hsinyu Tsai, Naigang Wang, Geoffrey Burr, Liu Liu, Meng Wang

Sparse Mixture-of-Experts (MoE) allows scaling of language and vision models efficiently by activating only a small subset of experts per input. While this reduces computation, the large number of parameters still incurs substantial memory overhead during inference. Post-training quantization has been explored to address this issue. Because uniform quantization suffers from significant accuracy loss at low bit-widths, mixed-precision methods have been recently explored; however, they often require substantial computation for bit-width allocation and overlook the varying sensitivity of model performance to the quantization of different experts. We propose a theoretically grounded expert-wise mixed-precision strategy that assigns bit-width to each expert primarily based on their *change in router’s* $l_2$ *norm* during training. Experts with smaller changes are shown to capture less frequent but critical features, and model performance is more sensitive to the quantization of these experts, thus requiring higher precision. Furthermore, to avoid allocating experts to lower precision that inject high quantization noise, experts with large *maximum intra-neuron variance* are also allocated higher precision. Experiments on large-scale MoE models, including Switch Transformer and Mixtral, show that our method achieves higher accuracy than existing approaches, while also reducing inference cost and incurring only negligible overhead for bit-width assignment.

Deep Learning · Generative Models and Autoencoders

Chengyin Xu, Kaiyuan Chen, Xiao Li, Ke Shen, Chenggang Li

The escalating scale and cost of Large Language Models (LLMs) training necessitate accurate pre-training prediction of downstream task performance for comprehensive understanding of scaling properties. This is challenged by: 1) the emergence phenomenon, where unpredictable capabilities appearing suddenly at critical model scales; and 2) uneven task difficulty and inconsistent performance scaling patterns, leading to high metric variability. Current prediction methods lack accuracy and reliability. We propose a Clustering-On-Difficulty (COD) framework for downstream performance prediction. The COD framework clusters tasks by their difficulty scaling features, thereby constructing a more stable and predictable task subset that exhibits well-behaved scaling characteristics with the increase of compute budget. We adopt a performance scaling law to predict cluster-wise performance with theoretical support. Predictable subset performance acts as an intermediate predictor for the full evaluation set. We further derive a mapping function to accurately extrapolate the performance of the subset to the full set. Applied to an LLM with 70B parameters, COD achieved a 1.55\% average prediction error across eight key LLM benchmarks, thus providing actionable insights for scaling properties and training monitoring during LLM pre-training.

Deep Learning · Generative Models and Autoencoders

Chi Zhang, Xixi Hu, Bo Liu, Qiang Liu

The slow, sequential nature of autoregressive (AR) language models has driven the adoption of parallel decoding methods. However, these non-autoregressive models often sacrifice generation quality because they struggle to model the complex joint distribution of token sequences. To narrow this performance gap, we introduce Gumbel Distillation, a novel distillation technique that enables parallel decoders to learn this distribution effectively. Our method leverages the Gumbel-Max trick to create a deterministic mapping from a latent Gumbel noise space to the output tokens of a high-performing AR teacher. As a model-agnostic technique, Gumbel Distillation seamlessly integrates with diverse parallel decoding architectures, including MDLM and BD3-LM. Experiments on LM1B and OpenWebText show that Gumbel Distillation substantially improves the generation quality of parallel language models, achieving a 30.0% improvement in MAUVE Score and 10.5% in generative perplexity over MDLM trained on OpenWebText dataset.

Deep Learning · Generative Models and Autoencoders

Shubham Parashar, Shurui Gui, Xiner Li, Hongyi Ling, Sushil Vemuri, Blake Olson, Eric Li, Yu Zhang, James Caverlee, Dileep Kalathil 等

We aim to improve the reasoning capabilities of language models via reinforcement learning with verifiable rewards (RLVR). Recent RLVR post-trained models like DeepSeek-R1 have demonstrated reasoning abilities on mathematical and coding tasks. However, prior studies suggest that using RLVR alone to improve reasoning on inherently difficult tasks is less effective due to sparse rewards. Here, we draw inspiration from curriculum learning and propose to schedule tasks from easy to hard (E2H), allowing LLMs to build reasoning skills gradually. Our method is termed E2H Reasoner. Empirically, we observe that, although easy tasks are important initially, fading them out through appropriate scheduling is essential in preventing overfitting. Theoretically, we establish convergence guarantees for E2H Reasoner within an approximate policy iteration framework. We derive finite-sample complexity bounds and show that when tasks are appropriately decomposed and conditioned, learning through curriculum stages requires fewer total samples than direct learning. Experiments across diverse datasets and models demonstrate that E2H Reasoner substantially enhances LLM reasoning. Code is available at - https://github.com/divelab/E2H-Reasoning

Deep Learning · Generative Models and Autoencoders

Yixuan Ren, Hanyu Wang, Bo He, Hao Chen, Abhinav Shrivastava

We present NeRV-Diffusion, an implicit latent video diffusion model that synthesizes videos via generating neural network weights. The generated weights can be arranged as the parameters of a convolutional network, forming an implicit neural representation (INR) and decoding into videos with frame indices as the input. Our framework consists of two stages: First, a hypernetwork-based tokenizer that encodes raw videos from pixel space to neural parameter space, and the bottleneck latent serves as INR weights to decode; Second, an implicit diffusion transformer that denoises on the latent INR weights. In contrast to traditional video tokenizers that compress videos into frame-wise feature maps, NeRV-Diffusion generates a video as a compact dedicated neural network. This continuous holistic video representation obviates temporal cross-frame attentions while preserving flexible temporal interpolability. The INR decoder and weight latent feature sublinear complexity overhead regarding video resolution and length increase with additional upsampling layers. To enable Gaussian-distributed neural weights with high expressiveness, we reuse the bottleneck latent across all INR layers, as well as reform its weight modulation, upsampling connection and input coordinates. We also introduce SNR-adaptive loss weighting and scheduled sampling for effective training of the implicit diffusion model. NeRV-Diffusion reaches superior video synthesis quality over previous INR-based models and comparable performance to most recent state-of-the-art non-implicit models on real-world video benchmarks including UCF-101 and Kinetics-600. It also achieves outstanding decoding and generation efficiency when scaling up to high-resolution and long videos.

Theory · Interpretability and Visualization

Lucas Prieto, Edward Stevinson, Melih Barsbey, Tolga Birdal, Pedro Mediano

A central idea in mechanistic interpretability is that neural networks represent more features than they have dimensions, arranging them in superposition to form an over-complete basis. This framing has been influential, motivating dictionary learning approaches such as sparse autoencoders. However, superposition has mostly been studied in idealized settings where features are sparse and uncorrelated. In these settings, superposition is typically understood as introducing interference that must be minimized geometrically and filtered out by non-linearities such as ReLUs, yielding local structures like regular polytopes. We show that this account is incomplete for realistic data by introducing Bag-of-Words Superposition (BOWS), a controlled setting to encode binary bag-of-words representations of internet text in superposition. Using BOWS, we find that when features are correlated, interference can be constructive rather than just noise to be filtered out. This is achieved by arranging features according to their co-activation patterns, making interference between active features constructive, while still using ReLUs to avoid false positives. We show that this kind of arrangement is more prevalent in models trained with weight decay and naturally gives rise to semantic clusters and cyclical structures which have been observed in real language models yet were not explained by the standard picture of superposition. Code for this paper can be found at: https://github.com/LucasPrietoAl/correlations-feature-geometry.

Computer Vision · Vision Models & Multimodal

Vardaan Pahuja, Samuel Stevens, Alyson East, Sydne Record, Yu Su

Morphological traits are physical characteristics of biological organisms that provide vital clues on how organisms interact with their environment. Yet extracting these traits remains a slow, expert-driven process, limiting their use in large-scale ecological studies. A major bottleneck is the absence of high-quality datasets linking biological images to trait-level annotations. In this work, we demonstrate that sparse autoencoders trained on foundation-model features yield monosemantic, spatially grounded neurons that consistently activate on meaningful morphological parts. Leveraging this property, we introduce a trait annotation pipeline that localizes salient regions and uses vision-language prompting to generate interpretable trait descriptions. Using this approach, we construct Bioscan-Traits, a dataset of 80K trait annotations spanning 19K insect images from BIOSCAN-5M. Human evaluation confirms the biological plausibility of the generated morphological descriptions. We assess design sensitivity through a comprehensive ablation study, systematically varying key design choices and measuring their impact on the quality of the resulting trait descriptions. By annotating traits with a modular pipeline rather than prohibitively expensive manual efforts, we offer a scalable way to inject biologically meaningful supervision into foundation models, enable large-scale morphological analyses, and bridge the gap between ecological relevance and machine-learning practicality.