Zhouhan Lin

Cheng Deng, Tianhang Zhang, Zhongmou He et al. · meta-ai, mila

Large language models (LLMs) have achieved great success in general domains of natural language processing. In this paper, we bring LLMs to the realm of geoscience with the objective of advancing research and applications in this field. To this end, we present the first-ever LLM in geoscience, K2, alongside a suite of resources developed to further promote LLM research within geoscience. For instance, we have curated the first geoscience instruction tuning dataset, GeoSignal, which aims to align LLM responses to geoscience-related user queries. Additionally, we have established the first geoscience benchmark, GeoBench, to evaluate LLMs in the context of geoscience. In this work, we experiment with a complete recipe to adapt a pre-trained general-domain LLM to the geoscience domain. Specifically, we further train the LLaMA-7B model on 5.5B tokens of geoscience text corpus, including over 1 million pieces of geoscience literature, and utilize GeoSignal's supervised data to fine-tune the model. Moreover, we share a protocol that can efficiently gather domain-specific data and construct domain-supervised data, even in situations where manpower is scarce. Meanwhile, we equip K2 with the abilities of using tools to be a naive geoscience aide. Experiments conducted on the GeoBench demonstrate the effectiveness of our approach and datasets on geoscience knowledge understanding and utilization.We open-source all the training data and K2 model checkpoints at https://github.com/davendw49/k2.

Jiexing Qi, Jingyao Tang, Ziwei He et al. · meta-ai, mila

Relational structures such as schema linking and schema encoding have been validated as a key component to qualitatively translating natural language into SQL queries. However, introducing these structural relations comes with prices: they often result in a specialized model structure, which largely prohibits using large pretrained models in text-to-SQL. To address this problem, we propose RASAT: a Transformer seq2seq architecture augmented with relation-aware self-attention that could leverage a variety of relational structures while inheriting the pretrained parameters from the T5 model effectively. Our model can incorporate almost all types of existing relations in the literature, and in addition, we propose introducing co-reference relations for the multi-turn scenario. Experimental results on three widely used text-to-SQL datasets, covering both single-turn and multi-turn scenarios, have shown that RASAT could achieve state-of-the-art results across all three benchmarks (75.5% EX on Spider, 52.6% IEX on SParC, and 37.4% IEX on CoSQL).

Yunchong Song, Chenghu Zhou, Xinbing Wang et al. · meta-ai, mila

Most graph neural networks follow the message passing mechanism. However, it faces the over-smoothing problem when multiple times of message passing is applied to a graph, causing indistinguishable node representations and prevents the model to effectively learn dependencies between farther-away nodes. On the other hand, features of neighboring nodes with different labels are likely to be falsely mixed, resulting in the heterophily problem. In this work, we propose to order the messages passing into the node representation, with specific blocks of neurons targeted for message passing within specific hops. This is achieved by aligning the hierarchy of the rooted-tree of a central node with the ordered neurons in its node representation. Experimental results on an extensive set of datasets show that our model can simultaneously achieve the state-of-the-art in both homophily and heterophily settings, without any targeted design. Moreover, its performance maintains pretty well while the model becomes really deep, effectively preventing the over-smoothing problem. Finally, visualizing the gating vectors shows that our model learns to behave differently between homophily and heterophily settings, providing an explainable graph neural model.

Yue Guan, Zhengyi Li, Jingwen Leng et al. · meta-ai, mila

Transformer architecture has become the de-facto model for many machine learning tasks from natural language processing and computer vision. As such, improving its computational efficiency becomes paramount. One of the major computational inefficiency of Transformer-based models is that they spend the identical amount of computation throughout all layers. Prior works have proposed to augment the Transformer model with the capability of skimming tokens to improve its computational efficiency. However, they suffer from not having effectual and end-to-end optimization of the discrete skimming predictor. To address the above limitations, we propose the Transkimmer architecture, which learns to identify hidden state tokens that are not required by each layer. The skimmed tokens are then forwarded directly to the final output, thus reducing the computation of the successive layers. The key idea in Transkimmer is to add a parameterized predictor before each layer that learns to make the skimming decision. We also propose to adopt reparameterization trick and add skim loss for the end-to-end training of Transkimmer. Transkimmer achieves 10.97x average speedup on GLUE benchmark compared with vanilla BERT-base baseline with less than 1% accuracy degradation.

Yusheng Huang, Jiexing Qi, Xinbing Wang et al. · meta-ai, mila

Various tasks are reformulated as multi-label classification problems, in which the binary cross-entropy (BCE) loss is frequently utilized for optimizing well-designed models. However, the vanilla BCE loss cannot be tailored for diverse tasks, resulting in a suboptimal performance for different models. Besides, the imbalance between redundant negative samples and rare positive samples could degrade the model performance. In this paper, we propose an effective Asymmetric Polynomial Loss (APL) to mitigate the above issues. Specifically, we first perform Taylor expansion on BCE loss. Then we ameliorate the coefficients of polynomial functions. We further employ the asymmetric focusing mechanism to decouple the gradient contribution from the negative and positive samples. Moreover, we validate that the polynomial coefficients can recalibrate the asymmetric focusing hyperparameters. Experiments on relation extraction, text classification, and image classification show that our APL loss can consistently improve performance without extra training burden.

Boyi Zeng, Yiqin Hao, Zitong Wang et al.

Self-attention selects information freely across the sequence, but across depth, Transformers merely add each layer's output to the residual stream, so later layers cannot selectively reuse earlier-layer representations. Recent cross-layer methods improve this flow but operate on hidden states outside attention, adding state beyond the key-value cache at inference--a cost that becomes increasingly salient as modern LLMs compress the cache with grouped-query and multi-head latent attention. We introduce Depth-Attention, which performs this selection inside the attention module itself: before a layer attends over the sequence, its query attends over the keys of earlier layers at the same token position and mixes their values into the value that self-attention then reads. Because Depth-Attention reuses the standard attention queries, keys, and value-cache slots, storing depth-mixed values in place of the original values, it adds no parameters and introduces no persistent inference state beyond the standard key-value cache--the same cache size as a vanilla decoder and less than hidden-state-based cross-layer methods. On Qwen3-style decoders at 1.5B and 3B parameters, Depth-Attention attains the lowest perplexity and the highest average downstream accuracy, improving over the vanilla Transformer by up to 2.3 accuracy points and surpassing strong cross-layer baselines in perplexity and average accuracy, while adding under 0.01% extra arithmetic FLOPs and no additional persistent inference state. The gains hold from 360M to 3B parameters and extend to looped Transformers.

Yi Xu, Luoyi Fu, Zhouhan Lin et al. · meta-ai, mila

Graph-to-text (G2T) generation and text-to-graph (T2G) triple extraction are two essential tasks for constructing and applying knowledge graphs. Existing unsupervised approaches turn out to be suitable candidates for jointly learning the two tasks due to their avoidance of using graph-text parallel data. However, they are composed of multiple modules and still require both entity information and relation type in the training process. To this end, we propose INFINITY, a simple yet effective unsupervised approach that does not require external annotation tools or additional parallel information. It achieves fully unsupervised graph-text mutual conversion for the first time. Specifically, INFINITY treats both G2T and T2G as a bidirectional sequence generation task by fine-tuning only one pretrained seq2seq model. A novel back-translation-based framework is then designed to automatically generate continuous synthetic parallel data. To obtain reasonable graph sequences with structural information from source texts, INFINITY employs reward-based training loss by leveraging the advantage of reward augmented maximum likelihood. As a fully unsupervised framework, INFINITY is empirically verified to outperform state-of-the-art baselines for G2T and T2G tasks.

Zhixin Guo, Minyxuan Yan, Jiexing Qi et al. · meta-ai, mila

Pre-trained language models (PLM) have achieved remarkable advancement in table-to-text generation tasks. However, the lack of labeled domain-specific knowledge and the topology gap between tabular data and text make it difficult for PLMs to yield faithful text. Low-resource generation likewise faces unique challenges in this domain. Inspired by how humans descript tabular data with prior knowledge, we suggest a new framework: PromptMize, which targets table-to-text generation under few-shot settings. The design of our framework consists of two aspects: a prompt planner and a knowledge adapter. The prompt planner aims to generate a prompt signal that provides instance guidance for PLMs to bridge the topology gap between tabular data and text. Moreover, the knowledge adapter memorizes domain-specific knowledge from the unlabelled corpus to supply essential information during generation. Extensive experiments and analyses are investigated on three open domain few-shot NLG datasets: human, song, and book. Compared with previous state-of-the-art approaches, our model achieves remarkable performance in generating quality as judged by human and automatic evaluations.

Shengyuan Hou, Jushi Kai, Haotian Xue et al. · gatech, meta-ai

Recent works have revealed that Transformers are implicitly learning the syntactic information in its lower layers from data, albeit is highly dependent on the quality and scale of the training data. However, learning syntactic information from data is not necessary if we can leverage an external syntactic parser, which provides better parsing quality with well-defined syntactic structures. This could potentially improve Transformer's performance and sample efficiency. In this work, we propose a syntax-guided localized self-attention for Transformer that allows directly incorporating grammar structures from an external constituency parser. It prohibits the attention mechanism to overweight the grammatically distant tokens over close ones. Experimental results show that our model could consistently improve translation performance on a variety of machine translation datasets, ranging from small to large dataset sizes, and with different source languages.

Ning Shi, Bin Tang, Bo Yuan et al. · meta-ai, mila

Text editing, such as grammatical error correction, arises naturally from imperfect textual data. Recent works frame text editing as a multi-round sequence tagging task, where operations -- such as insertion and substitution -- are represented as a sequence of tags. While achieving good results, this encoding is limited in flexibility as all actions are bound to token-level tags. In this work, we reformulate text editing as an imitation game using behavioral cloning. Specifically, we convert conventional sequence-to-sequence data into state-to-action demonstrations, where the action space can be as flexible as needed. Instead of generating the actions one at a time, we introduce a dual decoders structure to parallel the decoding while retaining the dependencies between action tokens, coupled with trajectory augmentation to alleviate the distribution shift that imitation learning often suffers. In experiments on a suite of Arithmetic Equation benchmarks, our model consistently outperforms the autoregressive baselines in terms of performance, efficiency, and robustness. We hope our findings will shed light on future studies in reinforcement learning applying sequence-level action generation to natural language processing.

Jiexing Qi, Shuhao Li, Zhixin Guo et al. · meta-ai, mila

Real-world data usually exhibits a long-tailed distribution,with a few frequent labels and a lot of few-shot labels. The study of institution name normalization is a perfect application case showing this phenomenon. There are many institutions worldwide with enormous variations of their names in the publicly available literature. In this work, we first collect a large-scale institution name normalization dataset LoT-insts1, which contains over 25k classes that exhibit a naturally long-tailed distribution. In order to isolate the few-shot and zero-shot learning scenarios from the massive many-shot classes, we construct our test set from four different subsets: many-, medium-, and few-shot sets, as well as a zero-shot open set. We also replicate several important baseline methods on our data, covering a wide range from search-based methods to neural network methods that use the pretrained BERT model. Further, we propose our specially pretrained, BERT-based model that shows better out-of-distribution generalization on few-shot and zero-shot test sets. Compared to other datasets focusing on the long-tailed phenomenon, our dataset has one order of magnitude more training data than the largest existing long-tailed datasets and is naturally long-tailed rather than manually synthesized. We believe it provides an important and different scenario to study this problem. To our best knowledge, this is the first natural language dataset that focuses on long-tailed and open-set classification problems.

Siyuan Huang, Yunchong Song, Jiayue Zhou et al.

Attention mechanisms have made significant strides in graph learning, yet they still exhibit notable limitations: local attention faces challenges in capturing long-range information due to the inherent problems of the message-passing scheme, while global attention cannot reflect the hierarchical neighborhood structure and fails to capture fine-grained local information. In this paper, we propose a novel multi-hop graph attention mechanism, named Subtree Attention (STA), to address the aforementioned issues. STA seamlessly bridges the fully-attentional structure and the rooted subtree, with theoretical proof that STA approximates the global attention under extreme settings. By allowing direct computation of attention weights among multi-hop neighbors, STA mitigates the inherent problems in existing graph attention mechanisms. Further we devise an efficient form for STA by employing kernelized softmax, which yields a linear time complexity. Our resulting GNN architecture, the STAGNN, presents a simple yet performant STA-based graph neural network leveraging a hop-aware attention strategy. Comprehensive evaluations on ten node classification datasets demonstrate that STA-based models outperform existing graph transformers and mainstream GNNs. The code is available at https://github.com/LUMIA-Group/SubTree-Attention.

Bo Xue, Yunchong Song, Fanghao Shao et al.

Standard autoregressive language models generate text token-by-token from a fixed vocabulary, inducing a tree-structured state space when viewing token sampling as an action, which limits flexibility and expressiveness. Recent work introduces dynamic vocabulary by sampling retrieved text spans but overlooks that the same sentence can be composed of spans of varying lengths, lacking explicit modeling of the directed acyclic graph (DAG) state space. This leads to restricted exploration of compositional paths and is biased toward the chosen path. Generative Flow Networks (GFlowNets) are powerful for efficient exploring and generalizing over state spaces, particularly those with a DAG structure. However, prior GFlowNets-based language models operate at the token level and remain confined to tree-structured spaces, limiting their potential. In this work, we propose Flow of SpanS (FOSS), a principled GFlowNets framework for span generation. FoSS constructs a dynamic span vocabulary by segmenting the retrieved text flexibly, ensuring a DAG-structured state space, which allows GFlowNets to explore diverse compositional paths and improve generalization. With specialized reward models, FoSS generates diverse, high-quality text. Empirically, FoSS improves MAUVE scores by up to 12.5% over Transformer on text generation and achieves 3.5% gains on knowledge-intensive tasks, consistently outperforming state-of-the-art methods. Scaling experiments further demonstrate FoSS benefits from larger models, more data, and richer retrieval corpora, retaining its advantage over strong baselines.

Zhouhan Lin, Cheng Deng, Le Zhou et al.

Large language models (LLMs) have achieved huge success for their general knowledge and ability to solve a wide spectrum of tasks in natural language processing (NLP). Due to their impressive abilities, LLMs have shed light on potential inter-discipline applications to foster scientific discoveries of a specific domain by using artificial intelligence (AI for science, AI4S). In the meantime, utilizing NLP techniques in geoscience research and practice is wide and convoluted, contributing from knowledge extraction and document classification to question answering and knowledge discovery. In this work, we take the initial step to leverage LLM for science, through a rather straightforward approach. We try to specialize an LLM into geoscience, by further pre-training the model with a vast amount of texts in geoscience, as well as supervised fine-tuning (SFT) the resulting model with our custom collected instruction tuning dataset. These efforts result in a model GeoGalactica consisting of 30 billion parameters. To our best knowledge, it is the largest language model for the geoscience domain. More specifically, GeoGalactica is from further pre-training of Galactica. We train GeoGalactica over a geoscience-related text corpus containing 65 billion tokens, preserving as the largest geoscience-specific text corpus. Then we fine-tune the model with 1 million pairs of instruction-tuning data consisting of questions that demand professional geoscience knowledge to answer. In this technical report, we will illustrate in detail all aspects of GeoGalactica, including data collection, data cleaning, base model selection, pre-training, SFT, and evaluation. We open-source our data curation tools and the checkpoints of GeoGalactica during the first 3/4 of pre-training.

Dingyi Rong, Wenzhuo Zheng, Bozitao Zhong et al.

Accurate prediction of enzyme function is crucial for elucidating biological mechanisms and driving innovation across various sectors. Existing deep learning methods tend to rely solely on either sequence data or structural data and predict the EC number as a whole, neglecting the intrinsic hierarchical structure of EC numbers. To address these limitations, we introduce MAPred, a novel multi-modality and multi-scale model designed to autoregressively predict the EC number of proteins. MAPred integrates both the primary amino acid sequence and the 3D tokens of proteins, employing a dual-pathway approach to capture comprehensive protein characteristics and essential local functional sites. Additionally, MAPred utilizes an autoregressive prediction network to sequentially predict the digits of the EC number, leveraging the hierarchical organization of EC classifications. Evaluations on benchmark datasets, including New-392, Price, and New-815, demonstrate that our method outperforms existing models, marking a significant advance in the reliability and granularity of protein function prediction within bioinformatics.

Boyi Zeng, Lizheng Wang, Yuncong Hu et al.

Protecting the copyright of large language models (LLMs) has become crucial due to their resource-intensive training and accompanying carefully designed licenses. However, identifying the original base model of an LLM is challenging due to potential parameter alterations. In this study, we introduce HuRef, a human-readable fingerprint for LLMs that uniquely identifies the base model without interfering with training or exposing model parameters to the public. We first observe that the vector direction of LLM parameters remains stable after the model has converged during pretraining, with negligible perturbations through subsequent training steps, including continued pretraining, supervised fine-tuning, and RLHF, which makes it a sufficient condition to identify the base model. The necessity is validated by continuing to train an LLM with an extra term to drive away the model parameters' direction and the model becomes damaged. However, this direction is vulnerable to simple attacks like dimension permutation or matrix rotation, which significantly change it without affecting performance. To address this, leveraging the Transformer structure, we systematically analyze potential attacks and define three invariant terms that identify an LLM's base model. Due to the potential risk of information leakage, we cannot publish invariant terms directly. Instead, we map them to a Gaussian vector using an encoder, then convert it into a natural image using StyleGAN2, and finally publish the image. In our black-box setting, all fingerprinting steps are internally conducted by the LLMs owners. To ensure the published fingerprints are honestly generated, we introduced Zero-Knowledge Proof (ZKP). Experimental results across various LLMs demonstrate the effectiveness of our method. The code is available at https://github.com/LUMIA-Group/HuRef.

Boyi Zeng, Shixiang Song, Siyuan Huang et al.

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. For language modeling tasks, pondering language models achieve performance comparable to vanilla models with twice the number of parameters. On 9 downstream benchmarks, our pondering-enhanced Pythia models significantly outperform the official Pythia models. Notably, PonderingPythia-2.8B surpasses Pythia-6.9B, and PonderingPythia-1B is comparable to TinyLlama-1.1B, which is trained on 10 times more data. The code is available at https://github.com/LUMIA-Group/PonderingLM.

Yuliang Liu, Junjie Lu, Zhaoling Chen et al. · tsinghua

Current approaches for training Process Reward Models (PRMs) often involve breaking down responses into multiple reasoning steps using rule-based techniques, such as using predefined placeholder tokens or setting the reasoning step's length into a fixed size. These approaches overlook the fact that specific words do not typically mark true decision points in a text. To address this, we propose AdaptiveStep, a method that divides reasoning steps based on the model's confidence in predicting the next word. This division method provides more decision-making information at each step, enhancing downstream tasks, such as reward model learning. Moreover, our method does not require manual annotation. We demonstrate its effectiveness through experiments with AdaptiveStep-trained PRMs in mathematical reasoning and code generation tasks. Experimental results indicate that the outcome PRM achieves state-of-the-art Best-of-N performance, surpassing greedy search strategy with token-level value-guided decoding, while also reducing construction costs by over 30% compared to existing open-source PRMs. In addition, we provide a thorough analysis and case study on the PRM's performance, transferability, and generalization capabilities.

Yusheng Huang, Zhouhan Lin

Multimodal information extraction is attracting research attention nowadays, which requires aggregating representations from different modalities. In this paper, we present the Intra- and Inter-Sample Relationship Modeling (I2SRM) method for this task, which contains two modules. Firstly, the intra-sample relationship modeling module operates on a single sample and aims to learn effective representations. Embeddings from textual and visual modalities are shifted to bridge the modality gap caused by distinct pre-trained language and image models. Secondly, the inter-sample relationship modeling module considers relationships among multiple samples and focuses on capturing the interactions. An AttnMixup strategy is proposed, which not only enables collaboration among samples but also augments data to improve generalization. We conduct extensive experiments on the multimodal named entity recognition datasets Twitter-2015 and Twitter-2017, and the multimodal relation extraction dataset MNRE. Our proposed method I2SRM achieves competitive results, 77.12% F1-score on Twitter-2015, 88.40% F1-score on Twitter-2017, and 84.12% F1-score on MNRE.

Yuliang Liu, Yunchong Song, Yixuan Wang et al.

We propose Next Concept Prediction (NCP), a generative pretraining paradigm built on top of Next Token Prediction (NTP). NCP predicts discrete concepts that span multiple tokens, thereby forming a more challenging pretraining objective. Our model, ConceptLM, quantizes hidden states using Vector Quantization and constructs a concept vocabulary. It leverages both NCP and NTP to drive parameter updates and generates a concept to guide the generation of the following tokens. We train ConceptLM from scratch at scales ranging from 70M to 1.5B parameters with up to 300B training data, including Pythia and GPT-2 backbones. Results on 13 benchmarks show that NCP yields consistent performance gains over traditional token-level models. Furthermore, continual pretraining experiments on an 8B-parameter Llama model indicate that NCP can further improve an NTP-trained model. Our analysis suggests that NCP leads to more powerful language models by introducing a harder pretraining task, providing a promising path toward better language modeling.

Yusheng Huang, Zhouhan Lin

Document-level relation extraction (DocRE) is a task that focuses on identifying relations between entities within a document. However, existing DocRE models often overlook the correlation between relations and lack a quantitative analysis of relation correlations. To address this limitation and effectively capture relation correlations in DocRE, we propose a relation graph method, which aims to explicitly exploit the interdependency among relations. Firstly, we construct a relation graph that models relation correlations using statistical co-occurrence information derived from prior relation knowledge. Secondly, we employ a re-weighting scheme to create an effective relation correlation matrix to guide the propagation of relation information. Furthermore, we leverage graph attention networks to aggregate relation embeddings. Importantly, our method can be seamlessly integrated as a plug-and-play module into existing models. Experimental results demonstrate that our approach can enhance the performance of multi-relation extraction, highlighting the effectiveness of considering relation correlations in DocRE.

Yifan Liu, Yu Fang, Zhouhan Lin

Video-to-speech (V2S) synthesis, the task of generating speech directly from silent video input, is inherently more challenging than other speech synthesis tasks due to the need to accurately reconstruct both speech content and speaker characteristics from visual cues alone. Recently, audio-visual pre-training has eliminated the need for additional acoustic hints in V2S, which previous methods often relied on to ensure training convergence. However, even with pre-training, existing methods continue to face challenges in achieving a balance between acoustic intelligibility and the preservation of speaker-specific characteristics. We analyzed this limitation and were motivated to introduce DiVISe (Direct Visual-Input Speech Synthesis), an end-to-end V2S model that predicts Mel-spectrograms directly from video frames alone. Despite not taking any acoustic hints, DiVISe effectively preserves speaker characteristics in the generated audio, and achieves superior performance on both objective and subjective metrics across the LRS2 and LRS3 datasets. Our results demonstrate that DiVISe not only outperforms existing V2S models in acoustic intelligibility but also scales more effectively with increased data and model parameters. Code and weights can be found at https://github.com/PussyCat0700/DiVISe.

Yixuan Wang, Qingyu Shi, Jiayu Zhou et al.

The growing context length of Large Language Models (LLMs) enlarges the Key-Value (KV) cache, limiting deployment in resource-limited environments. Prior training-free approaches for KV cache compression typically rely on low-rank approximation or scalar quantization, which fail to simultaneously achieve high compression ratios and high reconstruction fidelity. We propose VQKV, a novel, training-free method introducing vector quantization (VQ) to obtain highly compressed KV representations while preserving high model fidelity, allowing for the representation of thousands of floating-point values with just a few integer indices. As a result, VQKV achieves an 82.8\% compression ratio on LLaMA3.1-8B while retaining 98.6\% of the baseline performance on LongBench and enabling 4.3x longer generation length on the same memory footprint.

Junzhe Shen, Jieru Zhao, Ziwei He et al.

We study why continuous diffusion language models (DLMs) have lagged behind discrete diffusion approaches despite their appealing continuous generative dynamics. Under a controlled token--recovery study, we identify token rounding, the final projection from denoised embeddings to tokens, as a primary bottleneck. Building on these insights, we propose CoDAR (Continuous Diffusion with Contextual AutoRegressive Decoder), a two--stage framework that keeps diffusion entirely continuous in an embedding space while learning a strong, context--conditional discretizer: an autoregressive Transformer decoder that cross--attends to the denoised embedding sequence and performs contextualized rounding to tokens. Experiments on LM1B and OpenWebText demonstrate that CoDAR substantially improves generation quality over latent diffusion and becomes competitive with strong discrete DLMs, while exposing a simple decoder--temperature knob to navigate the fluency--diversity trade off.

Boyi Zeng, Yiqin Hao, He Li et al.

Scaling large language models by increasing parameters and training data is increasingly constrained by limited high-quality corpora and rising communication costs. This work explores an alternative axis: increasing per-token computation without expanding parameters, by internalizing latent Chain-of-Thought (CoT) into pretraining. We propose Pretraining with Token-Level Adaptive Latent CoT (adaptive latent CoT), where the model generates a variable-length latent CoT trajectory before emitting each token -- allocating longer trajectories to difficult tokens and shorter (or even zero) trajectories to easy ones. Importantly, this behavior emerges naturally from one-stage pretraining on general text and reduces computation in both training and inference via token-wise adaptive halting. Experiments with Llama architectures show that adaptive latent CoT consistently improves language modeling perplexity and broad downstream accuracy, even with fewer training FLOPs than prior recurrent baselines.

Wenhao Zhao, Qiran Zou, Rushi Shah et al.

Vector quantization is a technique in machine learning that discretizes continuous representations into a set of discrete vectors. It is widely employed in tokenizing data representations for large language models, diffusion models, and other generative models. Despite its prevalence, the characteristics and behaviors of vector quantization in generative models remain largely underexplored. In this study, we systematically investigate the issue of collapses in vector quantization, where collapsed representations are observed across discrete codebook tokens and continuous latent embeddings. By leveraging both synthetic and real datasets, we identify the severity of each type of collapses and triggering conditions. Our analysis reveals that random initialization and limited encoder capacity result in tokens collapse and embeddings collapse. Building on these findings, we propose potential solutions aimed at mitigating each collapse. To the best of our knowledge, this is the first comprehensive study examining representation collapsing problems in vector quantization.

Lin Chen, Samuel Drapeau, Fanghao Shao et al.

Generative Flow Network (GFlowNet) objectives implicitly fix an equal mixing of forward and backward policies, potentially constraining the exploration-exploitation trade-off during training. By further exploring the link between GFlowNets and Markov chains, we establish an equivalence between GFlowNet objectives and Markov chain reversibility, thereby revealing the origin of such constraints, and provide a framework for adapting Markov chain properties to GFlowNets. Building on these theoretical findings, we propose $α$-GFNs, which generalize the mixing via a tunable parameter $α$. This generalization enables direct control over exploration-exploitation dynamics to enhance mode discovery capabilities, while ensuring convergence to unique flows. Across various benchmarks, including Set, Bit Sequence, and Molecule Generation, $α$-GFN objectives consistently outperform previous GFlowNet objectives, achieving up to a $10 \times$ increase in the number of discovered modes.

Boyi Zeng, Lin Chen, Ziwei He et al.

Protecting the intellectual property of large language models (LLMs) is crucial, given the substantial resources required for their training. Consequently, there is an urgent need for both model owners and third parties to determine whether a suspect LLM is trained from scratch or derived from an existing base model. However, the intensive post-training processes that models typically undergo-such as supervised fine-tuning, extensive continued pretraining, reinforcement learning, multi-modal extension, pruning, and upcycling-pose significant challenges to reliable identification. In this work, we propose a training-free fingerprinting method based on weight matrices. We leverage the Linear Assignment Problem (LAP) and an unbiased Centered Kernel Alignment (CKA) similarity to neutralize the effects of parameter manipulations, yielding a highly robust and high-fidelity similarity metric. On a comprehensive testbed of 60 positive and 90 negative model pairs, our method demonstrates exceptional robustness against all six aforementioned post-training categories while exhibiting a near-zero risk of false positives. By achieving perfect scores on all classification metrics, our approach establishes a strong basis for reliable model lineage verification. Moreover, the entire computation completes within 30s on an NVIDIA 3090 GPU. The code is available at https://github.com/LUMIA-Group/AWM.

Ziwei He, Meng Yang, Minwei Feng et al.

The transformer model is known to be computationally demanding, and prohibitively costly for long sequences, as the self-attention module uses a quadratic time and space complexity with respect to sequence length. Many researchers have focused on designing new forms of self-attention or introducing new parameters to overcome this limitation, however a large portion of them prohibits the model to inherit weights from large pretrained models. In this work, the transformer's inefficiency has been taken care of from another perspective. We propose Fourier Transformer, a simple yet effective approach by progressively removing redundancies in hidden sequence using the ready-made Fast Fourier Transform (FFT) operator to perform Discrete Cosine Transformation (DCT). Fourier Transformer is able to significantly reduce computational costs while retain the ability to inherit from various large pretrained models. Experiments show that our model achieves state-of-the-art performances among all transformer-based models on the long-range modeling benchmark LRA with significant improvement in both speed and space. For generative seq-to-seq tasks including CNN/DailyMail and ELI5, by inheriting the BART weights our model outperforms the standard BART and other efficient models. Our code is publicly available at https://github.com/LUMIA-Group/FourierTransformer

Xuekai Zhu, Biqing Qi, Kaiyan Zhang et al.

While large language models (LLMs) excel in various natural language processing tasks, their huge size and the inaccessibility of parameters present challenges for practical deployment. Previous studies try to distill task-specific ability from LLMs to smaller models, using data synthesis and chain-of-thought (CoT) fine-tuning. However, synthetic CoT data often contains faulty reasoning, which deteriorates the quality of distillation, especially in reasoning capabilities. In this work, we propose Program-aided Distillation (PaD), which introduces reasoning programs to suppress the errors in distilled data, and thus achieves better distillation quality for reasoning tasks. In PaD, we utilize the reasoning program to substitute the CoT, allowing automated error checking of synthetic data. Further, through error injecting and further training, the small distilling model could iteratively self-refine the reasoning. Moreover, we conduct a step-wise beam search by step-by-step verifying to acquire more exact reasoning chains. We evaluate PaD on arithmetic reasoning, symbolic reasoning, and general ability. Experimental results demonstrate that smaller models using PaD can not only outperform certain LLMs~(e.g., LLaMA-1 13B) but also achieve strong improvement over baselines with a significantly smaller scale of parameters and data. The source code is publicly available at https://github.com/Xuekai-Zhu/pad.

Shixiang Song, He Li, Zitong Wang et al.

Test-time scaling via recurrent/iterative Transformers enables large language models to spend more computation at inference, but most pretrained recurrent LMs run a fixed number of iterations, wasting compute on easy tokens and lacking token-wise adaptivity. Following the core idea of Adaptive Computation Time(ACT) and Early Exit(EE), we propose AdaPonderLM, a self-supervised recurrent language model that learns token-wise early exiting during pretraining without manually tuned per-token/per-layer pruning ratios. AdaPonderLM uses iteration-specific MLP gates with a monotonic halting mask to decide when each token stops recurring, and introduces a KV reuse mechanism that reuses cached key/value states for halted tokens, ensuring train--test consistency and practical acceleration. Across Pythia backbones from 70M to 410M (pretraining) and up to 2.8B (continued pretraining), AdaPonderLM reduces inference compute at about 10% while maintaining comparable language modeling perplexity and competitive downstream accuracy. Our analysis shows the learned gates allocate more computation to high-NLL (hard) tokens, exhibiting adaptive computation time behavior in a fully self-supervised setting. Meanwhile, under iso-FLOPs, the learned halting policy consistently outperforms fixed pruning, showing AdaPonderLM allocates compute to the right tokens rather than just reducing average depth.

Wenhao Zhao, Qiran Zou, Zhouhan Lin et al.

Vector Quantization (VQ) has become the cornerstone of tokenization for many multimodal Large Language Models and diffusion synthesis. However, existing VQ paradigms suffer from a fundamental conflict: they enforce discretization before the encoder has captured the underlying data manifold. We term this phenomenon Premature Discretization. To resolve this, we propose Progressive Quantization (ProVQ), which incorporates the dynamics of quantization hardness as a fundamental yet previously overlooked axis in VQ training. By treating quantization as a curriculum that smoothly anneals from a continuous latent space to a discrete one, ProVQ effectively guides the codebook toward the well-expanded manifolds. Extensive experimental results demonstrate the broad effectiveness of ProVQ across diverse modalities. We report improved reconstruction and generative performance on the ImageNet-1K and ImageNet-100 benchmarks, highlighting the ProVQ's boost for generative modeling. Furthermore, ProVQ proves highly effective for modeling complex biological sequences, establishing a new performance ceiling for protein structure tokenization on the StrutTokenBench leaderboard.

He Li, Feichen Song, Boyi Zeng et al.

Test-time scaling has shown that allocating more additional computation at inference can improve generation quality, motivating a natural follow-up question: where should this computation be spent? Building on this insight, we introduce PonderLM-3, a pretraining framework for token-wise adaptive pondering that learns to selectively allocate additional computation under purely self-supervised objectives, built on top of the PonderLM-2 backbone. This makes additional inference computation an allocatable per-token resource, so tokens receive more computation only when it is beneficial, rather than paying a uniform extra cost. To make this allocation learnable while maintaining train-inference consistency, PonderLM-3 injects a differentiable attention mask during pretraining and pairs it with a matching hard pruning rule at inference. PonderLM-3 defines a stronger Pareto frontier: compared with existing recursive or adaptive baselines, it achieves lower pretraining perplexity at equal inference FLOPs. On downstream benchmarks, PonderLM-3 attains comparable performance to fixed-step PonderLM-2 under the same maximum number of additional computation steps, while using fewer inference FLOPs in practice. Overall, PonderLM-3 provides an end-to-end differentiable and train-inference consistent framework for token-wise adaptive computation, enabling additional inference compute to be allocated where it is most useful rather than paid uniformly by every token.

Jushi Kai, Tianhang Zhang, Hai Hu et al.

Large language models (LLMs) demonstrate great performance in text generation. However, LLMs are still suffering from hallucinations. In this work, we propose an inference-time method, Self-Highlighted Hesitation (SH2), to help LLMs decode more truthfully. SH2 is based on a simple fact rooted in information theory that for an LLM, the tokens predicted with lower probabilities are prone to be more informative than others. Our analysis shows that the tokens assigned with lower probabilities by an LLM are more likely to be closely related to factual information, such as nouns, proper nouns, and adjectives. Therefore, we propose to ''highlight'' the factual information by selecting the tokens with the lowest probabilities and concatenating them to the original context, thus forcing the model to repeatedly read and hesitate on these tokens before generation. During decoding, we also adopt contrastive decoding to emphasize the difference in the output probabilities brought by the hesitation. Experimental results demonstrate that our SH2, requiring no additional data or models, can effectively help LLMs elicit factual knowledge and distinguish hallucinated contexts. Significant and consistent improvements are achieved by SH2 for LLaMA-7b, LLaMA2-7b and Mistral-7b on multiple hallucination tasks.

Xuekai Zhu, Daixuan Cheng, Hengli Li et al. · tsinghua

Model collapse in synthetic data indicates that iterative training on self-generated data leads to a gradual decline in performance. With the proliferation of AI models, synthetic data will fundamentally reshape the web data ecosystem. Future GPT-$\{n\}$ models will inevitably be trained on a blend of synthetic and human-produced data. In this paper, we focus on two questions: what is the impact of synthetic data on language model training, and how to synthesize data without model collapse? We first pre-train language models across different proportions of synthetic data, revealing a negative correlation between the proportion of synthetic data and model performance. We further conduct statistical analysis on synthetic data to uncover distributional shift phenomenon and over-concentration of n-gram features. Inspired by the above findings, we propose token editing on human-produced data to obtain semi-synthetic data. As a proof of concept, we theoretically demonstrate that token-level editing can prevent model collapse, as the test error is constrained by a finite upper bound. We conduct extensive experiments on pre-training from scratch, continual pre-training, and supervised fine-tuning. The results validate our theoretical proof that token-level editing improves model performance.

Xuekai Zhu, Daixuan Cheng, Dinghuai Zhang et al. · stanford, tsinghua

We propose FlowRL: matching the full reward distribution via flow balancing instead of maximizing rewards in large language model (LLM) reinforcement learning (RL). Recent advanced reasoning models adopt reward-maximizing methods (\eg, PPO and GRPO), which tend to over-optimize dominant reward signals while neglecting less frequent but valid reasoning paths, thus reducing diversity. In contrast, we transform scalar rewards into a normalized target distribution using a learnable partition function, and then minimize the reverse KL divergence between the policy and the target distribution. We implement this idea as a flow-balanced optimization method that promotes diverse exploration and generalizable reasoning trajectories. We conduct experiments on math and code reasoning tasks: FlowRL achieves a significant average improvement of $10.0\%$ over GRPO and $5.1\%$ over PPO on math benchmarks, and performs consistently better on code reasoning tasks. These results highlight reward distribution-matching as a key step toward efficient exploration and diverse reasoning in LLM reinforcement learning.

Zhixin Guo, Jianping Zhou, Jiexing Qi et al.

The sheer volume of scientific experimental results and complex technical statements, often presented in tabular formats, presents a formidable barrier to individuals acquiring preferred information. The realms of scientific reasoning and content generation that adhere to user preferences encounter distinct challenges. In this work, we present a new task for generating fluent and logical descriptions that match user preferences over scientific tabular data, aiming to automate scientific document analysis. To facilitate research in this direction, we construct a new challenging dataset CTRLSciTab consisting of table-description pairs extracted from the scientific literature, with highlighted cells and corresponding domain-specific knowledge base. We evaluated popular pre-trained language models to establish a baseline and proposed a novel architecture outperforming competing approaches. The results showed that large models struggle to produce accurate content that aligns with user preferences. As the first of its kind, our work should motivate further research in scientific domains.

Yunchong Song, Siyuan Huang, Xinbing Wang et al.

Graph pooling compresses graph information into a compact representation. State-of-the-art graph pooling methods follow a hierarchical approach, which reduces the graph size step-by-step. These methods must balance memory efficiency with preserving node information, depending on whether they use node dropping or node clustering. Additionally, fixed pooling ratios or numbers of pooling layers are predefined for all graphs, which prevents personalized pooling structures from being captured for each individual graph. In this work, inspired by bottom-up grammar induction, we propose an efficient graph parsing algorithm to infer the pooling structure, which then drives graph pooling. The resulting Graph Parsing Network (GPN) adaptively learns personalized pooling structure for each individual graph. GPN benefits from the discrete assignments generated by the graph parsing algorithm, allowing good memory efficiency while preserving node information intact. Experimental results on standard benchmarks demonstrate that GPN outperforms state-of-the-art graph pooling methods in graph classification tasks while being able to achieve competitive performance in node classification tasks. We also conduct a graph reconstruction task to show GPN's ability to preserve node information and measure both memory and time efficiency through relevant tests.

Jiaqi Cao, Jiarui Wang, Rubin Wei et al.

Large Language Models (LLMs) have shown strong abilities in general language tasks, yet adapting them to specific domains remains a challenge. Current method like Domain Adaptive Pretraining (DAPT) requires costly full-parameter training and suffers from catastrophic forgetting. Meanwhile, Retrieval-Augmented Generation (RAG) introduces substantial inference latency due to expensive nearest-neighbor searches and longer context. This paper introduces Memory Decoder, a plug-and-play pretrained memory that enables efficient domain adaptation without changing the original model's parameters. Memory Decoder employs a small transformer decoder that learns to imitate the behavior of an external non-parametric retriever. Once trained, Memory Decoder can be seamlessly integrated with any pretrained language model that shares the same tokenizer, requiring no model-specific modifications. Experimental results demonstrate that Memory Decoder enables effective adaptation of various Qwen and Llama models to three distinct specialized domains: biomedicine, finance, and law, reducing perplexity by an average of 6.17 points. Overall, Memory Decoder introduces a novel paradigm centered on a specially pretrained memory component designed for domain-specific adaptation. This memory architecture can be integrated in a plug-and-play manner, consistently enhancing performance across multiple models within the target domain.

Jushi Kai, Boyi Zeng, Yixuan Wang et al.

Frequency-domain compression has proven effective in reducing redundancies for spatial signals. In this work, we propose FreqKV, a novel frequency domain key-value (KV) compression technique that enables efficient context window extension for decoder-only large language models (LLMs). Our approach is motivated by a key observation that, in the frequency domain, the energy distribution of the KV cache is predominantly concentrated in low-frequency components. By discarding high-frequency components, we achieve efficient compression of the KV cache with minimal information loss. FreqKV iteratively compresses the increasing KV cache to a fixed size in the frequency domain, allowing models to process lengthy contexts efficiently. Introducing no additional parameters or architectural modifications, FreqKV is applicable to both fine-tuning and inference. With minimal fine-tuning, LLMs can learn to leverage the limited cache that is compressed in the frequency domain and extend the context window. Experiments on a range of long context language modeling and understanding tasks demonstrate the efficiency and effectiveness of the proposed method.

Runzhe Zhang, Letian Chen, Wenpeng Zhang et al.

We present FlowLM, a flow matching language model transformed from pre-trained diffusion language models via efficient fine-tuning. By re-aligning the curved sampling trajectories of diffusion models into straight-line flows, FlowLM enables high quality few-step generation that rivals or even outperforms the quality of 2,000-step diffusion sampling with very few training epochs. Remarkably, finetuned FlowLM reaches performance saturation with only half as many training epochs as training from scratch, both approaches greatly outperforming the original diffusion model, thereby validating our method. Furthermore, we validate a more effective training objective for flow matching: predicting clean data to consistently guide the sampling process towards the true data distribution. Empirical results demonstrate that our approach is highly effective for high-quality, few-step text generation.

Yedan Shen, Kaixin Wu, Yuechen Ding et al.

Generative retrieval (GR) has revolutionized document retrieval with the advent of large language models (LLMs), and LLM-based GR is gradually being adopted by the industry. Despite its remarkable advantages and potential, LLM-based GR suffers from hallucination and generates documents that are irrelevant to the query in some instances, severely challenging its credibility in practical applications. We thereby propose an optimized GR framework designed to alleviate retrieval hallucination, which integrates knowledge distillation reasoning in model training and incorporate decision agent to further improve retrieval precision. Specifically, we employ LLMs to assess and reason GR retrieved query-document (q-d) pairs, and then distill the reasoning data as transferred knowledge to the GR model. Moreover, we utilize a decision agent as post-processing to extend the GR retrieved documents through retrieval model and select the most relevant ones from multi perspectives as the final generative retrieval result. Extensive offline experiments on real-world datasets and online A/B tests on Fund Search and Insurance Search in Alipay demonstrate our framework's superiority and effectiveness in improving search quality and conversion gains.

Siyuan Huang, Zhiyuan Ma, Jintao Du et al.

RAG systems rely on rerankers to identify relevant documents. However, fine-tuning these models remains challenging due to the scarcity of annotated query-document pairs. Existing distillation-based approaches suffer from training-inference misalignment and fail to capture interdependencies among candidate documents. To overcome these limitations, we reframe the reranking process as an attention-mask problem and propose Gumbel Reranking, an end-to-end training framework for rerankers aimed at minimizing the training-inference gap. In our approach, reranker optimization is reformulated as learning a stochastic, document-wise Top-$k$ attention mask using the Gumbel Trick and Relaxed Top-$k$ Sampling. This formulation enables end-to-end optimization by minimizing the overall language loss. Experiments across various settings consistently demonstrate performance gains, including a 10.4\% improvement in recall on HotpotQA for distinguishing indirectly relevant documents.

Yunchong Song, Jushi Kai, Liming Lu et al.

Transformers face a quadratic bottleneck in attention when scaling to long contexts. Recent approaches introduce recurrent memory to extend context beyond the current window, yet these often face a fundamental trade-off between theoretical principles and practical scalability. To address this, we introduce Elastic Memory, a novel memory architecture grounded in the HiPPO framework for online function approximation. Elastic Memory treats historical sequence as samples from continuous signals, applying optimal online compression to encode them into a fixed-size memory state. For retrieval, we propose a flexible \textit{polynomial sampling} mechanism that reconstructs a history summary from this compressed state. Elastic Memory consistently outperformed baselines on long-context (32k+) datasets across three domains. With equal parameters, it beat Memorizing Transformer by 16x memory and outperformed Melodi at all memory sizes, even when Melodi had 30% more parameters. When scaling model size, Elastic Memory stayed ahead of all baselines and was significantly faster than Melodi at 4x size. Furthermore, its decoupled design allows for injecting inductive biases at test-time to boost performance.

Beiya Dai, Yuliang Liu, Daozheng Xue et al.

Next-token prediction (NTP) is the cornerstone of modern large language models (LLMs) pretraining, driving their unprecedented capabilities in text generation, reasoning, and instruction following. However, the token-level prediction limits the model's capacity to capture higher-level semantic structures and long-range contextual relationships. To overcome this limitation, we introduce \textbf{ContextLM}, a framework that augments standard pretraining with an inherent \textbf{next-context prediction} objective. This mechanism trains the model to learn predictive representations of multi-token contexts, leveraging error signals derived from future token chunks. Crucially, ContextLM achieves this enhancement while remaining fully compatible with the standard autoregressive, token-by-token evaluation paradigm (e.g., perplexity). Extensive experiments on the GPT2 and Pythia model families, scaled up to $1.5$B parameters, show that ContextLM delivers consistent improvements in both perplexity and downstream task performance. Our analysis indicates that next-context prediction provides a scalable and efficient pathway to stronger language modeling, yielding better long-range coherence and more effective attention allocation with minimal computational overhead.

Junjie Lu, Yuliang Liu, Chaofeng Qu et al.

Current approaches for strengthening LLM reasoning tend to introduce a training bias toward human-like reasoning trajectories. In step-wise preference optimization, in particular, dependence on human or higher-capacity model annotations for intermediate steps limits exploration of alternative, non-human-like reasoning paths and thus constrains achievable performance. Furthermore, through a small-scale pilot study, we observed that in approximately 75% of cases, the model's first erroneous step occurs after the lowest-confidence point. This suggests that guiding the model at its lowest-confidence point before an error provides more accurate supervision than locating the first explicit error. In this paper, we propose Confidence-Guided Reasoning Path Preference Optimization (CGPO), a method that leverages a confidence signal to identify points of maximal uncertainty in the model's reasoning process and applies self-generated, non-human-like reasoning-path guidance to mitigate trajectory drift. Our experiments span diverse models applied to both code and mathematical reasoning tasks. The results show that, with the same amount of training data, our method using data generated by a small model can achieve better performance in most cases compared with approaches using data generated by a strong model or human-annotated.

Boyi Zeng, He Li, Shixiang Song et al.

The remarkable success of Chain-of-Thought (CoT), which enhances performance by scaling generation steps at test-time, inspires us to ask: can we leverage a similar scaling of computational steps during pretraining to improve the generation of each individual token? To address this, we propose a novel pre-training methodology: Pretraining Language Models with Latent Thoughts (PonderLM-2). Our approach pretrains a language model (LM) to first generate an intermediate latent thought-the last hidden state of the current position-which is then used as input to predict the actual subsequent token. This additional computational step enables the LM to refine its prediction within unconstrained continuous space. Our experiments demonstrate that, at an identical inference cost, a LM that generates one additional latent thought per token outperforms a standard model with double the parameters. For instance, our PonderLM-2-Pythia-1.4B, pretrained on 300B tokens from the Pile, significantly surpasses the vanilla Pythia-2.8B trained on the same data on both language modeling and a range of general downstream tasks. Furthermore, increasing the number of latent thoughts generated before each actual token-forming a chain analogous to CoT-consistently improves the model's performance.

Chang Su, Dengliang Shi, Siyuan Huang et al.

Large language models (LLMs) have increasingly been explored as powerful text embedders. Existing LLM-based text embedding approaches often leverage the embedding of the final token, typically a reserved special token such as [EOS]. However, these tokens have not been intentionally trained to capture the semantics of the whole context, limiting their capacity as text embeddings, especially for retrieval and re-ranking tasks. We propose to add a new training stage before contrastive learning to enrich the semantics of the final token embedding. This stage employs bidirectional generative reconstruction tasks, namely EBQ2D (Embedding-Based Query-to-Document) and EBD2Q (Embedding-Based Document-to-Query), which interleave to anchor the [EOS] embedding and reconstruct either side of Query-Document pairs. Experimental results demonstrate that our additional training stage significantly improves LLM performance on the Massive Text Embedding Benchmark (MTEB), achieving new state-of-the-art results across different LLM base models and scales.

Huanyu Wang, Jushi Kai, Haoli Bai et al.

Vision-Language Models (VLMs) typically replace the predefined image placeholder token (<image>) in textual instructions with visual features from an image encoder, forming the input to a backbone Large Language Model (LLM). However, the large number of vision tokens significantly increases the context length, leading to high computational overhead and inference latency. While previous efforts mitigate this by selecting only important visual features or leveraging learnable queries to reduce token count, they often compromise performance or introduce substantial extra costs. In response, we propose Fourier-VLM, a simple yet efficient method that compresses visual representations in the frequency domain. Our approach is motivated by the observation that vision features output from the vision encoder exhibit concentrated energy in low-frequency components. Leveraging this, we apply a low-pass filter to the vision features using a two-dimensional Discrete Cosine Transform (DCT). Notably, the DCT is efficiently computed via the Fast Fourier Transform (FFT) operator with a time complexity of $\mathcal{O}(n\log n)$, minimizing the extra computational cost while introducing no additional parameters. Extensive experiments across various image-based benchmarks demonstrate that Fourier-VLM achieves competitive performance with strong generalizability across both LLaVA and Qwen-VL architectures. Crucially, it reduce inference FLOPs by up to 83.8% and boots generation speed by 31.2% compared to LLaVA-v1.5, highlighting the superior efficiency and practicality.

Rubin Wei, Jiaqi Cao, Jiarui Wang et al.

Modern approaches to enhancing Large Language Models' factual accuracy and knowledge utilization face a fundamental trade-off: non-parametric retrieval-augmented generation (RAG) provides flexible access to external knowledge but suffers from high inference latency and shallow integration, while parametric fine-tuning methods like LoRA risk catastrophic forgetting and degraded general capabilities. In this work, we propose MLP Memory, a lightweight parametric module that learns to internalize retrieval patterns without explicit document access. By pretraining an MLP to imitate a $k$NN retriever's behavior on the entire pretraining dataset, we create a differentiable memory component that captures the benefits of retrieval-based knowledge access in a fully parametric form. Our architecture integrates this pretrained MLP Memory with Transformer decoders through simple probability interpolation, yielding 17.5\% and 24.1\% scaling gains on WikiText-103 and Web datasets, respectively. It further achieves 12.3\% relative improvement on five question-answering benchmarks and 5.2 points absolute gain across nine general NLP tasks, while reducing hallucinations by up to 10 points on HaluEval. Moreover, MLP Memory delivers 2.5$\times$ faster inference than RAG with superior accuracy. Our findings show that learning retrieval patterns parametrically bridges the gap between efficient inference and effective knowledge access, offering a practical alternative to both RAG and fine-tuning approaches.