Yuandong Tian

Zichang Liu, Jue Wang, Tri Dao et al. · eth-zurich

Large language models (LLMs) with hundreds of billions of parameters have sparked a new wave of exciting AI applications. However, they are computationally expensive at inference time. Sparsity is a natural approach to reduce this cost, but existing methods either require costly retraining, have to forgo LLM's in-context learning ability, or do not yield wall-clock time speedup on modern hardware. We hypothesize that contextual sparsity, which are small, input-dependent sets of attention heads and MLP parameters that yield approximately the same output as the dense model for a given input, can address these issues. We show that contextual sparsity exists, that it can be accurately predicted, and that we can exploit it to speed up LLM inference in wall-clock time without compromising LLM's quality or in-context learning ability. Based on these insights, we propose DejaVu, a system that uses a low-cost algorithm to predict contextual sparsity on the fly given inputs to each layer, along with an asynchronous and hardware-aware implementation that speeds up LLM inference. We validate that DejaVu can reduce the inference latency of OPT-175B by over 2X compared to the state-of-the-art FasterTransformer, and over 6X compared to the widely used Hugging Face implementation, without compromising model quality. The code is available at https://github.com/FMInference/DejaVu.

Randall Balestriero, Mark Ibrahim, Vlad Sobal et al. · meta-ai

Self-supervised learning, dubbed the dark matter of intelligence, is a promising path to advance machine learning. Yet, much like cooking, training SSL methods is a delicate art with a high barrier to entry. While many components are familiar, successfully training a SSL method involves a dizzying set of choices from the pretext tasks to training hyper-parameters. Our goal is to lower the barrier to entry into SSL research by laying the foundations and latest SSL recipes in the style of a cookbook. We hope to empower the curious researcher to navigate the terrain of methods, understand the role of the various knobs, and gain the know-how required to explore how delicious SSL can be.

Zhenyu Zhang, Ying Sheng, Tianyi Zhou et al.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H$_2$). Through a comprehensive investigation, we find that (i) the emergence of H$_2$ is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H$_2$O), a KV cache eviction policy that dynamically retains a balance of recent and H$_2$ tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H$_2$O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29$\times$, 29$\times$, and 3$\times$ on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9$\times$. The code is available at https://github.com/FMInference/H2O.

Tongzhou Wang, Simon S. Du, Antonio Torralba et al. · mit

The ability to separate signal from noise, and reason with clean abstractions, is critical to intelligence. With this ability, humans can efficiently perform real world tasks without considering all possible nuisance factors.How can artificial agents do the same? What kind of information can agents safely discard as noises? In this work, we categorize information out in the wild into four types based on controllability and relation with reward, and formulate useful information as that which is both controllable and reward-relevant. This framework clarifies the kinds information removed by various prior work on representation learning in reinforcement learning (RL), and leads to our proposed approach of learning a Denoised MDP that explicitly factors out certain noise distractors. Extensive experiments on variants of DeepMind Control Suite and RoboDesk demonstrate superior performance of our denoised world model over using raw observations alone, and over prior works, across policy optimization control tasks as well as the non-control task of joint position regression.

Kevin Yang, Dan Klein, Asli Celikyilmaz et al. · berkeley, meta-ai

We propose Reinforcement Learning from Contrastive Distillation (RLCD), a method for aligning language models to follow principles expressed in natural language (e.g., to be more harmless) without using human feedback. RLCD creates preference pairs from two contrasting model outputs, one using a positive prompt designed to encourage following the given principles, and one using a negative prompt designed to encourage violating them. Using two different prompts causes model outputs to be more differentiated on average, resulting in cleaner preference labels in the absence of human annotations. We then use the preference pairs to train a preference model, which is in turn used to improve a base unaligned language model via reinforcement learning. Empirically, RLCD outperforms RLAIF (Bai et al., 2022b) and context distillation (Huang et al., 2022) baselines across three diverse alignment tasks--harmlessness, helpfulness, and story outline generation--and when using both 7B and 30B model scales for simulating preference data.

Tianhao Wu, Weizhe Yuan, Olga Golovneva et al. · meta-ai

Large Language Models (LLMs) are rapidly surpassing human knowledge in many domains. While improving these models traditionally relies on costly human data, recent self-rewarding mechanisms (Yuan et al., 2024) have shown that LLMs can improve by judging their own responses instead of relying on human labelers. However, existing methods have primarily focused on improving model responses rather than judgment capabilities, resulting in rapid saturation during iterative training. To address this issue, we introduce a novel Meta-Rewarding step to the self-improvement process, where the model judges its own judgements and uses that feedback to refine its judgment skills. Surprisingly, this unsupervised approach improves the model's ability to judge {\em and} follow instructions, as demonstrated by a win rate improvement of Llama-3-8B-Instruct from 22.9% to 39.4% on AlpacaEval 2, and 20.6% to 29.1% on Arena-Hard. These results strongly suggest the potential for self-improving models without human supervision.

Guangxuan Xiao, Yuandong Tian, Beidi Chen et al.

Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a "sink" even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2x speedup. Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.

Danqing Wang, Kevin Yang, Hanlin Zhu et al. · cmu

Recent research has increasingly focused on evaluating large language models' (LLMs) alignment with diverse human values and preferences, particularly for open-ended tasks like story generation. Traditional evaluation metrics rely heavily on lexical similarity with human-written references, often showing poor correlation with human judgments and failing to account for alignment with the diversity of human preferences. To address these challenges, we introduce PerSE, an interpretable evaluation framework designed to assess alignment with specific human preferences. It is tuned to infer specific preferences from an in-context personal profile and evaluate the alignment between the generated content and personal preferences. PerSE enhances interpretability by providing detailed comments and fine-grained scoring, facilitating more personalized content generation. Our 13B LLaMA-2-based PerSE shows a 15.8% increase in Kendall correlation and a 13.7% rise in accuracy with zero-shot reviewers compared to GPT-4. It also outperforms GPT-4 by 46.01% in Kendall correlation on new domains, indicating its transferability.

Daochen Zha, Louis Feng, Qiaoyu Tan et al.

We study embedding table placement for distributed recommender systems, which aims to partition and place the tables on multiple hardware devices (e.g., GPUs) to balance the computation and communication costs. Although prior work has explored learning-based approaches for the device placement of computational graphs, embedding table placement remains to be a challenging problem because of 1) the operation fusion of embedding tables, and 2) the generalizability requirement on unseen placement tasks with different numbers of tables and/or devices. To this end, we present DreamShard, a reinforcement learning (RL) approach for embedding table placement. DreamShard achieves the reasoning of operation fusion and generalizability with 1) a cost network to directly predict the costs of the fused operation, and 2) a policy network that is efficiently trained on an estimated Markov decision process (MDP) without real GPU execution, where the states and the rewards are estimated with the cost network. Equipped with sum and max representation reductions, the two networks can directly generalize to any unseen tasks with different numbers of tables and/or devices without fine-tuning. Extensive experiments show that DreamShard substantially outperforms the existing human expert and RNN-based strategies with up to 19% speedup over the strongest baseline on large-scale synthetic tables and our production tables. The code is available at https://github.com/daochenzha/dreamshard

Daochen Zha, Louis Feng, Bhargav Bhushanam et al.

Embedding learning is an important technique in deep recommendation models to map categorical features to dense vectors. However, the embedding tables often demand an extremely large number of parameters, which become the storage and efficiency bottlenecks. Distributed training solutions have been adopted to partition the embedding tables into multiple devices. However, the embedding tables can easily lead to imbalances if not carefully partitioned. This is a significant design challenge of distributed systems named embedding table sharding, i.e., how we should partition the embedding tables to balance the costs across devices, which is a non-trivial task because 1) it is hard to efficiently and precisely measure the cost, and 2) the partition problem is known to be NP-hard. In this work, we introduce our novel practice in Meta, namely AutoShard, which uses a neural cost model to directly predict the multi-table costs and leverages deep reinforcement learning to solve the partition problem. Experimental results on an open-sourced large-scale synthetic dataset and Meta's production dataset demonstrate the superiority of AutoShard over the heuristics. Moreover, the learned policy of AutoShard can transfer to sharding tasks with various numbers of tables and different ratios of the unseen tables without any fine-tuning. Furthermore, AutoShard can efficiently shard hundreds of tables in seconds. The effectiveness, transferability, and efficiency of AutoShard make it desirable for production use. Our algorithms have been deployed in Meta production environment. A prototype is available at https://github.com/daochenzha/autoshard

Kevin Yang, Yuandong Tian, Nanyun Peng et al.

We consider the problem of automatically generating longer stories of over two thousand words. Compared to prior work on shorter stories, long-range plot coherence and relevance are more central challenges here. We propose the Recursive Reprompting and Revision framework (Re3) to address these challenges by (a) prompting a general-purpose language model to construct a structured overarching plan, and (b) generating story passages by repeatedly injecting contextual information from both the plan and current story state into a language model prompt. We then revise by (c) reranking different continuations for plot coherence and premise relevance, and finally (d) editing the best continuation for factual consistency. Compared to similar-length stories generated directly from the same base model, human evaluators judged substantially more of Re3's stories as having a coherent overarching plot (by 14% absolute increase), and relevant to the given initial premise (by 20%).

Yuandong Tian, Yiping Wang, Zhenyu Zhang et al.

We propose Joint MLP/Attention (JoMA) dynamics, a novel mathematical framework to understand the training procedure of multilayer Transformer architectures. This is achieved by integrating out the self-attention layer in Transformers, producing a modified dynamics of MLP layers only. JoMA removes unrealistic assumptions in previous analysis (e.g., lack of residual connection) and predicts that the attention first becomes sparse (to learn salient tokens), then dense (to learn less salient tokens) in the presence of nonlinear activations, while in the linear case, it is consistent with existing works that show attention becomes sparse over time. We leverage JoMA to qualitatively explains how tokens are combined to form hierarchies in multilayer Transformers, when the input tokens are generated by a latent hierarchical generative model. Experiments on models trained from real-world dataset (Wikitext2/Wikitext103) and various pre-trained models (OPT, Pythia) verify our theoretical findings. Code can be found in https://github.com/facebookresearch/luckmatters/tree/yuandong3.

Ajay Jaiswal, Yifan Wang, Lu Yin et al.

Large Language Models' (LLMs) weight matrices can often be expressed in low-rank form with potential to relax memory and compute resource requirements. Unlike prior efforts that focus on developing novel matrix decompositions, in this work we study the non-uniform low-rank properties of weight matrices in LLMs through the lens of stabilizing gradient subspace. First, we provide a theoretical framework to understand the stabilization of gradient subspaces through Hessian analysis. Second, we empirically establish an important relationship between gradient dynamics and low-rank expressiveness of weight matrices. Our findings reveal that different LLM components exhibit varying levels of converged low-rank structures, necessitating variable rank reduction across them to minimize drop in performance due to compression. Drawing on this result, we present Weight Low-Rank Projection(WeLore) that unifies weight compression and memory-efficient fine-tuning into one, in a data-agnostic and one-shot manner. When used as a compression technique, WeLore categorizes weight matrices into Low-rank Components (LRCs) and Non-Low-rank Components (N-LRCs) and suitably encodes them for minimum performance loss. Our gradient dynamics perspective illustrates that LRCs tend to have better fine-tuning capabilities and their standalone fine-tuning can closely mimic and sometimes outperform the training loss trajectory and performance of full fine-tuning with notable memory and compute footprint reduction. Codes are available at https://github.com/VITA-Group/WeLore.

Minghao Xu, Yuanfan Guo, Yi Xu et al.

Modeling spatial relationship in the data remains critical across many different tasks, such as image classification, semantic segmentation and protein structure understanding. Previous works often use a unified solution like relative positional encoding. However, there exists different kinds of spatial relations, including short-range, medium-range and long-range relations, and modeling them separately can better capture the focus of different tasks on the multi-range relations (e.g., short-range relations can be important in instance segmentation, while long-range relations should be upweighted for semantic segmentation). In this work, we introduce the EurNet for Efficient multi-range relational modeling. EurNet constructs the multi-relational graph, where each type of edge corresponds to short-, medium- or long-range spatial interactions. In the constructed graph, EurNet adopts a novel modeling layer, called gated relational message passing (GRMP), to propagate multi-relational information across the data. GRMP captures multiple relations within the data with little extra computational cost. We study EurNets in two important domains for image and protein structure modeling. Extensive experiments on ImageNet classification, COCO object detection and ADE20K semantic segmentation verify the gains of EurNet over the previous SoTA FocalNet. On the EC and GO protein function prediction benchmarks, EurNet consistently surpasses the previous SoTA GearNet. Our results demonstrate the strength of EurNets on modeling spatial multi-relational data from various domains. The implementations of EurNet for image modeling are available at https://github.com/hirl-team/EurNet-Image . The implementations for other applied domains/tasks will be released soon.

Shouyuan Chen, Sherman Wong, Liangjian Chen et al.

We present Position Interpolation (PI) that extends the context window sizes of RoPE-based pretrained LLMs such as LLaMA models to up to 32768 with minimal fine-tuning (within 1000 steps), while demonstrating strong empirical results on various tasks that require long context, including passkey retrieval, language modeling, and long document summarization from LLaMA 7B to 65B. Meanwhile, the extended model by Position Interpolation preserve quality relatively well on tasks within its original context window. To achieve this goal, Position Interpolation linearly down-scales the input position indices to match the original context window size, rather than extrapolating beyond the trained context length which may lead to catastrophically high attention scores that completely ruin the self-attention mechanism. Our theoretical study shows that the upper bound of interpolation is at least $\sim 600 \times$ smaller than that of extrapolation, further demonstrating its stability. Models extended via Position Interpolation retain its original architecture and can reuse most pre-existing optimization and infrastructure.

Xiao Wang, Haoqi Fan, Yuandong Tian et al.

Many recent self-supervised frameworks for visual representation learning are based on certain forms of Siamese networks. Such networks are conceptually symmetric with two parallel encoders, but often practically asymmetric as numerous mechanisms are devised to break the symmetry. In this work, we conduct a formal study on the importance of asymmetry by explicitly distinguishing the two encoders within the network -- one produces source encodings and the other targets. Our key insight is keeping a relatively lower variance in target than source generally benefits learning. This is empirically justified by our results from five case studies covering different variance-oriented designs, and is aligned with our preliminary theoretical analysis on the baseline. Moreover, we find the improvements from asymmetric designs generalize well to longer training schedules, multiple other frameworks and newer backbones. Finally, the combined effect of several asymmetric designs achieves a state-of-the-art accuracy on ImageNet linear probing and competitive results on downstream transfer. We hope our exploration will inspire more research in exploiting asymmetry for Siamese representation learning.

Zhengyao Jiang, Tianjun Zhang, Michael Janner et al.

Planning-based reinforcement learning has shown strong performance in tasks in discrete and low-dimensional continuous action spaces. However, planning usually brings significant computational overhead for decision-making, and scaling such methods to high-dimensional action spaces remains challenging. To advance efficient planning for high-dimensional continuous control, we propose Trajectory Autoencoding Planner (TAP), which learns low-dimensional latent action codes with a state-conditional VQ-VAE. The decoder of the VQ-VAE thus serves as a novel dynamics model that takes latent actions and current state as input and reconstructs long-horizon trajectories. During inference time, given a starting state, TAP searches over discrete latent actions to find trajectories that have both high probability under the training distribution and high predicted cumulative reward. Empirical evaluation in the offline RL setting demonstrates low decision latency which is indifferent to the growing raw action dimensionality. For Adroit robotic hand manipulation tasks with high-dimensional continuous action space, TAP surpasses existing model-based methods by a large margin and also beats strong model-free actor-critic baselines.

Taoan Huang, Aaron Ferber, Yuandong Tian et al.

Integer Linear Programs (ILPs) are powerful tools for modeling and solving a large number of combinatorial optimization problems. Recently, it has been shown that Large Neighborhood Search (LNS), as a heuristic algorithm, can find high quality solutions to ILPs faster than Branch and Bound. However, how to find the right heuristics to maximize the performance of LNS remains an open problem. In this paper, we propose a novel approach, CL-LNS, that delivers state-of-the-art anytime performance on several ILP benchmarks measured by metrics including the primal gap, the primal integral, survival rates and the best performing rate. Specifically, CL-LNS collects positive and negative solution samples from an expert heuristic that is slow to compute and learns a new one with a contrastive loss. We use graph attention networks and a richer set of features to further improve its performance.

Arman Zharmagambetov, Brandon Amos, Aaron Ferber et al.

Recent works in learning-integrated optimization have shown promise in settings where the optimization problem is only partially observed or where general-purpose optimizers perform poorly without expert tuning. By learning an optimizer $\mathbf{g}$ to tackle these challenging problems with $f$ as the objective, the optimization process can be substantially accelerated by leveraging past experience. The optimizer can be trained with supervision from known optimal solutions or implicitly by optimizing the compound function $f\circ \mathbf{g}$. The implicit approach may not require optimal solutions as labels and is capable of handling problem uncertainty; however, it is slow to train and deploy due to frequent calls to optimizer $\mathbf{g}$ during both training and testing. The training is further challenged by sparse gradients of $\mathbf{g}$, especially for combinatorial solvers. To address these challenges, we propose using a smooth and learnable Landscape Surrogate $M$ as a replacement for $f\circ \mathbf{g}$. This surrogate, learnable by neural networks, can be computed faster than the solver $\mathbf{g}$, provides dense and smooth gradients during training, can generalize to unseen optimization problems, and is efficiently learned via alternating optimization. We test our approach on both synthetic problems, including shortest path and multidimensional knapsack, and real-world problems such as portfolio optimization, achieving comparable or superior objective values compared to state-of-the-art baselines while reducing the number of calls to $\mathbf{g}$. Notably, our approach outperforms existing methods for computationally expensive high-dimensional problems.

Youwei Liang, Kevin Stone, Ali Shameli et al.

Finding the optimal pass sequence of compilation can lead to a significant reduction in program size and/or improvement in program efficiency. Prior works on compilation pass ordering have two major drawbacks. They either require an excessive budget (in terms of compilation steps) at compile time or fail to generalize to unseen programs. In this paper, for code-size reduction tasks, we propose a novel pipeline to find program-dependent pass sequences within 45 compilation calls. It first identifies a coreset of 50 pass sequences via greedy optimization of a submodular function, and then learns a policy with Graph Neural Network (GNN) to pick the optimal sequence by predicting the normalized values of the pass sequences in the coreset. Despite its simplicity, our pipeline outperforms the default -Oz flag by an average of 4.7% over a large collection (4683) of unseen code repositories from diverse domains across 14 datasets. In comparison, previous approaches like reinforcement learning on the raw pass sequence space may take days to train due to sparse reward, and may not generalize well in held-out ones from different domains. Our results demonstrate that existing human-designed compiler flags can be improved with a simple yet effective technique that transforms the raw action space into a small one with denser rewards.

Aaron Ferber, Taoan Huang, Daochen Zha et al.

Optimization problems with nonlinear cost functions and combinatorial constraints appear in many real-world applications but remain challenging to solve efficiently compared to their linear counterparts. To bridge this gap, we propose $\textbf{SurCo}$ that learns linear $\underline{\text{Sur}}$rogate costs which can be used in existing $\underline{\text{Co}}$mbinatorial solvers to output good solutions to the original nonlinear combinatorial optimization problem. The surrogate costs are learned end-to-end with nonlinear loss by differentiating through the linear surrogate solver, combining the flexibility of gradient-based methods with the structure of linear combinatorial optimization. We propose three $\texttt{SurCo}$ variants: $\texttt{SurCo}-\texttt{zero}$ for individual nonlinear problems, $\texttt{SurCo}-\texttt{prior}$ for problem distributions, and $\texttt{SurCo}-\texttt{hybrid}$ to combine both distribution and problem-specific information. We give theoretical intuition motivating $\texttt{SurCo}$, and evaluate it empirically. Experiments show that $\texttt{SurCo}$ finds better solutions faster than state-of-the-art and domain expert approaches in real-world optimization problems such as embedding table sharding, inverse photonic design, and nonlinear route planning.

Tengyu Xu, Eryk Helenowski, Karthik Abinav Sankararaman et al.

Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off of multiple and/or sometimes conflicting objectives). Applying RLHF for MTL currently requires careful tuning of the weights for reward model and data combinations. This is often done via human intuition and does not generalize. In this work, we introduce a novel post-training paradigm which we called Constrained Generative Policy Optimization (CGPO). The core of CGPO is Mixture of Judges (MoJ) with cost-efficient constrained policy optimization with stratification, which can identify the perfect blend in RLHF in a principled manner. It shows strong empirical results with theoretical guarantees, does not require extensive hyper-parameter tuning, and is plug-and-play in common post-training pipelines. Together, this can detect and mitigate reward hacking behaviors while reaching a pareto-optimal point across an extremely large number of objectives. Our empirical evaluations demonstrate that CGPO significantly outperforms standard RLHF algorithms like PPO and DPO across various tasks including general chat, STEM questions, instruction following, and coding. Specifically, CGPO shows improvements of 7.4% in AlpacaEval-2 (general chat), 12.5% in Arena-Hard (STEM & reasoning), and consistent gains in other domains like math and coding. Notably, PPO, while commonly used, is prone to severe reward hacking in popular coding benchmarks, which CGPO successfully addresses. This breakthrough in RLHF not only tackles reward hacking and extreme multi-objective optimization challenges but also advances the state-of-the-art in aligning general-purpose LLMs for diverse applications.

Kevin Yang, Dan Klein, Nanyun Peng et al.

We propose the Detailed Outline Control (DOC) framework for improving long-range plot coherence when automatically generating several-thousand-word-long stories. DOC consists of two complementary components: a detailed outliner and a detailed controller. The detailed outliner creates a more detailed, hierarchically structured outline, shifting creative burden from the main drafting procedure to the planning stage. The detailed controller ensures the more detailed outline is still respected during generation by controlling story passages to align with outline details. In human evaluations of automatically generated stories, DOC substantially outperforms a strong Re3 baseline (Yang et al., 2022) on plot coherence (22.5% absolute gain), outline relevance (28.2%), and interestingness (20.7%). Humans also judged DOC to be much more controllable in an interactive generation setting.

Zhenyu Zhang, Ajay Jaiswal, Lu Yin et al.

Training Large Language Models (LLMs) is memory-intensive due to the large number of parameters and associated optimization states. GaLore, a recent method, reduces memory usage by projecting weight gradients into a low-rank subspace without compromising performance. However, GaLore relies on time-consuming Singular Value Decomposition (SVD) operations to identify the subspace, and the frequent subspace updates lead to significant training time overhead. Moreover, GaLore offers minimal improvements in accuracy and efficiency compared to LoRA in more accessible fine-tuning scenarios. To address these limitations, we introduce Q-Galore, a novel approach that substantially reduces memory usage by combining quantization and low-rank projection, surpassing the benefits of GaLore. Our method is based on two key observations: (i) the gradient subspace exhibits diverse properties, with some layers converging early in training while others are subject to frequent changes; (ii) the projection matrices are highly resilient to low-bit quantization. Leveraging these insights, Q-GaLore adaptively updates the gradient subspace based on its convergence statistics, achieving comparable performance while significantly reducing the number of SVD operations. We maintain the projection matrices in INT4 format and weights in INT8 format, incorporating stochastic rounding to capture accumulated gradient information. This approach enables a high-precision training trajectory using only low-precision weights. We demonstrate that Q-GaLore achieves highly competitive performance with exceptional memory efficiency. At pre-training, Q-GaLore facilitates training a LLaMA-7B model from scratch on a single NVIDIA RTX 4060 Ti with only 16 GB memory. At fine-tuning, it reduces memory consumption by up to 50% compared to LoRA and GaLore, while consistently outperforming QLoRA at the same memory cost.

Yun Joon Soh, Hanxian Huang, Yuandong Tian et al.

Supporting longer context for Large Language Models (LLM) is a promising direction to advance LLMs. As training a model for a longer context window is computationally expensive, many alternative solutions, such as Retrieval Augmented Generation (RAG), have been used. However, most existing RAG methods adopt embedding-based retrieval that falls short on long contexts. To address such challenges, we propose an attention-based retrieval technique, You Only Use Reactive Attention slice (YOURA). YOURA leverages a novel retrieval heuristic called reaction score to rank the relevance of each sentence in the input context with the query sentence. Intuitively, we measure how the per-token attention score "reacts" to the query and greedily retrieves the most reactive sentences. Internally, YOURA generates a token-indexed vector (called reaction vector) for the whole input context. To map each sentence to the token-indexed vector, we propose an Embedding-Agnostic Sentence Yield (EASY), a best-effort token wiggling algorithm. We evaluate our retrieval technique on three open-source pre-trained LLM models across six LongBench QA datasets. Our technique achieves up to 30% vLLM inference throughput improvement for serving long-context queries with a nearly identical quality score to the simple yet effective truncate-middle approach.

Yuandong Tian

While the empirical success of self-supervised learning (SSL) heavily relies on the usage of deep nonlinear models, existing theoretical works on SSL understanding still focus on linear ones. In this paper, we study the role of nonlinearity in the training dynamics of contrastive learning (CL) on one and two-layer nonlinear networks with homogeneous activation $h(x) = h'(x)x$. We have two major theoretical discoveries. First, the presence of nonlinearity can lead to many local optima even in 1-layer setting, each corresponding to certain patterns from the data distribution, while with linear activation, only one major pattern can be learned. This suggests that models with lots of parameters can be regarded as a \emph{brute-force} way to find these local optima induced by nonlinearity. Second, in the 2-layer case, linear activation is proven not capable of learning specialized weights into diverse patterns, demonstrating the importance of nonlinearity. In addition, for 2-layer setting, we also discover \emph{global modulation}: those local patterns discriminative from the perspective of global-level patterns are prioritized to learn, further characterizing the learning process. Simulation verifies our theoretical findings.

Taoan Huang, Aaron Ferber, Yuandong Tian et al.

Large Neighborhood Search (LNS) is a popular heuristic algorithm for solving combinatorial optimization problems (COP). It starts with an initial solution to the problem and iteratively improves it by searching a large neighborhood around the current best solution. LNS relies on heuristics to select neighborhoods to search in. In this paper, we focus on designing effective and efficient heuristics in LNS for integer linear programs (ILP) since a wide range of COPs can be represented as ILPs. Local Branching (LB) is a heuristic that selects the neighborhood that leads to the largest improvement over the current solution in each iteration of LNS. LB is often slow since it needs to solve an ILP of the same size as input. Our proposed heuristics, LB-RELAX and its variants, use the linear programming relaxation of LB to select neighborhoods. Empirically, LB-RELAX and its variants compute as effective neighborhoods as LB but run faster. They achieve state-of-the-art anytime performance on several ILP benchmarks.

Andrew Cohen, Weiping Dou, Jiang Zhu et al.

Linear Partial Differential Equations (PDEs) govern the spatial-temporal dynamics of physical systems that are essential to building modern technology. When working with linear PDEs, designing a physical system for a specific outcome is difficult and costly due to slow and expensive explicit simulation of PDEs and the highly nonlinear relationship between a system's configuration and its behavior. In this work, we prove a parametric form that certain physical quantities in the Fourier domain must obey in linear PDEs, named the CZP (Constant-Zeros-Poles) framework. Applying CZP to antenna design, an industrial application using linear PDEs (i.e., Maxwell's equations), we derive a sample-efficient parametric surrogate model that directly predicts its scattering coefficients without explicit numerical PDE simulation. Combined with a novel image-based antenna representation and an attention-based neural network architecture, CZP outperforms baselines by 10% to 25% in terms of test loss and also is able to find 2D antenna designs verifiable by commercial software with $33\%$ greater success than baselines, when coupled with sequential search techniques like reinforcement learning.

Aaron Ferber, Arman Zharmagambetov, Taoan Huang et al.

Deep generative models like GAN and VAE have shown impressive results in generating unconstrained objects like images. However, many design settings arising in industrial design, material science, computer graphics and more require that the generated objects satisfy hard combinatorial constraints or meet objectives in addition to modeling a data distribution. To address this, we propose GenCO, a generative framework that guarantees constraint satisfaction throughout training by leveraging differentiable combinatorial solvers to enforce feasibility. GenCO imposes the generative loss on provably feasible solutions rather than intermediate soft solutions, meaning that the deep generative network can focus on ensuring the generated objects match the data distribution without having to also capture feasibility. This shift enables practitioners to enforce hard constraints on the generated outputs during end-to-end training, enabling assessments of their feasibility and introducing additional combinatorial loss components to deep generative training. We demonstrate the effectiveness of our approach on a variety of generative combinatorial tasks, including game level generation, map creation for path planning, and photonic device design, consistently demonstrating its capability to yield diverse, high-quality solutions that verifiably adhere to user-specified combinatorial properties.

Anselm Paulus, Arman Zharmagambetov, Chuan Guo et al.

Large Language Models (LLMs) are vulnerable to jailbreaking attacks that lead to generation of inappropriate or harmful content. Manual red-teaming requires a time-consuming search for adversarial prompts, whereas automatic adversarial prompt generation often leads to semantically meaningless attacks that do not scale well. In this paper, we present a novel method that uses another LLM, called AdvPrompter, to generate human-readable adversarial prompts in seconds. AdvPrompter, which is trained using an alternating optimization algorithm, generates suffixes that veil the input instruction without changing its meaning, such that the TargetLLM is lured to give a harmful response. Experimental results on popular open source TargetLLMs show highly competitive results on the AdvBench and HarmBench datasets, that also transfer to closed-source black-box LLMs. We also show that training on adversarial suffixes generated by AdvPrompter is a promising strategy for improving the robustness of LLMs to jailbreaking attacks.

Hanshi Sun, Zhuoming Chen, Xinyu Yang et al.

With large language models (LLMs) widely deployed in long content generation recently, there has emerged an increasing demand for efficient long-sequence inference support. However, key-value (KV) cache, which is stored to avoid re-computation, has emerged as a critical bottleneck by growing linearly in size with the sequence length. Due to the auto-regressive nature of LLMs, the entire KV cache will be loaded for every generated token, resulting in low utilization of computational cores and high latency. While various compression methods for KV cache have been proposed to alleviate this issue, they suffer from degradation in generation quality. We introduce TriForce, a hierarchical speculative decoding system that is scalable for long sequence generation. This approach leverages the original model weights and dynamic sparse KV cache via retrieval as a draft model, which serves as an intermediate layer in the hierarchy and is further speculated by a smaller model to reduce its drafting latency. TriForce not only facilitates impressive speedups for Llama2-7B-128K, achieving up to 2.31$\times$ on an A100 GPU but also showcases scalability in handling even longer contexts. For the offloading setting on two RTX 4090 GPUs, TriForce achieves 0.108s/token$\unicode{x2014}$only half as slow as the auto-regressive baseline on an A100, which attains 7.78$\times$ on our optimized offloading system. Additionally, TriForce performs 4.86$\times$ than DeepSpeed-Zero-Inference on a single RTX 4090 GPU. TriForce's robustness is highlighted by its consistently outstanding performance across various temperatures. The code is available at https://github.com/Infini-AI-Lab/TriForce.

Hanlin Zhu, Andrew Cohen, Danqing Wang et al.

Story plots, while short, carry most of the essential information of a full story that may contain tens of thousands of words. We study the problem of automatic generation of story plots, which includes story premise, character descriptions, plot outlines, etc. To generate a single engaging plot, existing plot generators (e.g., DOC (Yang et al., 2022a)) require hundreds to thousands of calls to LLMs (e.g., OpenAI API) in the planning stage of the story plot, which is costly and takes at least several minutes. Moreover, the hard-wired nature of the method makes the pipeline non-differentiable, blocking fast specialization and personalization of the plot generator. In this paper, we propose three models, $\texttt{OpenPlot}$, $\texttt{E2EPlot}$ and $\texttt{RLPlot}$, to address these challenges. $\texttt{OpenPlot}$ replaces expensive OpenAI API calls with LLaMA2 (Touvron et al., 2023) calls via careful prompt designs, which leads to inexpensive generation of high-quality training datasets of story plots. We then train an end-to-end story plot generator, $\texttt{E2EPlot}$, by supervised fine-tuning (SFT) using approximately 13000 story plots generated by $\texttt{OpenPlot}$. $\texttt{E2EPlot}$ generates story plots of comparable quality to $\texttt{OpenPlot}$, and is > 10$\times$ faster (1k tokens in only 30 seconds on average). Finally, we obtain $\texttt{RLPlot}$ that is further fine-tuned with RLHF on several different reward models for different aspects of story quality, which yields 60.0$\%$ winning rate against $\texttt{E2EPlot}$ along the aspect of suspense and surprise.

Zhuoming Chen, Ranajoy Sadhukhan, Zihao Ye et al. · uw

Large language models (LLMs) with long context windows have gained significant attention. However, the KV cache, stored to avoid re-computation, becomes a bottleneck. Various dynamic sparse or TopK-based attention approximation methods have been proposed to leverage the common insight that attention is sparse. In this paper, we first show that TopK attention itself suffers from quality degradation in certain downstream tasks because attention is not always as sparse as expected. Rather than selecting the keys and values with the highest attention scores, sampling with theoretical guarantees can provide a better estimation for attention output. To make the sampling-based approximation practical in LLM generation, we propose MagicPIG, a heterogeneous system based on Locality Sensitive Hashing (LSH). MagicPIG significantly reduces the workload of attention computation while preserving high accuracy for diverse tasks. MagicPIG stores the LSH hash tables and runs the attention computation on the CPU, which allows it to serve longer contexts and larger batch sizes with high approximation accuracy. MagicPIG can improve decoding throughput by up to $5\times$ across various GPU hardware and achieve 54ms decoding latency on a single RTX 4090 for Llama-3.1-8B-Instruct model with a context of 96k tokens. The code is available at https://github.com/Infini-AI-Lab/MagicPIG.

Yichao Fu, Xuewei Wang, Yuandong Tian et al.

Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional model training or hyperparameter tuning and can be seamlessly integrated into existing serving frameworks. We evaluate DeepConf across a variety of reasoning tasks and the latest open-source models, including Qwen 3 and GPT-OSS series. Notably, on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9% accuracy and reduces generated tokens by up to 84.7% compared to full parallel thinking.

Weizhe Yuan, Jane Yu, Song Jiang et al.

Scaling reasoning capabilities beyond traditional domains such as math and coding is hindered by the lack of diverse and high-quality questions. To overcome this limitation, we introduce a scalable approach for generating diverse and challenging reasoning questions, accompanied by reference answers. We present NaturalReasoning, a comprehensive dataset comprising 2.8 million questions that span multiple domains, including STEM fields (e.g., Physics, Computer Science), Economics, Social Sciences, and more. We demonstrate the utility of the questions in NaturalReasoning through knowledge distillation experiments which show that NaturalReasoning can effectively elicit and transfer reasoning capabilities from a strong teacher model. Furthermore, we demonstrate that NaturalReasoning is also effective for unsupervised self-training using external reward models or self-rewarding. To foster future work, we publicly release NaturalReasoning at https://huggingface.co/datasets/facebook/natural_reasoning.

DiJia Su, Sainbayar Sukhbaatar, Michael Rabbat et al. · meta-ai

In cognition theory, human thinking is governed by two systems: the fast and intuitive System 1 and the slower but more deliberative System 2. Analogously, Large Language Models (LLMs) can operate in two reasoning modes: outputting only the solutions (\emph{fast mode}) or both the reasoning chain and the final solution (\emph{slow mode}). We present \dualformer, a single Transformer model that seamlessly integrates both the fast and slow reasoning modes by training on randomized reasoning traces, where different parts of the traces are strategically dropped during training. At inference time, \dualformer can be easily configured to execute in either fast or slow mode, or automatically decide which mode to engage (\emph{auto mode}). It outperforms baselines in both performance and computational efficiency across all three modes: (1) in slow mode, \dualformer achieves $97.6\%$ optimal rate on unseen $30 \times 30$ maze tasks, surpassing the \searchformer baseline ($93.3\%$) trained on data with complete reasoning traces, with $45.5\%$ fewer reasoning steps; (2) in fast mode, \dualformer achieves $80\%$ optimal rate, significantly outperforming the Solution-Only model trained on solution-only data, which has an optimal rate of only $30\%$; (3) in auto mode, \dualformer achieves $96.6\%$ optimal rate with $59.9\%$ fewer steps than \searchformer. Moreover, \dualformer produces more diverse reasoning traces than \searchformer{}. For math reasoning problems, our techniques have also achieved improved performance with LLM fine-tuning, demonstrating its generalization beyond task-specific models. We open source our code at https://github.com/facebookresearch/dualformer.

Hanqing Zhu, Zhenyu Zhang, Hanxian Huang et al.

Reinforcement Learning with Verifiable Rewards (RLVR) reliably improves the reasoning performance of large language models, yet it appears to modify only a small fraction of parameters. We revisit this paradox and show that sparsity is a surface artifact of a model-conditioned optimization bias: for a fixed pretrained model, updates consistently localize to preferred parameter regions, highly consistent across runs and largely invariant to datasets and RL recipes. We mechanistically explain these dynamics with a Three-Gate Theory: Gate I (KL Anchor) imposes a KL-constrained update; Gate II (Model Geometry) steers the step off principal directions into low-curvature, spectrum-preserving subspaces; and Gate III (Precision) hides micro-updates in non-preferred regions, making the off-principal bias appear as sparsity. We then validate this theory and, for the first time, provide a parameter-level characterization of RLVR's learning dynamics: RLVR learns off principal directions in weight space, achieving gains via minimal spectral drift, reduced principal-subspace rotation, and off-principal update alignment. In contrast, SFT targets principal weights, distorts the spectrum, and even lags RLVR. Together, these results provide the first parameter-space account of RLVR's training dynamics, revealing clear regularities in how parameters evolve. Crucially, we show that RL operates in a distinct optimization regime from SFT, so directly adapting SFT-era parameter-efficient fine-tuning (PEFT) methods can be flawed, as evidenced by our case studies on advanced sparse fine-tuning and LoRA variants. We hope this work charts a path toward a white-box understanding of RLVR and the design of geometry-aware, RLVR-native learning algorithms, rather than repurposed SFT-era heuristics.

Yujie Zhao, Boqin Yuan, Junbo Huang et al.

Large Language Models (LLMs) are deployed as autonomous agents in increasingly complex applications, where enabling long-horizon memory is critical for achieving strong performance. However, a significant gap exists between practical applications and current evaluation standards for agent memory: existing benchmarks primarily focus on dialogue-centric, human-agent interactions. In reality, agent memory consists of a continuous stream of agent-environment interactions that are primarily composed of machine-generated representations. To bridge this gap, we introduce AMA-Bench (Agent Memory with Any length), which evaluates long-horizon memory for LLMs in real agentic applications. It features two key components: (1) a set of real-world agentic trajectories across representative agentic applications, paired with expert-curated QA, and (2) a set of synthetic agentic trajectories that scale to arbitrary horizons, paired with rule-based QA. Our comprehensive study shows that existing memory systems underperform on AMA-Bench primarily because they lack causality and objective information and are constrained by the lossy nature of similarity-based retrieval employed by many memory systems. To address these limitations, we propose AMA-Agent, an effective memory system featuring a causality graph and tool-augmented retrieval. Our results demonstrate that AMA-Agent achieves 57.22% average accuracy on AMA-Bench, surpassing the strongest memory system baselines by 11.16%.

Mingchen Zhuge, Changsheng Zhao, Haozhe Liu et al.

We propose a new frontier: Neural Computers (NCs) -- an emerging machine form that unifies computation, memory, and I/O in a learned runtime state. Unlike conventional computers, which execute explicit programs, agents, which act over external execution environments, and world models, which learn environment dynamics, NCs aim to make the model itself the running computer. Our long-term goal is the Completely Neural Computer (CNC): the mature, general-purpose realization of this emerging machine form, with stable execution, explicit reprogramming, and durable capability reuse. As an initial step, we study whether early NC primitives can be learned solely from collected I/O traces, without instrumented program state. Concretely, we instantiate NCs as video models that roll out screen frames from instructions, pixels, and user actions (when available) in CLI and GUI settings. These implementations show that learned runtimes can acquire early interface primitives, especially I/O alignment and short-horizon control, while routine reuse, controlled updates, and symbolic stability remain open. We outline a roadmap toward CNCs around these challenges. If overcome, CNCs could establish a new computing paradigm beyond today's agents, world models, and conventional computers.

Hanlin Zhu, Baihe Huang, Shaolun Zhang et al.

Auto-regressive large language models (LLMs) show impressive capacities to solve many complex reasoning tasks while struggling with some simple logical reasoning tasks such as inverse search: when trained on '$A \to B$' (e.g., 'Tom is the parent of John'), LLM fails to directly conclude '$B \gets A$' (e.g., 'John is the child of Tom') during inference even if the two sentences are semantically identical, which is known as the 'reversal curse'. In this paper, we theoretically analyze the reversal curse via the training dynamics of (stochastic) gradient descent for two auto-regressive models: (1) a bilinear model that can be viewed as a simplification of a one-layer transformer; (2) one-layer transformers under certain assumptions. Our analysis reveals that for both models, the reversal curse is a consequence of the (effective) model weights 'asymmetry', i.e., the increase of weights from a token $A$ to token $B$ during training does not necessarily cause the increase of the weights from $B$ to $A$, which is caused by the training dynamics under certain choice of loss function and the optimization space of model parameters. Moreover, our analysis can be naturally applied to other logical reasoning tasks such as chain-of-thought (COT), which provides a new perspective different from previous work that focuses on expressivity. Finally, we conduct experiments to validate our theory on multi-layer transformers under different settings. Our code is available at https://github.com/marlo-z/reversal_curse_analysis/.

Yixin Liu, Yue Yu, DiJia Su et al.

Reasoning LLMs-as-Judges, which can benefit from inference-time scaling, provide a promising path for extending the success of reasoning models to non-verifiable domains where the output correctness/quality cannot be directly checked. However, while reasoning judges have shown better performance on static evaluation benchmarks, their effectiveness in actual policy training has not been systematically examined. Therefore, we conduct a rigorous study to investigate the actual impact of non-reasoning and reasoning judges in reinforcement-learning-based LLM alignment. Our controlled synthetic setting, where a "gold-standard" judge (gpt-oss-120b) provides preference annotations to train smaller judges, reveals key differences between non-reasoning and reasoning judges: non-reasoning judges lead to reward hacking easily, while reasoning judges can lead to policies that achieve strong performance when evaluated by the gold-standard judge. Interestingly, we find that the reasoning-judge-trained policies achieve such strong performance by learning to generate highly effective adversarial outputs that can also score well on popular benchmarks such as Arena-Hard by deceiving other LLM-judges. Combined with our further analysis, our study highlights both important findings and room for improvements for applying (reasoning) LLM-judges in non-verifiable LLM post-training.

Jiawei Zhao, Zhenyu Zhang, Beidi Chen et al.

Training Large Language Models (LLMs) presents significant memory challenges, predominantly due to the growing size of weights and optimizer states. Common memory-reduction approaches, such as low-rank adaptation (LoRA), add a trainable low-rank matrix to the frozen pre-trained weight in each layer, reducing trainable parameters and optimizer states. However, such approaches typically underperform training with full-rank weights in both pre-training and fine-tuning stages since they limit the parameter search to a low-rank subspace and alter the training dynamics, and further, may require full-rank warm start. In this work, we propose Gradient Low-Rank Projection (GaLore), a training strategy that allows full-parameter learning but is more memory-efficient than common low-rank adaptation methods such as LoRA. Our approach reduces memory usage by up to 65.5% in optimizer states while maintaining both efficiency and performance for pre-training on LLaMA 1B and 7B architectures with C4 dataset with up to 19.7B tokens, and on fine-tuning RoBERTa on GLUE tasks. Our 8-bit GaLore further reduces optimizer memory by up to 82.5% and total training memory by 63.3%, compared to a BF16 baseline. Notably, we demonstrate, for the first time, the feasibility of pre-training a 7B model on consumer GPUs with 24GB memory (e.g., NVIDIA RTX 4090) without model parallel, checkpointing, or offloading strategies.

Jian Xie, Kai Zhang, Jiangjie Chen et al. · microsoft-research

Planning has been part of the core pursuit for artificial intelligence since its conception, but earlier AI agents mostly focused on constrained settings because many of the cognitive substrates necessary for human-level planning have been lacking. Recently, language agents powered by large language models (LLMs) have shown interesting capabilities such as tool use and reasoning. Are these language agents capable of planning in more complex settings that are out of the reach of prior AI agents? To advance this investigation, we propose TravelPlanner, a new planning benchmark that focuses on travel planning, a common real-world planning scenario. It provides a rich sandbox environment, various tools for accessing nearly four million data records, and 1,225 meticulously curated planning intents and reference plans. Comprehensive evaluations show that the current language agents are not yet capable of handling such complex planning tasks-even GPT-4 only achieves a success rate of 0.6%. Language agents struggle to stay on task, use the right tools to collect information, or keep track of multiple constraints. However, we note that the mere possibility for language agents to tackle such a complex problem is in itself non-trivial progress. TravelPlanner provides a challenging yet meaningful testbed for future language agents.

Shibo Hao, Sainbayar Sukhbaatar, DiJia Su et al. · meta-ai

Large language models (LLMs) are typically constrained to reason in the language space, where they express the reasoning process through a chain-of-thought (CoT) to solve complex problems. However, the language space may not always be optimal for reasoning. Most word tokens primarily ensure textual coherence and are not essential for reasoning, while some critical tokens require complex planning and pose challenges to LLMs. To explore the potential of reasoning beyond language, we introduce a new paradigm called Coconut (Chain of Continuous Thought). Coconut utilizes the last hidden state of the LLM as a representation of the reasoning state, termed "continuous thought." Instead of decoding this state into words, we feed it back to the model as the next input embedding directly in the continuous space. This latent reasoning paradigm enables an advanced reasoning pattern, where continuous thoughts can encode multiple alternative next steps, allowing the model to perform a breadth-first search (BFS) rather than committing prematurely to a single deterministic path as in CoT. Coconut outperforms CoT on logical reasoning tasks that require substantial search during planning and achieves a better trade-off between accuracy and efficiency.

Changsheng Zhao, Ernie Chang, Zechun Liu et al.

The paradigm shift in large language models (LLMs) from instinctive responses to chain-of-thought (CoT) reasoning has fueled two prevailing assumptions: (1) reasoning capabilities only emerge in sufficiently large models, and (2) such capabilities require training on massive datasets. While the first assumption has already been challenged by recent sub-billion-parameter reasoning models such as Qwen3-0.6B and DeepSeek distilled variants, the second remains largely unquestioned. In this work, we revisit the necessity of scaling to extremely large corpora (>10T tokens) for reasoning emergence. By carefully curating and resampling open-source datasets that we identify as beneficial under our designed metrics, we demonstrate that strong reasoning abilities can emerge with far less data. Specifically, we show that only ~2T tokens of high-quality data are sufficient, and pre-training with 4.2T tokens on the dataset resampled from these ~2T tokens, followed by a established post-training procedure, enables the development of MobileLLM-R1, a series of sub-billion-parameter reasoning models that substantially outperform prior models trained on fully open-sourced data. For example, MobileLLM-R1-950M achieves an AIME score of 15.5, compared to just 0.6 for OLMo-2-1.48B and 0.3 for SmolLM-2-1.7B. Remarkably, despite being trained on only 11.7% of the tokens compared to Qwen3's proprietary 36T-token corpus for pretraining, MobileLLM-R1-950M matches or surpasses Qwen3-0.6B across multiple reasoning benchmarks. To facilitate further research in this direction, we have released the complete training recipe, data sources, data mixing ratio, and model checkpoints, together with the key insights obtained throughout this study.

Zechun Liu, Changsheng Zhao, Forrest Iandola et al.

This paper addresses the growing need for efficient large language models (LLMs) on mobile devices, driven by increasing cloud costs and latency concerns. We focus on designing top-quality LLMs with fewer than a billion parameters, a practical choice for mobile deployment. Contrary to prevailing belief emphasizing the pivotal role of data and parameter quantity in determining model quality, our investigation underscores the significance of model architecture for sub-billion scale LLMs. Leveraging deep and thin architectures, coupled with embedding sharing and grouped-query attention mechanisms, we establish a strong baseline network denoted as MobileLLM, which attains a remarkable 2.7%/4.3% accuracy boost over preceding 125M/350M state-of-the-art models. Additionally, we propose an immediate block-wise weight-sharing approach with no increase in model size and only marginal latency overhead. The resultant models, denoted as MobileLLM-LS, demonstrate a further accuracy enhancement of 0.7%/0.8% than MobileLLM 125M/350M. Moreover, MobileLLM model family shows significant improvements compared to previous sub-billion models on chat benchmarks, and demonstrates close correctness to LLaMA-v2 7B in API calling tasks, highlighting the capability of small models for common on-device use cases.

Ranajoy Sadhukhan, Sheng Cao, Harry Dong et al.

Fine-grained sparsity promises higher parametric capacity without proportional per-token compute, but often suffers from training instability, load balancing, and communication overhead. We introduce STEM (Scaling Transformers with Embedding Modules), a static, token-indexed approach that replaces the FFN up-projection with a layer-local embedding lookup while keeping the gate and down-projection dense. This removes runtime routing, enables CPU offload with asynchronous prefetch, and decouples capacity from both per-token FLOPs and cross-device communication. Empirically, STEM trains stably despite extreme sparsity. It improves downstream performance over dense baselines while reducing per-token FLOPs and parameter accesses (eliminating roughly one-third of FFN parameters). STEM learns embedding spaces with large angular spread which enhances its knowledge storage capacity. More interestingly, this enhanced knowledge capacity comes with better interpretability. The token-indexed nature of STEM embeddings allows simple ways to perform knowledge editing and knowledge injection in an interpretable manner without any intervention in the input text or additional computation. In addition, STEM strengthens long-context performance: as sequence length grows, more distinct parameters are activated, yielding practical test-time capacity scaling. Across 350M and 1B model scales, STEM delivers up to ~3--4% accuracy improvements overall, with notable gains on knowledge and reasoning-heavy benchmarks (ARC-Challenge, OpenBookQA, GSM8K, MMLU). Overall, STEM is an effective way of scaling parametric memory while providing better interpretability, better training stability and improved efficiency.

Daochen Zha, Louis Feng, Liang Luo et al.

Sharding a large machine learning model across multiple devices to balance the costs is important in distributed training. This is challenging because partitioning is NP-hard, and estimating the costs accurately and efficiently is difficult. In this work, we explore a "pre-train, and search" paradigm for efficient sharding. The idea is to pre-train a universal and once-for-all neural network to predict the costs of all the possible shards, which serves as an efficient sharding simulator. Built upon this pre-trained cost model, we then perform an online search to identify the best sharding plans given any specific sharding task. We instantiate this idea in deep learning recommendation models (DLRMs) and propose NeuroShard for embedding table sharding. NeuroShard pre-trains neural cost models on augmented tables to cover various sharding scenarios. Then it identifies the best column-wise and table-wise sharding plans with beam search and greedy grid search, respectively. Experiments show that NeuroShard significantly and consistently outperforms the state-of-the-art on the benchmark sharding dataset, achieving up to 23.8% improvement. When deployed in an ultra-large production DLRM with multi-terabyte embedding tables, NeuroShard achieves 11.6% improvement in embedding costs over the state-of-the-art, which translates to 6.6% end-to-end training throughput improvement. To facilitate future research of the "pre-train, and search" paradigm in ML for Systems, we open-source our code at https://github.com/daochenzha/neuroshard

Kevin Yang, Tianjun Zhang, Chris Cummins et al.

Path planning, the problem of efficiently discovering high-reward trajectories, often requires optimizing a high-dimensional and multimodal reward function. Popular approaches like CEM and CMA-ES greedily focus on promising regions of the search space and may get trapped in local maxima. DOO and VOOT balance exploration and exploitation, but use space partitioning strategies independent of the reward function to be optimized. Recently, LaMCTS empirically learns to partition the search space in a reward-sensitive manner for black-box optimization. In this paper, we develop a novel formal regret analysis for when and why such an adaptive region partitioning scheme works. We also propose a new path planning method LaP3 which improves the function value estimation within each sub-region, and uses a latent representation of the search space. Empirically, LaP3 outperforms existing path planning methods in 2D navigation tasks, especially in the presence of difficult-to-escape local optima, and shows benefits when plugged into the planning components of model-based RL such as PETS. These gains transfer to highly multimodal real-world tasks, where we outperform strong baselines in compiler phase ordering by up to 39% on average across 9 tasks, and in molecular design by up to 0.4 on properties on a 0-1 scale. Code is available at https://github.com/yangkevin2/neurips2021-lap3.

Tianjun Zhang, Paria Rashidinejad, Jiantao Jiao et al.

In online reinforcement learning (RL), efficient exploration remains particularly challenging in high-dimensional environments with sparse rewards. In low-dimensional environments, where tabular parameterization is possible, count-based upper confidence bound (UCB) exploration methods achieve minimax near-optimal rates. However, it remains unclear how to efficiently implement UCB in realistic RL tasks that involve non-linear function approximation. To address this, we propose a new exploration approach via \textit{maximizing} the deviation of the occupancy of the next policy from the explored regions. We add this term as an adaptive regularizer to the standard RL objective to balance exploration vs. exploitation. We pair the new objective with a provably convergent algorithm, giving rise to a new intrinsic reward that adjusts existing bonuses. The proposed intrinsic reward is easy to implement and combine with other existing RL algorithms to conduct exploration. As a proof of concept, we evaluate the new intrinsic reward on tabular examples across a variety of model-based and model-free algorithms, showing improvements over count-only exploration strategies. When tested on navigation and locomotion tasks from MiniGrid and DeepMind Control Suite benchmarks, our approach significantly improves sample efficiency over state-of-the-art methods. Our code is available at https://github.com/tianjunz/MADE.