Wentao Guo

Yucheng Lu, Wentao Guo, Christopher De Sa

Random reshuffling, which randomly permutes the dataset each epoch, is widely adopted in model training because it yields faster convergence than with-replacement sampling. Recent studies indicate greedily chosen data orderings can further speed up convergence empirically, at the cost of using more computation and memory. However, greedy ordering lacks theoretical justification and has limited utility due to its non-trivial memory and computation overhead. In this paper, we first formulate an example-ordering framework named herding and answer affirmatively that SGD with herding converges at the rate $O(T^{-2/3})$ on smooth, non-convex objectives, faster than the $O(n^{1/3}T^{-2/3})$ obtained by random reshuffling, where $n$ denotes the number of data points and $T$ denotes the total number of iterations. To reduce the memory overhead, we leverage discrepancy minimization theory to propose an online Gradient Balancing algorithm (GraB) that enjoys the same rate as herding, while reducing the memory usage from $O(nd)$ to just $O(d)$ and computation from $O(n^2)$ to $O(n)$, where $d$ denotes the model dimension. We show empirically on applications including MNIST, CIFAR10, WikiText and GLUE that GraB can outperform random reshuffling in terms of both training and validation performance, and even outperform state-of-the-art greedy ordering while reducing memory usage over $100\times$.

Haoran Zhang, Youjin Wang, Yi Duan et al.

World models compress rich sensory streams into compact latent codes that anticipate future observations. We let separate agents acquire such models from distinct viewpoints of the same environment without any parameter sharing or coordination. After training, their internal representations exhibit a striking emergent property: the two latent spaces are related by an approximate linear isometry, enabling transparent translation between them. This geometric consensus survives large viewpoint shifts and scant overlap in raw pixels. Leveraging the learned alignment, a classifier trained on one agent can be ported to the other with no additional gradient steps, while distillation-like migration accelerates later learning and markedly reduces total compute. The findings reveal that predictive learning objectives impose strong regularities on representation geometry, suggesting a lightweight path to interoperability among decentralized vision systems. The code is available at https://anonymous.4open.science/r/Social-JEPA-5C57.

Yann Hicke, Abhishek Masand, Wentao Guo et al. · amazon-science

(Tack et al., 2023) organized the shared task hosted by the 18th Workshop on Innovative Use of NLP for Building Educational Applications on generation of teacher language in educational dialogues. Following the structure of the shared task, in this study, we attempt to assess the generative abilities of large language models in providing informative and helpful insights to students, thereby simulating the role of a knowledgeable teacher. To this end, we present an extensive evaluation of several benchmarking generative models, including GPT-4 (few-shot, in-context learning), fine-tuned GPT-2, and fine-tuned DialoGPT. Additionally, to optimize for pedagogical quality, we fine-tuned the Flan-T5 model using reinforcement learning. Our experimental findings on the Teacher-Student Chatroom Corpus subset indicate the efficacy of GPT-4 over other fine-tuned models, measured using BERTScore and DialogRPT. We hypothesize that several dataset characteristics, including sampling, representativeness, and dialog completeness, pose significant challenges to fine-tuning, thus contributing to the poor generalizability of the fine-tuned models. Finally, we note the need for these generative models to be evaluated with a metric that relies not only on dialog coherence and matched language modeling distribution but also on the model's ability to showcase pedagogical skills.

Pei Yang, Wanyi Chen, Tongyun Yang et al.

LLM routing matters most in long-horizon applications such as coding agents, deep research systems, and computer-use agents, where a single user request triggers many model calls. Routing each call to the cheapest sufficient model can cut costs without sacrificing quality, yet existing router benchmarks evaluate routers only on one-shot prompts. They never expose the router-visible prefix at an intermediate agent step, never test whether a cheaper replacement preserves downstream task success, and often rely on online LLM judges at evaluation time. We introduce TwinRouterBench, a step-level routing benchmark with two tracks. The static track provides 970 router-visible prefixes from 520 instances across SWE-bench, BFCL, mtRAG, QMSum, and PinchBench, each paired with an execution-verified target tier estimated under a released downgrade-and-cascade protocol; scoring is deterministic arithmetic over tier labels, trajectory membership, and token costs, with no online evaluator-side LLM judge. The dynamic track supplies a harness that runs routers on the full 500-case SWE-bench Verified suite; in this paper we report a 100-case held-out evaluation disjoint from the static SWE supervision split. At each LLM call the router selects a concrete model from a locked pool, and success is measured by official task resolution and realized API spend. The two tracks support fast offline iteration followed by end-to-end validation under live agent execution. Code and data are available at https://github.com/CommonstackAI/TwinRouterBench.

Lin Yao, Wentao Guo, Zhen Wang et al.

Single-step retrosynthesis (SSR) in organic chemistry is increasingly benefiting from deep learning (DL) techniques in computer-aided synthesis design. While template-free DL models are flexible and promising for retrosynthesis prediction, they often ignore vital 2D molecular information and struggle with atom alignment for node generation, resulting in lower performance compared to the template-based and semi-template-based methods. To address these issues, we introduce Node-Aligned Graph-to-Graph (NAG2G), a transformer-based template-free DL model. NAG2G combines 2D molecular graphs and 3D conformations to retain comprehensive molecular details and incorporates product-reactant atom mapping through node alignment which determines the order of the node-by-node graph outputs process in an auto-regressive manner. Through rigorous benchmarking and detailed case studies, we have demonstrated that NAG2G stands out with its remarkable predictive accuracy on the expansive datasets of USPTO-50k and USPTO-FULL. Moreover, the model's practical utility is underscored by its successful prediction of synthesis pathways for multiple drug candidate molecules. This not only proves NAG2G's robustness but also its potential to revolutionize the prediction of complex chemical synthesis processes for future synthetic route design tasks.

A. Feder Cooper, Wentao Guo, Khiem Pham et al.

Recent research on online Gradient Balancing (GraB) has revealed that there exist permutation-based example orderings for SGD that are guaranteed to outperform random reshuffling (RR). Whereas RR arbitrarily permutes training examples, GraB leverages stale gradients from prior epochs to order examples -- achieving a provably faster convergence rate than RR. However, GraB is limited by design: while it demonstrates an impressive ability to scale-up training on centralized data, it does not naturally extend to modern distributed ML workloads. We therefore propose Coordinated Distributed GraB (CD-GraB), which uses insights from prior work on kernel thinning to translate the benefits of provably faster permutation-based example ordering to distributed settings. With negligible overhead, CD-GraB exhibits a linear speedup in convergence rate over centralized GraB and outperforms distributed RR on a variety of benchmark tasks.

Wentao Guo, Mayank Mishra, Xinle Cheng et al.

Mixture of Experts (MoE) models have emerged as the de facto architecture for scaling up language models without significantly increasing the computational cost. Recent MoE models demonstrate a clear trend towards high expert granularity (smaller expert intermediate dimension) and higher sparsity (constant number of activated experts with higher number of total experts), which improve model quality per FLOP. However, fine-grained MoEs suffer from increased activation memory footprint and reduced hardware efficiency due to higher IO costs, while sparser MoEs suffer from wasted computations due to padding in Grouped GEMM kernels. In response, we propose a memory-efficient algorithm to compute the forward and backward passes of MoEs with minimal activation caching for the backward pass. We also design GPU kernels that overlap memory IO with computation benefiting all MoE architectures. Finally, we propose a novel "token rounding" method that minimizes the wasted compute due to padding in Grouped GEMM kernels. As a result, our method SonicMoE reduces activation memory by 45% and achieves a 1.86x compute throughput improvement on Hopper GPUs compared to ScatterMoE's BF16 MoE kernel for a fine-grained 7B MoE. Concretely, SonicMoE on 64 H100s achieves a training throughput of 213 billion tokens per day comparable to ScatterMoE's 225 billion tokens per day on 96 H100s for a 7B MoE model training with FSDP-2 using the lm-engine codebase. Under high MoE sparsity settings, our tile-aware token rounding algorithm yields an additional 1.16x speedup on kernel execution time compared to vanilla top-$K$ routing while maintaining similar downstream performance. We open-source all our kernels to enable faster MoE model training.

Tao Yu, Wentao Guo, Jianan Canal Li et al.

In this paper, we introduce MCTensor, a library based on PyTorch for providing general-purpose and high-precision arithmetic for DL training. MCTensor is used in the same way as PyTorch Tensor: we implement multiple basic, matrix-level computation operators and NN modules for MCTensor with identical PyTorch interface. Our algorithms achieve high precision computation and also benefits from heavily-optimized PyTorch floating-point arithmetic. We evaluate MCTensor arithmetic against PyTorch native arithmetic for a series of tasks, where models using MCTensor in float16 would match or outperform the PyTorch model with float32 or float64 precision.

Shang Xiang, Lin Yao, Zhen Wang et al.

The field of computer-aided synthesis planning (CASP) has seen rapid advancements in recent years, achieving significant progress across various algorithmic benchmarks. However, chemists often encounter numerous infeasible reactions when using CASP in practice. This article delves into common errors associated with CASP and introduces a product prediction model aimed at enhancing the accuracy of single-step models. While the product prediction model reduces the number of single-step reactions, it integrates multiple single-step models to maintain the overall reaction count and increase reaction diversity. Based on manual analysis and large-scale testing, the product prediction model, combined with the multi-model ensemble approach, has been proven to offer higher feasibility and greater diversity.

Blake Gella, Wei Wu, Yuhao Yin et al.

Recommender systems trained on user interaction data are susceptible to behavioral intensity imbalance--a systematic distortion arising from heterogeneous engagement patterns across users. This imbalance skews feedback signals such that observed interactions no longer faithfully reflect true preferences, causing models to disproportionately amplify signals from highly active users while underrepresenting others, which ultimately degrades recommendation quality and robustness at scale. To address this issue, we propose a nonparametric contrastive percentile approximation framework, PEARL, that models relative preference signals instead of absolute engagement magnitudes. Building upon relative advantage debiasing, PEARL leverages real contrastive interaction samples to approximate percentile relationships directly, without relying on auxiliary distribution estimation models. We provide theoretical justification demonstrating that such pairwise comparisons yield unbiased estimates of percentile-based preference signals. For broader applicability, we introduce a prediction-based bootstrapping mechanism for percentile smoothing to handle sparse and discrete feedback, alongside a generalized value-weighted formulation and a co-training strategy to enhance both modeling flexibility and representation learning. Extensive offline experiments demonstrate that PEARL effectively mitigates behavioral bias and consistently improves recommendation performance across multiple ranking targets. Deployed in a production livestream platform with a combined user base of billions, online A/B testing confirms substantial real-world gains: +2.10% Watch Duration, +0.80% Consumption Amount, +1.49% Interaction Rate, and -6.91% Report Rate.

Xinhang Yuan, Zexi Huang, Anjia Cao et al.

Modern recommender systems rely heavily on ID-based collaborative filtering: each item is represented by a unique ID embedding that accumulates collaborative signals from user interactions. Livestreaming recommendation, however, faces a unique challenge in this paradigm: a live room typically broadcasts for only tens of minutes, so its item ID remains poorly learned in a persistent cold-start state and ID-centric ranking models fail to generalize. We present FLUID, the first framework to fully retire the candidate-side item ID from a production-scale livestreaming ranker. FLUID couples a cross-domain multimodal encoder, jointly trained on short videos and livestreams to produce discrete hierarchical codes (LUCID), with a late-fusion, ID-free design that injects slice-level and room-level LUCID as independent tokens, stabilized by a staged warmup under online incremental training. Deployed on our industrial livestreaming recommenders with a cross-platform combined user base of over one billion globally, FLUID delivers significant online gains of +0.55% Quality Watch Duration, +2.05% Cold-Start Room Views, and +0.05% Active Hours.

Bob Junyi Zou, Sai Li, Tianyun Sun et al.

A fundamental challenge in recommender systems is balancing reliability for Low-Active Users (LAUs) with diversity for High-Active Users (HAUs). The key to this balance lies in quantifying model uncertainty, which approximates the risk of prediction errors and reveals the limits of the model's current knowledge. On large-scale short-video and livestream platforms, model uncertainty can warn of low-quality recommendations that may lead to disengagement of LAUs and at the same time identify opportunities to diversify content recommendation for HAUs. To leverage this dichotomy, we introduce a unified, production-ready framework that calibrates uncertainty to drive differentiated strategies. Specifically, we implement a model-uncertainty-based risk-averse deboosting policy for LAUs to suppress unreliable recommendations, while employing a risk-seeking Upper Confidence Bound (UCB) strategy for HAUs to encourage exploration. Validated on a major livestream platform, our framework demonstrates significant improvements in retention (active hours) and satisfaction (quality watch time ratio) for LAUs as well as remarkable increases in interest diversity and category coverage for HAUs, proving the value of uncertainty-aware recommendation in industrial settings.

Ruifeng Zhang, Zexi Huang, Zikai Wang et al.

Accurately capturing feature interactions is essential in recommender systems, and recent trends show that scaling up model capacity could be a key driver for next-level predictive performance. While prior work has explored various model architectures to capture multi-granularity feature interactions, relatively little attention has been paid to efficient feature handling and scaling model capacity without incurring excessive inference latency. In this paper, we address this by presenting Zenith, a scalable and efficient ranking architecture that learns complex feature interactions with minimal runtime overhead. Zenith is designed to handle a few high-dimensional Prime Tokens with Token Fusion and Token Boost modules, which exhibits superior scaling laws compared to other state-of-the-art ranking methods, thanks to its improved token heterogeneity. Its real-world effectiveness is demonstrated by deploying the architecture to TikTok Live, a leading online livestreaming platform that attracts billions of users globally. Our A/B test shows that Zenith achieves +1.05%/-1.10% in online CTR AUC and Logloss, and realizes +9.93% gains in Quality Watch Session / User and +8.11% in Quality Watch Duration / User.

Wentao Guo, Andrew Wang, Bradon Thymes et al.

We introduce the problem of ranking with slot constraints, which can be used to model a wide range of application problems -- from college admission with limited slots for different majors, to composing a stratified cohort of eligible participants in a medical trial. We show that the conventional Probability Ranking Principle (PRP) can be highly sub-optimal for slot-constrained ranking problems, and we devise a new ranking algorithm, called MatchRank. The goal of MatchRank is to produce rankings that maximize the number of filled slots if candidates are evaluated by a human decision maker in the order of the ranking. In this way, MatchRank generalizes the PRP, and it subsumes the PRP as a special case when there are no slot constraints. Our theoretical analysis shows that MatchRank has a strong approximation guarantee without any independence assumptions between slots or candidates. Furthermore, we show how MatchRank can be implemented efficiently. Beyond the theoretical guarantees, empirical evaluations show that MatchRank can provide substantial improvements over a range of synthetic and real-world tasks.

Bohan Lyu, Yucheng Yang, Siqiao Huang et al.

Modern AI progress has been driven by ML methods that are generalizable across settings and scalable to larger regimes. As large language models demonstrate advanced capabilities in reasoning, coding, and engineering tasks, it is increasingly important to understand whether they can discover such methods rather than only apply existing ones. We introduce MLS-Bench, a benchmark for evaluating whether AI systems can invent generalizable and scalable ML methods. MLS-Bench contains 140 tasks across 12 domains, each requiring an agent to improve one targeted component of an ML system or algorithm and demonstrate that the improvement generalizes across controlled settings and scales. We find that current agents remain far from reliably surpassing human-designed methods, and that engineering-style tuning is easier for them than genuine method invention. We further study the effects of test-time scaling, adaptive compute allocation, and context provision on agents' discovery performance, together with case studies of their behavior. Our analyses suggest that the bottleneck is not only in proposing new methods, but also in the scientific insight needed to plan, validate, and scale claims about them. More search, compute, or context alone does not remove this bottleneck. We build and maintain a community platform for cumulative and comparable iteration, and release the data and code at https://mls-bench.com.

Jianan Canal Li, Yimeng Zeng, Wentao Guo

We propose cKAM, cyclical Kernel Adaptive Metropolis, which incorporates a cyclical stepsize scheme to allow control for exploration and sampling. We show that on a crafted bimodal distribution, existing Adaptive Metropolis type algorithms would fail to converge to the true posterior distribution. We point out that this is because adaptive samplers estimates the local/global covariance structure using past history of the chain, which will lead to adaptive algorithms be trapped in a local mode. We demonstrate that cKAM encourages exploration of the posterior distribution and allows the sampler to escape from a local mode, while maintaining the high performance of adaptive methods.

Yitian Wang, Fanmeng Wang, Angxiao Yue et al.

Modeling peptide cyclization is critical for the virtual screening of candidate peptides with desirable physical and pharmaceutical properties. This task is challenging because a cyclic peptide often exhibits diverse, ring-shaped conformations, which cannot be well captured by deterministic prediction models derived from linear peptide folding. In this study, we propose MuCO (Multi-stage Conformation Optimization), a generative peptide cyclization method that models the distribution of cyclic peptide conformations conditioned on the corresponding linear peptide. In principle, MuCO decouples the peptide cyclization task into three stages: topology-aware backbone design, generative side-chain packing, and physics-aware all-atom optimization, thereby generating and optimizing conformations of cyclic peptides in a coarse-to-fine manner. This multi-stage framework enables an efficient parallel sampling strategy for conformation generation and allows for rapid exploration of diverse, low-energy conformations. Experiments on the large-scale CPSea dataset demonstrate that MuCO significantly outperforms state-of-the-art methods in consistently in physical stability, structural diversity, secondary structure recovery, and computational efficiency, making it a promising computational tool for exploring and designing cyclic peptides.

Wentao Guo, Charlie Ruan, Claire Zhou

Photo Rater is a computer vision project that uses neural networks to help photographers select the best photo among those that are taken based on the same scene. This process is usually referred to as "culling" in photography, and it can be tedious and time-consuming if done manually. Photo Rater utilizes three separate neural networks to complete such a task: one for general image quality assessment, one for classifying whether the photo is blurry (either due to unsteady hands or out-of-focusness), and one for assessing general aesthetics (including the composition of the photo, among others). After feeding the image through each neural network, Photo Rater outputs a final score for each image, ranking them based on this score and presenting it to the user.

Fanmeng Wang, Wentao Guo, Qi Ou et al.

Polymer conformation generation is a critical task that enables atomic-level studies of diverse polymer materials. While significant advances have been made in designing conformation generation methods for small molecules and proteins, these methods struggle to generate polymer conformations due to their unique structural characteristics. Meanwhile, the scarcity of polymer conformation datasets further limits the progress, making this important area largely unexplored. In this work, we propose PolyConf, a pioneering tailored polymer conformation generation method that leverages hierarchical generative models to unlock new possibilities. Specifically, we decompose the polymer conformation into a series of local conformations (i.e., the conformations of its repeating units), generating these local conformations through an autoregressive model, and then generating their orientation transformations via a diffusion model to assemble them into the complete polymer conformation. Moreover, we develop the first benchmark with a high-quality polymer conformation dataset derived from molecular dynamics simulations to boost related research in this area. The comprehensive evaluation demonstrates that PolyConf consistently outperforms existing conformation generation methods, thus facilitating advancements in polymer modeling and simulation.

Guojiang Zhao, Sihang Li, Zixiang Lu et al.

Large Language Models(LLMs) have demonstrated remarkable performance across various domains, yet their capabilities in molecular reasoning remain insufficiently explored. Current approaches tend to rely heavily on general-purpose prompting, which lacks domain-specific molecular semantics, while those that use fine-tuning strategies often face challenges with interpretability and reasoning depth. To address these issues, we introduce MolReasoner, a two-stage framework designed to transition LLMs from memorization towards chemical reasoning. First, we propose Mol-SFT, which initializes the model's reasoning abilities via synthetic Chain-of-Thought(CoT) samples generated by GPT-4o and verified for chemical accuracy. Subsequently, Mol-RL applies reinforcement learning with specialized reward functions designed explicitly to align chemical structures with linguistic descriptions, thereby enhancing molecular reasoning capabilities. Our approach notably enhances interpretability, improving the model 's molecular understanding and enabling better generalization. Extensive experiments demonstrate that MolReasoner outperforms existing methods, and marking a significant shift from memorization-based outputs to robust chemical reasoning.

Fanmeng Wang, Shan Mei, Wentao Guo et al.

Polymers, macromolecules formed from covalently bonded monomers, underpin countless technologies and are indispensable to modern life. While deep learning is advancing polymer science, existing methods typically represent the whole polymer solely through monomer-level descriptors, overlooking the global structural information inherent in polymer conformations, which ultimately limits their practical performance. Moreover, this field still lacks a universal foundation model that can effectively support diverse downstream tasks, thereby severely constraining progress. To address these challenges, we introduce PolyConFM, the first polymer foundation model that unifies polymer modeling and design through conformation-centric generative pretraining. Recognizing that each polymer conformation can be decomposed into a sequence of local conformations (i.e., those of its repeating units), we pretrain PolyConFM under the conditional generation paradigm, reconstructing these local conformations via masked autoregressive (MAR) modeling and further generating their orientation transformations to recover the corresponding polymer conformation. Besides, we construct the first high-quality polymer conformation dataset via molecular dynamics simulations to mitigate data sparsity, thereby enabling conformation-centric pretraining. Experiments demonstrate that PolyConFM consistently outperforms representative task-specific methods on diverse downstream tasks, equipping polymer science with a universal and powerful tool.

Yide Ran, Wentao Guo, Jingwei Sun et al.

Federated Learning enables collaborative fine-tuning of Large Language Models (LLMs) across decentralized Non-Independent and Identically Distributed (Non-IID) clients, but such models' massive parameter sizes lead to significant memory and communication challenges. This work introduces Meerkat, a sparse zeroth-order optimization (ZO) method designed for federated LLM fine-tuning. By limiting fine-tuning to a transferable, static, extremely sparse subset of parameters, Meerkat achieves remarkable communication efficiency, enabling cost-effective high-frequency synchronization. With theoretical analysis and experiments, we show that this high-frequency communication effectively mitigates Non-IID data challenges and leads to superior performance compared to full-parameter ZO. Furthermore, experiment results show that Meerkat outperforms existing sparsity baselines with better performance at the same communication frequency. To further handle Non-IID drift, Meerkat leverages traceable local updates and forms a virtual path for each client. This virtual path mechanism reveals the GradIP phenomenon: the inner products between LLM pre-training gradients maintained by server and client gradients estimated via ZO converges for extreme Non-IID clients but oscillates for IID ones. This distinct behavior provides a signal for identifying clients with extreme data heterogeneity. Using this signal, Meerkat-vp is proposed to analyze GradIP trajectories to identify extreme Non-IID clients and applies early stopping to enhance aggregated model quality. Experiments confirm that Meerkat and Meerkat-vp significantly improve the efficiency and effectiveness of ZO federated LLM fine-tuning.

Yuheng Wu, Wentao Guo, Zirui Liu et al.

This paper investigates the emergence of Theory-of-Mind (ToM) capabilities in large language models (LLMs) from a mechanistic perspective, focusing on the role of extremely sparse parameter patterns. We introduce a novel method to identify ToM-sensitive parameters and reveal that perturbing as little as 0.001% of these parameters significantly degrades ToM performance while also impairing contextual localization and language understanding. To understand this effect, we analyze their interaction with core architectural components of LLMs. Our findings demonstrate that these sensitive parameters are closely linked to the positional encoding module, particularly in models using Rotary Position Embedding (RoPE), where perturbations disrupt dominant-frequency activations critical for contextual processing. Furthermore, we show that perturbing ToM-sensitive parameters affects LLM's attention mechanism by modulating the angle between queries and keys under positional encoding. These insights provide a deeper understanding of how LLMs acquire social reasoning abilities, bridging AI interpretability with cognitive science. Our results have implications for enhancing model alignment, mitigating biases, and improving AI systems designed for human interaction.

Fanmeng Wang, Wentao Guo, Minjie Cheng et al.

Polymers are high-molecular-weight compounds constructed by the covalent bonding of numerous identical or similar monomers so that their 3D structures are complex yet exhibit unignorable regularity. Typically, the properties of a polymer, such as plasticity, conductivity, bio-compatibility, and so on, are highly correlated with its 3D structure. However, existing polymer property prediction methods heavily rely on the information learned from polymer SMILES sequences (P-SMILES strings) while ignoring crucial 3D structural information, resulting in sub-optimal performance. In this work, we propose MMPolymer, a novel multimodal multitask pretraining framework incorporating polymer 1D sequential and 3D structural information to encourage downstream polymer property prediction tasks. Besides, considering the scarcity of polymer 3D data, we further introduce the "Star Substitution" strategy to extract 3D structural information effectively. During pretraining, in addition to predicting masked tokens and recovering clear 3D coordinates, MMPolymer achieves the cross-modal alignment of latent representations. Then we further fine-tune the pretrained MMPolymer for downstream polymer property prediction tasks in the supervised learning paradigm. Experiments show that MMPolymer achieves state-of-the-art performance in downstream property prediction tasks. Moreover, given the pretrained MMPolymer, utilizing merely a single modality in the fine-tuning phase can also outperform existing methods, showcasing the exceptional capability of MMPolymer in polymer feature extraction and utilization.

Wentao Guo, Jikai Long, Yimeng Zeng et al.

Zeroth-order optimization (ZO) is a memory-efficient strategy for fine-tuning Large Language Models using only forward passes. However, the application of ZO fine-tuning in memory-constrained settings such as mobile phones and laptops is still challenging since full precision forward passes are infeasible. In this study, we address this limitation by integrating sparsity and quantization into ZO fine-tuning of LLMs. Specifically, we investigate the feasibility of fine-tuning an extremely small subset of LLM parameters using ZO. This approach allows the majority of un-tuned parameters to be quantized to accommodate the constraint of limited device memory. Our findings reveal that the pre-training process can identify a set of "sensitive parameters" that can guide the ZO fine-tuning of LLMs on downstream tasks. Our results demonstrate that fine-tuning 0.1% sensitive parameters in the LLM with ZO can outperform the full ZO fine-tuning performance, while offering wall-clock time speedup. Additionally, we show that ZO fine-tuning targeting these 0.1% sensitive parameters, combined with 4 bit quantization, enables efficient ZO fine-tuning of an Llama2-7B model on a GPU device with less than 8 GiB of memory and notably reduced latency.

Yan Yan, Wentao Guo, Meng Zhao et al.

With the transition from people's traditional `brick-and-mortar' shopping to online mobile shopping patterns in web 2.0 $\mathit{era}$, the recommender system plays a critical role in E-Commerce and E-Retails. This is especially true when designing this system for more than $\mathbf{236~million}$ daily active users. Ranking strategy, the key module of the recommender system, needs to be precise, accurate, and responsive for estimating customers' intents. We propose a dynamic ranking paradigm, named as DNN-MAB, that is composed of a pairwise deep neural network (DNN) $\mathit{pre}$-ranker connecting a revised multi-armed bandit (MAB) dynamic $\mathit{post}$-ranker. By taking into account of explicit and implicit user feedbacks such as impressions, clicks, conversions, etc. DNN-MAB is able to adjust DNN $\mathit{pre}$-ranking scores to assist customers locating items they are interested in most so that they can converge quickly and frequently. To the best of our knowledge, frameworks like DNN-MAB have not been discussed in the previous literature to either E-Commerce or machine learning audiences. In practice, DNN-MAB has been deployed to production and it easily outperforms against other state-of-the-art models by significantly lifting the gross merchandise volume (GMV) which is the objective metrics at JD.