Yichao Fu

Yichao Fu, Siqi Zhu, Runlong Su et al.

In Large Language Model (LLM) inference, the output length of an LLM request is typically regarded as not known a priori. Consequently, most LLM serving systems employ a simple First-come-first-serve (FCFS) scheduling strategy, leading to Head-Of-Line (HOL) blocking and reduced throughput and service quality. In this paper, we reexamine this assumption -- we show that, although predicting the exact generation length of each request is infeasible, it is possible to predict the relative ranks of output lengths in a batch of requests, using learning to rank. The ranking information offers valuable guidance for scheduling requests. Building on this insight, we develop a novel scheduler for LLM inference and serving that can approximate the shortest-job-first (SJF) schedule better than existing approaches. We integrate this scheduler with the state-of-the-art LLM serving system and show significant performance improvement in several important applications: 2.8x lower latency in chatbot serving and 6.5x higher throughput in synthetic data generation. Our code is available at https://github.com/hao-ai-lab/vllm-ltr.git

Gabriele Oliaro, Yichao Fu, May Jiang et al.

LLM-based agents for GPU kernel generation are advancing rapidly, yet their progress is fundamentally constrained by the benchmarks they optimize against. Existing benchmarks are poorly aligned with production inference frameworks: they evaluate kernels on a single GPU with synthetic inputs, ignore the surrounding compilation stack, and reward replicating known optimizations rather than discovering new ones. The resulting reward signals are misleading: agents learn to generate kernels that score well in sandboxes but introduce interface incompatibilities, compilation-stack conflicts, and silent correctness degradation when integrated into real systems. We introduce FastKernels, a kernel benchmark built around a minimal set of 46 representative architectures spanning 8 categories, whose kernels collectively subsume those of 96.2% (409/425) of HuggingFace Transformers architectures. FastKernels doubles as a minimalistic, production-grade inference framework that runs at parity with hardened systems such as vLLM and SGLang on mainstream LLM serving and substantially exceeds upstream references on under-served architectures; each task's interface mirrors the corresponding module in the state-of-the-art library for its architecture family, enabling direct deployment of optimized kernels into production codebases. Evaluating state-of-the-art kernel agents on FastKernels, we find that even the strongest agent achieves only 0.94$\times$ aggregate speedup over production baselines, with weaker agents at $0.78\times$ and $0.53\times$ -- confirming that benchmark-production misalignment is a critical bottleneck for the field. We release FastKernels as a stepping stone toward kernel agents whose benchmark gains translate directly into production throughput improvements. Code is available at https://github.com/Snowflake-AI-Research/fastkernels

Shangde Gao, Yichao Fu, Ke Liu et al.

Knowledge amalgamation (KA) aims to learn a compact student model to handle the joint objective from multiple teacher models that are are specialized for their own tasks respectively. Current methods focus on coarsely aligning teachers and students in the common representation space, making it difficult for the student to learn the proper decision boundaries from a set of heterogeneous teachers. Besides, the KL divergence in previous works only minimizes the probability distribution difference between teachers and the student, ignoring the intrinsic characteristics of teachers. Therefore, we propose a novel Contrastive Knowledge Amalgamation (CKA) framework, which introduces contrastive losses and an alignment loss to achieve intra-class cohesion and inter-class separation.Contrastive losses intra- and inter- models are designed to widen the distance between representations of different classes. The alignment loss is introduced to minimize the sample-level distribution differences of teacher-student models in the common representation space.Furthermore, the student learns heterogeneous unsupervised classification tasks through soft targets efficiently and flexibly in the task-level amalgamation. Extensive experiments on benchmarks demonstrate the generalization capability of CKA in the amalgamation of specific task as well as multiple tasks. Comprehensive ablation studies provide a further insight into our CKA.

Lanxiang Hu, Siqi Kou, Yichao Fu et al.

Multi-token generation has emerged as a promising paradigm for accelerating transformer-based large model inference. Recent efforts primarily explore diffusion Large Language Models (dLLMs) for parallel decoding to reduce inference latency. To achieve AR-level generation quality, many techniques adapt AR models into dLLMs to enable parallel decoding. However, they suffer from limited speedup compared to AR models due to a pretrain-to-posttrain mismatch. Specifically, the masked data distribution in post-training deviates significantly from the real-world data distribution seen during pretraining, and dLLMs rely on bidirectional attention, which conflicts with the causal prior learned during pretraining and hinders the integration of exact KV cache reuse. To address this, we introduce Jacobi Forcing, a progressive distillation paradigm where models are trained on their own generated parallel decoding trajectories, smoothly shifting AR models into efficient parallel decoders while preserving their pretrained causal inference property. The models trained under this paradigm, Jacobi Forcing Model, achieves 3.8x wall-clock speedup on coding and math benchmarks with minimal loss in performance. Based on Jacobi Forcing Models' trajectory characteristics, we introduce multi-block decoding with rejection recycling, which enables up to 4.5x higher token acceptance count per iteration and nearly 4.0x wall-clock speedup, effectively trading additional compute for lower inference latency. Our code is available at https://github.com/hao-ai-lab/JacobiForcing.

Yichao Fu, Peter Bailis, Ion Stoica et al.

Autoregressive decoding of large language models (LLMs) is memory bandwidth bounded, resulting in high latency and significant wastes of the parallel processing power of modern accelerators. Existing methods for accelerating LLM decoding often require a draft model (e.g., speculative decoding), which is nontrivial to obtain and unable to generalize. In this paper, we introduce Lookahead decoding, an exact, parallel decoding algorithm that accelerates LLM decoding without needing auxiliary models or data stores. It allows trading per-step log(FLOPs) to reduce the number of total decoding steps, is more parallelizable on single or multiple modern accelerators, and is compatible with concurrent memory-efficient attention (e.g., FlashAttention). Our implementation of Lookahead decoding can speed up autoregressive decoding by up to 1.8x on MT-bench and 4x with strong scaling on multiple GPUs in code completion tasks. Our code is avialable at https://github.com/hao-ai-lab/LookaheadDecoding

Dong Huang, Qingwen Bu, Yahao Qing et al.

Due to the widespread application of deep neural networks~(DNNs) in safety-critical tasks, deep learning testing has drawn increasing attention. During the testing process, test cases that have been fuzzed or selected using test metrics are fed into the model to find fault-inducing test units (e.g., neurons and feature maps, activating which will almost certainly result in a model error) and report them to the DNN developer, who subsequently repair them~(e.g., retraining the model with test cases). Current test metrics, however, are primarily concerned with the neurons, which means that test cases that are discovered either by guided fuzzing or selection with these metrics focus on detecting fault-inducing neurons while failing to detect fault-inducing feature maps. In this work, we propose DeepFeature, which tests DNNs from the feature map level. When testing is conducted, DeepFeature will scrutinize every internal feature map in the model and identify vulnerabilities that can be enhanced through repairing to increase the model's overall performance. Exhaustive experiments are conducted to demonstrate that (1) DeepFeature is a strong tool for detecting the model's vulnerable feature maps; (2) DeepFeature's test case selection has a high fault detection rate and can detect more types of faults~(comparing DeepFeature to coverage-guided selection techniques, the fault detection rate is increased by 49.32\%). (3) DeepFeature's fuzzer also outperforms current fuzzing techniques and generates valuable test cases more efficiently.

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.

Ke Liu, Shangde Gao, Yichao Fu et al.

Medical image segmentation is inherently influenced by data uncertainty, arising from ambiguous boundaries in medical scans and inter-observer variability in diagnosis. To address this challenge, previous works formulated the multi-rater medical image segmentation task, where multiple experts provide separate annotations for each image. However, existing models are typically constrained to either generate diverse segmentation that lacks expert specificity or to produce personalized outputs that merely replicate individual annotators. We propose Probabilistic modeling of multi-rater medical image Segmentation (ProSeg) that simultaneously enables both diversification and personalization. Specifically, we introduce two latent variables to model expert annotation preferences and image boundary ambiguity. Their conditional probabilistic distributions are then obtained through variational inference, allowing segmentation outputs to be generated by sampling from these distributions. Extensive experiments on both the nasopharyngeal carcinoma dataset (NPC) and the lung nodule dataset (LIDC-IDRI) demonstrate that our ProSeg achieves a new state-of-the-art performance, providing segmentation results that are both diverse and expert-personalized. Code can be found in https://github.com/AI4MOL/ProSeg.

Dong Huang, Qingwen Bu, Yichao Fu et al.

Deep Neural Networks~(DNNs) have been widely deployed in software to address various tasks~(e.g., autonomous driving, medical diagnosis). However, they could also produce incorrect behaviors that result in financial losses and even threaten human safety. To reveal the incorrect behaviors in DNN and repair them, DNN developers often collect rich unlabeled datasets from the natural world and label them to test the DNN models. However, properly labeling a large number of unlabeled datasets is a highly expensive and time-consuming task. To address the above-mentioned problem, we propose NSS, Neuron Sensitivity guided test case Selection, which can reduce the labeling time by selecting valuable test cases from unlabeled datasets. NSS leverages the internal neuron's information induced by test cases to select valuable test cases, which have high confidence in causing the model to behave incorrectly. We evaluate NSS with four widely used datasets and four well-designed DNN models compared to SOTA baseline methods. The results show that NSS performs well in assessing the test cases' probability of fault triggering and model improvement capabilities. Specifically, compared with baseline approaches, NSS obtains a higher fault detection rate~(e.g., when selecting 5\% test case from the unlabeled dataset in MNIST \& LeNet1 experiment, NSS can obtain 81.8\% fault detection rate, 20\% higher than baselines).

Wensong Bai, Chao Zhang, Yichao Fu et al.

In this paper, we propose the first fully push-forward-based distributional reinforcement learning algorithm, named PACER, which consists of a distributional critic, a stochastic actor and a sample-based encourager. Specifically, the push-forward operator is leveraged in both the critic and actor to model the return distributions and stochastic policies respectively, enabling them with equal modeling capability and thus enhancing the synergetic performance. Since it is infeasible to obtain the density function of the push-forward policies, novel sample-based regularizers are integrated in the encourager to incentivize efficient exploration and alleviate the risk of trapping into local optima. Moreover, a sample-based stochastic utility value policy gradient is established for the push-forward policy update, which circumvents the explicit demand of the policy density function in existing REINFORCE-based stochastic policy gradient. As a result, PACER fully utilizes the modeling capability of the push-forward operator and is able to explore a broader class of the policy space, compared with limited policy classes used in existing distributional actor critic algorithms (i.e. Gaussians). We validate the critical role of each component in our algorithm with extensive empirical studies. Experimental results demonstrate the superiority of our algorithm over the state-of-the-art.

Yichao Fu, Junda Chen, Siqi Zhu et al.

Test-time reasoning algorithms such as chain-of-thought, self-consistency, and MCTS enhance LLM problem-solving but can wastefully generate many tokens without improving accuracy. At the same time, we observe that these algorithms exhibit answer stabilization: their intermediate solutions often cease to change after a certain point, and further investment of compute does not change their final answer. To quantify this phenomenon, we introduce Certaindex, an algorithm-agnostic metric measuring this evolving stability, signaling when further computation is unlikely to alter the final result. Certaindex is lightweight, can accelerate reasoning program inference via early exit, and further enables dynamic token allocation, gang scheduling, and many opportunities when integrated with real-world LLM serving systems. To quantify real-world benefits, we built Certaindex as a scheduler into Dynasor, our reasoning-aware LLM serving system, and demonstrate up to 50% compute savings and 3.3x higher throughput in real workloads with no accuracy drop. Our code is available at https://github.com/hao-ai-lab/Dynasor.git

Yichao Fu, Rui Ge, Zelei Shao et al.

Reasoning models excel by generating long chain-of-thoughts, but decoding the resulting thousands of tokens is slow. Token-level speculative decoding (SD) helps, but its benefit is capped, because the chance that an entire $γ$-token guess is correct falls exponentially as $γ$ grows. This means allocating more compute for longer token drafts faces an algorithmic ceiling -- making the speedup modest and hardware-agnostic. We raise this ceiling with Lookahead Reasoning, which exploits a second, step-level layer of parallelism. Our key insight is that reasoning models generate step-by-step, and each step needs only to be semantically correct, not exact token matching. In Lookahead Reasoning, a lightweight draft model proposes several future steps; the target model expands each proposal in one batched pass, and a verifier keeps semantically correct steps while letting the target regenerate any that fail. Token-level SD still operates within each reasoning step, so the two layers of parallelism multiply. We show Lookahead Reasoning lifts the peak speedup of SD both theoretically and empirically. Across GSM8K, AIME, and other benchmarks, Lookahead Reasoning improves the speedup of SD from 1.4x to 2.1x while preserving answer quality, and its speedup scales better with additional GPU throughput. Our code is available at https://github.com/hao-ai-lab/LookaheadReasoning

Rui Ge, Yichao Fu, Yuyang Qian et al.

Large language models are increasingly deployed as autonomous agents that must plan, act, and recover from mistakes through long-horizon interaction with environments that provide rich feedback. However, prevailing outcome-driven post-training methods (e.g., RL with verifiable rewards) primarily optimize final success signals, leaving rich environment feedback underutilized. Consequently, they often lead to distribution sharpening: the policy becomes better at reproducing a narrow set of already-successful behaviors, while failing to improve the feedback-grounded agency needed to expand problem-solving capacity (e.g., Pass@k) in long-horizon settings. To address this, we propose LEAFE (Learning Feedback-Grounded Agency from Reflective Experience), a framework that internalizes recovery agency from reflective experience. Specifically, during exploration, the agent summarizes environment feedback into actionable experience, backtracks to earlier decision points, and explores alternative branches with revised actions. We then distill these experience-guided corrections into the model through supervised fine-tuning, enabling the policy to recover more effectively in future interactions. Across a diverse set of interactive coding and agentic tasks under fixed interaction budgets, LEAFE consistently improves Pass@1 over the base model and achieves higher Pass@k than outcome-driven baselines (GRPO) and experience-based methods such as Early Experience, with gains of up to 14% on Pass@128.

Haoran You, Yichao Fu, Zheng Wang et al.

Autoregressive Large Language Models (LLMs) have achieved impressive performance in language tasks but face two significant bottlenecks: (1) quadratic complexity in the attention module as the number of tokens increases, and (2) limited efficiency due to the sequential processing nature of autoregressive LLMs during generation. While linear attention and speculative decoding offer potential solutions, their applicability and synergistic potential for enhancing autoregressive LLMs remain uncertain. We conduct the first comprehensive study on the efficacy of existing linear attention methods for autoregressive LLMs, integrating them with speculative decoding. We introduce an augmentation technique for linear attention that ensures compatibility with speculative decoding, enabling more efficient training and serving of LLMs. Extensive experiments and ablation studies involving seven existing linear attention models and five encoder/decoder-based LLMs consistently validate the effectiveness of our augmented linearized LLMs. Notably, our approach achieves up to a 6.67 reduction in perplexity on the LLaMA model and up to a 2$\times$ speedup during generation compared to prior linear attention methods. Codes and models are available at https://github.com/GATECH-EIC/Linearized-LLM.

Haoran You, Yipin Guo, Yichao Fu et al.

Large language models (LLMs) have shown impressive performance on language tasks but face challenges when deployed on resource-constrained devices due to their extensive parameters and reliance on dense multiplications, resulting in high memory demands and latency bottlenecks. Shift-and-add reparameterization offers a promising solution by replacing costly multiplications with hardware-friendly primitives in both the attention and multi-layer perceptron (MLP) layers of an LLM. However, current reparameterization techniques require training from scratch or full parameter fine-tuning to restore accuracy, which is resource-intensive for LLMs. To address this, we propose accelerating pretrained LLMs through post-training shift-and-add reparameterization, creating efficient multiplication-free models, dubbed ShiftAddLLM. Specifically, we quantize each weight matrix into binary matrices paired with group-wise scaling factors. The associated multiplications are reparameterized into (1) shifts between activations and scaling factors and (2) queries and adds according to the binary matrices. To reduce accuracy loss, we present a multi-objective optimization method to minimize both weight and output activation reparameterization errors. Additionally, based on varying sensitivity across layers to reparameterization, we develop an automated bit allocation strategy to further reduce memory usage and latency. Experiments on five LLM families and eight tasks consistently validate the effectiveness of ShiftAddLLM, achieving average perplexity improvements of 5.6 and 22.7 points at comparable or lower latency compared to the most competitive quantized LLMs at 3 and 2 bits, respectively, and more than 80% memory and energy reductions over the original LLMs. Codes and models are available at https://github.com/GATECH-EIC/ShiftAddLLM.

Shangde Gao, Yichao Fu, Ke Liu et al.

Recently, many foundation models for medical image analysis such as MedSAM, SwinUNETR have been released and proven to be useful in multiple tasks. However, considering the inherent heterogeneity and inhomogeneity of real-world medical data, directly applying these models to specific medical image segmentation tasks often leads to negative domain shift effects, which can severely weaken the model's segmentation capabilities. To this end, we propose an adaptive amalgamation knowledge framework that aims to train a versatile foundation model to handle the joint goals of multiple expert models, each specialized for a distinct task. Specifically, we first train an nnUNet-based expert model for each task, and reuse the pre-trained SwinUNTER as the target foundation model. Then, the input data for all challenging tasks are encoded in the foundation model and the expert models, respectively, and their backbone features are jointly projected into the adaptive amalgamation layer. Within the hidden layer, the hierarchical attention mechanisms are designed to achieve adaptive merging of the target model to the hidden layer feature knowledge of all experts, which significantly reduces the domain shift arising from the inter-task differences. Finally, the gold amalgamated features and the prompt features are fed into the mask decoder to obtain the segmentation results. Extensive experiments conducted in these challenging tasks demonstrate the effectiveness and adaptability of our foundation model for real-world medical image segmentation.

Yu-Yang Qian, Hao-Cong Wu, Yichao Fu et al.

Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides verification feedback that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative "draft commits-feedback provides-draft adapts" evolving loop, which precisely matches the online learning paradigm. Motivated by this connection, we propose OnlineSpec, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in dynamic regret minimization, we establish a formal link between online learning performance and speculative system's acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and three foundation models.