Many applications today provide users with multiple auto-complete drafts as they type, including GitHub's code completion, Gmail's smart compose, and Apple's messaging auto-suggestions.

Large reasoning models (LRMs) often cost significant key-value (KV) cache overhead, due to their linear growth with the verbose chain-of-thought (CoT) reasoning process. This costs both memory and throughput bottleneck limiting their efficient deployment. Towards reducing KV cache size during inference, we first investigate the effectiveness of existing KV cache eviction methods for CoT reasoning. Interestingly, we find that due to unstable token-wise scoring and the reduced effective KV budget caused by padding tokens, state-of-the-art (SoTA) eviction methods fail to maintain accuracy in the multi-batch setting. Additionally, these methods often generate longer sequences than the original model, as semantic-unaware token-wise eviction leads to repeated revalidation during reasoning. To address these issues, we present \textbf{SkipKV}, a \textbf{\textit{training-free}} KV compression method for selective \textit{eviction} and \textit{generation} operating at a coarse-grained sentence-level sequence removal for efficient CoT reasoning. In specific, it introduces a \textit{sentence-scoring metric} to identify and remove highly similar sentences while maintaining semantic coherence. To suppress redundant generation, SkipKV dynamically adjusts a steering vector to update the hidden activation states during inference enforcing the LRM to generate concise response. Extensive evaluations on multiple reasoning benchmarks demonstrate the effectiveness of SkipKV in maintaining up to $\mathbf{26.7}\%$ improved accuracy compared to the alternatives, at a similar compression budget. Additionally, compared to SoTA, SkipKV yields up to $\mathbf{1.6}\times$ fewer generation length while improving throughput up to $\mathbf{1.7}\times$.

Reinforcement Learning (RL) has become critical for advancing modern Large Language Models (LLMs), yet existing synchronous RL systems face severe performance bottlenecks. The rollout phase, which dominates end-to-end iteration time, suffers from substantial long-tail latency and poor resource utilization due to inherent workload imbalance. We present Seer, a novel online context learning system that addresses these challenges by exploiting previously overlooked similarities in output lengths and generation patterns among requests sharing the same prompt. Seer introduces three key techniques: divided rollout for dynamic load balancing, context-aware scheduling, and adaptive grouped speculative decoding. Together, these mechanisms substantially reduce long-tail latency and improve resource efficiency during rollout. Evaluations on production-grade RL workloads demonstrate that Seer improves end-to-end rollout throughput by 74% to 97% and reduces long-tail latency by 75% to 93% compared to state-of-the-art synchronous RL systems, significantly accelerating RL training iterations.

Diffusion language models (DLMs) have emerged as a promising alternative to the long-dominant autoregressive (AR) paradigm, offering a parallelable decoding process that could yield greater efficiency. Yet, in practice, current open-source DLMs often underperform their AR counterparts in speed, limiting their real-world utility. This work presents a systematic study of DLM efficiency, identifying key issues in prior evaluation methods. Through empirical benchmarking and a theoretical analysis, we demonstrate that AR models generally achieve higher throughput, while DLMs consistently lag. We also investigate acceleration strategies, finding that techniques like dual cache and parallel decoding mainly offer gains at small batch sizes, with their benefits diminishing upon scaling. Our findings underscore the necessity of robust evaluation methods and improved acceleration strategies to advance research on DLMs.

LLMs have become the mainstream approaches to code generation. Existing LLMs mainly employ autoregressive generation, i.e. generating code token-by-token from left to right. However, the underlying autoregressive generation has two limitations in code generation. First, autoregressive LLMs only generate a token at each step, showing low efficiency in practice. Second, programming is a non-sequential process involving back-and-forth editing, while autoregressive LLMs only employ the left-to-right generation order. These two intrinsic limitations hinder the further development of LLMs in code generation. Recently, diffusion LLMs have emerged as a promising alternative. Diffusion LLMs address the above limitations with two advances, including multi-token prediction (i.e. generating multiple tokens at each step) and flexible generation order (i.e. flexibly determining which positions to generate tokens). However, there is no systematic study exploring diffusion LLMs in code generation. To bridge the knowledge gap, we present the first empirical study of diffusion LLMs for code generation. Our study involves 9 representative diffusion LLMs and conduct experiments on 4 widely used benchmarks. Based on the results, we summarize the following findings. (1) Existing diffusion LLMs are competitive with autoregressive LLMs with similar sizes. (2) Diffusion LLMs have a stronger length extrapolation ability than autoregressive LLMs and perform better in long code understanding. (3) We explore factors impacting the effectiveness and efficiency of diffusion LLMs, and provide practical guidance. (4) We discuss several promising further directions to improve diffusion LLMs on code generation. We open-source all source code, data, and results to facilitate the following research. The code is publicly available at https://github.com/zhangyitonggg/dllm4code.

Obtaining high-quality outputs from Large Language Models (LLMs) often depends upon the choice of a sampling-based decoding strategy to probabilistically choose the next token at each generation step. While a variety of such sampling methods have been proposed, their performance can be sensitive to the selection of hyperparameters which may require different settings depending upon the generation task and temperature configuration. In this work, we introduce $p$-less sampling: an information-theoretic approach to sampling which dynamically sets a truncation threshold at each decoding step based on the entire token probability distribution. Unlike existing methods, $p$-less sampling has no hyperparameters and consistently produces high-quality outputs as temperature increases. We provide theoretical perspectives on $p$-less sampling to ground our proposed method and conduct experiments to empirically validate its effectiveness across a range of math, logical reasoning, and creative writing tasks. Our results demonstrate how $p$-less sampling consistently outperforms existing sampling approaches while exhibiting much less degradation in text quality at higher temperature values. We further show how $p$-less achieves greater inference-time efficiency than alternative methods through lower average token sampling times and shorter generation lengths, without sacrificing accuracy. Finally, we provide analyses to highlight the benefits of $p$-less through qualitative examples, case studies, and diversity assessments. The code is available at https://github.com/ryttry/p-less .

Reinforcement learning (RL) has become central to training large language models (LLMs), yet the field lacks predictive scaling methodologies comparable to those established for pre-training. Despite rapidly rising compute budgets, there is no principled understanding of how to evaluate algorithmic improvements for scaling RL compute. We present the first large-scale systematic study, amounting to more than 400,000 GPU-hours, that defines a principled framework for analyzing and predicting RL scaling in LLMs. We fit sigmoidal compute-performance curves for RL training and ablate a wide range of common design choices to analyze their effects on asymptotic performance and compute efficiency. We observe: (1) Not all recipes yield similar asymptotic performance, (2) Details such as loss aggregation, normalization, curriculum, and off-policy algorithm primarily modulate compute efficiency without materially shifting the asymptote, and (3) Stable, scalable recipes follow predictable scaling trajectories, enabling extrapolation from smaller-scale runs. Combining these insights, we propose a best-practice recipe, ScaleRL, and demonstrate its effectiveness by successfully scaling and predicting validation performance on a single RL run scaled up to 100,000 GPU-hours. Our work provides both a scientific framework for analyzing scaling in RL and a practical recipe that brings RL training closer to the predictability long achieved in pre-training.

Large Language Model (LLM) inference has emerged as a fundamental paradigm. In real-world scenarios, variations in output length cause severe workload imbalance in the decode phase, particularly for long-output reasoning tasks. Existing systems, such as PD disaggregation architectures, rely on static prefill-to-decode scheduling, which often results in SLO violations and OOM failures under evolving decode workloads. In this paper, we propose ARES, an adaptive decoding rescheduling system powered by length prediction to anticipate future workloads. Our core contributions include: (1) A lightweight and continuous LLM-native prediction method that leverages LLM hidden state to model remaining generation length with high precision (reducing MAE by 49.42%) and low overhead (cutting predictor parameters by 93.28%); (2) A rescheduling solution in decode phase with : A dynamic balancing mechanism that integrates current and predicted workloads, reducing P99 TPOT by 74.77% and achieving up to 2.24 times higher goodput.

Concise reasoning in large language models seeks to generate only essential intermediate steps needed to arrive at a final answer, thereby alleviating issues of "over-thinking". Most proposed approaches hinge on carefully hand-crafted heuristics, struggling to balance concision with performance, often failing to adapt across domains and model scales. In this work, we address these challenges by introducing a principled and pragmatic strategy, performance-aware length updating (P ALU). As a principled algorithm, P ALU formulates concise reasoning as a constrained optimization problem, minimizing response length subject to a performance constraint, and then applies Lagrangian optimization to convert it into a tractable unconstrained problem. As a pragmatic solution, P ALU streamlines complicated update rules through three approximations: (i) estimating performance with off-policy rollouts, (ii) truncating the Lagrange multiplier to two extremes, and (iii) replacing gradient-based updates with quantile-driven length adjustments. Furthermore, P ALU is demonstrated to adapt across both domain (logic, STEM and math) and model scale (1.5B, 7B, 14B) entrenching the algorithm as a practical and effective concise reasoning approach. Reasoning, requiring large language models (LLMs) to work through intermediate steps before producing a final answer, substantially improves performance on complex tasks such as mathematics (Jaech et al., 2024; Shao et al., 2024), programming (Lambert et al., 2024), and value alignment (Guo et al., 2025). Y et this benefit is often accompanied by overthinking: redundant self-reflection, backtracking, and validation (Chen et al., 2024; Zhang et al., 2024; Fatemi et al., 2025). These limitations inflate inference costs and hampers user experience, motivating the need for concise reasoning--the production of only the essential steps required to reach a correct answer.