Jiahui Sheng, Le Li, Limin Lin, Jiahao Li, Wansen Feng, Hong Gu, Shuhan Chen, Xiaorun Li 等 9 位作者
Portrait video relighting aims to modify the illumination of a portrait video while preserving photorealism and temporal stability. The task becomes particularly challenging under dynamic lighting, where illumination direction, intensity, and color vary over time. To tackle this problem, we introduce FlashRelight, a novel two-stage framework for dynamic portrait video relighting. FlashRelight first suppresses the source illumination through delighting, and then synthesizes the target dynamic lighting on an illumination-neutral portrait video. FlashRelight supports three complementary control modalities: envmap sequences for explicit frame-level lighting control, dynamic lighting text for semantic illumination editing, and lighting reference videos for visual lighting transfer. Training follows a low-cost hybrid strategy that combines static one-light-at-a-time (OLAT) data with generated dynamic-lighting portrait videos, covering both controlled illumination variation and realistic portrait motion. Extensive experiments show that FlashRelight achieves temporally coherent and photorealistic dynamic relighting while enabling flexible cross-video lighting transfer. Code is available at https://github.com/vivoCameraResearch/FlashRelight
Hybrid language models interleave a few softmax-attention layers with many linear-attention or state-space layers, so every layer keeps a memory of earlier tokens: a key-value cache or a fixed-size recurrent state. Serving them spends more compute and high-bandwidth memory on these memories than the output requires. The prefill runs every layer for every prompt token, and every decode step loads and stores whole recurrent states, which on Qwen3.8-Flash-Next take as many bytes per request as the attention cache of thousands of tokens. We find two properties of trained hybrids that locate this excess. First, writes are easier than reads: writing a deep layer's memory of a token needs little of the deep computation that predicting the next token needs. Trained copies of the deep layers' write paths, which we call emitters, therefore write a prompt's deep memories from the residual stream of a shallower anchor layer, and all prompt tokens but the last run as few as 63% of the layers. Second, recurrent states have a state sink: most of their energy lies in a rank-one term along a fixed direction of each head, and the rest is a low-rank working memory. SEED builds on both: it makes the prefill shallow in depth and the decode state shallow in representation. Each recurrent state is stored as its sink plus low-rank factors, which the Gated DeltaNet, Kimi Delta Attention and Mamba-2 updates extend exactly; we prove that the open-loop error of the periodic compressions back to low rank is bounded independently of their number. SEED keeps the original weights frozen, compresses no attention cache, and runs all layers for every generated token. On six checkpoints of five production hybrids of 30B to over 300B parameters, it is nearly lossless on eleven official exams. Served with SGLang on NVIDIA GB300 GPUs on an agentic-coding trace, it raises the per-GPU throughput of Qwen3.8-Flash-Next 1.2-1.4x.
Coding agents must deliver correct software while respecting users' expertise, preferences, and working conditions. Existing benchmarks increasingly incorporate interaction, yet personalized requirement fulfillment remains underexplored. We introduce PersonaCodeBench, which turns user characteristics into verifiable obligations in persona-driven multi-turn coding sessions. Our persona pool spans roles, technical abilities, interaction preferences, and practical constraints, yielding task-grounded evolved requirements for implementation, permissions, testing, explanation, and delivery. At runtime, a controller governs user simulator responses and requirement disclosure, using online requirement evaluation to provide feedback and advance the interaction. Requirements emerge progressively, with their content, disclosure conditions, and acceptance criteria fixed before evaluation. A completion verifier and a personalization verifier separately assess native correctness and persona fulfillment. Joint pass requires both adherence to the declared budget and compliance with validity conditions. We evaluate all major frontier agents, establish a comprehensive leaderboard, and conduct detailed experimental analyses to validate effectiveness. PersonaCodeBench supports reproducible evaluation of coding agents' ability to meet explicit, personalized user expectations alongside coding objectives.
Video generation has progressed from isolated clips to coherent short-form audiovisual sequences, and the bottleneck is shifting from rendering to semantic control. This report summarizes the observations and insights on multimodal understanding gained from successive generations of Kling. We argue that generation quality is the projection of multimodal understanding through the entire data--conditioning--alignment--evaluation chain, and we examine this chain module by module. Kling-MLLM, a shared multimodal foundation model that unifies image, video, audio, synchronized audio--video, and interleaved inputs, is specialized for each downstream module. Because a DiT is trained on far fewer training samples than a language model, we treat captioning as a semantic contract that maximizes description density, and extend it to reference-grounded captioning that makes reference bindings explicit. Prompt enhancement translates user intent into this caption distribution: basic PE is derived from captions of compelling videos and refined with text-level and generator-in-the-loop reinforcement learning, while agentic PE refines the story and retrieves compelling examples. We show that a reliable text encoder comes from perceptual training rather than from stronger language ability, and that candidates can be screened from text alone before end-to-end training. Reward modeling pairs a general reward model for explainable, localized critiques with a decision reward model for fast multidimensional scoring, treating perceptual bandwidth as a first-class design parameter. Finally, we review task-specific automatic evaluation, focusing on protocols that expose scoring shortcuts and help trace failures to their upstream source. We hope these lessons offer practical guidance for building the understanding stack behind controllable video generation.
Qi Wen, Borui Xu, Hussein Amro, Hongshi Tan, Hanmin Lin, Chaoyi Ruan, Shanghao Liu, Gang Yang 等 14 位作者
Vision-language-action (VLA) models offer a promising path toward general-purpose robot manipulation. While inference latency can affect task success in a constantly evolving physical world, the relationship between latency and task success rate remains poorly understood. In this paper, we introduce VLALagBench, a comprehensive benchmark for studying VLA inference latency sensitivity across task types, object speeds, execution modes (synchronous and asynchronous), and VLA policies. Using VLALagBench, we find that even a modest inference latency of 100 ms can nearly halve task success on moving targets, with sensitivity varying substantially across tasks and policies. These differences reflect task-specific requirements for precision and continuous motion. By overlapping inference with action execution, asynchronous execution can recover some of the lost success at the cost of more frequent inference requests. Building on these findings, we develop an empirical model that predicts task success from limited calibration data, achieving root-mean-square prediction errors of only 5.7-7.8% across various VLA policies. Given an inference system's latency and per-robot throughput, these predictions guide the selection of system-specific VLA algorithm configurations to maximize task success. We believe VLALagBench and these findings provide a solid foundation for future robotics algorithm-system co-design.
Siyang Wu, Weizhi Zhang, Chen Yang, Esther Xu Fei, Jorin Overwiening, Xu Pan, Jing Tang, Yixuan He 等 37 位作者
Scientific agents collaborate with researchers whose expertise and preferences differ, but benchmarks usually hold the collaborator fixed. We introduce PersonaScienceBench: 200 tasks from five benchmarks across four scientific domains. Simulated researchers provide context and feedback according to persona profiles, while task objectives and grading remain fixed. Agents encounter researcher preferences only through conversation. Across eighteen modelconfigurations, task performance, fact recovery, and self-repetition reveal different strengths. Detailed GPT 6 Astra analyses show distinct agent responses across researcher profiles; interaction traces illustrate targeted clar ification and workflow revision. These behaviors can differ even when final scores match. PersonaScienceBench therefore evaluates how agents collaborate with diverse researchers alongside whether they complete the task.
Cuckoo filters are a widely adopted approximate membership query (AMQ) data structure that supports insertions, lookups, and deletions with low memory overhead. At high load factors, however, insertion performance degrades sharply due to the random-walk eviction process, which may require a large number of element relocations. Recent work on bucketized cuckoo hash tables has shown that \emph{fresh-priority eviction}, which preferentially evicts elements whose alternate bucket location has not yet been used, achieves $O(\ell m)$ total expected evictions for bucket size~$\ell$, an improvement over the standard random walk. In this paper, we show that this technique transfers to \emph{cuckoo filters} despite their use of XOR-based partial-key hashing. The key insight is that a fresh element's fingerprint, although stored in the table, is not examined by the eviction policy, so when the fingerprint length satisfies $k = Ω(\log m)$, the alternate bucket hash remains a near-uniform random variable, a property we call \emph{latent randomness}. We prove that under this condition the corruption graph remains subcritical and the total number of evictions is $O(\ell m \log m)$; we conjecture the tighter $O(\ell m)$ bound holds via the full amortized analysis of Bell and Kuszmaul. We further propose encoding the fresh/kicked state implicitly by the element's position within its bucket, requiring no additional per-element metadata. Experimental results on a filter with $2^{22}$ slots and $8$-bit fingerprints demonstrate $1.2\times$ to $2.4\times$ fewer evictions at load factors between $94\%$ and $99\%$ for bucket size $\ell = 8$, and a $1.73\times$ reduction in cumulative evictions to full capacity.
Huanyue Liao, Seok Min Lim, Yan Ling Lim, Shi Min Cheng, Kellin Pelrine, Chen Hui Ong
AI companion chatbots are increasingly used by vulnerable individuals, including adolescents in suicidal crisis, for emotional support. Existing safety evaluations focus on base models, leaving gaps in product-level assessment, crisis-stage sensitivity, and comparisons between companion apps and general-purpose LLMs. We introduce MIRROR-Eval (Multi-turn Interaction Risk Review: Observing suicidal-crisis Responses), a dataset of 960 annotated 20-turn conversations across eight deployed platform configurations, six clinically grounded personas, and two crisis stages: ideation and explicit planning. An LLM-as-a-Judge pipeline applies an expert-reviewed rubric covering user risk severity, chatbot risk awareness, and intervention behavior. We observe a three-tier harm structure: companion apps (Character.AI and Replika) and Grok 4.3 form a high-harm tier; Gemini 3.1 Pro is mid-range; GPT-5.5 and Claude Sonnet 5 show substantially lower harm. At planning, companion apps show increased glorification and normalisation, advancement toward harmful action, and harmful methods and instructions. Crisis recognition and empathetic language do not ensure safe responses. These findings highlight the need for product-level safety evaluation across crisis stages. Dataset and code: https://huggingface.co/datasets/imda-biztech/MIRROR-Eval. Warning: this paper includes examples that may be offensive or harmful.
Document image binarization must preserve thin, broken, and low-contrast strokes against degraded backgrounds; Mamba-based state space models have so far been explored only preliminarily for it. We measure what state-space propagation does to document features and find that it does not attenuate foreground responses: every region gains amplitude, but the background gain is 1.8 times the stroke gain, compressing the foreground-background amplitude contrast from 1.86 to 1.03. This compression scales monotonically with the effective receptive field, suggesting it is not specific to state space models. Guided by this, we propose DeepMine-Mamba, whose Anti-Dilution Gate computes its modulation from the propagation-induced feature difference and confines it to stroke regions by a learned stroke prior. Under a strict leave-one-year-out protocol on DIBCO/H-DIBCO with no external training data it is competitive with recent transformer- and gated-convolution baselines; because published protocols differ and are often unstated, we report protocol-controlled evidence separately from cross-paper comparison. Ablations under a fixed objective show that the gate recovers stroke-background discriminability and improves all four metrics, including the two absent from the objective.
World action models (WAMs) predict the visual consequences of the actions they generate, which makes test-time selection a natural way to improve a fixed policy: imagine the futures of candidate actions and execute the best one. The benefit of selection is bounded by the selectable headroom, how much better the best candidate truly is, but the standard estimate, which executes each candidate once and keeps the best outcome, conflates candidate quality with the randomness of everything that follows. We introduce a denoised counterfactual protocol, based on bitwise-reproducible branching, repeated continuations, held-out oracles, and common random numbers, that measures the headroom a selector can realize, and use its measurements to build a fail-gated selector. Across four WAMs, two benchmarks, over 100,000 branched rollouts, and pre-registered tests, we find that (i) the standard estimate overstates the selectable headroom of native samples: bounds from 17 continuations per candidate place its selectable share between 24% and 66% for three WAMs; (ii) structured action perturbations create up to 16 points of selectable headroom; and (iii) in both WAMs with action-conditioned imagination, imagined futures carry candidate-outcome information with no detectable loss relative to real futures, information that the value head of Cosmos Policy leaves unused. Deviating from the policy chunk only when a readout of the WAM predicts failure, the selector raises closed-loop task success from 76.9% to 81.6% on held-out episodes of training tasks, with the largest gain where the policy fails most often, and from 62.6% to 69.8% on unseen tasks.
Generalist robot policies and language-model agents provide complementary capabilities, but completing a task also requires adapting their use to the consequences of physical interaction. We study execution as a locus of adaptation: feedback guides the current action and supplies evidence for revising subsequent decisions under fixed model weights. We introduce RoboDoctor, a training-free framework that connects diagnosis, grounded action, and persistent experience across three revision targets. The harness allocates control between policy and agent; intervention determines when to assist while preserving useful policy behavior; and operation refines physical methods through observed effects. Linked experience records retain the reason for a decision, the attempted method, its outcome, and conditions for reuse. On a curated archive of 54 RoboDojo-Sim task entries, RoboDoctor achieves a five-dimension mean score of 57.10, with a native success rate of 49.44%. Recorded interactions illustrate strategy changes, selective correction, and grasp revision. These findings characterize execution-time adaptation in the evaluated system; controlled trials remain necessary to isolate the contribution of cross-episode experience.
Long-horizon embodied navigation may be interrupted by a higher-priority instruction before the current task sequence is complete. The agent must finish its current atomic task, execute the inserted task, and resume the remaining sequence from the updated environment state, where its viewpoint and available navigation candidates may have changed. To evaluate this capability, we introduce PIFR benchmark, a novel task for runtime task preemption and resumption in embodied navigation. PIFR includes single, repeated, and nested interruption settings,with stage-specific metrics for inserted-task completion, resumption of the original task, and completion of the full sequence. Evaluations show that recovery performance varies across interruption conditions, while aggregate success can obscure differences in interruption handling and resumption. We further propose a lightweight candidate-ranking baseline to improve target re-grounding after interruption. Together, PIFR and the baseline provide a systematic framework for studying interruption handling and task recovery in long-horizon embodied navigation.
Computer-use agents (CUAs) are commonly evaluated against fixed task specifications, but users may omit requirements, add constraints, or require approval before consequential actions. Completing the benchmark instruction does not guarantee serving the person who asked. We introduce PersonaCUABench, which holds 74 tasks across 11 domains fixed and varies only who asks. A persona agent, playing one of 44 personas in four groups, rewrites each instruction in that user's own words and answers the CUA's questions, while the CUA sees only the dialogue, never the profile or the original instruction. Across nine backbones (3,330 runs), mean task completion falls by 36% relative to the original requests. The losses have two sources: in 59.9% of cells that lose points, the agent engages with what the user asked for and still gets it wrong, and 11.2% are stops the user requested that the checker scores as failures. Even when a persona request is fully completed, the user gives the top satisfaction rating in only 62.7% of cells, against 84.1% for the original requests. CUAs should therefore be evaluated not only on task completion but also on whether they meet the requirements and constraints users express. Code: https://github.com/MatrAIx-ai/PersonaCUABench. Project page: https://matraix.ai/research/persona-cua.html
Liangliang Dong Jilin University, Lianlei Shan, Shuaimin Li
While Chain-of-Thought (CoT) reasoning empowers Large Language Models (LLMs) to tackle complex tasks, its reliance on discrete token decoding imposes an inherent Discreteness Bottleneck, limiting expressiveness within a restricted vocabulary space. Existing continuous reasoning approaches, such as SoftCoT, mitigate this but typically rely on external auxiliary models, resulting in complex deployment and fractured inference pipelines. To address these challenges, we propose Self-SoftCoT, a self-contained framework that enables a frozen LLM to internally generate and consume latent thoughts without external assistants. By establishing a single-stream “Thinking → Speaking” closed-loop, we decouple latent planning from explicit generation. Furthermore, we adopt Group Sequence Policy Optimization (GSPO) to stabilize learning and employ Position-Aware Independent Projection to mitigate representation homogenization. Experimental results on five reasoning benchmarks demonstrate that our method significantly improves the reasoning performance of frozen LLMs. Specifically, our Qwen2.5-based model uses only N=2 soft tokens to outperform the SoftCoT baseline (N=4), improving the average accuracy from 75.06% to 78.42%. Similarly, LLaMA-3.1 performance increases from 70.52% to 74.55%.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Kunpeng Kang, Shuaimin Li, Kaiyuan Zhang, Luyang Zhang, Jiasheng Si Qilu University of Technology (Shandong Academy of Sciences), Bing Xu, Kehai Chen Harbin Institute of Technology (Shenzhen), Muyun Yang 等 9 位作者
Word sense disambiguation (WSD) is a foundational task in natural language processing. Recent research has reformulated WSD for large language models (LLMs) as a generative task, where the model produces a definition to convey the intended meaning of an ambiguous word in context.In practice, most existing approaches implement this formulation through straightforward supervised fine-tuning, which tends to prioritize superficial context-to-gloss memorization over true contextual sense discrimination, leading to degraded performance on less frequent senses (LFS), particularly in unseen settings.To address this issue, we propose WSDPO, a training framework for generative WSD with chain-of-thought (CoT) and preference optimization. WSDPO consists of three stages: (1) disambiguation-aware CoT construction, which produces training data containing explicit disambiguation steps for the later stage;(2) disambiguation-guided supervised fine-tuning, which explicitly trains the model to discriminate word sense before generating the final definition; and(3) preference-based optimization, which further strengthens the model’s ability to generate sense-faithful definitions by optimizing it using preference pairs constructed from multiple sampled CoT outputs.Extensive experiments across benchmark datasets and multiple backbone LLMs demonstrate that WSDPO achieves substantial performance gains on rare and unseen settings, and exhibits strong generalization in standard evaluation settings.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zongmin Li, Jian Su A*STAR, Farah Benamara, Aixin Sun Nanyang Technological University
Large language models (LLMs) are often assumed to contain "safety regions” - parameter subsets whose modification directly influences safety behaviors. We conduct a systematic evaluation of four safety region identification methods spanning different parameter granularities, from individual weights to entire Transformer layers, across four families of backbone LLMs with varying sizes. Using ten safety identification datasets, we find that the identified safety regions exhibit only low to moderate overlap, as measured by IoU. The overlap drops significantly when the safety regions are further refined using utility datasets (i.e. non-harmful queries). These results suggest that current techniques fail to reliably identify a stable, dataset-agnostic safety region.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yihong Dong, Zhaoyu Ma, Xue Jiang, Zhiyuan Fan Northwest Polytechnical University Xi'an, Jiaru Qian, Yongmin Li Peking University, Jianha Xiao, Zhi Jin Peking University 等 9 位作者
Diffusion language models (DLMs) are emerging as a compelling alternative to the dominant autoregressive paradigm, offering inherent advantages in parallel generation and bidirectional context modeling. However, for the tasks with strict structural constraints such as code generation, DLMs face a critical trade-off between inference speed and output quality, where accelerating generation by reducing sampling steps often leads to catastrophic performance collapse.We find that the fundamental reasons are: 1) the generation difficulty is uneven in the structured sequence decoding steps, making DLM’s static acceleration strategy suboptimal; 2) the context of tokens generated by DLM evolves continuously, causing early high-confidence predictions to turn into irreversible errors.In this paper, we introduce efficient Sampling with Adaptive acceleration and Backtracking Enhanced Remasking (i.e., Saber), a novel training-free sampling algorithm for DLMs that the first to improve both inference speed and output quality in code generation. Saber dynamically adjusts the number of tokens unmasked per step based on the model’s evolving confidence, and utilizes a backtracking mechanism to revert tokens whose confidence drops as new context emerges, with its effectiveness supported by theoretical analysis.Extensive experiments on multiple mainstream code generation benchmarks show that Saber boosts Pass@1 accuracy by an average of 1.9% over mainstream DLM sampling methods, while achieving an average 251.4% inference speedup. By leveraging the inherent advantages of DLMs, our work significantly narrows the performance gap with autoregressive models in code generation.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Speech codecs provide an important interface between continuous speech signals and large language models. An ideal codec for speech language models should not only preserve acoustic information but also capture rich semantic information. However, existing codecs struggle to balance these objectives at low bitrates. We propose \textbf{XY-Tokenizer}, a low-bitrate speech codec (around 1 kbps) trained with a structured multi-stage, multi-task strategy that aligns discrete speech representations with text while preserving fine-grained acoustic details for reconstruction. This design explicitly mitigates the semantic–acoustic conflict observed in prior low-bitrate codecs. Experiments show that XY-Tokenizer achieves stronger semantic alignment than representative semantic-distillation codecs such as SpeechTokenizer and Mimi, while maintaining high-quality speech reconstruction across both clean and out-of-distribution conditions. Furthermore, XY-Tokenizer consistently outperforms existing low-bitrate codecs in LLM-based speech understanding and generation tasks, demonstrating its effectiveness as a general-purpose speech representation for speech–language modeling.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jiahao Tang Central South University, Henry Hengyuan Zhao, Lijian Wu, Zijian Zhang, Yifei Tao, Dongxing Mao Central South University, Yang Wan, Jingru Tan Central South University 等 11 位作者
We introduce Chart2Code, a new benchmark for evaluating the natural language to chart code generation capabilities of large multimodal models (LMMs). Chart2Code is explicitly designed from a user-driven perspective, capturing diverse real-world scenarios and progressively increasing task difficulty. It consists of three levels: Level 1 (Chart Reproduction) reproduces charts from a reference figure and user query; Level 2 (Chart Editing) involves complex modifications such as changing chart types or adding elements; and Level 3 (Long-Table to Chart Generation) requires models to transform long, unprocessed tables into faithful charts following user instructions. To our knowledge, this is the first hierarchical benchmark that reflects practical chart2code usage while systematically scaling task complexity. In total, Chart2Code contains 2,186 tasks across 22 chart types, paired with multi-level evaluation metrics that assess both code correctness and the visual fidelity of rendered charts. We benchmark 29 state-of-the-art (SoTA) LMMs, including both proprietary and the latest open-source models such as GPT-5.2, Qwen3-VL, InternVL3/3.5, MiMo-VL, and Seed-1.6-VL. Experimental results demonstrate that even the SoTA model GPT-5.2 averages 72.21 on code-based evaluation and only 33.41 on chart-quality assessment across the editing tasks, underscoring the difficulty of Chart2Code. We anticipate this benchmark will drive advances in multimodal reasoning and foster the development of more robust and general-purpose LMMs.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Mengshi Chen, Yuxiang Sun, Tengchao Li Shanghai Jiaotong University, Jianwei Wang University of New South Wales, Kai Wang Shanghai Jiao Tong University, Xuemin Lin Shanghai Jiaotong University, Ying Zhang University of Technology Sydney (UTS), Wenjie Zhang the university of new south wales
Data preparation is a critical step in enhancing the usability of tabular data and thus boosts downstream data-driven tasks. Traditional methods often face challenges in capturing the intricate relationships within tables and adapting to the tasks involved. Recent advances in Language Models (LMs), especially in Large Language Models (LLMs), offer new opportunities to automate and support tabular data preparation. However, why LMs suit tabular data preparation (i.e., how their capabilities match task demands) and how to use them effectively across phases still remain to be systematically explored. In this survey, we systematically analyze the role of LMs in enhancing tabular data preparation processes, focusing on four core phases: data acquisition, integration, cleaning, and transformation. For each phase, we present an integrated analysis of how LMs can be combined with other components for different preparation tasks, highlight key advancements, and outline prospective pipelines.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Inference providers must maintain availability for many LLMs, including low-volume but essential models, making resource efficiency increasingly important as token prices fall. Analysis of production traces reveals a dynamic bursty-group pattern in which sets of models become active together and shift over time; existing space- and time-sharing approaches lack principled mechanisms to adapt to this variability, forcing trade-offs between SLO adherence and efficiency. We observe that elastic memory allocation can unify spatial and temporal sharing. Based on this insight, we have developed Prism, a memory-centric LLM co-serving framework that applies memory ballooning to reclaim memory across models and support both forms of sharing under a single scheme. Prism's balloon driver, referred to as kvcached, has been open-sourced at https://github.com/ovg-project/kvcached, and deployed in production environments across 10K+ GPUs.
Coach not crutch: Evidence that AI can improve writing skill despite reducing effort
Benjamin Lira, Todd Rogers, Daniel G. Goldstein, Lyle Ungar, Angela L. Duckworth
AI 速览：多项预注册实验表明，使用AI写作工具虽降低练习时的努力程度，却能通过“示例教学”提升无AI写作测试中的表现。
In a series of highly-powered empirical studies, we examine the intuition that by sparing effort, using AI inevitably hinders learning. First, in a nationally representative survey of young adults, the majority expressed the view that using AI makes people lazier and less capable. Next, in a random-assignment experiment, we gave participants a tutorial on best practices in professional writing, then provided one group with access to an AI writing tool and asked another to practice writing on their own. Those who practiced with AI indeed exerted less effort while practicing -- yet wrote better cover letters in no-AI writing tests. In a second experiment with more rigorous control conditions, access to AI improved writing more than either googling cover letter examples and tips or receiving personalized feedback on their practice letters from experienced human editors. A third experiment explained these learning gains by showing that AI can teach by example: participants who viewed a cover letter that had been revised by the AI tool but did no further practice improved their writing as much as those who practiced writing with the original AI tool. Collectively, these pre-registered experiments suggest that AI can exert opposing effects on effort and learning rate -- making it possible in at least some cases to work less and learn more.
Chaofan Tao, Qian Liu, Longxu Dou, Niklas Muennighoff, Zhongwei Wan, Ping Luo, Min Lin, Ngai Wong
Research on scaling large language models (LLMs) has primarily focused on model parameters and training data size, overlooking the role of vocabulary size. We investigate how vocabulary size impacts LLM scaling laws by training models ranging from 33M to 3B parameters on up to 500B characters with various vocabulary configurations. We propose three complementary approaches for predicting the compute-optimal vocabulary size: IsoFLOPs analysis, derivative estimation, and parametric fit of the loss function. Our approaches converge on the conclusion that the optimal vocabulary size depends on the compute budget, with larger models requiring larger vocabularies. Most LLMs, however, use insufficient vocabulary sizes. For example, we predict that the optimal vocabulary size of Llama2-70B should have been at least 216K, 7 times larger than its vocabulary of 32K. We validate our predictions empirically by training models with 3B parameters across different FLOPs budgets. Adopting our predicted optimal vocabulary size consistently improves downstream performance over commonly used vocabulary sizes. By increasing the vocabulary size from the conventional 32K to 43K, we improve performance on ARC-Challenge from 29.1 to 32.0 with the same 2.3e21 FLOPs. Our work highlights the importance of jointly considering tokenization and model scaling for efficient pre-training. The code and demo are available at https://github.com/sail-sg/scaling-with-vocab and https://hf.co/spaces/sail/scaling-with-vocab-demo.