arXiv subjects

Ce Zhang

Publications and source records attributed to Ce Zhang.

At least 55 records · Page 3Linked to original sources

arXiv2025

Hold Onto That Thought: Assessing KV Cache Compression On Reasoning

Large language models (LLMs) have demonstrated remarkable performance on long-context tasks, but are often bottlenecked by memory constraints. Namely, the KV cache, which is used to significantly speed up attention computations, grows linearly with context length. A suite of compression algorithms has been introduced to alleviate cache growth by evicting unimportant tokens. However, several popular strategies are targeted towards the prefill phase, i.e., processing long prompt context, and their performance is rarely assessed on reasoning tasks requiring long decoding. In particular, short but complex prompts, such as those in benchmarks like GSM8K and MATH500, often benefit from multi-step reasoning and self-reflection, resulting in thinking sequences thousands of tokens long. In this work, we benchmark the performance of several popular compression strategies on long-reasoning tasks. For the non-reasoning Llama-3.1-8B-Instruct, we determine that no singular strategy fits all, and that performance is heavily influenced by dataset type. However, we discover that H2O and our decoding-enabled variant of SnapKV are dominant strategies for reasoning models, indicating the utility of heavy-hitter tracking for reasoning traces. We also find that eviction strategies at low budgets can produce longer reasoning traces, revealing a tradeoff between cache size and inference costs.

cs.CL↗

arXiv2025

Beat the long tail: Distribution-Aware Speculative Decoding for RL Training

Reinforcement learning(RL) post-training has become essential for aligning large language models (LLMs), yet its efficiency is increasingly constrained by the rollout phase, where long trajectories are generated token by token. We identify a major bottleneck:the long-tail distribution of rollout lengths, where a small fraction of long generations dominates wall clock time and a complementary opportunity; the availability of historical rollouts that reveal stable prompt level patterns across training epochs. Motivated by these observations, we propose DAS, a Distribution Aware Speculative decoding framework that accelerates RL rollouts without altering model outputs. DAS integrates two key ideas: an adaptive, nonparametric drafter built from recent rollouts using an incrementally maintained suffix tree, and a length aware speculation policy that allocates more aggressive draft budgets to long trajectories that dominate makespan. This design exploits rollout history to sustain acceptance while balancing base and token level costs during decoding. Experiments on math and code reasoning tasks show that DAS reduces rollout time up to 50% while preserving identical training curves, demonstrating that distribution-aware speculative decoding can significantly accelerate RL post training without compromising learning quality.

cs.LG↗

arXiv2025

TiDAR: Think in Diffusion, Talk in Autoregression

Diffusion language models hold the promise of fast parallel generation, while autoregressive (AR) models typically excel in quality due to their causal structure aligning naturally with language modeling. This raises a fundamental question: can we achieve a synergy with high throughput, higher GPU utilization, and AR level quality? Existing methods fail to effectively balance these two aspects, either prioritizing AR using a weaker model for sequential drafting (speculative decoding), leading to lower drafting efficiency, or using some form of left-to-right (AR-like) decoding logic for diffusion, which still suffers from quality degradation and forfeits its potential parallelizability. We introduce TiDAR, a sequence-level hybrid architecture that drafts tokens (Thinking) in Diffusion and samples final outputs (Talking) AutoRegressively - all within a single forward pass using specially designed structured attention masks. This design exploits the free GPU compute density, achieving a strong balance between drafting and verification capacity. Moreover, TiDAR is designed to be serving-friendly (low overhead) as a standalone model. We extensively evaluate TiDAR against AR models, speculative decoding, and diffusion variants across generative and likelihood tasks at 1.5B and 8B scales. Thanks to the parallel drafting and sampling as well as exact KV cache support, TiDAR outperforms speculative decoding in measured throughput and surpasses diffusion models like Dream and Llada in both efficiency and quality. Most notably, TiDAR is the first architecture to close the quality gap with AR models while delivering 4.71x to 5.91x more tokens per second.

cs.CL↗

arXiv2025

Search for light Dark Sectors with GeV Muon Beams

Sub-GeV light dark matter often requires new light mediators, such as a dark $Z$ boson in the $L_μ- L_τ$ gauge theory. We study the search potential for such a $Z^\prime$ boson via the process $μe^- \to μe^- X$, with $X$ decaying invisibly, in a muon on-target experiment using a high-intensity 1-10 GeV muon beam from facilities such as HIAF-HIRIBL. Events are identified by the scattered muon and electron from the target using silicon strip detectors in a single-station telescope system. Backgrounds are suppressed through a trained boosted decision tree (BDT) classifier, and activity in downstream subdetectors remains low. This approach can probe a $Z^\prime$ boson in the 10 MeV mass range with improved sensitivity. Nearly three orders of magnitude improvement is achievable with a full multi-telescope station system employing a 160 GeV muon beam at CERN, such as in the MUonE experiment.

hep-ph↗

arXiv2025

DGL-RSIS: Decoupling Global Spatial Context and Local Class Semantics for Training-Free Remote Sensing Image Segmentation

The emergence of vision language models (VLMs) bridges the gap between vision and language, enabling multimodal understanding beyond traditional visual-only deep learning models. However, transferring VLMs from the natural image domain to remote sensing (RS) segmentation remains challenging due to the large domain gap and the diversity of RS inputs across tasks, particularly in open-vocabulary semantic segmentation (OVSS) and referring expression segmentation (RES). Here, we propose a training-free unified framework, termed DGL-RSIS, which decouples visual and textual representations and performs visual-language alignment at both local semantic and global contextual levels. Specifically, a Global-Local Decoupling (GLD) module decomposes textual inputs into local semantic tokens and global contextual tokens, while image inputs are partitioned into class-agnostic mask proposals. Then, a Local Visual-Textual Alignment (LVTA) module adaptively extracts context-aware visual features from the mask proposals and enriches textual features through knowledge-guided prompt engineering, achieving OVSS from a local perspective. Furthermore, a Global Visual-Textual Alignment (GVTA) module employs a global-enhanced Grad-CAM mechanism to capture contextual cues for referring expressions, followed by a mask selection module that integrates pixel-level activations into mask-level segmentation outputs, thereby achieving RES from a global perspective. Experiments on the iSAID (OVSS) and RRSIS-D (RES) benchmarks demonstrate that DGL-RSIS outperforms existing training-free approaches. Ablation studies further validate the effectiveness of each module. To the best of our knowledge, this is the first unified training-free framework for RS image segmentation, which effectively transfers the semantic capability of VLMs trained on natural images to the RS domain without additional training.

cs.CV↗

arXiv2025

Staircase Streaming for Low-Latency Multi-Agent Inference

Recent advances in large language models (LLMs) opened up new directions for leveraging the collective expertise of multiple LLMs. These methods, such as Mixture-of-Agents, typically employ additional inference steps to generate intermediate outputs, which are then used to produce the final response. While multi-agent inference can enhance response quality, it can significantly increase the time to first token (TTFT), posing a challenge for latency-sensitive applications and hurting user experience. To address this issue, we propose staircase streaming for low-latency multi-agent inference. Instead of waiting for the complete intermediate outputs from previous steps, we begin generating the final response as soon as we receive partial outputs from these steps. Experimental results demonstrate that staircase streaming reduces TTFT by up to 93% while maintaining response quality.

cs.AI↗

arXiv2025

Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models

While recent Large Vision-Language Models (LVLMs) have shown remarkable performance in multi-modal tasks, they are prone to generating hallucinatory text responses that do not align with the given visual input, which restricts their practical applicability in real-world scenarios. In this work, inspired by the observation that the text-to-image generation process is the inverse of image-conditioned response generation in LVLMs, we explore the potential of leveraging text-to-image generative models to assist in mitigating hallucinations in LVLMs. We discover that generative models can offer valuable self-feedback for mitigating hallucinations at both the response and token levels. Building on this insight, we introduce self-correcting Decoding with Generative Feedback (DeGF), a novel training-free algorithm that incorporates feedback from text-to-image generative models into the decoding process to effectively mitigate hallucinations in LVLMs. Specifically, DeGF generates an image from the initial response produced by LVLMs, which acts as an auxiliary visual reference and provides self-feedback to verify and correct the initial response through complementary or contrastive decoding. Extensive experimental results validate the effectiveness of our approach in mitigating diverse types of hallucinations, consistently surpassing state-of-the-art methods across six benchmarks. Code is available at https://github.com/zhangce01/DeGF.

cs.CV↗

arXiv2025

Enhancing Visual Planning with Auxiliary Tasks and Multi-token Prediction

Visual Planning for Assistance (VPA) aims to predict a sequence of user actions required to achieve a specified goal based on a video showing the user's progress. Although recent advances in multimodal large language models (MLLMs) have shown promising results in video understanding, long-horizon visual planning remains a challenging problem. We identify two challenges in training large MLLMs for video-based planning tasks: (1) scarcity of procedural annotations, limiting the model's ability to learn procedural task dynamics effectively, and (2) inefficiency of next-token prediction objective to explicitly capture the structured action space for visual planning when compared to free-form, natural language. To tackle data scarcity, we introduce Auxiliary Task Augmentation. We design and train our model on auxiliary tasks relevant to long-horizon video-based planning (e.g., goal prediction) to augment the model's planning ability. To more explicitly model the structured action space unique to visual planning tasks, we leverage Multi-token Prediction, extending traditional next-token prediction by using multiple heads to predict multiple future tokens during training. Our approach, VideoPlan, achieves state-of-the-art VPA performance on the COIN and CrossTask datasets, surpassing prior methods by 7.3% and 3.4%, respectively, when predicting 3 future actions. We further extend our method to the challenging Ego4D Long-term Action Anticipation task, and show that it is on par with the state-of-the-art approaches despite not using specialized egocentric features. Code will be made available.

cs.CV↗

arXiv2025

Data Diversification Methods In Alignment Enhance Math Performance In LLMs

While recent advances in preference learning have enhanced alignment in human feedback, mathematical reasoning remains a persistent challenge. We investigate how data diversification strategies in preference optimization can improve the mathematical reasoning abilities of large language models (LLMs). We evaluate three common data generation methods: temperature sampling, Chain-of-Thought prompting, and Monte Carlo Tree Search (MCTS), and introduce Diversified-ThinkSolve (DTS), a novel structured approach that systematically decomposes problems into diverse reasoning paths. Our results show that with strategically diversified preference data, models can substantially improve mathematical reasoning performance, with the best approach yielding gains of 7.1% on GSM8K and 4.2% on MATH over the base model. Despite its strong performance, DTS incurs only a marginal computational overhead (1.03x) compared to the baseline, while MCTS is nearly five times more costly with lower returns. These findings demonstrate that structured exploration of diverse problem-solving methods creates more effective preference data for mathematical alignment than traditional approaches.

cs.AI↗

arXiv2025

ONLY: One-Layer Intervention Sufficiently Mitigates Hallucinations in Large Vision-Language Models

Recent Large Vision-Language Models (LVLMs) have introduced a new paradigm for understanding and reasoning about image input through textual responses. Although they have achieved remarkable performance across a range of multi-modal tasks, they face the persistent challenge of hallucination, which introduces practical weaknesses and raises concerns about their reliable deployment in real-world applications. Existing work has explored contrastive decoding approaches to mitigate this issue, where the output of the original LVLM is compared and contrasted with that of a perturbed version. However, these methods require two or more queries that slow down LVLM response generation, making them less suitable for real-time applications. To overcome this limitation, we propose ONLY, a training-free decoding approach that requires only a single query and a one-layer intervention during decoding, enabling efficient real-time deployment. Specifically, we enhance textual outputs by selectively amplifying crucial textual information using a text-to-visual entropy ratio for each token. Extensive experimental results demonstrate that our proposed ONLY consistently outperforms state-of-the-art methods across various benchmarks while requiring minimal implementation effort and computational cost. Code is available at https://github.com/zifuwan/ONLY.

cs.CV↗

arXiv2025

HAMburger: Accelerating LLM Inference via Token Smashing

The growing demand for efficient Large Language Model (LLM) inference requires a holistic optimization on algorithms, systems, and hardware. However, very few works have fundamentally changed the generation pattern: each token needs one forward pass and one KV cache. This can be sub-optimal because we found that LLMs are extremely capable of self-identifying the exact dose of information that a single KV cache can store, and many tokens can be generated confidently without global context. Based on this insight, we introduce HAMburger, a Hierarchically Auto-regressive Model that redefines resource allocation in LLMs by moving beyond uniform computation and storage per token during inference. Stacking a compositional embedder and a micro-step decoder in between a base LLM, HAMburger smashes multiple tokens into a single KV and generates several tokens per step. Additionally, HAMburger functions as a speculative decoding framework where it can blindly trust self-drafted tokens. As a result, HAMburger shifts the growth of KV cache and forward FLOPs from linear to sub-linear with respect to output length, and adjusts its inference speed based on query perplexity and output structure. Extensive evaluations show that HAMburger reduces the KV cache computation by up to 2$\times$ and achieves up to 2$\times$ TPS, while maintaining quality in both short- and long-context tasks. Our method explores an extremely challenging inference regime that requires both computation- and memory-efficiency with a hardware-agnostic design.

cs.CL↗

arXiv2025

InstructPart: Task-Oriented Part Segmentation with Instruction Reasoning

Large multimodal foundation models, particularly in the domains of language and vision, have significantly advanced various tasks, including robotics, autonomous driving, information retrieval, and grounding. However, many of these models perceive objects as indivisible, overlooking the components that constitute them. Understanding these components and their associated affordances provides valuable insights into an object's functionality, which is fundamental for performing a wide range of tasks. In this work, we introduce a novel real-world benchmark, InstructPart, comprising hand-labeled part segmentation annotations and task-oriented instructions to evaluate the performance of current models in understanding and executing part-level tasks within everyday contexts. Through our experiments, we demonstrate that task-oriented part segmentation remains a challenging problem, even for state-of-the-art Vision-Language Models (VLMs). In addition to our benchmark, we introduce a simple baseline that achieves a twofold performance improvement through fine-tuning with our dataset. With our dataset and benchmark, we aim to facilitate research on task-oriented part segmentation and enhance the applicability of VLMs across various domains, including robotics, virtual reality, information retrieval, and other related fields. Project website: https://zifuwan.github.io/InstructPart/.

cs.CV↗

arXiv2025

Spectral-Aware Global Fusion for RGB-Thermal Semantic Segmentation

Semantic segmentation relying solely on RGB data often struggles in challenging conditions such as low illumination and obscured views, limiting its reliability in critical applications like autonomous driving. To address this, integrating additional thermal radiation data with RGB images demonstrates enhanced performance and robustness. However, how to effectively reconcile the modality discrepancies and fuse the RGB and thermal features remains a well-known challenge. In this work, we address this challenge from a novel spectral perspective. We observe that the multi-modal features can be categorized into two spectral components: low-frequency features that provide broad scene context, including color variations and smooth areas, and high-frequency features that capture modality-specific details such as edges and textures. Inspired by this, we propose the Spectral-aware Global Fusion Network (SGFNet) to effectively enhance and fuse the multi-modal features by explicitly modeling the interactions between the high-frequency, modality-specific features. Our experimental results demonstrate that SGFNet outperforms the state-of-the-art methods on the MFNet and PST900 datasets.

cs.CV↗

arXiv2025

Speculative Prefill: Turbocharging TTFT with Lightweight and Training-Free Token Importance Estimation

Improving time-to-first-token (TTFT) is an essentially important objective in modern large language model (LLM) inference engines. Optimizing TTFT directly results in higher maximal QPS and meets the requirements of many critical applications. However, boosting TTFT is notoriously challenging since it is compute-bounded and the performance bottleneck shifts from the self-attention that many prior works focus on to the MLP part. In this work, we present SpecPrefill, a training free framework that accelerates the inference TTFT for both long and medium context queries based on the following insight: LLMs are generalized enough to preserve the quality given only a carefully chosen subset of prompt tokens. At its core, SpecPrefill leverages a lightweight model to speculate locally important tokens based on the context. These tokens, along with the necessary positional information, are then sent to the main model for processing. We evaluate SpecPrefill with a diverse set of tasks, followed by a comprehensive benchmarking of performance improvement both in a real end-to-end setting and ablation studies. SpecPrefill manages to serve Llama-3.1-405B-Instruct-FP8 with up to 7$\times$ maximal end-to-end QPS on real downstream tasks and 7.66$\times$ TTFT improvement.

cs.CL↗

arXiv2025

Improving Model Alignment Through Collective Intelligence of Open-Source LLMS

Building helpful and harmless large language models (LLMs) requires effective model alignment approach based on human instructions and feedback, which necessitates high-quality human-labeled data. Constructing such datasets is often expensive and hard to scale, and may face potential limitations on diversity and generalization. To address these challenges, we introduce Mixture of Agents Alignment (MoAA), that leverages the collective strengths of various language models to provide high-quality data for model alignment. By employing MoAA, we enhance both supervised fine-tuning and preference optimization, leading to improved performance compared to using a single model alone to generate alignment data (e.g. using GPT-4o alone). Evaluation results show that our approach can improve win rate of LLaMA-3.1-8B-Instruct from 19.5 to 48.3 on Arena-Hard and from 22.33 to 57.23 on AlpacaEval2, highlighting a promising direction for model alignment through this new scalable and diverse synthetic data recipe. Furthermore, we demonstrate that MoAA enables a self-improvement pipeline, where models finetuned on MoA-generated data surpass their own initial capabilities, providing evidence that our approach can push the frontier of open-source LLMs without reliance on stronger external supervision. Data and code will be released.

cs.CL↗

arXiv2025

Partition laser assembling technique

The advancement of micro/nanofabrication techniques with high throughput, efficiency, and flexibility is critical for fields like integrated photonics, biosensing, and medical diagnostics. This study presents Partition Laser Assembling (PLA), a novel laser technique for fabricating complex micro/nanostructures akin to puzzle pieces. By dividing the target patterns described by scalable vector graphics into partitions, any structures in each partition can be fabricated via structured lights as "light stamp" through spatial light modulation. Unlike traditional direct laser writing, PLA eliminates reliance on mechanical components, avoiding step-like artifacts and ensuring smoother fabrication of complex micro/nanostructures. By seamlessly assembling basic shapes, PLA achieves intricate structures like micro artworks and metalenses with unmatched precision and resolution. Leveraging two-photon fabrication, PLA guarantees high resolution and structural integrity, positioning it as a transformative tool for nanoscale 3D printing. With applications spanning research and industry, PLA paves the way for advanced optical devices, micro/nanofabrications, and next-gen manufacturing technologies.

physics.optics↗

arXiv2025

Scaling Instruction-Tuned LLMs to Million-Token Contexts via Hierarchical Synthetic Data Generation

Large Language Models (LLMs) struggle with long-context reasoning, not only due to the quadratic scaling of computational complexity with sequence length but also because of the scarcity and expense of annotating long-context data. There has been barely any open-source work that systematically ablates long-context data, nor is there any openly available instruction tuning dataset with contexts surpassing 100K tokens. To bridge this gap, we introduce a novel post-training synthetic data generation strategy designed to efficiently extend the context window of LLMs while preserving their general task performance. Our approach scalably extends to arbitrarily long context lengths, unconstrained by the length of available real-world data, which effectively addresses the scarcity of raw long-context data. Through a step-by-step rotary position embedding (RoPE) scaling training strategy, we demonstrate that our model, with a context length of up to 1M tokens, performs well on the RULER benchmark and InfiniteBench and maintains robust performance on general language tasks.

cs.CL↗

arXiv2025

Symmetry breaking of large-amplitude parametric oscillations in few-layer graphene nanomechanical resonators

Graphene nanomechanical resonators are well suited for the study of parametric oscillations. Their large frequency tunability and their pronounced nonlinearities enable an efficient modulation of their resonant frequencies. Here, we present measurements of the response of few-layer graphene nanomechanical resonators, each driven by a large parametric pump at frequency $2ω$ and a weak external drive at $ω$, where $ω$ is set near the mechanical resonant frequency $ω_0$. The pump actuates the resonator beyond the threshold for large-amplitude parametric oscillations, while the drive breaks the symmetry between the parametric phase states. By increasing and decreasing a gate voltage to detune $ω_0$ in the presence of the pump and the drive, we observe a double hysteresis in the response. The double hysteresis reveals the existence of two possible large-amplitude vibrational states whose phase difference is nearly $π$ radians. We deterministically prepare the resonator in either one of these states by cycling the gate voltage. We measure the stationary occupation probabilities of the two states in the presence of a white Gaussian force noise, and find that they strongly depend on the amplitude and on the phase of the external drive. The phase states of parametric oscillations with broken amplitude symmetry can be mapped to biased bi-modal degrees of freedom, such as Ising spins in an external magnetic field. Therefore, they hold promise as units of binary information.

cond-mat.mes-hall↗