arXiv subjects

Yifan Hao

Publications and source records attributed to Yifan Hao.

At least 19 recordsLinked to original sources

arXiv2026

Retrieval-Guided Fine-Tuning as Noisy Estimation: Risk bounds and Architectural Analysis

Retrieval-Guided Fine-Tuning (RAG-FT) incorporates retrieved data directly into the training objective, but the statistical consequences of noisy retrieval during training remain theoretically undercharacterized. We study this question by modeling RAG-FT as an estimation problem in a multi-task linear regression framework, using an OLS proxy for single-layer linear self-attention to obtain finite-sample risk bounds. Under homoscedastic retrieval noise, we show that retrieval failure decays exponentially with task separation relative to noise, and derive explicit finite-sample conditions under which RAG-FT achieves lower risk than both target-only and full-corpus training. We then introduce a Distance-Proportional Noise (DPN) model, in which retrieval quality degrades with rank, and compare two estimators under the same retrieval process: the OLS proxy and the literal, uniform-weight forward pass of linear self-attention. We prove that the attention estimator's bias diverges as $Θ(n^{2q})$ even under exact retrieval, while OLS risk remains $Θ(d/n)$ for every noise exponent $q>0$. These results locate the instability not in noisy retrieval itself, but in the fixed, unweighted aggregation of the literal LSA forward pass, which reweighting by reliability empirically removes. We validate the predicted rate separation through direct simulation of the DPN model.

cs.LG↗

arXiv2026

Task Vector Bases: A Unified and Scalable Framework for Compressed Task Arithmetic

Task arithmetic, representing downstream tasks through linear operations on task vectors, has emerged as a simple yet powerful paradigm for transferring knowledge across diverse settings. However, maintaining a large collection of task vectors introduces scalability challenges in both storage and computation. We propose Task Vector Bases, a framework compressing $T$ task vectors into $M < T$ basis vectors while preserving the functionality of task arithmetic. By representing each task vector as a structured linear combination of basis atoms, our approach supports standard operations such as addition, negation, as well as more advanced arithmetic ones. The framework is orthogonal to other efficiency-oriented improvements in task arithmetic and can be used in combination with them. We provide theoretical analysis showing that basis compression retains addition generalization guarantees and enables principled unlearning, with error bounds depending on reconstruction quality. Empirically, our proposed basis construction methods consistently outperform heuristic basis construction baselines and, in some cases, even surpass the performance of full task vector collections across diverse downstream applications while reducing storage and computational requirements. The code is available at https://github.com/uiuctml/TaskVectorBasis.

cs.LG↗

arXiv2026

Adaptive Layerwise Perturbation: Unifying Off-Policy Corrections for LLM RL

Off-policy problems such as policy staleness and training--inference mismatch have become a major bottleneck for training stability and further exploration in LLM RL. The distribution gap between the inference and updated policies grows because of the techniques to enhance inference efficiency, leading to heavy-tailed importance ratios. Heavy-tailed ratios arise when the policy is locally sharp, which further inflates gradients and can push updates outside the trust region. To address this, we propose Adaptive Layerwise Perturbation (ALP), which injects small learnable perturbations into the input hidden states of each layer during updates and uses the resulting perturbed policy as the numerator of the importance ratio against the unchanged inference policy in the objective. Intuitively, by adding controlled noise to intermediate representations, ALP prevents the updated policy from deviating too sharply from the inference policy and enlarges the policy family to cover inference-time mismatch noise. Hence, the flattened distribution can naturally tighten the gap between the updated and inference policies and reduce the tail of importance ratios, thus maintaining training stability. This is further validated empirically. Experiments on single-turn math and multi-turn tool-integrated reasoning tasks show that ALP not only improves final performance, but also avoids blow-up in the importance-ratio tail and KL spikes during iterative training, along with boosted exploration. Ablations show that representation-level perturbations across all layers are most effective, substantially outperforming partial-layer and logits-only variants.

cs.LG↗

arXiv2026

LKV: End-to-End Learning of Head-wise Budgets and Token Selection for LLM KV Cache Eviction

Long-context inference in Large Language Models (LLMs) is bottlenecked by the linear growth of Key-Value (KV) cache memory. Existing KV cache compression paradigms are fundamentally limited by heuristics: heuristic budgeting relies on statistical priors rather than task objectives, causing resource misallocation, while heuristic selection relies on coupled query-key interactions or static inductive biases (e.g., attention sinks). To address this limitation, we introduce LKV (Learned KV Eviction), which formulates KV compression as an end-to-end differentiable optimization problem. LKV integrates LKV-H to learn task-optimized global budgets, and LKV-T to derive intrinsic KV importance without materializing attention matrices. This design bypasses heuristic proxies, strictly aligning compression with task objectives. Extensive evaluations demonstrate that LKV achieves state-of-the-art performance on both LongBench and RULER benchmarks at high compression rates. In particular, on LongBench, LKV achieves near-lossless performance with only 15\% KV cache retention. Crucially, our analysis identifies learned budgeting as the dominant driver of fidelity, demonstrating that data-driven allocation is essential to overcome the limitations of hand-crafted heuristics.

cs.LG↗

arXiv2026

DA-Mamba: Learning Domain-Aware State Space Model for Global-Local Alignment in Domain Adaptive Object Detection

Domain Adaptive Object Detection (DAOD) aims to transfer detectors from a labeled source domain to an unlabeled target domain. Existing DAOD methods employ multi-granularity feature alignment to learn domain-invariant representations. However, the local connectivity of their CNN-based backbone and detection head restricts alignment to local regions, failing to extract global domain-invariant features. Although transformer-based DAOD methods capture global dependencies via attention mechanisms, their quadratic computational cost hinders practical deployment. To solve this, we propose DA-Mamba, a hybrid CNN-State Space Models (SSMs) architecture that combines the efficiency of CNNs with the linear-time long-range modeling capability of State Space Models (SSMs) to capture both global and local domain-invariant features. Specifically, we introduce two novel modules: Image-Aware SSM (IA-SSM) and Object-Aware SSM (OA-SSM). IA-SSM is integrated into the backbone to enhance global domain awareness, enabling image-level global and local alignment. OA-SSM is inserted into the detection head to model spatial and semantic dependencies among objects, enhancing instance-level alignment. Comprehensive experiments demonstrate that the proposed method can efficiently improve the cross-domain performance of the object detector.

cs.CV↗

arXiv2026

Hardwired-Neurons Language Processing Units as General-Purpose Cognitive Substrates

The rapid advancement of Large Language Models (LLMs) has established language as a core general-purpose cognitive substrate, driving the demand for specialized Language Processing Units (LPUs) tailored for LLM inference. To overcome the growing energy consumption of LLM inference systems, this paper proposes a Hardwired-Neurons Language Processing Unit (HNLPU), which physically hardwires LLM weight parameters into the computational fabric, achieving several orders of magnitude computational efficiency improvement by extreme specialization. However, a significant challenge still lies in the scale of modern LLMs. A straightforward hardwiring of gpt-oss 120 B would require fabricating photomask sets valued at over 6 billion dollars, rendering this straightforward solution economically impractical. Addressing this challenge, we propose the novel Metal-Embedding methodology. Instead of embedding weights in a 2D grid of silicon device cells, Metal-Embedding embeds weight parameters into the 3D topology of metal wires. This brings two benefits: (1) a 15x increase in density, and (2) 60 out of 70 photomask layers are homogeneous across chips, including all EUV photomasks. In total, Metal-Embedding reduced the photomask cost by 112x, bringing the Non-Recurring Engineering (NRE) cost of HNLPU into an economically viable range. Experimental results show that HNLPU achieved 249,960 tokens/s (5,555x/85x that of GPU/WSE), 36 tokens/J (1,047x/283x that of GPU/WSE), 13,232 mm2 total die area, $59.46 M-123.5 M estimated NRE at 5 nm technology. Analysis shows that HNLPU achieved 41.7-80.4x improvement in cost-effectiveness and 357x reduction in carbon footprint compared to OpenAI-scale H100 clusters, under an annual weight updating assumption.

cs.AR↗

arXiv2025

Towards Better Generalization via Distributional Input Projection Network

As overparameterized models become increasingly prevalent, training loss alone offers limited insight into generalization performance. While smoothness has been linked to improved generalization across various settings, directly enforcing smoothness in neural networks remains challenging. To address this, we introduce Distributional Input Projection Networks (DIPNet), a novel framework that projects inputs into learnable distributions at each layer. This distributional representation induces a smoother loss landscape with respect to the input, promoting better generalization. We provide theoretical analysis showing that DIPNet reduces both local smoothness measures and the Lipschitz constant of the network, contributing to improved generalization performance. Empirically, we validate DIPNet across a wide range of architectures and tasks, including Vision Transformers (ViTs), Large Language Models (LLMs), ResNet and MLPs. Our method consistently enhances test performance under standard settings, adversarial attacks, out-of-distribution inputs, and reasoning benchmarks. We demonstrate that the proposed input projection strategy can be seamlessly integrated into existing models, providing a general and effective approach for boosting generalization performance in modern deep learning.

cs.LG↗

arXiv2025

JT-Math: A Multi-Stage Framework for Advanced Mathematical Reasoning in Large Language Models

Mathematical reasoning is a cornerstone of artificial general intelligence and a primary benchmark for evaluating the capabilities of Large Language Models (LLMs). While state-of-the-art models show promise, they often falter when faced with complex problems that demand deep conceptual understanding and intricate, multi-step deliberation. To address this challenge, we introduce JT-Math-8B, a series of open-source models comprising base, instruct, and thinking versions, built upon a systematic, multi-stage optimization framework. Our pre-training corpus is a high-quality, 210B-token dataset curated through a dedicated data pipeline that uses model-based validation to ensure quality and diversity. The Instruct Model is optimized for direct, concise answers through Supervised Fine-Tuning (SFT) and a GRPO-based reinforcement learning (RL) method. The Thinking Model is trained for complex problem-solving using a Long Chain-of-Thought (Long CoT) approach, combining SFT with a novel, multi-stage RL curriculum that progressively increases task difficulty and context length up to 32K tokens. JT-Math-8B achieves state-of-the-art results among open-source models of similar size, surpassing prominent models like OpenAI's O1-mini and GPT-4o , and demonstrating superior performance on competition-level mathematics.

cs.CL↗

arXiv2025

Sycophancy in Vision-Language Models: A Systematic Analysis and an Inference-Time Mitigation Framework

Large Vision-Language Models (LVLMs) have shown significant capability in vision-language understanding. However, one critical issue that persists in these models is sycophancy, where models are unduly influenced by leading or deceptive prompts, resulting in biased outputs and hallucinations. Despite the rapid development of LVLMs, evaluating and mitigating sycophancy remains largely under-explored. In this work, we fill this gap by systematically analyzing sycophancy across multiple vision-language benchmarks and propose an inference-time mitigation framework. We curate leading queries and quantify the susceptibility of state-of-the-art LVLMs to prompt-induced bias, revealing consistent performance degradation and instability across models and tasks. Our analysis further uncovers model-specific behavioral traits, such as sentiment sensitivity and prediction polarity shifts under sycophancy. To mitigate these issues, we propose a training-free, model-agnostic framework that operates entirely at inference time. Our approach first employs a query neutralizer, leveraging an language model to suppress implicit sycophantic bias in user queries. We then introduce a sycophancy-aware contrastive decoding mechanism that dynamically recalibrates token-level output distributions by contrasting responses to neutralized and leading queries. Finally, an adaptive logits refinement module further modifies the contrasted logits by integrating both a adaptive plausibility filter and query sentiment scaler, ensuring coherent and robust generation. Extensive experiments demonstrate that this framework effectively mitigates sycophancy across all evaluated models, while maintaining performance on neutral prompts. Our results suggest that sycophancy in LVLMs is a general and urgent challenge, and that inference-time strategies offer a promising path toward trustworthy multimodal reasoning.

cs.AI↗

arXiv2025

QiMeng: Fully Automated Hardware and Software Design for Processor Chip

Processor chip design technology serves as a key frontier driving breakthroughs in computer science and related fields. With the rapid advancement of information technology, conventional design paradigms face three major challenges: the physical constraints of fabrication technologies, the escalating demands for design resources, and the increasing diversity of ecosystems. Automated processor chip design has emerged as a transformative solution to address these challenges. While recent breakthroughs in Artificial Intelligence (AI), particularly Large Language Models (LLMs) techniques, have opened new possibilities for fully automated processor chip design, substantial challenges remain in establishing domain-specific LLMs for processor chip design. In this paper, we propose QiMeng, a novel system for fully automated hardware and software design of processor chips. QiMeng comprises three hierarchical layers. In the bottom-layer, we construct a domain-specific Large Processor Chip Model (LPCM) that introduces novel designs in architecture, training, and inference, to address key challenges such as knowledge representation gap, data scarcity, correctness assurance, and enormous solution space. In the middle-layer, leveraging the LPCM's knowledge representation and inference capabilities, we develop the Hardware Design Agent and the Software Design Agent to automate the design of hardware and software for processor chips. Currently, several components of QiMeng have been completed and successfully applied in various top-layer applications, demonstrating significant advantages and providing a feasible solution for efficient, fully automated hardware/software design of processor chips. Future research will focus on integrating all components and performing iterative top-down and bottom-up design processes to establish a comprehensive QiMeng system.

cs.AR↗

arXiv2025

Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods

Supervised fine-tuning (SFT) on domain-specific data is the dominant approach for adapting foundation models to specialized tasks. However, it has been observed that SFT models tend to forget knowledge acquired during pretraining. In vision models, ensembling a pretrained model with its fine-tuned counterpart has been shown to mitigate this issue. In this work, we demonstrate that the same holds for language models, and, more strikingly, we observe an overadaptation phenomenon: the ensemble model not only retains general knowledge from the foundation model but also outperforms the fine-tuned model even on the fine-tuning domain itself. Despite the empirical success of ensembling, a theoretical understanding of its benefits remains underexplored. We develop a formal theoretical analysis of the overadaptation phenomenon. Ensembling mitigates this by balancing two primary sources of error: bias, caused by insufficient fine-tuning, and variance, introduced by overfitting to fine-tuning data. While regularization techniques aim to address this trade-off, we show that ensembling provides a more effective solution. We analyze this phenomenon in over-parameterized linear settings and demonstrate that interpolating between pretrained and fine-tuned weights significantly improves performance. These findings offer theoretical justification for the observed advantages of model ensembling, supported by empirical experiments consistent with our analysis.

cs.AI↗

arXiv2025

Transformers as Multi-task Learners: Decoupling Features in Hidden Markov Models

Transformer based models have shown remarkable capabilities in sequence learning across a wide range of tasks, often performing well on specific task by leveraging input-output examples. Despite their empirical success, a comprehensive theoretical understanding of this phenomenon remains limited. In this work, we investigate the layerwise behavior of Transformers to uncover the mechanisms underlying their multi-task generalization ability. Taking explorations on a typical sequence model, i.e, Hidden Markov Models, which are fundamental to many language tasks, we observe that: first, lower layers of Transformers focus on extracting feature representations, primarily influenced by neighboring tokens; second, on the upper layers, features become decoupled, exhibiting a high degree of time disentanglement. Building on these empirical insights, we provide theoretical analysis for the expressiveness power of Transformers. Our explicit constructions align closely with empirical observations, providing theoretical support for the Transformer's effectiveness and efficiency on sequence learning across diverse tasks.

cs.LG↗

arXiv2025

QiMeng-CPU-v2: Automated Superscalar Processor Design by Learning Data Dependencies

Automated processor design, which can significantly reduce human efforts and accelerate design cycles, has received considerable attention. While recent advancements have automatically designed single-cycle processors that execute one instruction per cycle, their performance cannot compete with modern superscalar processors that execute multiple instructions per cycle. Previous methods fail on superscalar processor design because they cannot address inter-instruction data dependencies, leading to inefficient sequential instruction execution. This paper proposes a novel approach to automatically designing superscalar processors using a hardware-friendly model called the Stateful Binary Speculation Diagram (State-BSD). We observe that processor parallelism can be enhanced through on-the-fly inter-instruction dependent data predictors, reusing the processor's internal states to learn the data dependency. To meet the challenge of both hardware-resource limitation and design functional correctness, State-BSD consists of two components: 1) a lightweight state-selector trained by the simulated annealing method to detect the most reusable processor states and store them in a small buffer; and 2) a highly precise state-speculator trained by the BSD expansion method to predict the inter-instruction dependent data using the selected states. It is the first work to achieve the automated superscalar processor design, i.e. QiMeng-CPU-v2, which improves the performance by about $380\times$ than the state-of-the-art automated design and is comparable to human-designed superscalar processors such as ARM Cortex A53.

cs.AR↗

arXiv2025

Optimizing Chain-of-Thought Reasoners via Gradient Variance Minimization in Rejection Sampling and RL

Chain-of-thought (CoT) reasoning in large language models (LLMs) can be formalized as a latent variable problem, where the model needs to generate intermediate reasoning steps. While prior approaches such as iterative reward-ranked fine-tuning (RAFT) have relied on such formulations, they typically apply uniform inference budgets across prompts, which fails to account for variability in difficulty and convergence behavior. This work identifies the main bottleneck in CoT training as inefficient stochastic gradient estimation due to static sampling strategies. We propose GVM-RAFT, a prompt-specific Dynamic Sample Allocation Strategy designed to minimize stochastic gradient variance under a computational budget constraint. The method dynamically allocates computational resources by monitoring prompt acceptance rates and stochastic gradient norms, ensuring that the resulting gradient variance is minimized. Our theoretical analysis shows that the proposed dynamic sampling strategy leads to accelerated convergence guarantees under suitable conditions. Experiments on mathematical reasoning show that GVM-RAFT achieves a 2-4x speedup and considerable accuracy improvements over vanilla RAFT. The proposed dynamic sampling strategy is general and can be incorporated into other reinforcement learning algorithms, such as GRPO, leading to similar improvements in convergence and test accuracy. Our code is available at https://github.com/RLHFlow/GVM.

cs.LG↗

arXiv2025

QiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a Neural-Symbolic Approach

Heterogeneous deep learning systems (DLS) such as GPUs and ASICs have been widely deployed in industrial data centers, which requires to develop multiple low-level tensor programs for different platforms. An attractive solution to relieve the programming burden is to transcompile the legacy code of one platform to others. However, current transcompilation techniques struggle with either tremendous manual efforts or functional incorrectness, rendering "Write Once, Run Anywhere" of tensor programs an open question. We propose a novel transcompiler, i.e., QiMeng-Xpiler, for automatically translating tensor programs across DLS via both large language models (LLMs) and symbolic program synthesis, i.e., neural-symbolic synthesis. The key insight is leveraging the powerful code generation ability of LLM to make costly search-based symbolic synthesis computationally tractable. Concretely, we propose multiple LLM-assisted compilation passes via pre-defined meta-prompts for program transformation. During each program transformation, efficient symbolic program synthesis is employed to repair incorrect code snippets with a limited scale. To attain high performance, we propose a hierarchical auto-tuning approach to systematically explore both the parameters and sequences of transformation passes. Experiments on 4 DLS with distinct programming interfaces, i.e., Intel DL Boost with VNNI, NVIDIA GPU with CUDA, AMD MI with HIP, and Cambricon MLU with BANG, demonstrate that QiMeng-Xpiler correctly translates different tensor programs at the accuracy of 95% on average, and the performance of translated programs achieves up to 2.0x over vendor-provided manually-optimized libraries. As a result, the programming productivity of DLS is improved by up to 96.0x via transcompiling legacy tensor programs.

cs.CL↗

arXiv2025

Object-Level Verbalized Confidence Calibration in Vision-Language Models via Semantic Perturbation

Vision-language models (VLMs) excel in various multimodal tasks but frequently suffer from poor calibration, resulting in misalignment between their verbalized confidence and response correctness. This miscalibration undermines user trust, especially when models confidently provide incorrect or fabricated information. In this work, we propose a novel Confidence Calibration through Semantic Perturbation (CSP) framework to improve the calibration of verbalized confidence for VLMs in response to object-centric queries. We first introduce a perturbed dataset where Gaussian noise is applied to the key object regions to simulate visual uncertainty at different confidence levels, establishing an explicit mapping between visual ambiguity and confidence levels. We further enhance calibration through a two-stage training process combining supervised fine-tuning on the perturbed dataset with subsequent preference optimization. Extensive experiments on popular benchmarks demonstrate that our method significantly improves the alignment between verbalized confidence and response correctness while maintaining or enhancing overall task performance. These results highlight the potential of semantic perturbation as a practical tool for improving the reliability and interpretability of VLMs.

cs.CV↗

arXiv2025

SPDiffusion: Semantic Protection Diffusion Models for Multi-concept Text-to-image Generation

Recent text-to-image models have achieved impressive results in generating high-quality images. However, when tasked with multi-concept generation creating images that contain multiple characters or objects, existing methods often suffer from semantic entanglement, including concept entanglement and improper attribute binding, leading to significant text-image inconsistency. We identify that semantic entanglement arises when certain regions of the latent features attend to incorrect concept and attribute tokens. In this work, we propose the Semantic Protection Diffusion Model (SPDiffusion) to address both concept entanglement and improper attribute binding using only a text prompt as input. The SPDiffusion framework introduces a novel concept region extraction method SP-Extraction to resolve region entanglement in cross-attention, along with SP-Attn, which protects concept regions from the influence of irrelevant attributes and concepts. To evaluate our method, we test it on existing benchmarks, where SPDiffusion achieves state-of-the-art results, demonstrating its effectiveness.

cs.CV↗

arXiv2025

AGON: Automated Design Framework for Customizing Processors from ISA Documents

Customized processors are attractive solutions for vast domain-specific applications due to their high energy efficiency. However, designing a processor in traditional flows is time-consuming and expensive. To address this, researchers have explored methods including the use of agile development tools like Chisel or SpinalHDL, high-level synthesis (HLS) from programming languages like C or SystemC, and more recently, leveraging large language models (LLMs) to generate hardware description language (HDL) code from natural language descriptions. However, each method has limitations in terms of expressiveness, correctness, and performance, leading to a persistent contradiction between the level of automation and the effectiveness of the design. Overall, how to automatically design highly efficient and practical processors with minimal human effort remains a challenge. In this paper, we propose AGON, a novel framework designed to leverage LLMs for the efficient design of out-of-order (OoO) customized processors with minimal human effort. Central to AGON is the nano-operator function (nOP function) based Intermediate Representation (IR), which bridges high-level descriptions and hardware implementations while decoupling functionality from performance optimization, thereby providing an automatic design framework that is expressive and efficient, has correctness guarantees, and enables PPA (Power, Performance, and Area) optimization. Experimental results show that superior to previous LLM-assisted automatic design flows, AGON facilitates designing a series of customized OoO processors that achieve on average 2.35 $\times$ speedup compared with BOOM, a general-purpose CPU designed by experts, with minimal design effort.

cs.AR↗