arXiv subjects

Li Shang

Publications and source records attributed to Li Shang.

At least 19 recordsLinked to original sources

arXiv2026

The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training

FP4 training promises substantial memory and compute savings for large language models, but remains fragile because blockwise quantization is dictated by extreme activation magnitudes, which inflate dynamic range and compress long-tail signals. We identify a counterintuitive source of this failure: dominant activation outliers are not merely arbitrary sparse events, but are largely induced by a coherent rank-one mean bias, whose direction aligns with the leading anisotropic spectral component. This mean component strengthens during training, is amplified and reshaped by attention and FFN operators, and increasingly dominates top activation magnitudes. Crucially, this discovery reveals that a seemingly complex outlier-suppression problem admits a truly simple solution: isolate the coherent mean before quantization. We therefore propose Averis, a mean-residual splitting quantization method that separates the mean component using only reductions and elementwise subtractions before FP4 quantization. Across Qwen3 0.6B Dense trained on 100B tokens and Qwen3 7B A1.5B MoE trained on 50B tokens, Averis enables robust W4A4G4 FP4 training, reducing BF16 loss gaps to 1.19%/0.81% versus 2.05%/1.10% for NVIDIA's recently released Hadamard-based outlier-smoothing method, while limiting downstream gaps to 0.89/0.71 points. With only 2.20% end-to-end overhead over vanilla NVFP4, about 30% of NVIDIA's Hadamard-based design, Averis provides a hardware-efficient path to stable low-bit LLM training. Complementary to Hadamard, Averis further reduces the Qwen3-0.6B loss and downstream gaps to 0.94% and 0.73 points when combined. Code is available at: https://anonymous.4open.science/r/averis-504D.

cs.LG↗

arXiv2026

SD-MoE: Spectral Decomposition for Effective Expert Specialization

Mixture-of-Experts (MoE) architectures scale Large Language Models via expert specialization induced by conditional computation. In practice, however, expert specialization often fails: some experts become functionally similar, while others functioning as de facto shared experts, limiting the effective capacity and model performance. In this work, we analysis from a spectral perspective on parameter and gradient spaces, uncover that (1) experts share highly overlapping dominant spectral components in their parameters, (2) dominant gradient subspaces are strongly aligned across experts, driven by ubiquitous low-rank structure in human corpus, and (3) gating mechanisms preferentially route inputs along these dominant directions, further limiting specialization. To address this, we propose Spectral-Decoupled MoE (SD-MoE), which decomposes both parameter and gradient in the spectral space. SD-MoE improves performance across downstream tasks, enables effective expert specialization, incurring minimal additional computation, and can be seamlessly integrated into a wide range of existing MoE architectures, including Qwen and DeepSeek.

cs.LG↗

arXiv2026

Multi-Head Attention as a Source of Catastrophic Forgetting in MoE Transformers

Mixture-of-Experts (MoE) architectures are often considered a natural fit for continual learning because sparse routing should localize updates and reduce interference, yet MoE Transformers still forget substantially even with sparse, well-balanced expert utilization. We attribute this gap to a pre-routing bottleneck: multi-head attention concatenates head-specific signals into a single post-attention router input, forcing routing to act on co-occurring feature compositions rather than separable head channels. We show that this router input simultaneously encodes multiple separately decodable semantic and structural factors with uneven head support, and that different feature compositions induce weakly aligned parameter-gradient directions; as a result, routing maps many distinct compositions to the same route. We quantify this collision effect via a route-wise effective composition number $N_{eff}$ and find that higher $N_{eff}$ is associated with larger old-task loss increases after continual training. Motivated by these findings, we propose MH-MoE, which performs head-wise routing over sub-representations to increase routing granularity and reduce composition collisions. On TRACE with Qwen3-0.6B/8B, MH-MoE effectively mitigates forgetting, reducing BWT on Qwen3-0.6B from 11.2% (LoRAMoE) to 4.5%.

cs.LG↗

arXiv2026

Dispelling the Curse of Singularities in Neural Network Optimizations

This work investigates the optimization instability of deep neural networks from a less-explored yet insightful perspective: the emergence and amplification of singularities in the parametric space. Our analysis reveals that parametric singularities inevitably grow with gradient updates and further intensify alignment with representations, leading to increased singularities in the representation space. We show that the gradient Frobenius norms are bounded by the top singular values of the weight matrices, and as training progresses, the mutually reinforcing growth of weight and representation singularities, termed the curse of singularities, relaxes these bounds, escalating the risk of sharp loss explosions. To counter this, we propose Parametric Singularity Smoothing (PSS), a lightweight, flexible, and effective method for smoothing the singular spectra of weight matrices. Extensive experiments across diverse datasets, architectures, and optimizers demonstrate that PSS mitigates instability, restores trainability even after failure, and improves both training efficiency and generalization.

cs.LG↗

arXiv2026

Spectra: Rethinking Optimizers for LLMs Under Spectral Anisotropy

Gradient signals in LLM training are highly anisotropic: recurrent linguistic structure concentrates energy into a small set of dominant spectral directions, while context specific information resides in a long tail. We show that this spike tail separation persists throughout training, with the spike occupying only about 1.5% of directions yet dominating optimizer statistics. This dominance suppresses tail learning by contracting tail updates through second moment normalization and tightening the globally stable learning rate bound. Motivated by this analysis, we propose Spectra, a spike aware optimizer that suppresses the dominant low rank spike subspace without amplifying the noise sensitive spectral tail. Spectra tracks the spike subspace via cached, warm started power iteration and applies low rank spectral shaping with negligible overhead and substantially reduced optimizer state memory. On LLaMA3 8B trained on 50B tokens, Spectra reaches the same target loss 30% faster than AdamW, reduces per step end to end overhead by 0.7%, cuts optimizer state memory by 49.25%, and improves average downstream accuracy by 1.62%. Compared to Muon, Spectra is 5.1x faster in optimizer processing time, achieves a lower final loss, and improves average accuracy by 0.66%.

cs.LG↗

arXiv2026

LLM-Assisted Op-Amp Behavioral-Level Design via Agentic Human-Mimicking Reasoning

This paper proposes White-Op, an operational amplifier (op-amp) behavioral-level parameter design framework assisted by the human-mimicking reasoning of large language model agents. A symbolic reasoning-numerical solving decoupled paradigm is adopted: the agent performs step-by-step symbolic reasoning and formulates the design as a white-box optimization problem, which is then solved programmatically, verified via simulation, and refined iteratively. To guide this symbolic design process, implicit human reasoning mechanisms are formalized into explicit steps of introducing hypothetical constraints during transfer function simplification, pole-zero extraction and position regulation, converting design heuristics into mathematical formulations. A programming mapping protocol then standardizes the translation from symbolic designs to executable programs. Finally, a causality-driven refinement loop enables the agent to trace simulation-theory mismatches back to specific symbolic reasoning steps and make targeted corrections iteratively until convergence. Experiments on 9 op-amp topologies demonstrate that White-Op achieves interpretable behavioral-level designs with an average of 8.52\% theoretical prediction error and retains circuit functionality after transistor-level mapping for all topologies, whereas black-box baselines fail in 5 to 7 topologies. White-Op is open-sourced at https://github.com/zhchenfdu/whiteop.

cs.AI↗

arXiv2025

Bridging the Initialization Gap: A Co-Optimization Framework for Mixed-Size Global Placement

Global placement is a critical step with high computational complexity in VLSI physical design. Modern analytical placers formulate the placement problem as a nonlinear optimization, where initialization strongly affects both convergence behavior and final placement quality. However, existing initialization methods exhibit a trade-off: area-aware initializers account for cell areas but are computationally expensive and can dominate total runtime, while fast point-based initializers ignore cell area, leading to a modeling gap that impairs convergence and solution quality. We propose a lightweight co-optimization framework that bridges this initialization gap through two strategies. First, an area-hint refinement initializer incorporates heuristic cell area information into a signed graph signal by augmenting the netlist graph with virtual nodes and negative-weight edges, yielding an area-aware and spectrally smooth placement initialization. Second, a macro-schedule placement procedure progressively restores area constraints, enabling a smooth transition from the refined initializer to the full area-aware objective and producing high-quality placement results. We evaluate the framework on macro-heavy ISPD2005 academic benchmarks and two real-world industrial designs across two technology nodes (12 cases in total). Experimental results show that our method consistently improves half-perimeter wirelength (HPWL) over point-based initializers in 11 out of 12 cases, achieving up to 2.2% HPWL reduction, while running approximately 100 times faster than the state-of-the-art area-aware initializer.

eess.SP↗

arXiv2025

Metis: Training LLMs with FP4 Quantization

This work identifies anisotropy in the singular value spectra of parameters, activations, and gradients as the fundamental barrier to low-bit training of large language models (LLMs). These spectra are dominated by a small fraction of large singular values, inducing wide numerical ranges that cause quantization bias and severe spectral distortion, ultimately degrading training performance. This work presents Metis, a spectral-domain quantization framework that partitions anisotropic spectra into narrower sub-distributions for independent quantization, thereby reducing errors and preserving spectral structure. To minimize overhead, Metis leverages two key properties of the dominant spectral subspace: preservation via sparsely random sampling and preservation via random projection, reducing decomposition cost to a negligible level. On LLaMA-3 8B trained with 100B tokens, Metis enables robust W4A4G4 training with FP4 quantization of weights, activations, and gradients, yielding only a 0.4% training loss gap and a 0.1% degradation in downstream accuracy relative to BF16. Beyond matching BF16 fidelity, Metis also surpasses our implementation of Nvidia's recently announced (yet to be publicly released) FP4 recipe, consistently achieving lower loss and higher downstream accuracy while incurring significantly lower computational overhead. The code implementation for Metis is available at: https://anonymous.4open.science/r/Metis-quantization-644B.

cs.LG↗

arXiv2025

Dataset Construction for Training LLM to Learn Analog Circuit Knowledge

This paper constructs a textual dataset for training large language models (LLMs) to learn analog circuit knowledge and customizes LLM training techniques. For dataset construction, high-quality textbooks are collected and decomposed into fine-grained learning nodes, which are then used to construct structured question-thinking-solution-answer (QTSA) quadruples using a multi-agent framework to capture both final answers and thought processes. The resulting dataset consists of 7.26M tokens of unlabeled data for continual pre-training (CPT) and 112.65M tokens of labeled data for supervised fine-tuning (SFT). We customize the training techniques including initial model selection, training paradigms, regularization techniques, and practical implementation references. Instruct models are identified as suitable training initialization points, an SFT-centric training paradigm is established (finding that CPT provides marginal benefits compared with SFT due to imbalanced data distribution), and SFT with KL divergence regularization can achieve a 2.71 percentage-point improvement over SFT alone. A practical training implementation method is provided for resource-constrained scenarios. Experiments demonstrate that the dataset and training techniques enhance LLMs' analog circuit knowledge. The trained 32B instruct model achieves 84.59% accuracy on the AMSBench-TQA benchmark, showing a 15.67 percentage-point improvement over the initial model. The trained model also shows capability in the operational amplifier design task based on the Atelier framework.

cs.AR↗

arXiv2025

Bidirectional Knowledge Distillation for Enhancing Sequential Recommendation with Large Language Models

Large language models (LLMs) have demonstrated exceptional performance in understanding and generating semantic patterns, making them promising candidates for sequential recommendation tasks. However, when combined with conventional recommendation models (CRMs), LLMs often face challenges related to high inference costs and static knowledge transfer methods. In this paper, we propose a novel mutual distillation framework, LLMD4Rec, that fosters dynamic and bidirectional knowledge exchange between LLM-centric and CRM-based recommendation systems. Unlike traditional unidirectional distillation methods, LLMD4Rec enables iterative optimization by alternately refining both models, enhancing the semantic understanding of CRMs and enriching LLMs with collaborative signals from user-item interactions. By leveraging sample-wise adaptive weighting and aligning output distributions, our approach eliminates the need for additional parameters while ensuring effective knowledge transfer. Extensive experiments on real-world datasets demonstrate that LLMD4Rec significantly improves recommendation accuracy across multiple benchmarks without increasing inference costs. This method provides a scalable and efficient solution for combining the strengths of both LLMs and CRMs in sequential recommendation systems.

cs.IR↗

arXiv2025

LLM-Based User Simulation for Low-Knowledge Shilling Attacks on Recommender Systems

Recommender systems (RS) are increasingly vulnerable to shilling attacks, where adversaries inject fake user profiles to manipulate system outputs. Traditional attack strategies often rely on simplistic heuristics, require access to internal RS data, and overlook the manipulation potential of textual reviews. In this work, we introduce Agent4SR, a novel framework that leverages Large Language Model (LLM)-based agents to perform low-knowledge, high-impact shilling attacks through both rating and review generation. Agent4SR simulates realistic user behavior by orchestrating adversarial interactions, selecting items, assigning ratings, and crafting reviews, while maintaining behavioral plausibility. Our design includes targeted profile construction, hybrid memory retrieval, and a review attack strategy that propagates target item features across unrelated reviews to amplify manipulation. Extensive experiments on multiple datasets and RS architectures demonstrate that Agent4SR outperforms existing low-knowledge baselines in both effectiveness and stealth. Our findings reveal a new class of emergent threats posed by LLM-driven agents, underscoring the urgent need for enhanced defenses in modern recommender systems.

cs.IR↗

arXiv2025

The Power of Graph Signal Processing for Chip Placement Acceleration

Placement is a critical task with high computation complexity in VLSI physical design. Modern analytical placers formulate the placement objective as a nonlinear optimization task, which suffers a long iteration time. To accelerate and enhance the placement process, recent studies have turned to deep learning-based approaches, particularly leveraging graph convolution networks (GCNs). However, learning-based placers require time- and data-consuming model training due to the complexity of circuit placement that involves large-scale cells and design-specific graph statistics. This paper proposes GiFt, a parameter-free technique for accelerating placement, rooted in graph signal processing. GiFt excels at capturing multi-resolution smooth signals of circuit graphs to generate optimized placement solutions without the need for time-consuming model training, and meanwhile significantly reduces the number of iterations required by analytical placers. Experimental results show that GiFt significantly improving placement efficiency, while achieving competitive or superior performance compared to state-of-the-art placers. In particular, compared to DREAMPlace, the recently proposed GPU-accelerated analytical placer, GF-Placer improves total runtime over 45%.

cs.LG↗

arXiv2025

Unbiased Collaborative Filtering with Fair Sampling

Recommender systems leverage extensive user interaction data to model preferences; however, directly modeling these data may introduce biases that disproportionately favor popular items. In this paper, we demonstrate that popularity bias arises from the influence of propensity factors during training. Building on this insight, we propose a fair sampling (FS) method that ensures each user and each item has an equal likelihood of being selected as both positive and negative instances, thereby mitigating the influence of propensity factors. The proposed FS method does not require estimating propensity scores, thus avoiding the risk of failing to fully eliminate popularity bias caused by estimation inaccuracies. Comprehensive experiments demonstrate that the proposed FS method achieves state-of-the-art performance in both point-wise and pair-wise recommendation tasks. The code implementation is available at https://github.com/jhliu0807/Fair-Sampling.

cs.IR↗

arXiv2025

AgentCF++: Memory-enhanced LLM-based Agents for Popularity-aware Cross-domain Recommendations

LLM-based user agents, which simulate user interaction behavior, are emerging as a promising approach to enhancing recommender systems. In real-world scenarios, users' interactions often exhibit cross-domain characteristics and are influenced by others. However, the memory design in current methods causes user agents to introduce significant irrelevant information during decision-making in cross-domain scenarios and makes them unable to recognize the influence of other users' interactions, such as popularity factors. To tackle this issue, we propose a dual-layer memory architecture combined with a two-step fusion mechanism. This design avoids irrelevant information during decision-making while ensuring effective integration of cross-domain preferences. We also introduce the concepts of interest groups and group-shared memory to better capture the influence of popularity factors on users with similar interests. Comprehensive experiments validate the effectiveness of AgentCF++. Our code is available at https://github.com/jhliu0807/AgentCF-plus.

cs.IR↗

arXiv2025

Improving LLM-powered Recommendations with Personalized Information

Due to the lack of explicit reasoning modeling, existing LLM-powered recommendations fail to leverage LLMs' reasoning capabilities effectively. In this paper, we propose a pipeline called CoT-Rec, which integrates two key Chain-of-Thought (CoT) processes -- user preference analysis and item perception analysis -- into LLM-powered recommendations, thereby enhancing the utilization of LLMs' reasoning abilities. CoT-Rec consists of two stages: (1) personalized information extraction, where user preferences and item perception are extracted, and (2) personalized information utilization, where this information is incorporated into the LLM-powered recommendation process. Experimental results demonstrate that CoT-Rec shows potential for improving LLM-powered recommendations. The implementation is publicly available at https://github.com/jhliu0807/CoT-Rec.

cs.IR↗

arXiv2024

Oracle-guided Dynamic User Preference Modeling for Sequential Recommendation

Sequential recommendation methods can capture dynamic user preferences from user historical interactions to achieve better performance. However, most existing methods only use past information extracted from user historical interactions to train the models, leading to the deviations of user preference modeling. Besides past information, future information is also available during training, which contains the ``oracle'' user preferences in the future and will be beneficial to model dynamic user preferences. Therefore, we propose an oracle-guided dynamic user preference modeling method for sequential recommendation (Oracle4Rec), which leverages future information to guide model training on past information, aiming to learn ``forward-looking'' models. Specifically, Oracle4Rec first extracts past and future information through two separate encoders, then learns a forward-looking model through an oracle-guiding module which minimizes the discrepancy between past and future information. We also tailor a two-phase model training strategy to make the guiding more effective. Extensive experiments demonstrate that Oracle4Rec is superior to state-of-the-art sequential methods. Further experiments show that Oracle4Rec can be leveraged as a generic module in other sequential recommendation methods to improve their performance with a considerable margin.

cs.IR↗

arXiv2024

Denoising Reuse: Exploiting Inter-frame Motion Consistency for Efficient Video Latent Generation

Video generation using diffusion-based models is constrained by high computational costs due to the frame-wise iterative diffusion process. This work presents a Diffusion Reuse MOtion (Dr. Mo) network to accelerate latent video generation. Our key discovery is that coarse-grained noises in earlier denoising steps have demonstrated high motion consistency across consecutive video frames. Following this observation, Dr. Mo propagates those coarse-grained noises onto the next frame by incorporating carefully designed, lightweight inter-frame motions, eliminating massive computational redundancy in frame-wise diffusion models. The more sensitive and fine-grained noises are still acquired via later denoising steps, which can be essential to retain visual qualities. As such, deciding which intermediate steps should switch from motion-based propagations to denoising can be a crucial problem and a key tradeoff between efficiency and quality. Dr. Mo employs a meta-network named Denoising Step Selector (DSS) to dynamically determine desirable intermediate steps across video frames. Extensive evaluations on video generation and editing tasks have shown that Dr. Mo can substantially accelerate diffusion models in video tasks with improved visual qualities.

cs.CV↗

arXiv2024

Physically Aware Synthesis Revisited: Guiding Technology Mapping with Primitive Logic Gate Placement

A typical VLSI design flow is divided into separated front-end logic synthesis and back-end physical design (PD) stages, which often require costly iterations between these stages to achieve design closure. Existing approaches face significant challenges, notably in utilizing feedback from physical metrics to better adapt and refine synthesis operations, and in establishing a unified and comprehensive metric. This paper introduces a new Primitive logic gate placement guided technology MAPping (PigMAP) framework to address these challenges. With approximating technology-independent spatial information, we develop a novel wirelength (WL) driven mapping algorithm to produce PD-friendly netlists. PigMAP is equipped with two schemes: a performance mode that focuses on optimizing the critical path WL to achieve high performance, and a power mode that aims to minimize the total WL, resulting in balanced power and performance outcomes. We evaluate our framework using the EPFL benchmark suites with ASAP7 technology, using the OpenROAD tool for place-and-route. Compared with OpenROAD flow scripts, performance mode reduces delay by 14% while increasing power consumption by only 6%. Meanwhile, power mode achieves a 3% improvement in delay and a 9% reduction in power consumption.

cs.LO↗