arXiv subjects

Hongyu Lu

Publications and source records attributed to Hongyu Lu.

At least 19 recordsLinked to original sources

arXiv2026

Chiral Graviton Modes in Non-Abelian lattice Fractional Quantum Hall states

Synthetic quantum matter provides a highly tunable route to fractional quantum Hall physics beyond the constraints of conventional electronic materials. However, previous theoretical studies have mostly focused on their ground state properties. It remains unclear to what extent such platforms could reveal key excitation properties of fractional quantum Hall states. Here, we study charge-neutral collective excitations in a non-abelian lattice fractional quantum Hall state realized in the bosonic Harper-Hofstadter model at unity filling factior, realizing a Moore-Read ground state. Combining full exact diagonalization, band-projected exact diagonalization, and matrix-product-state simulations, we demonstrate the existence of a long-lived chiral graviton mode, probed by chiral 3-body correlators, for the first time on lattice non-Abelian states. The graviton signal is topological sector-independent and could be observed via geometric quenches in small open droplets directly relevant to current cold-atom experiments, while other neutral modes, such as the magnetoroton and neutral fermion, are less resolved at presently achievable volumes.

cond-mat.quant-gas↗

arXiv2026

Beyond Fully Random Masking: Attention-Guided Denoising and Optimization for Diffusion Language Models

Diffusion large language models (dLLMs) offer an efficient alternative to autoregressive models through parallel decoding, yet existing post-training methods largely rely on random masking strategies that overlook intrinsic token dependencies. In this work, we present an empirical analysis of attention in dLLMs and show that tokens attending more strongly to unmasked context exhibit greater generation stability and play a critical role in reasoning. Motivated by these findings, we propose AGDO, an attention-guided denoising and optimization framework that aligns both training and optimization with attention-derived dependencies. AGDO determines the denoising order based on attention structure and emphasizes attention-critical tokens during supervised fine-tuning and reinforcement learning. Experiments on mathematical and coding benchmarks demonstrate that AGDO consistently improves reasoning performance, outperforming state-of-the-art post-training methods for dLLMs.

cs.CL↗

arXiv2026

EvoCut: Multi-Layer Evolution-Aware Visual Token Compression for Efficient Large Vision-Language Models

Large vision-language models (LVLMs) achieve strong performance on image and video understanding tasks, but their inference efficiency is constrained by the large number of visual tokens produced by vision encoders. Most existing visual token compression methods estimate token importance from attention scores or representation properties at specific layers, overlooking how visual tokens evolve across the vision encoder. Such layer-specific criteria may provide incomplete importance estimates and limit performance preservation after compression. To address this issue, we analyze layer-wise visual token evolution directions and observe that tokens form multiple group evolution directions across vision-encoder layers. Our analysis further shows that informative tokens tend to exhibit persistent deviations from common group evolution directions. Based on this observation, we propose EvoCut, a training-free and attention-free visual token compression method that estimates token importance from multi-layer evolution deviation. Experimental results show that EvoCut can retain only 11.1\% of the visual tokens on LLaVA-1.5-7B while preserving 94.4\% of the average performance, demonstrating its effectiveness in balancing efficiency and accuracy.

cs.CV↗

arXiv2026

LRCP: Low-Rank Compressibility Guided Visual Token Pruning for Efficient LVLMs

Large vision-language models (LVLMs) achieve strong multimodal understanding, but their inference cost grows rapidly with the number of visual tokens, especially for high-resolution images and long videos. Existing attention-based methods estimate token importance from attention scores, which may introduce positional bias, while representation-based methods reduce visual redundancy based on feature relations or reconstruction errors, overlooking the global structure of the visual token set. In this paper, we revisit visual token compression from the perspective of low-rank compressibility. Across models and datasets, we observe that visual token representations exhibit a pronounced low-rank structure, with a dominant subspace that remains stable even after a large fraction of tokens is randomly removed. Motivated by this finding, we propose LRCP, a training-free compression framework that first estimates the dominant low-rank subspace of visual tokens via PCA, and then scores each token by its projection residual onto this subspace, retaining tokens that are poorly explained by the low-rank background. Extensive experiments show that LRCP achieves superior results, preserving 94.7% of the original image-understanding performance with an 88.9% token reduction and 97.8% of the average video-understanding accuracy with an 87.5% token reduction.

cs.CV↗

arXiv2026

The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes

On-policy distillation (OPD) and on-policy self-distillation (OPSD) have emerged as promising post-training methods for large language models, offering dense token-level supervision on trajectories sampled from the model's own policy. However, existing results on their effectiveness remain mixed: while OP(S)D has shown promise in system prompt and knowledge internalization, recent studies also report instability and degradation. In this work, we present a comprehensive empirical study of when OPD and OPSD work, when they fail, and why. We find that OPD on mathematical reasoning is highly sensitive to teacher choice and loss formulation, whereas OPSD fails in our tested settings due to test-time absence of instance-specific privileged information (PI). In contrast, OPSD is effective when PI represents a shared latent rule, such as a system prompt or alignment preference. We identify three failure mechanisms: (1) distribution mismatch between teacher and student caused by conditioning on student-generated prefixes, (2) optimization instability from biased TopK reverse-KL gradients, and (3) an OPSD-specific limitation where the student learns a PI-free policy that aggregates PI-conditioned teachers, which is insufficient when PI is instance-specific. We further show that stop-gradient TopK objectives, RLVR-adapted teachers, and SFT-stabilized students mitigate these failures.

cs.AI↗

arXiv2026

SA-CAISR: Stage-Adaptive and Conflict-Aware Incremental Sequential Recommendation

Sequential recommendation (SR) aims to predict a user's next action by learning from their historical interaction sequences. In real-world applications, these models require periodic updates to adapt to new interactions and evolving user preferences. While incremental learning methods facilitate these updates, they face significant challenges. Replay-based approaches incur high memory and computational costs, and regularization-based methods often struggle to discard outdated or conflicting knowledge. To overcome these challenges, we propose SA-CAISR, a Stage-Adaptive and Conflict-Aware Incremental Sequential Recommendation framework. As a buffer-free framework, SA-CAISR operates using only the old model and new data, directly addressing the high costs of replay-based techniques. SA-CAISR introduces a novel Fisher-weighted knowledge-screening mechanism that dynamically identifies outdated knowledge by estimating parameter-level conflicts between the old model and new data, selectively removing obsolete knowledge while preserving compatible historical patterns. This dynamic balance between stability and adaptability allows our method to achieve state-of-the-art performance in incremental SR. Specifically, SA-CAISR improves Recall@20 by 2.0% on average across datasets, while reducing memory usage by 97.5% and training time by 46.9% compared to the best baseline. This efficiency allows real-world systems to rapidly update user profiles with minimal computational overhead, ensuring more timely and accurate recommendations.

cs.IR↗

arXiv2026

Collective excitations in chiral spin liquid: chiral roton and long-wavelength nematic mode

Chiral spin liquid (CSL) is a magnetic analogue of the fractional quantum Hall (FQH) liquid. Collective excitations play a vital role in shaping our understanding of these exotic quantum phases of matter and their quantum phase transitions. While the magneto-roton and long-wavelength chiral graviton modes in the FQH liquids have been extensively explored, the collective excitations of CSLs remain elusive. Here we explore the collective excitations in the SU(2) symmetric CSL phase of the spin-1/2 square-lattice $J_1-J_2-J_\chi$ model, where an intriguing quantum phase diagram was recently revealed. Combining exact diagonalization and time-dependent variational principle calculations, we observe two spin-singlet collective modes: a chiral p-wave (low-energy) roton mode at finite momentum and a elliptically polarized d-wave (higher-energy) nematic mode at zero momentum, both of which are prominent across the CSL phase. Such exotic modes exhibit fingerprints distinct from those of FQH liquids, and to the best of our knowledge, are reported for the first time. By tuning $J_2$, we find the nematic mode to be pronouncedly soft, together with the spin-triplet two-spinon bound states, potentially promoting strong nematic and spin stripe instabilities. Our work paves the way for further understanding CSL from the dynamical perspective and provides new spectroscopic signatures for future experiments of CSL candidates.

cond-mat.str-el↗

arXiv2026

GenCI: Generative Modeling of User Interest Shift via Cohort-based Intent Learning for CTR Prediction

Click-through rate (CTR) prediction plays a pivotal role in online advertising and recommender systems. Despite notable progress in modeling user preferences from historical behaviors, two key challenges persist. First, exsiting discriminative paradigms focus on matching candidates to user history, often overfitting to historically dominant features and failing to adapt to rapid interest shifts. Second, a critical information chasm emerges from the point-wise ranking paradigm. By scoring each candidate in isolation, CTR models discard the rich contextual signal implied by the recalled set as a whole, leading to a misalignment where long-term preferences often override the user's immediate, evolving intent. To address these issues, we propose GenCI, a generative user intent framework that leverages semantic interest cohorts to model dynamic user preferences for CTR prediction. The framework first employs a generative model, trained with a next-item prediction (NTP) objective, to proactively produce candidate interest cohorts. These cohorts serve as explicit, candidate-agnostic representations of a user's immediate intent. A hierarchical candidate-aware network then injects this rich contextual signal into the ranking stage, refining them with cross-attention to align with both user history and the target item. The entire model is trained end-to-end, creating a more aligned and effective CTR prediction pipeline. Extensive experiments on three widely used datasets demonstrate the effectiveness of our approach.

cs.IR↗

arXiv2026

Chiral Graviton Modes in Fermionic Fractional Chern Insulators

Chiral graviton modes are hallmark collective excitations of Fractional Quantum Hall (FQH) liquids. However, their existence on the lattice, where continuum symmetries that protect them from decay are lost, is still an open and urgent question, especially considering the recent advances in the realization of Fractional Chern Insulators (FCI) in transition metal dichalcogenides and rhombohedral pentalayer graphene. Here we present a comprehensive theoretical and numerical study of graviton-modes in fermionic FCI, and thoroughly demonstrate their existence. We first derive a lattice stress tensor operator in the context of the fermionic Harper-Hofstadter(HH) model which captures the graviton in the flat band limit. Importantly, we discover that such lattice stress-tensor operators are deeply connected to lattice quadrupolar density correlators, readily generalizable to generic Chern bands. We then explicitly show the adiabatic connection between FQH and FCI chiral graviton modes by interpolating from a low flux HH model to a Checkerboard lattice model that hosts a topological flat band. In particular, using state-of-the-art matrix product state and exact diagonalization simulations, we provide strong evidence that chiral graviton modes are long-lived excitations in FCIs despite the lack of continuous symmetries and the scattering with a two-magnetoroton continuum. By means of a careful finite-size analysis, we show that the lattice generates a finite but small intrinsic decay rate for the graviton mode. We discuss the relevance of our results for the exploration of graviton modes in FCI phases realized in solid state settings, as well as cold atom experiments.

cond-mat.str-el↗

arXiv2025

Entropy-Guided Token Dropout: Training Autoregressive Language Models with Limited Domain Data

As access to high-quality, domain-specific data grows increasingly scarce, multi-epoch training has become a practical strategy for adapting large language models (LLMs). However, autoregressive models often suffer from performance degradation under repeated data exposure, where overfitting leads to a marked decline in model capability. Through empirical analysis, we trace this degradation to an imbalance in learning dynamics: predictable, low-entropy tokens are learned quickly and come to dominate optimization, while the model's ability to generalize on high-entropy tokens deteriorates with continued training. To address this, we introduce EntroDrop, an entropy-guided token dropout method that functions as structured data regularization. EntroDrop selectively masks low-entropy tokens during training and employs a curriculum schedule to adjust regularization strength in alignment with training progress. Experiments across model scales from 0.6B to 8B parameters show that EntroDrop consistently outperforms standard regularization baselines and maintains robust performance throughout extended multi-epoch training. These findings underscore the importance of aligning regularization with token-level learning dynamics when training on limited data. Our approach offers a promising pathway toward more effective adaptation of LLMs in data-constrained domains.

cs.CL↗

arXiv2025

Bosonic Laughlin and Moore-Read states from non-Chern flat bands

The rapid advances in the study of fractional Chern insulators (FCIs) raise a fundamental question: while initially discovered in flat Chern bands motivated by their topological equivalence to Landau levels, is single- particle band topology actually a prerequisite for these many-body topological orders emergent at fractional fillings? Here, we numerically demonstrate bosonic FCIs in two types of non-Chern flat bands in honeycomb lattices, using exact diagonalization and density matrix renormalization group calculations. In a gapless flat band with a singular band touching, we observe a Laughlin state at half filling, stabilized by onsite interactions from the hard-core limit down to arbitrarily small strength. Furthermore, we report the first example of a non- Abelian FCI in a non-Chern band system: a Moore-Read state at $\nu$ = 1 filling of the same singular flat band with hard-core bosons. Under lattice parameters that realize a gapped trivial band (C = 0) of exact flatness, we also find the Laughlin FCI of soft-core bosons in the isolated band limit where onsite interaction is much smaller than the band gap. In this case, the FCI forms as interacting bosons spontaneously avoid the peaks in quantum metric and Berry curvature, preferentially occupying Brillouin zone region with relatively uniform quantum geometry. Our work significantly expands the landscape for (non-)Abelian FCIs and broadens the understanding of their formation beyond the Chern band paradigm.

cond-mat.str-el↗

arXiv2025

Experience-Guided Reflective Co-Evolution of Prompts and Heuristics for Automatic Algorithm Design

Combinatorial optimization problems are traditionally tackled with handcrafted heuristic algorithms, which demand extensive domain expertise and significant implementation effort. Recent progress has highlighted the potential of automatic heuristics design powered by large language models (LLMs), enabling the automatic generation and refinement of heuristics. These approaches typically maintain a population of heuristics and employ LLMs as mutation operators to evolve them across generations. While effective, such methods often risk stagnating in local optima. To address this issue, we propose the Experience-Guided Reflective Co-Evolution of Prompt and Heuristics (EvoPH) for automatic algorithm design, a novel framework that integrates the island migration model with the elites selection algorithm to simulate diverse heuristics populations. In EvoPH, prompts are co-evolved with heuristic algorithms, guided by performance feedback. We evaluate our framework on two problems, i.e., Traveling Salesman Problem and Bin Packing Problem. Experimental results demonstrate that EvoPH achieves the lowest relative error against optimal solutions across both datasets, advancing the field of automatic algorithm design with LLMs.

cs.AI↗

arXiv2025

Probing Non-Fermi-Liquid Behaviour of Composite Fermi Liquid via Efficient Thermal Simulations

The physics of two-dimensional electron gas in a perpendicular magnetic field, i.e., the quantum Hall system, is remarkably rich. At half filling of the lowest Landau level, it has been predicted that "composite fermions"---emergent quasiparticles consisting of an electron attached to two magnetic flux quanta---experience zero net magnetic field and form a Fermi sea, dubbed composite Fermi liquid (CFL). However, despite its seemingly simple appearance, CFL is a strongly correlated quantum many-body state in disguise, and solving it in a controlled manner is extremely difficult, to the extent that the thermodynamic properties of CFL remain largely unknown. In this work, we perform state-of-the-art thermal tensor network simulations of the $\nu=1/2$ Landau level system and observe low-temperature power-law behaviour of the specific heat, signaling the gapless nature of CFL. More importantly, the power is extracted to be close to $2/3$, clearly deviating from the ordinary linear-$T$ behaviour of Fermi liquid, suggesting coupling between the CFs and the dynamical emergent gauge field and thereby revealing the quantum many-body nature of the CFL state.

cond-mat.str-el↗

arXiv2025

Interaction-induced nematic Dirac semimetal from quadratic band touching: A constrained-path quantum Monte Carlo study

Electronic systems with quadratic band touchings, commonly found in two- and three-dimensional materials such as Bernal-stacked bilayer graphene, kagome metals, HgTe, and pyrochlore iridates, have attracted significant interest concerning the role of interactions in shaping their electronic properties. However, even in the simplest model of spinless fermions on a two-dimensional checkerboard lattice, the quantum phase diagram as a function of nearest-neighbor interaction remains under debate. We employ constrained-path quantum Monte Carlo simulations (CP-QMC) simulations to investigate the problem using a two-dimensional torus geometry. We cross-validate our results on small lattices by comparing them with density-matrix renormalization group calculations, finding quantitative agreement. In particular, we implement an improved optimization scheme within the CP-QMC simulations, enabling the identification of a bond-nematic Dirac semimetal phase that was found in tensor-network studies on cylindrical geometries, but remains inaccessible to Hartree-Fock mean-field methods. The CP-QMC approach makes it possible to establish the emergence of this phase in a geometry that preserves lattice rotational symmetry and permits extrapolation to the thermodynamic limit. Our results show that the quantum phase diagram of spinless fermions on the checkerboard lattice with nearest-neighbor repulsion features three interaction-induced phases at half filling: a quantum anomalous Hall insulator at weak coupling, a bond-nematic Dirac semimetal at intermediate coupling, and a site-nematic insulator at strong coupling.

cond-mat.str-el↗

arXiv2025

Fractional Chern insulator states in an isolated flat band of zero Chern number

A flat band with Chern number $C=0$, and well isolated from the rest of Hilbert space by a gap much larger than interaction strength, is a context that has not been regarded as relevant for fractional quantum Hall physics. In this work, we demonstrate the emergence of the fractional Chern insulator (FCI) states in such a trivial flat band, using large-scale exact diagonalization (ED) and infinite density matrix renormalization group (iDMRG) simulations. The $C=0$ isolated flat band is hosted by an anisotropic fluxed dice lattice. Both the quantum metric and Berry curvature of the $C=0$ flat band have a sharp peak at the $\Gamma$ point, whereas in the rest of the Brillouin zone (BZ) they mimic the quantum geometry of the lowest Landau level. We consider nearest-neighbor repulsion that is weak enough to ensure the isolated-band limit is always satisfied. From the projected ED simulations at $\nu_\mathrm{F}=2/3$ electron filling of the flat band (i.e. $1/3$ hole filling), we find the unexpected FCI with 3-fold ground-state degeneracy and $\sigma_\mathrm{H}=-1/3 (e^2/h)$. The momentum space carrier distribution shows that the quantum metric peak tends to push the interacting holes away from $\Gamma$ point towards the BZ regions with the nearly ``ideal'' quantum geometry, underlying the formation of FCI in the $C=0$ flat band. Besides, when tuning the single-particle anisotropy such that the quantum geometry of the $C=0$ flat band becomes less sharp around $\Gamma$, we find the ground state becomes a charge density wave with tripled unit cell at $\nu_\mathrm{F}=2/3$. Our two-band iDMRG simulations further corroborate the FCI in the isolated $C=0$ flat band, demonstrating in such parameter regime the fractionally quantized charge pumping upon flux insertion as well as the momentum-resolved entanglement spectrum characteristic of the $1/3$ Laughlin state.

cond-mat.mes-hall↗

arXiv2025

Generic integer and fractional quantum anomalous Hall crystals from interaction-driven band folding

Among the extensive studies of fractional quantum anomalous Hall (FQAH) states, there recently appears a growing interest in the topological states with coexisting charge density wave (CDW) orders. Such states are referred to as Hall crystals. However, compared to those with integer Hall conductivities, the FQAH crystal (FQAHC) is still elusive even at the level of microscopic model. In this work, we numerically study a topological flat-band model on triangular lattice with spinless fermions. At fractional filling of the Chern band, the nearest-neighbor interaction leads to a commensurate and topologically trivial CDW state. Interestingly, the folded mini-band above the CDW gap is non-trivial, and we focus on the doping of it without any projection. A series of (F)QAHC states at (fractional) integer fillings of this mini-band are discovered and some FQAHC state might even exist in less "ideal" conditions. The ground-state degeneracies of such (F)QAHC states are enlarged by the CDW degeneracy and the Hall conductivities -- determined by the fillings of the mini-band -- are different from the fillings of the original Chern band. We also study the thermodynamics of an FQAHC state and find a compressible CDW phase at intermediate temperatures, which might serve as a precursor of lower temperature FQAHC phase. Moreover, we numerically demonstrate that such a generic scheme of doping CDW-folded topological mini-band could be applied to bosonic systems, broadening the platforms of Hall-crystal physics and motivating its exploration in quantum moire and cold-atom systems.

cond-mat.str-el↗

arXiv2025

Slow Thinking for Sequential Recommendation

To develop effective sequential recommender systems, numerous methods have been proposed to model historical user behaviors. Despite the effectiveness, these methods share the same fast thinking paradigm. That is, for making recommendations, these methods typically encodes user historical interactions to obtain user representations and directly match these representations with candidate item representations. However, due to the limited capacity of traditional lightweight recommendation models, this one-step inference paradigm often leads to suboptimal performance. To tackle this issue, we present a novel slow thinking recommendation model, named STREAM-Rec. Our approach is capable of analyzing historical user behavior, generating a multi-step, deliberative reasoning process, and ultimately delivering personalized recommendations. In particular, we focus on two key challenges: (1) identifying the suitable reasoning patterns in recommender systems, and (2) exploring how to effectively stimulate the reasoning capabilities of traditional recommenders. To this end, we introduce a three-stage training framework. In the first stage, the model is pretrained on large-scale user behavior data to learn behavior patterns and capture long-range dependencies. In the second stage, we design an iterative inference algorithm to annotate suitable reasoning traces by progressively refining the model predictions. This annotated data is then used to fine-tune the model. Finally, in the third stage, we apply reinforcement learning to further enhance the model generalization ability. Extensive experiments validate the effectiveness of our proposed method.

cs.IR↗

arXiv2025

Universal Item Tokenization for Transferable Generative Recommendation

Recently, generative recommendation has emerged as a promising paradigm, attracting significant research attention. The basic framework involves an item tokenizer, which represents each item as a sequence of codes serving as its identifier, and a generative recommender that predicts the next item by autoregressively generating the target item identifier. However, in existing methods, both the tokenizer and the recommender are typically domain-specific, limiting their ability for effective transfer or adaptation to new domains. To this end, we propose UTGRec, a Universal item Tokenization approach for transferable Generative Recommendation. Specifically, we design a universal item tokenizer for encoding rich item semantics by adapting a multimodal large language model (MLLM). By devising tree-structured codebooks, we discretize content representations into corresponding codes for item tokenization. To effectively learn the universal item tokenizer on multiple domains, we introduce two key techniques in our approach. For raw content reconstruction, we employ dual lightweight decoders to reconstruct item text and images from discrete representations to capture general knowledge embedded in the content. For collaborative knowledge integration, we assume that co-occurring items are similar and integrate collaborative signals through co-occurrence alignment and reconstruction. Finally, we present a joint learning framework to pre-train and adapt the transferable generative recommender across multiple domains. Extensive experiments on four public datasets demonstrate the superiority of UTGRec compared to both traditional and generative recommendation baselines.

cs.IR↗