arXiv subjects

Tiantian Zhang

Publications and source records attributed to Tiantian Zhang.

At least 19 recordsLinked to original sources

arXiv2026

Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus

We present the first systematic study of Massive activations (MAs) in layer-interleaved HLA LLMs and uncover two architecture-aligned morphologies: MAs consistently spike immediately before full attention layers, forming pre-attention spikes (PAS), and can persist through intervening linear attention layers, giving rise to inter-spike plateaus (ISP). As full attention becomes denser, successive PAS become increasingly connected through ISP, ultimately recovering the stable MA morphology of full attention LLMs. We establish the recurrence of this organization across five linear attention architectures, six hybridization configurations, five data domains, and representative open-source hybrid models spanning 1.2B to 397B total parameters. Controlled pretraining of GDN-based hybrids at scales up to 1.3B shows that both morphologies emerge early and respond asymmetrically to output gating: full attention output gating strongly attenuates their absolute magnitudes without eliminating their layerwise organization, whereas removing GDN gates yields comparatively modest amplification. Mechanistically, our systematic-outlier analysis supports a shared lifecycle account governed by the timing of MA cancellation. PAS follows a localized write-sink-cancel process, while the extended persistence of ISP is consistent with delayed cancellation. At the full attention limit, this account recovers the stable MA morphology characteristic of full attention LLMs. Our code is available at https://github.com/StartLuxLabs/Massive-Activations-HLA.

cs.CL↗

arXiv2026

Beyond Isolation: Unlocking Reinforcement Learning Component Synergy for Sample-Efficient Continuous Control

Reinforcement learning systems are significantly more complex than other machine learning paradigms due to inherent properties, causing RL system design to jointly account for many tightly coupled factors. Despite advances in individual algorithmic components, their functional interdependencies remain underexplored: do they exhibit mutual synergy or counterproductive interference? To bridge this gap, we conduct a systematic investigation and find that the efficacy of different components exhibits significant task-dependency, and naively stacking state-of-the-art techniques does not necessarily yield performance gains; instead, it often triggers emergent challenges, such as compounded non-stationarity. Building upon these findings, we distill a suite of actionable insights into the principled coordination of these components. Guided by these insights, we propose ROSER, an RL framework that coordinates three critical dimensions: Model-based Representation, Optimization Stability, and Experience Replay. Across diverse continuous-control benchmarks, ROSER consistently outperforms vanilla baselines and achieves 17.60% gains over naive stack. Our findings underscore the necessity of a holistic perspective in RL system design and paves the way for developing sample-efficient agents.

cs.LG↗

arXiv2026

Annular Majorana mode in a superconducting topological insulator

When the surface states of a topological insulator becomes superconducting, topological superconductivity can be obtained, and each vortex on the surface can host one single Majorana zero-energy mode which is usually a wave packet decaying exponentially off the vortex core. Here, we predict stable Majorana zero-energy mode whose wave function is ring-shape, dubbed as annular Majorana mode, in the superconducting vortex in topological insulators respecting $3$-fold or $6$-fold rotational symmetry. Such topological insulators are featured with a single nonlinear Dirac cone located at $\barΓ$ or three linear Dirac cones at $\bar{\text{M}}$ in the surface Brillouin zone. The annular Majorana mode originates from the effective chiral $f$-wave superconductivity on the nonlinear Dirac cone in the former case and the interference of the effective chiral $p$-wave superconductivity on the three linear Dirac cones in the latter. In both cases, the annular Majorana mode is stabilized by the rotational symmetry and the winding number $3$ carried by the surface states. Candidate materials supporting the annular Majorana mode are predicted. Our work provides new insights into the topological superconductivity in superconducting topological insulators.

cond-mat.supr-con↗

arXiv2026

Advances in Phonons: From Band Topology to Phonon Chirality

Phonons, the quantized collective vibrations of a crystal lattice, are among the most fundamental bosonic excitations in condensed matter systems. They govern thermal transport, mediate electron-phonon coupling, and drive symmetry-breaking orders such as charge density waves and conventional superconductivity. Long regarded as spin-0 bosons characterized only by their vibrational frequencies and linear or circular polarization, phonons have recently been revealed to host a much richer internal structure. Recent advances in topological band theory and quantum geometry have shown that phonon eigenstates, encoded in both their eigenvalues and eigenvectors, can exhibit nontrivial topological and geometric properties. These developments have established topological and circularly polarized phonons as two major frontiers in phonon physics, motivating this review of recent theoretical and experimental advances. We present a unified framework for classifying phonon modes in both reciprocal and real space, encompassing symmetry-protected topological phases, topological invariants, and phonon polarization. We then examine their coexistence in PT-broken systems through Weyl phonons, highlighting the simultaneous emergence of topological and rotational chirality. Finally, we discuss outstanding challenges and future research directions, including the role of topology and quantum geometry in phonon-mediated interactions and the controlled manipulation of phonon angular momentum for prospective quantum technologies.

cond-mat.mtrl-sci↗

arXiv2026

Raman spectroscopy of van der Waals topological magnet GdGaI

We report polarization-resolved Raman spectroscopy of a van der Waals compound GdGaI that is a candidate for excitonic insulators. By combining the symmetry analysis with density functional theory calculations, we identify six Raman-active phonons. The spectra exhibit only the expected anharmonic hardening down to 4 K: no additional peaks, no soft modes, and no signatures of zone folding are observed. This result indicates that any lattice distortion is below our experimental sensitivity, supporting an electronically driven origin for the band reconstruction reported by angle-resolved photoemission spectroscopy rather than an electron-phonon-driven mechanism. Moreover, we observe a pronounced circular dichroism of the $A_{1g}$ modes under an out-of-plane magnetic field. Based on symmetry considerations, we attribute this dichroic response to chiral $A_{1g}$ phonons with opposite angular momenta generated by spin-phonon coupling in the time-reversal-broken state. The temperature evolution of the degree of circular polarization further suggests that circularly polarized Raman spectroscopy detects the emergence of short-range antiferromagnetic correlations. Our results highlight GdGaI as a promising platform in which excitonic order, magnetism, and circularly polarized phonons can be intertwined, and demonstrate that circular-polarization Raman provides a sensitive probe of spin-phonon coupling in excitonic systems.

cond-mat.mtrl-sci↗

arXiv2026

Mirror Horizon: Viable Path Entropy as a Measure of Bounded Reflection

Mirror Theory proposes that an intelligent system should be studied not only by what it represents, but by what coherent continuations it can sustain under repeated reflection. We make this claim operational through \emph{viable path entropy} (VPE), a finite-budget measure of verified continuation capacity. Given a mirror state, a rollout protocol, a verifier, and a mode map, VPE decomposes bounded capability into two parts: the probability of reaching a viable continuation and the diversity of verified continuation modes reached among successful rollouts. This paper restores the full theoretical scaffold behind the measure: intuition as local underdetermining constraint, taste as invariant-selecting pressure, reflection as taste-guided resolution of underdetermination, and geometry as the learned structure that makes future reflection stable. We then instantiate the theory in language-model reasoning experiments on GSM8K. Across Qwen2.5-Instruct models, 32 sampled rollouts per problem, and two reflection horizons, increasing the token budget from 96 to 160 substantially expands verified reachability, reduces zero-reachability, increases verified-mode entropy, and improves smoothed VPE. At 160 tokens, Qwen2.5-1.5B realizes the strongest mirror horizon among the tested models, even though Qwen2.5-3B has more parameters. This shows that mirror horizon is not parameter count, but accessible verified continuation capacity under a bounded reflection protocol. The result supports Mirror Theory as a measure-level account: capability is the structure of viable continuations made reachable, not merely one-shot accuracy or pass@k.

cs.LG↗

arXiv2026

Distribution Preference Optimization: A Fine-grained Perspective for LLM Unlearning

As Large Language Models (LLMs) demonstrate remarkable capabilities learned from vast corpora, concerns regarding data privacy and safety are receiving increasing attention. LLM unlearning, which aims to remove the influence of specific data while preserving overall model utility, is becoming an important research area. One of the mainstream unlearning classes is optimization-based methods, which achieve forgetting directly through fine-tuning, exemplified by Negative Preference Optimization (NPO). However, NPO's effectiveness is limited by its inherent lack of explicit positive preference signals. Attempts to introduce such signals by constructing preferred responses often necessitate domain-specific knowledge or well-designed prompts, fundamentally restricting their generalizability. In this paper, we shift the focus to the distribution-level, directly targeting the next-token probability distribution instead of entire responses, and derive a novel unlearning algorithm termed \textbf{Di}stribution \textbf{P}reference \textbf{O}ptimization (DiPO). We show that the requisite preference distribution pairs for DiPO, which are distributions over the model's output tokens, can be constructed by selectively amplifying or suppressing the model's high-confidence output logits, thereby effectively overcoming NPO's limitations. We theoretically prove the consistency of DiPO's loss function with the desired unlearning direction. Extensive experiments demonstrate that DiPO achieves a strong trade-off between model utility and forget quality. Notably, DiPO attains the highest forget quality on the TOFU benchmark, and maintains leading scalability and sustainability in utility preservation on the MUSE benchmark.

cs.LG↗

arXiv2026

General ab initio framework for electronic-order-induced lattice-dynamics symmetry breaking

Conventional \textit{ab initio} approaches are unable to describe phonon time-reversal symmetry ($\mathcal{T}$) breaking. Here, we develop an \textit{ab initio} framework, grounded in molecular Berry curvature (MBC) theory, that captures electronic-order-driven symmetry breaking in lattice dynamics. Using Co$_3$Sn$_2$S$_2$ as a model system, our \textit{ab initio} framework yields phonon spectra that break both $\mathcal{T}$ and mirror symmetries, quantitatively reproduce the observed phonon splittings observed in experiments, and reveal distinct microscopic origins for the $E_g$ and $E_u$ modes: $E_g$ splitting is governed by MBC and is accurately captured by our algorithm, whereas $E_u$ splitting is enhanced by the Fano resonance and matches the experimental data once the Fano-factor correction is included. Leveraging this algorithm, we predict several candidate materials with nonzero electronic-order-driven symmetry breaking in lattice dynamics, establishing a first-principles route to understand electron-phonon coupling, phonon magnetism, and related Hall-type lattice responses.

cond-mat.mtrl-sci↗

arXiv2026

Principled RL for Flow Matching Emerges from the Chunk-level Policy Optimization

Recent Progress in post-training flow matching for text-to-image (T2I) generation with Group Relative Policy Optimization (GRPO) has demonstrated strong potential. However, it is hindered by a critical limitation: inaccurate advantage attribution. In this work, we argue that aggregating consecutive steps into a coherent 'chunk' and shifting the policy optimization paradigm from GRPO's step level to the chunk level can effectively mitigate the negative impact of this issue. Building on this insight, we propose Group Chunking Policy Optimization (GCPO), the first chunk-level reinforcement learning approach for post-training flow matching. Extensive experiments demonstrate that GCPO achieves superior performance on both standard T2I benchmarks and preference alignment, with up to 43% relative gains over GRPO, highlighting the promise of chunk-level policy optimization. The code is available on https://github.com/xingzhejun/GCPO.

cs.CV↗

arXiv2026

Bridge Graphical Models: Coupling, Projection, and Current-Preserving Dynamics for Generative Modeling

Continuous-time generative models are often built from endpoint-conditioned bridges, but generation requires a different object: a non-anticipative Markov decoder that only observes the current state and time. We identify this bridge-to-decoder compression as a structural bottleneck shared by diffusion models, flow matching, rectified flow, Schrödinger bridges, and field-based generative models. We introduce the \emph{Markovization gap}, the time-integrated conditional variance of the bridge velocity given the Markov state. It is the MMSE of predicting endpoint-conditioned motion from the information available to a sampler, and it measures an irreducible loss incurred before any neural network is trained. To make this bottleneck comparable across model families, we define \emph{Bridge Graphical Models} (BGMs), which separate endpoint coupling, bridge law, Markovian projection, and current-preserving dynamics representation as independent design choices. The same formalism also represents Poisson and electrostatic models as field-line bridge kernels with a corresponding field-line Markovization gap. Across synthetic, latent, and pixel-space pilots on CIFAR-10 and Fashion-MNIST, a feature-space proxy gap estimated in minutes before training ranks design choices in the same direction as downstream training loss and FID under fixed architecture, bridge, sampler, and compute. These results support the Markovization gap as a pre-training diagnostic for bridge and coupling design.

cs.LG↗

arXiv2026

Semidirect Fourier Delta Attention: Phase-Controlled Delta Memory with Constructive Chunk-WY Kernels

Linear attention replaces softmax attention's growing KV cache with a fixed recurrent state, but this compression limits exact state tracking and long-context memory. We introduce \emph{Semidirect Fourier Delta Attention} (SFDA), a phase-controlled generalization of Kimi Delta Attention that replaces real diagonal decay with block-rotational Fourier control: \[ S_t=(I-β_t k_tk_t^*)Λ_tS_{t-1}+β_tk_tv_t^*, \qquad Λ_t=\diag(α_t\odot e^{iθ_t}). \] Our main result is a constructive chunk-WY factorization for products $A_t=Λ_t-u_tr_t^*$, giving \[ A_t\cdots A_1=Γ_t-Y_tM_tW_t^* \] with rank growth bounded inside fixed chunks. This yields an exact affine chunk transfer, formal stability and complexity bounds, and a compact characterization of phase-plus-low-rank memory. We verify the algebra numerically and show in toy state-tracking experiments that SFDA learns cyclic memory where the phase-disabled KDA baseline remains near chance. Fused kernels and large-scale language-model comparisons are left to future work.

cs.LG↗

arXiv2026

Magnetism and Topology from Circularly Polarized Phonon Floquet Engineering

We theoretically show that circularly polarized phonons induce electronic magnetization and drive a topological phase transition via phonon Floquet engineering. Considering the electronic states modulated by circularly polarized phonons on a honeycomb lattice, we show that such lattice dynamics generates an effective next-nearest-neighbor electron hopping, leading to a Haldane-type mass term. Circularly polarized phonon breaks time-reversal symmetry (TRS) and opens a gap at valley points, undergoing phase transition from a trivial insulator to a Chern insulator. Moreover, the orbital and spin magnetizations emerge due to the breaking of TRS. Our results show that circularly polarized phonons serve as an effective magnetic field to engineer magnetism and topology, offering new opportunities for phonon Floquet approaches.

cond-mat.mes-hall↗

arXiv2026

Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation

As the foundational architecture of modern machine learning, Transformers have driven remarkable progress across diverse AI domains. Despite their transformative impact, a persistent challenge across various Transformers is Attention Sink (AS), in which a disproportionate amount of attention is focused on a small subset of specific yet uninformative tokens. AS complicates interpretability, significantly affecting the training and inference dynamics, and exacerbates issues such as hallucinations. In recent years, substantial research has been dedicated to understanding and harnessing AS. However, a comprehensive survey that systematically consolidates AS-related research and offers guidance for future advancements remains lacking. To address this gap, we present the first survey on AS, structured around three key dimensions that define the current research landscape: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Our work makes a pivotal contribution by highlighting the key concepts and main trends in the field, guiding researchers through the evolution of AS-related studies. We envision this survey as a valuable resource, empowering researchers to effectively manage AS within the current Transformer paradigm, while simultaneously inspiring innovative advancements for the next generation of Transformers. The paper list of this work is available at https://github.com/ZunhaiSu/Awesome-Attention-Sink.

cs.LG↗

arXiv2026

Direct probing the quantum geometric tensor for bosonic collective excitations

The quantum geometric tensor (QGT), whose real and imaginary parts define the quantum metric and Berry curvature, encodes the intrinsic geometry of quantum states. While electronic QGT has recently become experimentally accessible and linked to diverse physical phenomena, its bosonic counterpart remains largely unexplored. Here we show that the dynamical structure factor encodes the momentum-space structure of bosonic wave functions and thereby provides direct access to the full bosonic QGT throughout the Brillouin zone. Applying this framework, we uncover clear geometric signatures in the twofold quadrupole-Weyl phonon of BaPtGe and the nodal-line magnon in Gd, and further generalize the formalism to multiband systems. Our results establish a general route to measuring (non-)Abelian quantum geometry in bosonic systems, a crucial step toward elucidating its impact on condensed matter phenomena.

cond-mat.mtrl-sci↗

arXiv2026

Topological phononics

Topological phononics extends the foundational concepts of topological condensed matter physics to the realm of lattice vibrations and classical mechanical waves, unlocking robust, defect-immune states and phenomena beyond the reach of conventional phononic engineering. This review provides a unified, systematic framework for understanding topological phonons across natural and artificial systems, spanning solid-state materials, acoustic/mechanical metamaterials, and non-Hermitian platforms. We cover the core theoretical principles -- from Berry curvature and symmetry-protected topological invariants to bulk-boundary correspondence -- alongside experimental advances in probing topological phonon states via inelastic scattering and momentum-resolved techniques for solid-state phonons as well as pump-probe measurements in acoustic/mechanical metamaterials. Key topics include Weyl/Dirac/nodal-line phonons in crystalline solids, symmetry-engineered topological phases in metamaterials, non-Hermitian effects (exceptional points, skin effect), and emergent directions such as Floquet engineering, synthetic dimensions, and real-space topological textures (skyrmions, merons). We also highlight technological applications in robust waveguides, on-chip surface-acoustic-wave devices, and acoustofluidics, while outlining future challenges and opportunities in quantum phononics, nonlinear topological phenomena, and interdisciplinary integration with photonics and electronics. This review serves as a comprehensive guide across physics, materials science, and engineering, bridging fundamental theory with cutting-edge experiments and innovations in topological phononics.

cond-mat.mtrl-sci↗

arXiv2026

Power Reinforcement Post-Training of Text-to-Image Models with Super-Linear Advantage Shaping

Recently, post-training methods based on reinforcement learning, with a particular focus on Group Relative Policy Optimization (GRPO), have emerged as the robust paradigm for further advancement of text-to-image (T2I) models. However, these methods are often prone to reward hacking, wherein models exploit biases in imperfect reward functions rather than yielding genuine performance gains. In this work, we identify that normalization could lead to miscalibration and directly removing the prompt-level standard deviation term yields an optimal policy ascent direction that is linear in the advantage but still limits the separation of genuine signals from noise. To mitigate the above issues, we propose Super-Linear Advantage Shaping (SLAS) by revisiting the functional update from an information geometry perspective. By extending the Fisher-Rao information metric with advantage-dependent weighting, SLAS introduces a non-linear geometric structure that reshapes the local policy space. This design relaxes constraints along high-advantage directions to amplify informative updates, while tightening those in low-advantage regions to suppress illusory gradients. In addition, batch-level normalization is applied to stabilize training under varying reward scales. Extensive evaluations demonstrate that SLAS consistently surpasses the DanceGRPO baseline across multiple backbones and benchmarks. In particular, it yields faster training dynamics, improved out-of-domain performance on GenEval and UniGenBench++, and enhanced robustness to model scaling, while mitigating reward hacking and preserving semantic and compositional fidelity in generations.

cs.CV↗

arXiv2026

Bridging Perception and Action: A Lightweight Multimodal Meta-Planner Framework for Robust Earth Observation Agents

Autonomous Earth Observation (EO) agents are transitioning from passive perception to complex, multi-step task execution. However, current architectures that integrate planning and execution within a single model often struggle with combinatorial complexity and reasoning errors in dynamic EO scenarios. To resolve these challenges, we propose the Lightweight Multimodal Meta-Planner (LMMP) framework. LMMP incorporates a dual-awareness mechanism that grounds strategic plans in both multimodal image features and high-level task semantics. Crucially, we introduce a Meta Task Library to inject remote sensing expert knowledge directly into the workflow, which standardizes domain logic and ensures plans are physically feasible. We further implement a two-stage training pipeline, initializing the Meta-Planner via expert-distilled Supervised Fine-Tuning and refining it through Direct Preference Optimization based on execution feedback. Extensive experiments on a dataset derived from EarthBench and ThinkGeo demonstrate that LMMP significantly improves tool-calling accuracy and task success rates. Moreover, the framework exhibits strong ``plug-and-play'' versatility, consistently enhancing the performance of diverse executor backbones across previously unseen EO missions.

cs.MA↗

arXiv2026

DDO-RM: Distribution-Level Policy Improvement after Reward Learning

Recent theory suggests that reward-model-first methods can be more sample-efficient than direct policy fitting when the reward function is statistically simpler than the induced policy. We propose DDO-RM, a finite-candidate decision-optimization method that converts reward scores into an explicit target distribution. Unlike PPO-based RLHF or DPO, DDO-RM performs a KL-regularized mirror-descent update to project the policy toward a reward-improved distribution over a candidate set. Preliminary experiments on Pythia-410M show that DDO-RM outperforms DPO in pair accuracy (0.52 to 0.56) and mean margin (0.13 to 0.53). Our framework provides a principled connection between reward learning and mirror-descent policy improvement.

stat.ML↗