arXiv subjects

Junpu Wang

Publications and source records attributed to Junpu Wang.

At least 19 recordsLinked to original sources

arXiv2026

TALAN: Task-Aligned Latent Adaptation Networks for Targeted Post-Training of Large Language Models

Targeted post-training aims to improve reasoning, math, and code without degrading strengths. Low-rank adapters are efficient but task-global; activation interventions are input-aware but often require separate probes, vectors, or inference-time steering. We introduce TALAN (Task-Aligned Latent Adaptation Networks), a sequence-conditioned latent side path inserted into a transformer's residual stream and co-trained with a low-rank adapter in one SFT loop. TALAN compresses the active sequence into latent memory, remixes it into token-level perturbations, and writes them back through a controlled residual update. It is configured along six axes: insertion location, memory size, mixer, writeback rule, trainability scope, and gradient scale. Across four Qwen3-family backbones and four STEM/code benchmarks, TALAN improves matched LoRA and DoRA baselines. With LoRA, it yields a +1.41 point cross-model mean gain, positive on all four backbones and non-negative on all 16 model-benchmark cells. With DoRA, it yields a +1.85 point mean gain, positive on all backbones and on 13 of 16 cells. Paired seed checks support positive average effects but show nontrivial variance, so we treat them as sensitivity checks. Cost is small: <1% trainable parameters relative to the backbone and 1.01-1.02x inference overhead versus matched LoRA. A Llama-3.2-1B transfer probe is also positive under LoRA and rsLoRA across seven paired seeds, supporting a transfer beyond Qwen. Internal-state analyses suggest TALAN is a small complementary activation intervention. The matched adapter update is 80-1,700x larger than the TALAN perturbation, yet their directions have near-zero cosine; per-layer measurements show this small orthogonal perturbation propagates and amplifies through depth. TALAN offers a practical platform for studying steerable activation-level adaptation within standard adapter-based post-training.

cs.LG↗

arXiv2024

Multi-feature Reconstruction Network using Crossed-mask Restoration for Unsupervised Industrial Anomaly Detection

Unsupervised anomaly detection using only normal samples is of great significance for quality inspection in industrial manufacturing. Although existing reconstruction-based methods have achieved promising results, they still face two problems: poor distinguishable information in image reconstruction and well abnormal regeneration caused by model under-regularization. To overcome the above issues, we convert the image reconstruction into a combination of parallel feature restorations and propose a multi-feature reconstruction network, MFRNet, using crossed-mask restoration in this paper. Specifically, a multi-scale feature aggregator is first developed to generate more discriminative hierarchical representations of the input images from a pre-trained model. Subsequently, a crossed-mask generator is adopted to randomly cover the extracted feature map, followed by a restoration network based on the transformer structure for high-quality repair of the missing regions. Finally, a hybrid loss is equipped to guide model training and anomaly estimation, which gives consideration to both the pixel and structural similarity. Extensive experiments show that our method is highly competitive with or significantly outperforms other state-of-the-arts on four public available datasets and one self-made dataset.

cs.CV↗

arXiv2022

Defect Transformer: An Efficient Hybrid Transformer Architecture for Surface Defect Detection

Surface defect detection is an extremely crucial step to ensure the quality of industrial products. Nowadays, convolutional neural networks (CNNs) based on encoder-decoder architecture have achieved tremendous success in various defect detection tasks. However, due to the intrinsic locality of convolution, they commonly exhibit a limitation in explicitly modeling long-range interactions, critical for pixel-wise defect detection in complex cases, e.g., cluttered background and illegible pseudo-defects. Recent transformers are especially skilled at learning global image dependencies but with limited local structural information necessary for detailed defect location. To overcome the above limitations, we propose an efficient hybrid transformer architecture, termed Defect Transformer (DefT), for surface defect detection, which incorporates CNN and transformer into a unified model to capture local and non-local relationships collaboratively. Specifically, in the encoder module, a convolutional stem block is firstly adopted to retain more detailed spatial information. Then, the patch aggregation blocks are used to generate multi-scale representation with four hierarchies, each of them is followed by a series of DefT blocks, which respectively include a locally position-aware block for local position encoding, a lightweight multi-pooling self-attention to model multi-scale global contextual relationships with good computational efficiency, and a convolutional feed-forward network for feature transformation and further location information learning. Finally, a simple but effective decoder module is proposed to gradually recover spatial details from the skip connections in the encoder. Extensive experiments on three datasets demonstrate the superiority and efficiency of our method compared with other CNN- and transformer-based networks.

cs.CV↗

arXiv2019

FedMD: Heterogenous Federated Learning via Model Distillation

Federated learning enables the creation of a powerful centralized model without compromising data privacy of multiple participants. While successful, it does not incorporate the case where each participant independently designs its own model. Due to intellectual property concerns and heterogeneous nature of tasks and data, this is a widespread requirement in applications of federated learning to areas such as health care and AI as a service. In this work, we use transfer learning and knowledge distillation to develop a universal framework that enables federated learning when each agent owns not only their private data, but also uniquely designed models. We test our framework on the MNIST/FEMNIST dataset and the CIFAR10/CIFAR100 dataset and observe fast improvement across all participating models. With 10 distinct participants, the final test accuracy of each model on average receives a 20% gain on top of what's possible without collaboration and is only a few percent lower than the performance each model would have obtained if all private datasets were pooled and made directly available for all participants.

cs.LG↗

arXiv2017

Exact Virasoro Blocks from Wilson Lines and Background-Independent Operators

Aspects of black hole thermodynamics and information loss can be derived as a consequence of Virasoro symmetry. To bolster the connection between Virasoro conformal blocks and AdS$_3$ quantum gravity, we study sl$(2)$ Chern-Simons Wilson line networks and revisit the idea that they compute a variety of CFT$_2$ observables, including Virasoro OPE blocks, exactly. We verify this in the semiclassical large central charge limit and to low orders in a perturbative $1/c$ expansion. Wilson lines connecting the boundary to points in the bulk play a natural role in bulk reconstruction. Because quantum gravity in AdS$_3$ is rigidly fixed by Virasoro symmetry, we argue that sl$(2)$ Wilson lines provide building blocks for background independent bulk reconstruction. In particular, we show explicitly that they automatically compute the uniformizing coordinates appropriate to any background state.

hep-th↗

arXiv2016

A Universe Without Dark Energy: Cosmic Acceleration from Dark Matter-Baryon Interactions

Cosmic acceleration is widely believed to require either a source of negative pressure (i.e., dark energy), or a modification of gravity, which necessarily implies new degrees of freedom beyond those of Einstein gravity. In this paper we present a third possibility, using only dark matter and ordinary matter. The mechanism relies on the coupling between dark matter and ordinary matter through an effective metric. Dark matter couples to an Einstein-frame metric, and experiences a matter-dominated, decelerating cosmology up to the present time. Ordinary matter couples to an effective metric that depends also on the DM density, in such a way that it experiences late-time acceleration. Linear density perturbations are stable and propagate with arbitrarily small sound speed, at least in the case of `pressure' coupling. Assuming a simple parametrization of the effective metric, we show that our model can successfully match a set of basic cosmological observables, including luminosity distance, BAO measurements, angular-diameter distance to last scattering {\it etc.} For the growth history of density perturbations, we find an intriguing connection between the growth factor and the Hubble constant. To get a growth history similar to the $Λ$CDM prediction, our model predicts a higher $H_0$, closer to the value preferred by direct estimates. On the flip side, we tend to overpredict the growth of structures whenever $H_0$ is comparable to the Planck preferred value. The model also tends to predict larger redshift-space distortions at low redshift than $Λ$CDM.

hep-th↗

arXiv2016

Degenerate Operators and the $1/c$ Expansion: Lorentzian Resummations, High Order Computations, and Super-Virasoro Blocks

One can obtain exact information about Virasoro conformal blocks by analytically continuing the correlators of degenerate operators. We argued in recent work that this technique can be used to explicitly resolve information loss problems in AdS$_3$/CFT$_2$. In this paper we use the technique to perform calculations in the small $1/c \propto G_N$ expansion: (1) we prove the all-orders resummation of logarithmic factors $\propto \frac{1}{c} \log z$ in the Lorentzian regime, demonstrating that $1/c$ corrections directly shift Lyapunov exponents associated with chaos, as claimed in prior work, (2) we perform another all-orders resummation in the limit of large $c$ with fixed $cz$, interpolating between the early onset of chaos and late time behavior, (3) we explicitly compute the Virasoro vacuum block to order $1/c^2$ and $1/c^3$, corresponding to $2$ and $3$ loop calculations in AdS$_3$, and (4) we derive the heavy-light vacuum blocks in theories with $\mathcal{N}=1,2$ superconformal symmetry.

hep-th↗

arXiv2016

On Information Loss in AdS$_3$/CFT$_2$

We discuss information loss from black hole physics in AdS$_3$, focusing on two sharp signatures infecting CFT$_2$ correlators at large central charge $c$: 'forbidden singularities' arising from Euclidean-time periodicity due to the effective Hawking temperature, and late-time exponential decay in the Lorentzian region. We study an infinite class of examples where forbidden singularities can be resolved by non-perturbative effects at finite $c$, and we show that the resolution has certain universal features that also apply in the general case. Analytically continuing to the Lorentzian regime, we find that the non-perturbative effects that resolve forbidden singularities qualitatively change the behavior of correlators at times $t \sim S_{BH}$, the black hole entropy. This may resolve the exponential decay of correlators at late times in black hole backgrounds. By Borel resumming the $1/c$ expansion of exact examples, we explicitly identify 'information-restoring' effects from heavy states that should correspond to classical solutions in AdS$_3$. Our results suggest a line of inquiry towards a more precise formulation of the gravitational path integral in AdS$_3$.

hep-th↗

arXiv2015

Hawking from Catalan

The Virasoro algebra determines all `graviton' matrix elements in AdS$_3$/CFT$_2$. We study the explicit exchange of any number of Virasoro gravitons between heavy and light CFT$_2$ operators at large central charge. These graviton exchanges can be written in terms of new on-shell tree diagrams, organized in a perturbative expansion in $h_H/c$, the heavy operator dimension divided by the central charge. The Virasoro vacuum conformal block, which is the sum of all the tree diagrams, obeys a differential recursion relation generalizing that of the Catalan numbers. We use this recursion relation to sum the on-shell diagrams to all orders, computing the Virasoro vacuum block. Extrapolating to large $h_H/c$ determines the Hawking temperature of a BTZ black hole in dual AdS$_3$ theories.

hep-th↗

arXiv2015

Eikonalization of Conformal Blocks

Classical field configurations such as the Coulomb potential and Schwarzschild solution are built from the t-channel exchange of many light degrees of freedom. We study the CFT analog of this phenomenon, which we term the `eikonalization' of conformal blocks. We show that when an operator $T$ appears in the OPE $\mathcal{O}(x) \mathcal{O}(0)$, then the large spin $\ell$ Fock space states $[TT \cdots T]_{\ell}$ also appear in this OPE with a computable coefficient. The sum over the exchange of these Fock space states in an $\langle \mathcal{O} \mathcal{O} \mathcal{O} \mathcal{O} \rangle$ correlator build the classical `$T$ field' in the dual AdS description. In some limits the sum of all Fock space exchanges can be represented as the exponential of a single $T$ exchange in the 4-pt correlator of $\mathcal{O}$. Our results should be useful for systematizing $1/\ell$ perturbation theory in general CFTs and simplifying the computation of large spin OPE coefficients. As examples we obtain the leading $\log \ell$ dependence of Fock space conformal block coefficients, and we directly compute the OPE coefficients of the simplest `triple-trace' operators.

hep-th↗

arXiv2014

Reheating constraints to inflationary models

Evidence from the BICEP2 experiment for a significant gravitational-wave background has focussed attention on inflaton potentials $V(ϕ) \propto ϕ^α$ with $α=2$ ("chaotic" or "$m^2ϕ^2$" inflation) or with smaller values of $α$, as may arise in axion-monodromy models. Here we show that reheating considerations may provide additional constraints to these models. The reheating phase preceding the radiation era is modeled by an effective equation-of-state parameter $w_{\rm re}$. The canonical reheating scenario is then described by $w_{\rm re}=0$. The simplest $α=2$ models are consistent with $w_{\rm re} = 0$ for values of $n_s$ well within the current $1σ$ range. Models with $α=1$ or $α=2/3$ require a more exotic reheating phase, with $-1/3 1/3$, unless $n_s$ is close to the lower limit of the $2σ$ range. For $m^2ϕ^2$ inflation and canonical reheating as a benchmark, we derive a relation $\log_{10}\left(T_{\rm re}/10^6\,{\rm GeV} \right) \simeq 2000\,(n_s-0.96)$ between the reheat temperature $T_{\rm re}$ and the scalar spectral index $n_s$. Thus, if $n_s$ is close to its central value, then $T_{\rm re}\lesssim 10^6$~GeV, just above the electroweak scale. If the reheat temperature is higher, as many theorists may prefer, then the scalar spectral index should be closer to $n_s\simeq0.965$ (at the pivot scale $k=0.05\,{\rm Mpc}^{-1}$), near the upper limit of the $1σ$ error range. Improved precision in the measurement of $n_s$ should allow $m^2ϕ^2$, axion-monodromy, and $ϕ^4$ models to be distinguished, even without precise measurement of $r$, and to test the $m^2ϕ^2$ expectation of $n_s\simeq0.965$.

astro-ph.CO↗

arXiv2014

An Effective Theory for Holographic RG Flows

We study the dilaton action induced by RG flows between holographic CFT fixed points. For this purpose we introduce a general bulk effective theory for the goldstone boson of the broken spacetime symmetry, providing an AdS analog of the EFT of Inflation. In two dimensions, we use the effective theory to compute the dilaton action, as well as the UV and IR conformal anomalies, without further assumptions. In higher dimensions we take a `slow-flow' limit analogous to the assumption of slow-roll in Inflation, and in this context we obtain the dilaton action, focusing on terms proportional to the difference of the A-type anomalies. We include Gauss-Bonnet terms in the gravitational action in order to verify that our method correctly differentiates between A-type and other anomalies.

hep-th↗

arXiv2014

Non-Trivial Checks of Novel Consistency Relations

Single-field perturbations satisfy an infinite number of consistency relations constraining the squeezed limit of correlation functions at each order in the soft momentum. These can be understood as Ward identities for an infinite set of residual global symmetries, or equivalently as Slavnov-Taylor identities for spatial diffeomorphisms. In this paper, we perform a number of novel, non-trivial checks of the identities in the context of slow-roll single field inflationary models with arbitrary sound speed. We focus for concreteness on identities involving 3-point functions with a soft external mode, and consider all possible scalar and tensor combinations for the hard-momentum modes. In all these cases, we check the consistency relations up to and including cubic order in the soft momentum. For this purpose, we compute for the first time the 3-point functions involving 2 scalars and 1 tensor, as well as 2 tensors and 1 scalar, for arbitrary sound speed.

hep-th↗

arXiv2013

The squeezed limit of the solid inflation three-point function

The recently proposed model of 'solid inflation' features a peculiar three-point function for scalar perturbations with an anisotropic, purely quadrupolar, squeezed limit. We confirm this result as well as the overall amplitude of the three point-function via an extremely simple computation, where we focus on the squeezed limit from the start and follow the standard logic adopted in deriving the consistency relations. Our system violates the consistency relations, but in the squeezed limit the three-point function can still be traded for a background-dependent two-point function, which is immediate to compute. Additionally, we use these simple methods to derive some new results - namely, certain squeezed limits of the three-point correlators involving vector and tensor perturbations as well.

hep-th↗

arXiv2013

Construction of the Conserved Non-linear Zeta via the Effective Action for Perfect Fluids

We consider the problem of how to construct the curvature perturbation $ζ$ to nonlinear levels, which is expected to evolve time independently on super-horizon scales; in particular we concentrate on the situation where the universe is dominated by a perfect fluid. We have used a low energy/long wavelength effective action to model the fluid sector. Different from previous work, our approach assumes neither the absence of vector and tensor perturbations nor ``local homogeneity and isotropy''. As a corollary, we also show that the nonlinearly defined graviton field $γ_{ij}$ is conserved outside the horizon in the same manner as $ζ$ is.

gr-qc↗

arXiv2012

Dissipation in the effective field theory for hydrodynamics: First order effects

We introduce dissipative effects in the effective field theory of hydrodynamics. We do this in a model-independent fashion by coupling the long-distance degrees of freedom explicitly kept in the effective field theory to a generic sector that "lives in the fluid", which corresponds physically to the microscopic constituents of the fluid. At linear order in perturbations, the symmetries, the derivative expansion, and the assumption that this microscopic sector is thermalized, allow us to characterize the leading dissipative effects at low frequencies via three parameters only, which correspond to bulk viscosity, shear viscosity, and--in the presence of a conserved charge--heat conduction. Using our methods we re-derive the Kubo relations for these transport coefficients.

hep-th↗

arXiv2012

Solid Inflation

We develop a cosmological model where primordial inflation is driven by a 'solid', defined as a system of three derivatively coupled scalar fields obeying certain symmetries and spontaneously breaking a certain subgroup of these. The symmetry breaking pattern differs drastically from that of standard inflationary models: time translations are unbroken. This prevents our model from fitting into the standard effective field theory description of adiabatic perturbations, with crucial consequences for the dynamics of cosmological perturbations. Most notably, non-gaussianities in the curvature perturbations are unusually large, with f_NL ~ 1/(ε.c_s^2), and have a novel shape: peaked in the squeezed limit, with anisotropic dependence on how the limit is approached. Other unusual features include the absence of adiabatic fluctuation modes during inflation---which does not impair their presence and near scale-invariance after inflation---and a slightly blue tilt for the tensor modes.

hep-th↗

arXiv2012

No-Go Theorems for Generalized Chameleon Field Theories

The chameleon, or generalizations thereof, is a light scalar that couple to matter with gravitational strength, but whose manifestation depends on the ambient matter density. A key feature is that the screening mechanism suppressing its effects in high-density environments is determined by the local scalar field value. Under very general conditions, we prove two theorems limiting its cosmological impact: i) the Compton wavelength of such a scalar can be at most Mpc at present cosmic density, which restricts its impact to non-linear scales; ii) the conformal factor relating Einstein- and Jordan-frame scale factors is essentially constant over the last Hubble time, which precludes the possibility of self-acceleration. These results imply that chameleon-like scalar fields have a negligible effect on the linear-scale growth history; theories that invoke a chameleon-like scalar to explain cosmic acceleration rely on a form of dark energy rather than a genuine modified gravity effect. Our analysis applies to a broad class of chameleon, symmetron and dilaton theories.

astro-ph.CO↗