arXiv subjects

Jialu Wang

Publications and source records attributed to Jialu Wang.

At least 19 recordsLinked to original sources

arXiv2026

Adaptive KappaSharp: Condition-Number Shaping for Preferential Bayesian Optimization

Preferential Bayesian optimization (PBO) optimizes objectives accessible only through pairwise user comparisons. The standard approach fits a Gaussian process surrogate for observed pairwise comparisons (PairwiseGP) using the Laplace approximation and selects queries with the Expected Utility of Best Option (EUBO) acquisition function. EUBO queries new candidates at each step, producing pairs that share no candidates with previous queries. Each such pair forms an isolated component in the comparison graph, removing one degree of freedom from the likelihood Hessian and making it rank-deficient. This deficiency is structural and cannot be resolved by changing the surrogate modeling approach. Existing approaches to remedy this issue either waste query budget by forcing comparisons to stay connected, or apply uniform regularization that also perturbs directions already well-constrained by the observed comparisons. We propose KappaSharp that enables a diagonal correction to the Hessian to reduce its condition number, with larger corrections where the prior uncertainty is higher. The correction is only applied in the model fitting step, not query selection. An adaptive variant of KappaSharp is also presented that activates the correction only when the surrogate is confident about recent comparisons, avoiding unnecessary corrections when the problem is well-conditioned. On 11 benchmarks (5--20 dimensions), including a 16-dimensional controller tuning problem in plasma medicine, Adaptive KappaSharp outperforms the standard PBO baseline, with up to +10.9% ($p{=}0.003$).

cs.LG↗

arXiv2026

SAF3R: Dynamic Sparse Attention for Feed-Forward 3D Reconstruction Transformers

Feed-forward 3D reconstruction (F3R) transformers have recently achieved remarkable success. However, scaling them to long image sequences remains challenging, as the quadratic complexity of cross-view global attention quickly becomes the dominant computational bottleneck. While recent efforts attempt to improve efficiency through compressed or sparse attention, they fail to fully exploit the inherent sparsity and dynamic behavior of global attention. In this work, we present a comprehensive analysis of global attention across multiple F3R transformers and reveal that attention patterns are highly heterogeneous, dynamic, and extremely sparse across layers and attention heads. Motivated by these findings, we propose SAF3R, a training-free dynamic sparse attention framework tailored to F3R transformers. SAF3R integrates tailored sparse attention mechanisms with offline head profiling and an efficient online adaptation strategy to match input-dependent attention behaviors. Extensive experiments demonstrate that SAF3R achieves high sparsity ratios while preserving camera pose estimation and 3D reconstruction quality, translating into substantial end-to-end speedup on F3R transformers compared to existing methods. Code is available at https://github.com/jndeng/SAF3R

cs.CV↗

arXiv2026

When Attribution Patching Lies: Diagnosis and a Second-Order Correction

A central goal of mechanistic interpretability is to identify which internal components causally drive a language model's behavior. Because these importance estimates serve as the evidence for identifying circuits, systematic errors can lead to the misidentification of the underlying mechanisms. While activation patching provides a gold-standard causal metric, its computational cost is prohibitive at scale. Practitioners instead rely on attribution patching, a gradient-based, first-order approximation whose reliability remains poorly understood. In this work, we characterize the source of this unreliability, demonstrating that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component. This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product (HVP) correction that eliminates the leading-order error with only one additional backward pass. In evaluations across five model families (124M-9B parameters) and both random-token and naturalistic (name-swap) perturbations, HVP is the only second-order correction feasible at larger scale, where standard baselines like Integrated Gradients become computationally prohibitive. In comparative experiments, a multi-step HVP variant matches or exceeds the accuracy of Integrated Gradients at significantly lower compute, outperforming prior second-order baselines. These improvements lead to higher-fidelity circuit recovery on standard benchmarks and support a Screen-Flag-Fix workflow that targets computational effort only toward the components flagged as unreliable.

cs.LG↗

arXiv2026

Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior

Multi-LLM systems use multiple language models to deliberate, judge each other's outputs, or coordinate as agents. Their value depends on the models producing measurably different conversational behaviors when given the same input. Prior offline studies recommend drawing one model per family for behavioral diversity, because LLMs prefer outputs from their own family when rating one another in isolation. Whether the same family label predicts behavior in interactive multi-LLM systems, the setting that real deployed systems use, has not been tested. We study this with a 940,000-chain 11-checkpoint corpus and a 1.6M-chain same-base Llama factorial. On our validated headline metric, hedging, a reasoning-distilled Llama checkpoint shifts by 18% depending on which same-base partner it replies to, more than any cross-family hedging gap in the controlled subset. Qwen, closed-API, and runtime checks suggest the pattern is not isolated, while repair and challenge analyses remain exploratory because their surface-cue detectors are weaker. Overall, the results identify post-training recipe as a first-class axis for multi-LLM panel composition and show that model family alone is an incomplete proxy for conversational diversity.

cs.CL↗

arXiv2026

UniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale

Modern recommendation systems predominantly train retrieval and ranking as separate models despite both increasingly relying on large transformers encoding the same user behavior data, duplicating parameters, compute, and serving cost. Prior work unifies the model architecture but not the full pipeline: input formats, training procedures, and serving stacks remain fragmented across stages. We present UniPinRec, which achieves full-stack unification of retrieval and ranking at Pinterest: one input format, one model, one training stage, deployed within existing serving infrastructure. A shared transformer encodes the user action sequence into candidate-independent representations that branch into retrieval (ANN dot-product) and ranking (cross-attention) via task-specific heads. Three ideas make this work: (1) Masked Action Modeling (MAM) eliminates interleaving, enabling weight sharing without doubling context length; (2) Blended training examples pair action sequences with feedview impression slates to satisfy both objectives jointly; (3) Cross-stage KV cache sharing reuses user-history computation from retrieval for ranking, reducing total FLOPs versus serving two independent models. Deployed in the Pinterest core surfaces, UniPinRec delivers approximately +1% online engagement lift while cutting end-to-end serving latency by 11.1% and lifting QPS by 63.6%. To our knowledge, this is the first full-stack unification of retrieval and ranking, covering inputs, model, training and serving, deployed in a production recommendation system.

cs.IR↗

arXiv2026

MasFACT: Continual Multi-Agent Topology Learning via Geometry-Aware Posterior Transfer

Multi-agent systems (MAS) powered by large language models (LLMs) have emerged as a powerful paradigm for complex problem solving, where performance critically depends on the underlying inter-agent communication topology. However, existing topology generation methods mainly optimize for isolated tasks, while real-world deployments involve streams of evolving tasks, requiring previously effective collaboration patterns to be retained and reused rather than rediscovered or overwritten. We identify a previously underexplored failure mode, \emph{topology forgetting}, in which adapting to new tasks shifts the topology generator away from communication structures required by earlier tasks. This issue stems from cross-task misalignment in both agent-level functional semantics and relational communication structures. To address this challenge, we propose \textbf{\textsc{MasFACT}}, a geometry-aware posterior transfer framework that preserves and reuses historical collaboration knowledge as transferable topology priors. We transfer these priors across task-specific agent spaces through Fused Gromov-Wasserstein optimal transport and perform PAC-Bayes-guided conservative posterior adaptation to balance task-specific plasticity with structural stability. Experiments across class-, domain-, and task-level continual settings demonstrate that \textsc{MasFACT} consistently improves average accuracy while reducing topology forgetting compared to strong topology generation and replay-based baselines, and can be seamlessly integrated with different MAS topology generators.

cs.LG↗

arXiv2026

Do Agents Repair When Challenged -- or Just Reply? Challenge, Repair, and Public Correction in a Deployed Agent Forum

As large language model (LLM) agents are deployed in public interactive settings, a key question is whether their communities can sustain challenge, repair, and public correction, or merely produce norm-like language. We compare Moltbook, a live deployed agent forum, with five matched Reddit communities by tracing a three-step mechanism: whether discussions create threaded exchange, whether challenges elicit a response, and whether correction becomes visible to the wider thread. Relative to Reddit, Moltbook discussions are roughly ten times less threaded, leaving far fewer chances for challenge and response. When challenges do occur, the original author almost never returns (1.2% vs. 40.9% on Reddit), multi-turn continuation is nearly absent (0.1% vs. 38.5%), and we detect no repairs under a shared conservative protocol. A non-challenge baseline within Reddit suggests this gap is linked to challenge, not simply deeper threading. These results indicate that social alignment depends not only on producing norm-aware language, but on sustaining the interactional processes through which communities teach, enforce, and revise norms. This matters for safety, because correction is increasingly decentralized, and for fairness, because communities differ in how they expect participants to engage with challenge.

cs.CY↗

arXiv2026

$L^p$-estimates for the wave equation with partial inverse-square potentials

This paper investigates $L^p$-estimates for solutions to the wave equation perturbed by a scaling-critical partial inverse-square potential. We study a model in which the singularity of the potential appears only in a subset of the variables, corresponding to the Schr\"{o}dinger operator $\mathcal{H}_a = -\Delta_x - \Delta_y + a/|x|^2$ on $\mathbb{R}^{2+n}$. Using spectral analysis, we establish the $L^p$-boundedness of the wave propagator $(1+\sqrt{\mathcal{H}_a})^{-\gamma} e^{it\sqrt{\mathcal{H}_a}}$ for a range of exponents $\gamma$ and $p$ satisfying $|1/p -1/2| < \gamma/(n+1)$. The key ingredients are the spectral measure kernel of the partial inverse-square operator $\mathcal{H}_a$ and the complex interpolation argument.

math.AP↗

arXiv2026

SPOT: Span-level Pause-of-Thought for Efficient and Interpretable Latent Reasoning in Large Language Models

Explicit Chain-of-Thought improves the reasoning performance of large language models but often incurs high inference cost due to verbose token-level traces. While recent approaches reduce this overhead via concise prompting or step pruning, they largely truncate what the model says rather than internalize what the model thinks. Latent reasoning offers a promising alternative by performing computation in the hidden space, yet prior methods face two critical challenges. Many existing approaches rely on rigid point-to-point alignment, forcing a latent token to approximate the final representation of a reasoning step, which can be insufficient to capture the dense, variable-length semantics of an entire reasoning segment. Furthermore, these methods often suffer from a lack of interpretability: latent states are commonly produced by unconstrained optimization or embedding mixing, yielding vectors that are difficult to decode or audit under the pretrained language head. We propose SPOT, a flexible framework that compresses explicit CoT into compact latent pause tokens without enforcing a fixed response template. At the core of SPOT is Span-level Semantic Alignment, a Sinkhorn optimal-transport objective that softly matches each pause token to the semantics of an entire reasoning segment, overcoming the rigidity of step-end alignment. To further improve interpretability, SPOT introduces a Frozen-Head Decoding Constraint that keeps latent states directly decodable as token distributions under the frozen pretrained LM head, enabling readable keyword interpretations of latent thoughts. Experiments on reasoning benchmarks demonstrate that SPOT improves accuracy by 2.3 points on average while reducing generated tokens by 37.5% and provides faithful semantic interpretations of the latent reasoning process.

cs.CL↗

arXiv2026

AgenticGEO: A Self-Evolving Agentic System for Generative Engine Optimization

Generative search engines represent a transition from traditional ranking-based retrieval to Large Language Model (LLM)-based synthesis, transforming optimization goals from ranking prominence towards content inclusion. Generative Engine Optimization (GEO), specifically, aims to maximize visibility and attribution in black-box summarized outputs by strategically manipulating source content. However, existing methods rely on static heuristics, single-prompt optimization, or engine preference rule distillation that is prone to overfitting. They cannot flexibly adapt to diverse content or the changing behaviors of generative engines. Moreover, effectively optimizing these strategies requires an impractical amount of interaction feedback from the engines. To address these challenges, we propose AgenticGEO, a self-evolving agentic framework formulating optimization as a content-conditioned control problem, which enhances intrinsic content quality to robustly adapt to the unpredictable behaviors of black-box engines. Unlike fixed-strategy methods, AgenticGEO employs a MAP-Elites archive to evolve diverse, compositional strategies. To mitigate interaction costs, we introduce a Co-Evolving Critic, a lightweight surrogate that approximates engine feedback for content-specific strategy selection and refinement, efficiently guiding both evolutionary search and inference-time planning. Through extensive in-domain and cross-domain experiments on two representative engines, AgenticGEO achieves state-of-the-art performance and demonstrates robust transferability, outperforming 14 baselines across 3 datasets. Our code and model are available at: https://github.com/AIcling/agentic_geo.

cs.AI↗

arXiv2026

Personalized Group Relative Policy Optimization for Heterogenous Preference Alignment

Despite their sophisticated general-purpose capabilities, Large Language Models (LLMs) often fail to align with diverse individual preferences because standard post-training methods, like Reinforcement Learning with Human Feedback (RLHF), optimize for a single, global objective. While Group Relative Policy Optimization (GRPO) is a widely adopted on-policy reinforcement learning framework, its group-based normalization implicitly assumes that all samples are exchangeable, inheriting this limitation in personalized settings. This assumption conflates distinct user reward distributions and systematically biases learning toward dominant preferences while suppressing minority signals. To address this, we introduce Personalized GRPO (P-GRPO), a novel alignment framework that decouples advantage estimation from immediate batch statistics. By normalizing advantages against preference-group-specific reward histories rather than the concurrent generation group, P-GRPO preserves the contrastive signal necessary for learning distinct preferences. We evaluate P-GRPO across diverse tasks and find that it consistently achieves faster convergence and higher rewards than standard GRPO, thereby enhancing its ability to recover and align with heterogeneous preference signals. Our results demonstrate that accounting for reward heterogeneity at the optimization level is essential for building models that faithfully align with diverse human preferences without sacrificing general capabilities.

cs.LG↗

arXiv2026

Cosmo3DFlow: Wavelet Flow Matching for Spatial-to-Spectral Compression in Reconstructing the Early Universe

Reconstructing the early universe from the evolved present-day universe is a challenging and computationally demanding problem in modern astrophysics. We devise a novel generative framework, Cosmo3DFlow, designed to address dimensionality and sparsity, the critical bottlenecks inherent in current state-of-the-art methods for cosmological inference. By integrating 3D Discrete Wavelet Transform (DWT) with flow matching, we effectively represent high-dimensional cosmological structures. The Wavelet Transform addresses the ``void problem'' by translating spatial emptiness into spectral sparsity. It decouples high-frequency details from low-frequency structures, and wavelet-space velocity fields facilitate stable ordinary differential equation (ODE) solvers with large step sizes. Using large-scale cosmological $N$-body simulations at $128^3$ resolution, we achieve up to $46\times$ faster sampling than diffusion models. Our results enable initial conditions to be sampled in seconds, compared to minutes for previous methods.

astro-ph.IM↗

arXiv2026

Large room temperature anomalous Nernst effect coupled with topological Nernst effect from incommensurate spin structure in a Kagome antiferromagnet

Kagome magnets exhibit a range of novel and nontrivial topological properties due to the strong interplay between topology and magnetism, which also extends to their thermoelectric applications. Recent advances in the study of magnetic topological materials have highlighted their intriguing anomalous Hall and thermoelectric effects, arising primarily from large intrinsic Berry curvature. Here, we report observation of a large room-temperature (RT) anomalous Nernst effects (ANE) of S_xy^A ~ 1.3 {\mu}V K^(-1) in the kagome antiferromagnet (AFM) ErMn6Sn6, which is comparable to the largest signals observed in known magnetic materials. Surprisingly, we further found that a significant topological Nernst signal at RT and peaking a maximum of approximately 0.2 {\mu}V K^(-1) at 180 K, exactly coupling with ANE in the spiral AFM state, originates from the real-space nonzero spin chirality caused by incommensurate spin structure. This study demonstrates a potential room-temperature thermoelectric application platform based on Nernst effect, and provides insights for discovering significant anomalous and topological transverse transport effects in the incommensurate AFM system.

cond-mat.str-el↗

arXiv2025

Large longitudinal and anomalous transverse Magneto-thermoelectric effect in kagome antiferromagnet FeGe

Topological Kagome magnets, characterized by nontrivial electronic band structures featuring flat band, Dirac cone and van Hove singularities, provide a new avenue for the realization of thermoelectric devices. Unlike the conventional longitudinal Seebeck effect, transverse thermoelectric (TE) effects like the Nernst effect have attracted growing interest due to their unique transverse geometry and potential advantages. Here, we report the observation of a significant transverse thermoelectric conductivity alpha A_zx of 15 A K-1m-1 at low temperatures, together with a pronounced anomalous Nernst effect in the Kagome antiferromagnet FeGe, which exhibits a charge density wave inside the antiferromagnetic (AFM) state. This value is the highest record among known AFM materials. Furthermore, the thermopower at 14 T increases by 102-104% around the canted-AFM (CAFM) transition temperature, Tcant, comparable to that of the well-known AFM thermoelectric materials. These effects are attributed to large Berry curvature arising from the non-collinear spin texture in FeGe, highlighting its potential for enhancing thermoelectric performance and its candidacy for magneto-TE applications in Kagome antiferromagnetic materials.

cond-mat.mtrl-sci↗

arXiv2025

SPECTRA: Revealing the Full Spectrum of User Preferences via Distributional LLM Inference

Large Language Models (LLMs) are increasingly used to model user preferences, with the typical output as a directly-generated ranked item list per user. However, this generative paradigm inherits the bias and opacity of autoregressive decoding. It over-emphasizes frequent (head) preferences and suppresses minority, long-tail ones. To address this, we propose SPECTRA (Softmax Probing for Extracted Category-level Token Readouts and Analysis), which treats the finetuned LLM as an implicit probabilistic model and probes its softmax to infer a probability distribution over semantically interpretable preference categories. We evaluate SPECTRA on MovieLens, Yelp, and a large-scale short-video platform. SPECTRA delivers (i) distributional alignment, reducing Jensen-Shannon divergence to the empirical preference distribution by 38 to 44 percent across public datasets; (ii) long-tail recovery with cross-user fairness, raising top-3 category exposure entropy by 23 percent on MovieLens and producing a larger gain on tail-preference users than on head-preference users; and (iii) downstream application value, with a 41 to 46 percent category-NDCG boost on MovieLens and Yelp, and a 7x improvement on long-tail category ranking on a large-scale deployment against a head-optimized production ranker.

cs.CL↗

arXiv2025

Anomalous Nernst Effect and Its Implications for Time-Reversal Symmetry Breaking in Kagome Metal ScV6Sn6

The nonmagnetic kagome metal ScV6Sn6 displays an unconventional charge order (CO) accompanied by signatures of an anomalous Hall effect, hidden magnetism, and multiple lattice instabilities. In this study, we report the observation of unconventional anomalous thermoelectric properties. Notably, unexpected anomalous transverse Nernst signals reach a peak value of ~4 {\mu}V/K near the TCDW ~92 K in ScV6Sn6, and these signals persist in the charge-ordered state as the temperature decreases to 10 K. Furthermore, both thermopower and thermal conductivity exhibit significant changes under magnetic fields, even in the nonmagnetic ground state. These observations strongly suggest the emergence of time-reversal symmetry breaking in ScV6Sn6, as supported by muon spin relaxation ({\mu}SR) measurements. While hidden magnetism represents the most plausible origin, alternative mechanisms involving orbital currents and chiral charge order remain possible.

cond-mat.str-el↗

arXiv2025

Anomalous Hall and Nernst effects in the Two-Dimensional ferromagnetic metal FePd2Te2

The transverse thermoelectric effect enables simpler, more flexible thermoelectric devices by generating electricity perpendicular to heat flow, offering promising solutions for waste heat recovery and solid-state cooling applications. Here, we report a striking observation of zero-field anomalous Hall effect (AHE) and anomalous Nernst effect (ANE) below TC in the two-dimensional metallic magnet FePd2Te2. The anomalous Nernst signal Syx^A peaks a maximum value of 0.15 {\mu}V/K at 100 K, much larger than that of conventional FM materials. Remarkably, the derived ratio alpha_ij/sigma_ij in FePd2Te2 approaches the fundamental limit of 86 {\mu}V/K. Our findings suggest a dominant Berry curvature contribution to the ANE. The observed giant zero-field anomalous Nernst response in 2D FePd2Te2 not only advances fundamental understanding of transverse thermoelectricity in layered magnets, but also provides this material as a promising candidate for practical thermoelectric spintronic applications.

cond-mat.str-el↗

arXiv2025

$L^p$-estimates for the wave equation with critical magnetic field in higher dimensions

In this paper, we study the $L^{p}$-estimates for the solution to the wave equation with a scaling-critical magnetic potential in Euclidean $R^N$ with $N\geq3$. Inspired by the work of \cite{L}, we show that the operators $(I+\mathcal{L}_{\mathbf{A}})^{-\gamma}e^{it\sqrt{\mathcal{L}_{\mathbf{A}}}}$ is bounded in $L^{p}(\mathbb{R}^{N})$ for $1 |1/p-1/2|$ and $t>0$, where $\mathcal{L}_{\mathbf{A}}$ is a magnetic Schr\"odinger operator. In particular, we derive the $L^{p}$-bounds for the sine wave propagator $\sin(t\sqrt{\mathcal{L}_{\mathbf{A}}})\mathcal{L}^{-\frac12}_{\mathbf{A}}$. The key ingredient is the $L^p\rightarrow L^p$ boundedness of the analytic operator family $f_{w,t}(\mathcal{L}_{\mathbf{A}})$.

math.AP↗