arXiv subjects

Cheng Chi

Publications and source records attributed to Cheng Chi.

At least 37 records · Page 2Linked to original sources

arXiv2026

On three open problems in zero-sum Ramsey numbers

Let $K_N^{(r)}$ denote the $N$-vertex complete $r$-uniform hypergraph. For an $r$-uniform hypergraph $H$ and an integer $k\geq2$, the $k$-color Ramsey number $R(H,k)$ is the least integer $N$ such that every $k$-edge-coloring of $K_N^{(r)}$ contains a monochromatic copy of $H$. When $k\mid\esize(H)$, the zero-sum Ramsey number $R(H,\mathbb Z_k)$ is the least integer $N$ such that every edge-labeling of $K_N^{(r)}$ by elements of $\mathbb Z_k$ contains a copy of $H$ whose edge labels sum to $0$ in $\mathbb Z_k$. We settle two conjectures and a problem concerning these two Ramsey numbers. First, Caro and Provstgaard proposed exact values for the zero-sum Ramsey numbers over $\mathbb Z_2$ of delta-systems with an even number of edges. We determine these numbers and thereby prove their conjecture. Second, for a forest $F$ with $m$ edges, let $tF$ denote the disjoint union of $t$ copies of $F$. Caro conjectured that $R(tF,\mathbb Z_{mt})=R(tF,2)$ for all sufficiently large $t$. We show that this conjecture does not hold for double stars. Caro also asked whether there exists a tree $T$ with $m$ edges such that $R(T,\mathbb Z_m)>R(T,2)$. We answer this question affirmatively by constructing an infinite family of such trees.

math.CO↗

arXiv2026

A note on zero-sum Ramsey numbers of complete graphs

For a graph $H$ with $3\mid e(H)$, the zero-sum Ramsey number $R(H,\Z_3)$ is the least integer $N$ such that every labeling of the edges of $K_N$ by elements of $\Z_3$ contains a copy of $H$ whose edge labels sum to zero. We determine the last previously unresolved infinite family in the complete-graph case modulo $3$. More precisely, we prove that $R(K_n,\Z_3)=n+3$ for every $n\ge 10$ satisfying $n\equiv 1\pmod 3$. Consequently, for $k\ge 1$, $R(K_{9k+7},\Z_3)=9k+10$, resolving a problem of Caro and Mifsud.

math.CO↗

arXiv2026

Robobench: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models as Embodied Brain

Building robots that can perceive, reason, and act in dynamic, unstructured environments remains a central challenge. Recent embodied systems often follow a dual-system paradigm, where System 2 performs high-level reasoning and System 1 handles low-level control. We refer to System 2 as the embodied brain, the cognitive core for decision-making in manipulation. Although evaluating this embodied brain is crucial, existing benchmarks mainly measure execution success or cover only limited aspects of high-level cognition and task realism. We introduce RoboBench, a benchmark for evaluating multimodal large language models (MLLMs) as embodied brains. RoboBench covers five dimensions: Instruction Comprehension, Perception Reasoning, Generalized Planning, Affordance Prediction, and Failure Analysis. It spans 14 capabilities, 25 tasks, and 6,092 QA pairs. To improve realism, it draws from large-scale real robotic data and in-house collection across diverse embodiments, attribute-rich objects, multi-view scenes, and memory-driven navigation. For planning, RoboBench introduces an MLLM-as-world-simulator framework that assesses whether predicted plans can achieve critical object-state changes under physical and visual constraints, enabling more faithful evaluation of long-horizon reasoning than symbolic matching. Experiments on 18 state-of-the-art MLLMs reveal persistent limitations in implicit instruction understanding, spatiotemporal reasoning, cross-scenario planning, fine-grained affordance understanding, and failure diagnosis. We further analyze how embodied cognitive abilities relate to downstream robotic control. RoboBench offers a comprehensive scaffold for quantifying high-level cognition and guiding next-generation MLLMs toward more robust robotic intelligence.

cs.RO↗

arXiv2026

Towards Spatial Trace with Reasoning in Vision-Language Models for Robotics

Spatial tracing, as a fundamental embodied interaction ability for robots, is inherently challenging as it requires multi-step metric-grounded reasoning compounded with complex spatial referring and real-world metric measurement. However, existing methods struggle with this compositional task. To this end, we propose RoboTracer, a 3D-aware VLM that first achieves both 3D spatial referring and measuring via a universal spatial encoder and a regression-supervised decoder to enhance scale awareness during supervised fine-tuning (SFT). Moreover, RoboTracer advances multi-step metric-grounded reasoning via reinforcement fine-tuning (RFT) with metric-sensitive process rewards, supervising key intermediate perceptual cues to accurately generate spatial traces. To support SFT and RFT training, we introduce TraceSpatial, a large-scale dataset of 30M QA pairs, spanning outdoor/indoor/tabletop scenes and supporting complex reasoning processes (up to 9 steps). We further present TraceSpatial-Bench, a challenging benchmark filling the gap to evaluate spatial tracing. Experimental results show that RoboTracer surpasses baselines in spatial understanding, measuring, and referring, with an average success rate of 79.1%, and also achieves SOTA performance on TraceSpatial-Bench by a large margin, exceeding Gemini-2.5-Pro by 36% accuracy. Notably, RoboTracer can be integrated with various control policies to execute long-horizon, dynamic tasks across diverse robots (UR5, G1 humanoid) in cluttered real-world scenes. Please see the project page at https://zhoues.github.io/RoboTracer.

cs.RO↗

arXiv2026

On Zero-sum Ramsey numbers of complete bipartite graphs

For an integer $q\ge 2$ and a graph $F$ satisfying $q\mid e(F)$, the zero-sum Ramsey number $R(F,\mathbb Z_q)$ is the least integer $n$ such that every edge-labeling $w\colon E(K_n)\to \mathbb Z_q$ contains a copy of $F$ whose edge-label sum is zero in $\mathbb Z_q$. Write $K_{s,t}$ for the complete bipartite graph with $s$ vertices on one side and $t$ vertices on the other side. We prove that for every $q\ge2$, there is an explicit threshold $S(q)$ such that $R(K_{s,qk},\mathbb Z_q)=s+qk$ for all $s\ge S(q)$ and all $k\ge1$. We also determine the zero-sum Ramsey number of $K_{s,3k}$ over $\mathbb Z_3$ for all $s\ge2$ and $k\ge1$. We prove that $R(K_{s,3k},\mathbb Z_3)=s+3k$, except when $s=2$ and $k\ge1$, or when $s\in\{3,4,5,7\}$ and $k=1$. In these exceptional cases, $R(K_{s,3k},\mathbb Z_3)=s+3k+1$. In particular, this shows that the threshold $S(q)$ is best possible for $q=3$.

math.CO↗

arXiv2026

New Extremal Ranges and Constructions of the Erdős--Kleitman Problem

For integers $n\ge s\ge2$, let $e(n,s)$ denote the maximum size of a family $\mathcal F\subseteq2^{[n]}$ with no $s$ pairwise disjoint members. The problem of determining $e(n,s)$, now called the Erdős--Kleitman problem, is the non-uniform analogue of the Erdős matching conjecture. We prove that for every fixed $m\ge3$, there exist constants $β_m$ and $δ_m$ such that for sufficiently large $s$, the extremal families for $e(ms+c,s)$ are \[ \mathcal P'(m,s,\ell;L'):=\binom{L'}m\cup\binom{[ms+c]}{\ge m+1} \] for some $L'$ with $\ell=s-c$ and $|L'|=m\ell-1$, when $β_m s^{(m-1)/m}\le c\le δ_m s$. This determines the extremal families in an unknown range when $\ell$ is large, complementing our earlier work on the range when $\ell$ is small. Moreover, for $m=3$, we sharpen this to the asymptotically optimal range. Let \[ t(s)=\frac{17-18s+\sqrt{49-852s+1284s^2}}{20}=0.8916\cdots s+O(1) \] We prove that $\mathcal P'(3,s,\ell;L')$ is the unique extremal family when $t(s)<\ell<s-((4/3)^{1/3}+o(1))s^{2/3}$. Note that the lower bound $t(s)$ of $\ell$ is exact, while the the constant $(4/3)^{1/3}$ in the upper bound of $\ell$ is best possible. Kupavskii and Sokolov introduced four candidate extremal families and conjectured that the value of $e(n,s)$ is the maximum of their sizes. We disprove this conjecture by constructing a new family $\mathcal R(m,s,\ell)$ that is larger than each of their four proposed candidates when $α_{\mathrm R}s^{1/2}\le c\le β_{\mathrm R}s^{(m-1)/m}$ for some constants $α_{\mathrm R}$ and $β_{\mathrm R}$. This also shows that the exponent $(m-1)/m$ in the first result is tight.

math.CO↗

arXiv2026

PIGEON: VLM-Driven Object Navigation via Points of Interest Selection

Object navigation in unseen indoor environments requires agents to perform semantic search under partial observability. Vision-language models (VLMs) provide strong semantic-spatial priors for this task, but how to interface them with robot navigation remains challenging: dense VLM inference is expensive, while abstracting environments into symbolic memories often separates high-level reasoning from the raw visual evidence that supports it. We propose we propose PIGEON (Point of Interest Guided Exploration for Object Navigation), a VLM-driven framework that formulates object navigation as raw-observation-grounded sparse decision problem. PIGEON introduces Points of Interest (PoIs) as sparse visual decision units that couple geometrically executable waypoints with raw egocentric observations. Rather than using VLMs as dense controllers or restricting them to frontier ranking, PIGEON enables VLMs to select among task-critical PoIs, including exploration frontiers, suspected target objects, traversable stairs, and floor-level summaries, while low-level planners execute continuous motion between them. This PoI interface further makes high-level navigation decisions verifiable, allowing us to develop an RLVR pipeline that improves local VLMs without manual Chain-of-Thought annotations. Extensive experiments on Habitat ObjectNav benchmarks show that PIGEON achieves state-of-the-art zero-shot performance, scales consistently with foundation model capacity, and transfers to Active Embodied Question Answering with only prompt modifications. Real-world deployments on physical robots further demonstrate its robustness and efficiency.

cs.RO↗

arXiv2026

Xiaomi Auto World Model: A Joint World Model Integrating Reconstruction and Generation for Autonomous Driving

This report presents a unified technical system addressing the two core capabilities of world models for autonomous driving: world representation and world generation. For world representation, we propose WorldRec, a feed-forward reconstruction architecture driven by sparse scene queries. WorldRec initializes structured queries in 3D space, leveraging them to aggregate cross-view, cross-temporal features, thereby naturally enforcing spatial consistency across frames and yielding compact yet high-fidelity 3D Gaussian scene representations. For world generation, we propose WorldGen, a two-stage training framework of bidirectional pretraining followed by causal fine-tuning through three progressive stages (Teacher Forcing, ODE distillation, and DMD), enabling high-quality online causal video generation in as few as 4 denoising steps. Building on both modules, we further introduce the JWM, which deeply integrates WorldRec and WorldGen to achieve synergistic gains in generation stability, cross-frame consistency, and visual fidelity, providing a solid foundation for closed-loop simulation, data synthesis, and end-to-end training in autonomous driving.

cs.CV↗

arXiv2026

On zero-sum Ramsey numbers of cycles and wheels

For an integer $q\ge 2$ and a graph $F$ with $q\mid e(F)$, let $R(F,\Z_q)$ be the least integer $n$ such that every edge-labeling $w\colon E(K_n)\to \Z_q$ contains a copy of $F$ whose edge-label sum is zero in $\Z_q$. Write $C_{qk}$ for the cycle on $qk$ vertices. We prove that $R(C_{qk},\Z_q)\le \max\{R(C_{2q},\Z_q),qk+q-1\}$ via an insertion argument rooted in the classic Erdős-Ginzburg-Ziv theorem. Combined with Pikhurko's result, we obtain $R(C_{qk},\Z_q)\le \max\{35q^2,qk+q-1\}$ for every $q\ge 3$. We also show that $R(C_{qk},\Z_q)\ge qk+q-1$ for odd $q\ge 3$. Hence, for every fixed odd $q\ge 3$ and every $k\ge 35q$, we obtain the exact value $R(C_{qk},\Z_q)=qk+q-1$. For even $q\ge 4$, the same method gives $qk+\frac q2-1\le R(C_{qk},\Z_q)\le \max\{35q^2,qk+q-1\}$, leaving an additive gap of order $q/2$ when $k$ is large. Moreover, for the case $q=3$, we prove that $R(C_{3k}, \mathbb{Z}_3) = 3k + 2$ for all $k \ge 2$. Extending our techniques beyond cycles, we also resolve the zero-sum Ramsey number for wheel graphs $W_m = C_m + K_1$, proving that $R(W_{3k}, \mathbb{Z}_3) = 3k + 1$ for all $k \ge 2$.

math.CO↗

arXiv2026

A solution to Frankl and Kupavskii's conjecture concerning Erdős-Kleitman matching problem

For integers $n\ge s\ge2$, let $e(n,s)$ be the maximum size of a family $\mathcal F\subseteq2^{[n]}$ with no $s$ pairwise disjoint members. The study of determining $e(n,s)$ is closely related to its uniform counterpart, the well-known Erdős matching conjecture. Frankl and Kupavskii conjectured an exact formula for $e((m+1)s-\ell,s)$ when $1\le \ell\le \lceil s/2\rceil$. We prove that for every fixed $m\ge3$ and sufficiently large $s$, the extremal families for $e((m+1)s-\ell,s)$ are $P(m,s,\ell;L)\coloneqq\{A\subseteq [n]\colon |A|+|A\cap L|\ge m+1\}$ for some $L$ with $|L|=\ell-1$ when $1\le \ell\le (\frac{m+1}{2m+1}-o(1))s$. In particular, this confirms the Frankl--Kupavskii conjecture for every fixed $m\ge3$ and all sufficiently large $s$. For $m=3$, we determine the whole range of $\ell$ for which $P(3,s,\ell;L)$ is extremal, generalizing a theorem of Kupavskii and Sokolov.

math.CO↗

arXiv2026

PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations

High-fidelity reconstruction of driving scenes is crucial for autonomous driving. While recent feedforward 3D Gaussian Splatting (3DGS) methods enable fast reconstruction, their per-pixel Gaussian prediction paradigm often suffers from multi-view inconsistency and layering artifacts. Moreover, existing methods often model dynamic instances via dense flow prediction, which lacks explicit cross-view correspondence and instance-level consistency. In this paper, we propose PointForward, a feedforward driving reconstruction framework through point-aligned representations. Unlike pixel-aligned methods, we initialize sparse 3D queries in world space and aggregate multi-view image information via spatial-temporal fusion onto these queries, enforcing explicit cross-view consistency in a single feedforward pass. To handle scene dynamics, we introduce scene graphs that explicitly organize moving instances during reconstruction. By leveraging 3D bounding boxes, our method enables instance-level motion propagation and temporally consistent dynamic representations. Extensive experiments demonstrate that PointForward achieves state-of-the-art performance on large-scale driving benchmarks. The code will be available upon the publication of the paper.

cs.CV↗

arXiv2026

Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models

Vision-Language-Action (VLA) models benefit from chain-of-thought (CoT) reasoning, but existing approaches incur high inference overhead and rely on discrete reasoning representations that mismatch continuous perception and control. We propose Latent Reasoning VLA (LaRA-VLA), a unified VLA framework that internalizes multi-modal CoT reasoning into continuous latent representations for embodied action. LaRA-VLA performs unified reasoning and prediction in latent space, eliminating explicit CoT generation at inference time and enabling efficient, action-oriented control. To realize latent embodied reasoning, we introduce a curriculum-based training paradigm that progressively transitions from explicit textual and visual CoT supervision to latent reasoning, and finally adapts latent reasoning dynamics to condition action generation. We construct two structured CoT datasets and evaluate LaRA-VLA on both simulation benchmarks and long-horizon real-robot manipulation tasks. Experimental results show that LaRA-VLA consistently outperforms state-of-the-art VLA methods while reducing inference latency by up to 90\% compared to explicit CoT-based approaches, demonstrating latent reasoning as an effective and efficient paradigm for real-time embodied control. Project Page: https://loveju1y.github.io/Latent-Reasoning-VLA/

cs.RO↗

arXiv2026

Anti-Ramsey numbers for cancellative configurations in p-graphs

We study edge-colorings of the complete $p$-graph on $n$ vertices that contain no three edges $A,B,C$ of distinct colors such that the symmetric difference of $A$ and $B$ is contained in $C$. For $p\ge3$ and $n\ge p+1$, we show that every such coloring contains at most $1+\floor{n/p}$ colors and characterize the extremal colorings, generalizing a theorem of Erdős, Simonovits and Sós. %\cite{erdos1975}. When $p=3$, the condition $A\triangle B\subseteq C$ implies $|A\triangle B|=2$, and the three edges necessarily form a copy of $F_4\coloneqq\{abc,abd,bcd\}$ or $F_5\coloneqq\{abc,abd,cde\}$. For $n\ge5$, we show that every rainbow $F_5$-free edge-coloring is rainbow cancellative. For rainbow $F_4$-free colorings, we construct colorings with $m(n)+1$ colors for all $n\ge4$, where $m(n)$ is the size of a maximum partial Steiner triple system of order $n$ and satisfies $m(n)=n^2/6+O(n)$, improving the linear lower bound by Budden and Stiles. %\cite{budden}. Moreover, for $n=2^s-1$, we obtain $\ar(n,F_4)\ge m(n)+n^2/42+o(n^2)=4n^2/21+o(n^2)$ via a construction based on independent sets in the Grassmann graph. We also prove that $\ar(n,F_4)\le (5n^2-8n)/21$ for $n\ge4$, improving the quadratic coefficient in the upper bound of Budden and Stiles from $1/4$ to $5/21$.

math.CO↗

arXiv2026

RoboCOIN: An Open-Sourced Bimanual Robotic Data Collection for Integrated Manipulation

Despite the critical role of bimanual manipulation in endowing robots with human-like dexterity, large-scale and diverse datasets remain scarce due to the significant hardware heterogeneity across bimanual robotic platforms. To bridge this gap, we introduce RoboCOIN, a large-scale multi-embodiment bimanual manipulation dataset comprising over 180,000 demonstrations collected from 15 distinct robotic platforms. Spanning 16 diverse environments-including residential, commercial, and industrial settings-the dataset features 421 bimanual tasks systematically categorized by 39 bimanual collaboration actions and 432 objects. A key innovation of our work is the hierarchical capability pyramid, which provides granular annotations ranging from trajectory-level concepts to segment-level subtasks and frame-level kinematics. Furthermore, we present CoRobot, an efficient data processing pipeline powered by the Robot Trajectory Markup Language (RTML), designed to facilitate quality assessment, automated annotation, and unified multi-embodiment and data management. Extensive experiments demonstrate the effectiveness of RoboCOIN in enhancing the performance of various bimanual manipulation models across a wide spectrum of robotic embodiments. The entire dataset and codebase are fully open-sourced, providing a valuable resource for advancing research in bimanual and multi-embodiment manipulation.

cs.RO↗

arXiv2026

The Turán number of the Cartesian product of a star and an edge

Let $C_k$ denote the cycle of length $k$, $S_t$ be a star with $t$ edges. And let $B_t$ be the graph consisting of $t$ copies of $C_4$ sharing one fixed edge. Equivalently, $B_t=K_2 \mathbin{\square} S_t$, which is the Cartesian product of a star with $t$ edges and an edge. Recently, Gao, Janzer, Liu and Xu [\textit{Israel J. Math. 269(2025)}] proved that the Turán number of $K_2\mathbin{\square} C_{2l}$ is $Θ(n^{\frac{3}{2}})$ for every $l\ge 4$. In this paper, we obtain upper and lower estimates for the Turán number of $B_t$ in both the general and bipartite settings for every $t\geq 2$. For the lower bound, we use random construction based on the extremal structure of $C_4$. These results imply that $\frac{1}{2\sqrt{2}}\leq \lim_{t\to \infty} \frac{\mathrm{ex}(n,B_t)}{\sqrt{t}}\leq \frac{1}{2}$, and $\frac{1}{4}\leq \lim_{t\to \infty} \frac{\mathrm{ex}_{bip}(n,B_t)}{\sqrt{t}}\leq \frac{1}{2\sqrt{2}}.$ In the case of $B_2$, we obtain sharper estimates. We show that the Turán number of $B_2$ is approximately between $(0.518+o(1))n^{\frac{3}{2}}$ and $(0.603+o(1))n^{\frac{3}{2}}$. And in the bipartite setting, it is approximately between $(0.385+o(1))n^{\frac{3}{2}}$ and $(0.468+o(1))n^{\frac{3}{2}}$. Moreover, in the bipartite setting, we give a more general result, which shows that for every tree $T$ with $t$ edges, the bipartite Turán number of $K_2\mathbin{\square}T$ is at most $\frac{\sqrt{t}}{2\sqrt{2}}(1+o(1))n^{\frac{3}{2}}$.

math.CO↗

arXiv2026

UniDA3D: A Unified Domain-Adaptive Framework for Multi-View 3D Object Detection

Camera-only 3D object detection is critical for autonomous driving, offering a cost-effective alternative to LiDAR based methods. In particular, multi-view 3D object detection has emerged as a promising direction due to its balanced trade-off between performance and cost. However, existing methods often suffer significant performance degradation under complex environmental conditions such as nighttime, fog, and rain, primarily due to their reliance on training data collected mostly in ideal conditions. To address this challenge, we propose UniDA3D, a unified domain-adaptive multi-view 3D object detector designed for robust perception under diverse adverse conditions. UniDA3D formulates nighttime, rainy, and foggy scenes as a unified multi target domain adaptation problem and leverages a novel query guided domain discrepancy mitigation (QDDM) module to align object features between source and target domains at both batch and global levels via query-centric adversarial and contrastive learning. Furthermore, we introduce a domain-adaptive teacher student training pipeline with an exponential-moving-average teacher and dynamically updated high-quality pseudo labels to enhance consistency learning and suppress background noise in unlabeled target domains. In contrast to prior approaches that require separate training for each condition, UniDA3D performs a single unified training process across multiple domains, enabling robust all-weather 3D perception. On a synthesized multi-view 3D benchmark constructed by generating nighttime, rainy, and foggy counterparts from nuScenes (nuScenes-Night, nuScenes-Rain, and nuScenes-Haze), UniDA3D consistently outperforms state of-the-art camera-only multi-view 3D detectors under extreme conditions, achieving substantial gains in mAP and NDS while maintaining real-time inference efficiency.

cs.CV↗

arXiv2026

PRM-as-a-Judge: A Dense Evaluation Paradigm for Fine-Grained Robotic Auditing

Current robotic evaluation is still largely dominated by binary success rates, which collapse rich execution processes into a single outcome and obscure critical qualities such as progress, efficiency, and stability. To address this limitation, we propose PRM-as-a-Judge, a dense evaluation paradigm that leverages Process Reward Models (PRMs) to audit policy execution directly from trajectory videos by estimating task progress from observation sequences. Central to this paradigm is the OPD (Outcome-Process-Diagnosis) metric system, which explicitly formalizes execution quality via a task-aligned progress potential. We characterize dense robotic evaluation through two axiomatic properties: macro-consistency, which requires additive and path-consistent aggregation, and micro-resolution, which requires sensitivity to fine-grained physical evolution. Under this formulation, potential-based PRM judges provide a natural instantiation of dense evaluation, with macro-consistency following directly from the induced scalar potential. We empirically validate the micro-resolution property using RoboPulse, a diagnostic benchmark specifically designed for probing micro-scale progress discrimination, where several trajectory-trained PRM judges outperform discriminative similarity-based methods and general-purpose foundation-model judges. Finally, leveraging PRM-as-a-Judge and the OPD metric system, we conduct a structured audit of mainstream policy paradigms across long-horizon tasks, revealing behavioral signatures and failure modes that are invisible to outcome-only metrics.

cs.RO↗

arXiv2026

LassoFlexNet: Flexible Neural Architecture for Tabular Data

Despite their dominance in vision and language, deep neural networks often underperform relative to tree-based models on tabular data. To bridge this gap, we incorporate five key inductive biases into deep learning: robustness to irrelevant features, axis alignment, localized irregularities, feature heterogeneity, and training stability. We propose \emph{LassoFlexNet}, an architecture that evaluates the linear and nonlinear marginal contribution of each input via Per-Feature Embeddings, and sparsely selects relevant variables using a Tied Group Lasso mechanism. Because these components introduce optimization challenges that destabilize standard proximal methods, we develop a \emph{Sequential Hierarchical Proximal Adaptive Gradient optimizer with exponential moving averages (EMA)} to ensure stable convergence. Across $52$ datasets from three benchmarks, LassoFlexNet matches or outperforms leading tree-based models, achieving up to a $10$\% relative gain, while maintaining Lasso-like interpretability. We substantiate these empirical results with ablation studies and theoretical proofs confirming the architecture's enhanced expressivity and structural breaking of undesired rotational invariance.

stat.ML↗