Bindu Reddy

Standard Basis Dimensions Encode Semantic Features Without Training

Transformer hidden states' standard basis dimensions directly encode semantic content through sign patterns (binary features) and magnitudes (confidence), eliminating the need for learned feature bases or training. Sign patterns alone achieve 72-93% top-5 accuracy and enable unsupervised discovery of 1500+ semantic features across model families via single-token caching, establishing that mechanistic interpretability requires no optimization beyond a single forward pass per vocabulary token.

Galaxy Inclination Determines Gas Flow Direction at Cosmic Noon

Using deep JWST/NIRSpec spectroscopy of 688 star-forming galaxies at z~3, this study reveals that galactic gas flows are fundamentally oriented perpendicular (outflows) or parallel (inflows) to disk geometry, consistent with theoretical predictions. Outflow velocity shows no correlation with galaxy properties like mass or SFR, but galaxies with detected outflows skew toward higher stellar mass and star formation rates, indicating selection effects. Critically, galaxies' inclination angle predicts gas flow signatures: face-on systems show stronger absorption and higher outflow velocities, while inclined disks preferentially exhibit inflow signatures—a pattern consistent with low-redshift observations and simulations.

Protoplume-to-Megaplume Transition Governs Linear Heat Transfer in Ultimate Rayleigh-Darcy Convection

DNS simulations of Rayleigh-Darcy convection reveal a distinct transition to an ultimate regime at Ra ≈ 4×10⁵ where Nusselt number maintains approximately linear scaling with Ra. This scaling persistence stems from progressive refinement of near-wall protoplumes that merge into megaplumes, with finer structures emerging at higher Ra despite shifting thermal dissipation from boundary layers to the bulk. The thermal boundary-layer thickness scaling (Ra⁻¹) combined with increased protoplume density generates efficient heat transport across the entire domain.

ShallowBench: A New Benchmark Revealing Generative Drug Design Weaknesses on Shallow Protein Pockets

Generative AI models for drug design excel with deep pockets but falter on shallow ones, exemplified by targets like KRAS and MYC. The newly introduced ShallowBench, containing 5,780 shallow-pocket targets from CrossDocked2020, evaluates these models, revealing their weaker performance on such targets. This highlights the need for novel architectures or loss functions to address these challenging binding sites.

Lighting-Aware Representation Learning for Improved Visual Understanding

This research introduces a lighting-aware representation learning framework that treats illumination variation as an explicit training signal, diverging from conventional invariance-focused data augmentation. By extending contrastive learning with an auxiliary objective for illumination-dependent variations, the model learns robust representations that maintain semantic consistency while adapting to lighting-induced structural changes. Evaluations on benchmark datasets show consistent performance improvements in downstream tasks like image classification and object detection compared to standard contrastive learning.

HLS-based FPGA Hardblock Programming via RTL Blackbox Abstraction

This work introduces a compiler-agnostic methodology enabling High-Level Synthesis (HLS) tools to directly target domain-specific FPGA hardblocks. By abstracting hardblocks as C-level operators with explicit latency and initiation-interval contracts, the approach allows HLS schedulers to optimize specialized hardware without manual RTL integration. Evaluations on a Tensor Slice-based architecture demonstrate that this method yields lower area-delay products and improved productivity-adjusted efficiency compared to baseline HLS and handwritten RTL, respectively, while enhancing accessibility of specialized FPGA architectures.

3d Summation-by-Parts scheme for Linear Wave Equations on Hyperboloidal Slices

Learning Long Range Spatio-Temporal Representations over Continuous Time Dynamic Graphs with State Space Models

Would you still call this Dax? Novel Visual References in VLMs and Humans

HoT-SSM:Higher-order Temporal Knowledge Graph Reasoning with State Space Models for Health Care

SEED: Semi-supervised Continual MalwarE Detection for Tackling ConcEpt Drift on a BuDget

AgentSociety: Incentivizing Agentic Social Intelligence

A Vertical Look at UAV Connectivity in the Wild: Cellular vs. Starlink, 3D Characterization, and Performance Prediction

A Preliminary Assessment of Midhaul Links at 140 GHz using Ray-Tracing

Supervised Latent Restructuring for Small-Data Quantum Learning in Plant Phenomics

Articulate but Wrong: Self-Review Failures in LLM-Based Code Modernization

AIGaitor: Privacy-preserving and cloud-free motion analysis for everyone, using edge computing

RISED: A Pre-Deployment Safety Evaluation Framework for Clinical AI Decision-Support Systems

No Stream Left Unscathed: The imprint of a host galaxy

Extensions of the Furstenberg-Sárközy theorem via the arithmetic level-$d$ inequality

SABER: A Scalable Action-Based Embodied Dataset for Real-World VLA Adaptation

Trajectory Supervision for Continual Tool-Use Learning in LLMs

Quantifying Rodda and Graham Gait Classification from 3D Makerless Kinematics derived from a Single-view Video in a Heterogeneous Pediatric Clinical Cohort

Rhamba: Region-Aware Hybrid Attention-Mamba Framework for Self-Supervised Learning in Resting-State fMRI

FinRAG-12B: A Production-Validated Recipe for Grounded Question Answering in Banking

LEAP: Layer-wise Exit-Aware Pretraining for Efficient Transformer Inference

SURGE: SuperBatch Unified Resource-efficient GPU Encoding for Heterogeneous Partitioned Data

Rhamba: Region-Aware Hybrid Attention-Mamba Framework for Self-Supervised Learning in Resting-State fMRI