arXiv subjects

Leo Liu

Publications and source records attributed to Leo Liu.

13 recordsLinked to original sources

arXiv2026

SportD: How do VLMs physically strategize?

Vision-language models (VLMs) can describe a scene, but can they act well within one? We study whether VLMs can make sound strategic decisions, using soccer as an objective testbed with quantifiably-valued actions. We introduce SportD, a dataset and evaluation consisting of 1415 decision scenarios across professional men's and women's soccer games, where a VLM observes the seconds before a decision and chooses the next action. Models only select the optimal action around 30% of the time, even less frequently than humans do. Furthermore, they exhibit a clear preference for safer actions, favoring lower-variance, lower-value choices that also make less physical progress toward goal. Frontier VLMs are better at estimating whether an action will succeed, placing the highest-success-probability action among their top choices in 83-92% of cases. Yet VLMs systematically conflate likelihood with value, assigning higher value to actions that are more likely to succeed ($\rho$=+0.30 to +0.52), despite no such relationship in the ground truth ($\rho$=-0.08). The conservatism therefore reflects a mis-calibration of value. Replacing a single deliberation sentence with one that steers toward risk lifts the frontier models towards the real players' skills. SportD opens a new direction for rigorously evaluating physical strategic decision-making in VLMs, showing that careful decomposition of their choices can reveal the mechanisms underlying systematic biases such as risk aversion.

cs.AI↗

arXiv2026

Parthenon Law: A Self-Evolving Legal-Agent Framework

As agents grow more capable, legal-domain LLM agents promise to turn document-heavy matters into reviewable work products -- yet reliable deployment faces three obstacles: no large-scale evidence on how today's strongest model-and-harness combinations behave on end-to-end legal matters; no agent architecture adapted to the legal vertical, only general-purpose harnesses; and, in a setting that keeps shifting with new facts, authorities, and deadlines, no mechanism for systems to learn from their own outcomes. We address each. A large-scale empirical study on Harvey LAB -- $12{,}510$ agent trajectories -- shows that even frontier agents remain far from completing matters in a single pass: per-criterion accuracy climbs with stronger models while strict matter completion stalls. We then introduce \textsc{Parthenon}, a self-evolving legal-agent framework that factors Model, Harness, Agent roles, legal Knowledge, deterministic Tools, and procedural Skills into auditable surfaces for source traceability, date and number grounding, deliverable compliance, and issue closure. Finally, an anti-leakage learning loop converts scored failures into task-agnostic edits to skills, tools, and knowledge, letting the system improve with experience -- as a firm refines its checklists and playbooks after each matter -- without touching model weights. Across our large-scale empirical analysis, \textsc{Parthenon} substantially improves the performance of state-of-the-art models and harnesses on legal-matter tasks.

cs.AI↗

arXiv2025

FSL-HDnn: A 40 nm Few-shot On-Device Learning Accelerator with Integrated Feature Extraction and Hyperdimensional Computing

This paper introduces FSL-HDnn, an energy-efficient accelerator that implements the end-to-end pipeline of feature extraction and on-device few-shot learning (FSL). The accelerator addresses fundamental challenges of on-device learning (ODL) for resource-constrained edge applications through two synergistic modules: a parameter-efficient feature extractor employing weight clustering and an FSL classifier based on hyperdimensional computing (HDC). The feature extractor exploits the weight clustering mechanism to reduce computational complexity, while the HDC-based FSL classifier eliminates gradient-based back propagation operations, enabling single-pass training with substantially reduced latency. Additionally, FSL-HDnn enables low-latency ODL and inference via two proposed optimization strategies, including an early-exit mechanism with branch feature extraction and batched single-pass training that improves hardware utilization. Measurement results demonstrate that our chip fabricated in a 40 nm CMOS process delivers superior training energy efficiency of 6 mJ/image and end-to-end training throughput of 28 images/s on a 10-way 5-shot FSL task. The end-to-end training latency is also reduced by 2x to 20.9x compared to state-of-the-art ODL chips.

cs.AR↗

arXiv2025

Clo-HDnn: A 4.66 TFLOPS/W and 3.78 TOPS/W Continual On-Device Learning Accelerator with Energy-efficient Hyperdimensional Computing via Progressive Search

Clo-HDnn is an on-device learning (ODL) accelerator designed for emerging continual learning (CL) tasks. Clo-HDnn integrates hyperdimensional computing (HDC) along with low-cost Kronecker HD Encoder and weight clustering feature extraction (WCFE) to optimize accuracy and efficiency. Clo-HDnn adopts gradient-free CL to efficiently update and store the learned knowledge in the form of class hypervectors. Its dual-mode operation enables bypassing costly feature extraction for simpler datasets, while progressive search reduces complexity by up to 61% by encoding and comparing only partial query hypervectors. Achieving 4.66 TFLOPS/W (FE) and 3.78 TOPS/W (classifier), Clo-HDnn delivers 7.77x and 4.85x higher energy efficiency compared to SOTA ODL accelerators.

cs.AR↗

arXiv2025

Wavenumber affects the lift of ray-inspired fins near a substrate

Rays and skates tend to have different fin kinematics depending on their proximity to a ground plane such as the seafloor. Near the ground, rays tend to be more undulatory (high wavenumber), while far from the ground, rays tend to be more oscillatory (low wavenumber). It is unknown whether these differences are driven by hydrodynamics or other biological pressures. Here we show that near the ground, the time-averaged lift on a ray-like fin is highly dependent on wavenumber. We support our claims using a ray-inspired robotic rig that can produce oscillatory and undulatory motions on the same fin. Potential flow simulations reveal that lift is always negative because quasisteady forces overcome wake-induced forces. Three-dimensional flow measurements demonstrate that oscillatory wakes are more disrupted by the ground than undulatory wakes. All these effects lead to a suction force toward the ground that is stronger and more destabilizing for oscillatory fins than undulatory fins. Our results suggest that wavenumber plays a role in the near-ground dynamics of ray-like fins, particularly in terms of dorsoventral accelerations. The fact that lower wavenumber is linked with stronger suction forces offers a new way to interpret the depth-dependent kinematics of rays and ray-inspired robots.

physics.flu-dyn↗

arXiv2024

Multimodal Alignment of Histopathological Images Using Cell Segmentation and Point Set Matching for Integrative Cancer Analysis

Histopathological imaging is vital for cancer research and clinical practice, with multiplexed Immunofluorescence (MxIF) and Hematoxylin and Eosin (H&E) providing complementary insights. However, aligning different stains at the cell level remains a challenge due to modality differences. In this paper, we present a novel framework for multimodal image alignment using cell segmentation outcomes. By treating cells as point sets, we apply Coherent Point Drift (CPD) for initial alignment and refine it with Graph Matching (GM). Evaluated on ovarian cancer tissue microarrays (TMAs), our method achieves high alignment accuracy, enabling integration of cell-level features across modalities and generating virtual H&E images from MxIF data for enhanced clinical interpretation.

eess.IV↗

arXiv2023

Personalization of CTC-based End-to-End Speech Recognition Using Pronunciation-Driven Subword Tokenization

Recent advances in deep learning and automatic speech recognition have improved the accuracy of end-to-end speech recognition systems, but recognition of personal content such as contact names remains a challenge. In this work, we describe our personalization solution for an end-to-end speech recognition system based on connectionist temporal classification. Building on previous work, we present a novel method for generating additional subword tokenizations for personal entities from their pronunciations. We show that using this technique in combination with two established techniques, contextual biasing and wordpiece prior normalization, we are able to achieve personal named entity accuracy on par with a competitive hybrid system.

cs.LG↗

arXiv2023

Acoustic Model Fusion for End-to-end Speech Recognition

Recent advances in deep learning and automatic speech recognition (ASR) have enabled the end-to-end (E2E) ASR system and boosted the accuracy to a new level. The E2E systems implicitly model all conventional ASR components, such as the acoustic model (AM) and the language model (LM), in a single network trained on audio-text pairs. Despite this simpler system architecture, fusing a separate LM, trained exclusively on text corpora, into the E2E system has proven to be beneficial. However, the application of LM fusion presents certain drawbacks, such as its inability to address the domain mismatch issue inherent to the internal AM. Drawing inspiration from the concept of LM fusion, we propose the integration of an external AM into the E2E system to better address the domain mismatch. By implementing this novel approach, we have achieved a significant reduction in the word error rate, with an impressive drop of up to 14.3% across varied test sets. We also discovered that this AM fusion approach is particularly beneficial in enhancing named entity recognition.

cs.SD↗

arXiv2023

AlerTiger: Deep Learning for AI Model Health Monitoring at LinkedIn

Data-driven companies use AI models extensively to develop products and intelligent business solutions, making the health of these models crucial for business success. Model monitoring and alerting in industries pose unique challenges, including a lack of clear model health metrics definition, label sparsity, and fast model iterations that result in short-lived models and features. As a product, there are also requirements for scalability, generalizability, and explainability. To tackle these challenges, we propose AlerTiger, a deep-learning-based MLOps model monitoring system that helps AI teams across the company monitor their AI models' health by detecting anomalies in models' input features and output score over time. The system consists of four major steps: model statistics generation, deep-learning-based anomaly detection, anomaly post-processing, and user alerting. Our solution generates three categories of statistics to indicate AI model health, offers a two-stage deep anomaly detection solution to address label sparsity and attain the generalizability of monitoring new models, and provides holistic reports for actionable alerts. This approach has been deployed to most of LinkedIn's production AI models for over a year and has identified several model issues that later led to significant business metric gains after fixing.

cs.LG↗

arXiv2022

Training Large-Vocabulary Neural Language Models by Private Federated Learning for Resource-Constrained Devices

Federated Learning (FL) is a technique to train models using data distributed across devices. Differential Privacy (DP) provides a formal privacy guarantee for sensitive data. Our goal is to train a large neural network language model (NNLM) on compute-constrained devices while preserving privacy using FL and DP. However, the DP-noise introduced to the model increases as the model size grows, which often prevents convergence. We propose Partial Embedding Updates (PEU), a novel technique to decrease noise by decreasing payload size. Furthermore, we adopt Low Rank Adaptation (LoRA) and Noise Contrastive Estimation (NCE) to reduce the memory demands of large models on compute-constrained devices. This combination of techniques makes it possible to train large-vocabulary language models while preserving accuracy and privacy.

cs.LG↗

arXiv2020

SNDCNN: Self-normalizing deep CNNs with scaled exponential linear units for speech recognition

Very deep CNNs achieve state-of-the-art results in both computer vision and speech recognition, but are difficult to train. The most popular way to train very deep CNNs is to use shortcut connections (SC) together with batch normalization (BN). Inspired by Self- Normalizing Neural Networks, we propose the self-normalizing deep CNN (SNDCNN) based acoustic model topology, by removing the SC/BN and replacing the typical RELU activations with scaled exponential linear unit (SELU) in ResNet-50. SELU activations make the network self-normalizing and remove the need for both shortcut connections and batch normalization. Compared to ResNet- 50, we can achieve the same or lower (up to 4.5% relative) word error rate (WER) while boosting both training and inference speed by 60%-80%. We also explore other model inference optimization schemes to further reduce latency for production use.

cs.LG↗

arXiv2020

Voice trigger detection from LVCSR hypothesis lattices using bidirectional lattice recurrent neural networks

We propose a method to reduce false voice triggers of a speech-enabled personal assistant by post-processing the hypothesis lattice of a server-side large-vocabulary continuous speech recognizer (LVCSR) via a neural network. We first discuss how an estimate of the posterior probability of the trigger phrase can be obtained from the hypothesis lattice using known techniques to perform detection, then investigate a statistical model that processes the lattice in a more explicitly data-driven, discriminative manner. We propose using a Bidirectional Lattice Recurrent Neural Network (LatticeRNN) for the task, and show that it can significantly improve detection accuracy over using the 1-best result or the posterior.

cs.CL↗

arXiv2020

Diffuser-Assisted Infrared Transit Photometry for Four Dynamically Interacting \textit{Kepler} Systems

We present ground-based infrared transit observations for four dynamically interacting \textit{Kepler} planets, including Kepler-29b, Kepler-36c, KOI-1783.01, and Kepler-177c, obtained using the Wide-field Infrared Camera on the Hale 200" telescope at Palomar Observatory. By utilizing an engineered diffuser and custom guiding software, we mitigate time-correlated telluric and instrumental noise sources in these observations. We achieve an infrared photometric precision comparable to or better than that of space-based observatories such as the \textit{Spitzer Space Telescope}, and detect transits with greater than 3$σ$ significance for all planets. For Kepler-177c ($J=13.9$) our measurement uncertainties are only $1.2\times$ the photon noise limit and 1.9 times better than the predicted photometric precision for \textit{Spitzer} IRAC photometry of this same target. We find that a single transit observation obtained $4-5$ years after the end of the original \textit{Kepler} mission can reduce dynamical mass uncertainties by as much as a factor of three for these systems. Additionally, we combine our new observations of KOI-1783.01 with information from the literature to confirm the planetary nature of this system. We discuss the implications of our new mass and radius constraints in the context of known exoplanets with low incident fluxes, and we note that Kepler-177c may be a more massive analog to the currently known super-puffs given its core mass (3.8$\pm0.9M_\Earth$) and large gas-to-core ratio (2.8$\pm0.7$). Our demonstrated infrared photometric performance opens up new avenues for ground-based observations of transiting exoplanets previously thought to be restricted to space-based investigation.

astro-ph.EP↗