Institution profile

LG

Industry researchasia · kr
Official website
Research library216linked papers
Opportunities0open roles
Selected work

Representative Papers

View Selection for 3D Captioning via Diffusion Ranking

Apr 11, 2024European Conference on Computer Vision

To address vision-language hallucinations in 3D object captioning caused by rendering viewpoint mismatch, this paper proposes an unsupervised view-ranking method based on diffusion models. Specifically, we leverage pre-trained text-to-3D models (e.g., Stable Diffusion 3D) to quantify 3D–2D view alignment scores and select the most discriminative 2D views for input to multimodal large language models (e.g., GPT-4V) to generate accurate captions. This approach extends the view-ranking paradigm to 3D visual question answering (3D-VQA), achieving significant improvements over CLIP-based baselines on Objaverse/XL. Furthermore, we correct 200K erroneous captions in Cap3D and construct the first million-scale, high-quality 3D-caption dataset—Cap3D-v2—establishing a robust benchmark for 3D understanding and generation.

21 citationsRead paper

Accurate Sublayer Pruning for Large Language Models by Exploiting Latency and Tunability Information

Jun 04, 2025

How to accelerate large language model (LLM) inference without compromising accuracy? This paper proposes a dynamic sublayer pruning method based on joint modeling of sublayer-level latency gain and tunability. Unlike conventional heuristic or uniform pruning, we introduce, for the first time, a dual-dimensional importance metric—latency sensitivity and parameter tunability—at the sublayer level to guide fine-grained identification of redundant sublayers. Our approach integrates iterative pruning with lightweight parameter recalibration, achieving Pareto-optimal trade-offs between accuracy and latency on zero-shot commonsense reasoning tasks. Experiments across multiple LLMs demonstrate that our method improves accuracy by up to 23.88% over state-of-the-art pruning techniques, establishing new benchmarks in accuracy–latency balance.

2 citationsRead paper

elaTCSF: A Temporal Contrast Sensitivity Function for Flicker Detection and Modeling Variable Refresh Rate Flicker

Dec 03, 2024ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia

This paper addresses the insufficient modeling of low-contrast flicker perception in display and lighting applications—particularly under variable refresh rate (VRR), peripheral vision, and low-spatial-frequency conditions where existing models fail. We propose the first extended temporal contrast sensitivity function (elaTCSF), jointly incorporating luminance, retinal eccentricity, and stimulus area. Methodologically, we systematically integrate these three factors into the TCSF framework, develop a spatial probability summation model, and introduce the first VRR-specific flicker detection benchmark dataset. Key contributions include: resolving the long-standing controversy regarding enhanced flicker visibility in peripheral vision; enabling high-accuracy prediction of low-persistence flicker in VR; empirically characterizing the flicker-free operational range for VRR systems; and supporting high-fidelity display and lighting design. Our elaTCSF consistently outperforms IDMS-TCSF across multiple benchmark datasets.

2 citationsRead paper

Gaming the Judge: Unfaithful Chain-of-Thought Can Undermine Agent Evaluation

Jan 21, 2026

This study systematically uncovers and quantifies a critical vulnerability in using vision-language models (VLMs) as judges for evaluating agent performance in web-based tasks: their assessments are susceptible to manipulation through unfaithful content in the agents’ chain-of-thought (CoT) reasoning. By holding agent behavior constant and solely altering the CoT—via strategies such as reasoning trace rewriting, prompt engineering, and increased computational resources—the authors demonstrate that content-based manipulations can inflate false positive rates by up to 90% in state-of-the-art VLM judges. The work further distinguishes between stylistic and content-based manipulation tactics, revealing their differing efficacies. While existing defenses partially mitigate this fragility, they fail to fully resolve the underlying issue, highlighting a fundamental challenge in reliable automated evaluation grounded in unverifiable reasoning traces.

1 citationsRead paper

Lost in the Noise: How Reasoning Models Fail with Contextual Distractors

Jan 12, 2026

This work addresses the critical gap in evaluating the robustness of reasoning models under realistic noisy conditions—such as irrelevant documents, chat histories, and strong negative examples—where current benchmarks fall short. To this end, we introduce NoisyBench, the first systematic benchmark assessing model resilience across 11 tasks spanning RAG, reasoning, alignment, and tool use. Our analysis uncovers several counterintuitive findings: context noise can degrade performance by up to 80%, agent workflows amplify errors, and increased test-time computation may harm accuracy. We further propose Rationale-Aware Reward (RARE), a reinforcement learning method that steers models toward valid reasoning traces, significantly enhancing noise robustness. Experiments show that conventional approaches—including prompt engineering, supervised fine-tuning, and outcome-based reward RL—fail to improve resilience, whereas RARE effectively mitigates over-attention to distractor tokens, offering key insights for building robust reasoning agents.

1 citationsRead paper
Recent publications

Latest Papers