arXiv subjects

Xinlong Li

Publications and source records attributed to Xinlong Li.

4 recordsLinked to original sources

arXiv2026

Let RGB Be the Language of Vision

This work introduces a unified formulation for vision models, where diverse forms of visual information beyond natural images, such as masks, depth maps, and other structured visual signals, are all represented as RGB images, while general visual tasks can be converted into a common RGB-to-RGB image editing problem. In this paradigm, different types of visual information internally share the same encoding and decoding architecture and parameters as natural images, enabling a single model to transfer across tasks through a unified visual interface, in a way analogous to how language models operate over text. We refer to this formulation as RGB In and RGB Out (RINO). Built upon a generic image editing backbone without task-specific fine-tuning, RINO demonstrates robust and competitive zero-shot performance on both dense understanding tasks such as segmentation and depth estimation (where we unify outputs as RGB), and dense-conditioned generation tasks such as pose-to-image generation (where we unify inputs as RGB). We hope this study provides useful insights toward general unified vision-language systems, where diverse visual tasks can be expressed, interpreted, and solved through a shared visual language. Code is available at https://github.com/yangtiming/RINO.

cs.CV↗

arXiv2024

Analyzing and Reducing Catastrophic Forgetting in Parameter Efficient Tuning

Existing research has shown that large language models (LLMs) exhibit remarkable performance in language understanding and generation. However, when LLMs are continuously fine-tuned on complex and diverse domain-specific downstream tasks, the inference performance on historical tasks decreases dramatically, which is known as a catastrophic forgetting problem. A trade-off needs to be kept between learning plasticity and memory stability. Plenty of existing works have explored strategies like memory replay, regularization and parameter isolation, but little is known about the geometric connection of various adjacent minima in the continual LLMs fine-tuning scenarios. In this work, we investigate the geometric connections of different minima through the lens of mode connectivity, which means different minima can be connected by a low-loss valley. Through extensive experiments, we uncover the mode connectivity phenomenon in the LLMs continual learning scenario and find that it can strike a balance between plasticity and stability. Building upon these findings, we propose a simple yet effective method called Interpolation-based LoRA (I-LoRA), which constructs a dual-memory experience replay framework based on LoRA parameter interpolations. Extensive experiments and analysis on eight domain-specific CL benchmarks demonstrate that I-LoRA consistently show significant improvement over the previous state-of-the-art approaches with up to $11\%$ performance gains, providing a strong baseline and insights for future research on the large language model continual learning problem. Our code is available at \url{https://github.com/which47/LLMCL}.

cs.LG↗

arXiv2019

Mid-infrared frequency comb with 6.7 W average power based on difference frequency generation

We report on the development of a high-power mid-infrared frequency comb with 100 MHz repetition rate and 100 fs pulse duration. Difference frequency generation is realized between two branches derived from an Er:fiber comb, amplified separately in Yb:fiber and Er:fiber amplifiers. Average powers of 6.7 W and 14.9 W are generated in the 2.9 $μ$m idler and 1.6 $μ$m signal, respectively. With high average power, excellent beam quality, and passive carrier-envelope phase stabilization, this light source is a promising platform for generating broadband frequency combs in the far infrared, visible, and deep ultraviolet.

physics.optics↗

arXiv2018

Ultrafast extreme ultraviolet photoemission without space charge

Time- and Angle-resolved photoelectron spectroscopy from surfaces can be used to record the dynamics of electrons and holes in condensed matter on ultrafast time scales. However, ultrafast photoemission experiments using extreme-ultraviolet (XUV) light have previously been limited by either space-charge effects, low photon flux, or limited tuning range. In this article, we describe space-charge-free XUV photoelectron spectroscopy experiments with up to 5 nA of average sample current using a tunable cavity-enhanced high-harmonic source operating at 88 MHz repetition rate. The source delivers $ > 10^{11}$ photons/s in isolated harmonics to the sample over a broad photon energy range from 18 to 37 eV with a spot size of $58 \times 100 \; μ$m$^2$. From photoelectron spectroscopy data, we place conservative upper limits on the XUV pulse duration and photon energy bandwidth of 93 fs and 65 meV, respectively. The high photocurrent, lack of space charge distortions of the photoelectron spectra, and excellent isolation of individual harmonic orders allow us to observe the laser-assisted photoelectric effect with sideband amplitudes as low as $6 \times 10^{-4}$, enabling time-resolved XUV photoemission experiments in a qualitatively new regime.

physics.optics↗