Vision-Language Model

VLMs (Vision Language Models) typically consist of a visual encoder (such as ViT), a language model, and a projection layer connecting the two, with representative models including GPT-4V/4o, Claude 3.5, the LLaVA series, and InternVL

RLVR is Reinforcement Learning with Verifiable Rewards, which guides model optimization by designing verifiable reward signals.

Common means of mitigating vision language model hallucinations include: providing more reference images or a candidate species list for comparison, supplementing with authoritative species databases via retrieval augmentation, explicitly giving confidence levels in the output, and having users participate in the final confirmation step.

A vision language model generating persuasive explanations and it truly accurately identifying a species are two different things; the model can fluently describe bird features while its species judgment may be wrong.

In vision-language model inference, one image usually produces about 576 tokens, and the prefill stage has high computing overhead.

Vision-language models are good at identifying "whether there is content" (blank, truncated, garbled), but not good at evaluating "whether the design looks good" (alignment, spacing, visual weight balance).

Hallucination is not a program bug, but an inherent byproduct of the model's working mechanism of probabilistically predicting the next word.

Most vision-language models process CT scans as a stack of flat images, sampling only a few slices or compressing volumetric data into a small number of tokens, causing critical spatial details to be lost before inference.

The UniEvo-VL project proposes enabling vision-language models to achieve self-improvement through self-distillation training, reducing reliance on external supervision signals.

Vision-language models tend to rewrite anomalous, incomplete, or unconventional text into linguistically smoother expressions during OCR, compromising the faithfulness of the transcription

The current mainstream technical approach in embodied intelligence is to use vision-language models (VLMs) as the robot's 'brain,' combined with low-level motion control to achieve end-to-end intelligent operation.

BI Visual Advisor is essentially a multimodal AI application based on image understanding, relying on the capabilities of a Vision-Language Model (VLM).

The researchers propose a simple, training-free framework that enables reasoning-capable vision-language models to perform iterative search, dynamic reasoning, and evidence accumulation.

MedProb is a lightweight probing framework that bypasses the free-text generation step by directly predicting multiple-choice medical visual question answering answers from frozen vision-language model representations.

The research team conducted a systematic comparison of 14 matched pairs of general VLMs and medical VLMs, finding that medical domain-specific adaptation did not consistently improve linear decodability.

Vision-language models are relatively mature in evaluating prompt adherence, but still have limitations in dimensions such as motion quality that require temporal understanding.

When endowing existing VLMs with generative capabilities, the understanding abilities accumulated during model pretraining are often impaired, resulting in catastrophic forgetting.

Tulip has launched a new product called Factory Playback, which connects cameras and uses vision-language models (VLMs) to inject on-site context into factory digital touchpoints.