---
id: "lib-2026-10-07-training-a-vlm-agent-to-operate-the-web"
type: "library"
title: "Training a VLM agent to operate the web"
title_original: "RL meets Browser: как мы обучили VLM-агента работать в вебе за пользователя"
lang: "RU"
org: "Yandex for ML (Practical ML Conf)"
source_type: "Talk"
date: "2026-10-07"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-07-training-a-vlm-agent-to-operate-the-web/"
original_url: "https://www.youtube.com/watch?v=Ic1GvfBOIHo"
tier: "C"
tier_name: "Weak signal"
tier_reason: "Only the title and conference abstract of this Yandex Practical ML Conf talk were available; the content was not reviewed, so it counts as a signal until a transcript is checked."
summary_basis: "metadata-only"
original_minutes: "40"
tags: ["ai technology", "computer use agents", "reinforcement learning"]
---

# Training a VLM agent to operate the web

Original title: RL meets Browser: как мы обучили VLM-агента работать в вебе за пользователя

Tier C (Weak signal): Only the title and conference abstract of this Yandex Practical ML Conf talk were available; the content was not reviewed, so it counts as a signal until a transcript is checked.

Summary based on the title and listing only; the content itself was not available.

## In short

A 40-minute Russian-language talk by Artemy Golikov, published on 7 October 2026 by Yandex for ML from the Practical ML Conf programme. The original was not accessible for this digest: no transcript was pulled, and the YouTube page returned only the title. According to the conference abstract as logged in our research lane, the talk describes multi-turn reinforcement learning for a computer-use agent built on a vision-language model (VLM) that operates in a live, non-deterministic browser on behalf of a user. No findings or figures can be reported until the talk is transcribed.

## What it reports

- Topic, from the title and abstract: training a VLM-based agent with reinforcement learning to carry out tasks in a real web browser for a user.
- Setting: a production computer-use agent rather than a research demo; the training environment is a live browser where pages change between runs.
- Format: conference talk, about 40 minutes, Yandex for ML channel, Practical ML Conf.
- No metrics, success rates, dataset sizes or costs were available in the material reviewed.

## How they got it

Not verifiable from the material available. The abstract indicates multi-turn RL in a live browser environment; reward design, task sets and evaluation method are not described in the metadata.

## Limits

- Digest based on title and abstract only; the content of the talk was not reviewed.
- Speaker affiliation and product context are inferred from the channel and conference, not confirmed from the talk.
- Any vendor or product claims made in the talk are unverified here.

## Why it matters for companies

Browser agents that act inside web applications are one of the main routes by which AI could take over routine clerical work in systems without APIs, such as HR portals, procurement tools or travel booking. Most public material on such agents covers demos; a first-hand account of training one for a live product is less common. Managers evaluating browser-automation vendors should ask the questions this talk is likely to address: how the agent copes with pages that change, and how success is measured. This entry should be revisited once a transcript is available.

## Key passages in translation

> RL meets Browser: как мы обучили VLM-агента работать в вебе за пользователя

RL meets Browser: how we trained a VLM agent to work on the web on the user's behalf. (Talk title; no further original text was accessible.)

Original: https://www.youtube.com/watch?v=Ic1GvfBOIHo