Technology / WORLD

Yonsei team tests a unified web-access interface for agents

Web$hell combines search, browser actions and revisiting earlier page states in one interface.

Reading sizeSources ↓Share story ↗
Conceptual illustration of technology, processors and connected information.Illustration
Symbolic illustration

Yonsei University researchers introduced Web$hell, an interface that brings web search, page interaction and revisiting previously gathered evidence into one system for AI agents. WebContract compiles pages into information and action objects; WebCache stores page states so agents can inspect earlier versions after an interaction.

In experiments, Web$hell used an average of 17.6 interaction turns per task on DeepWeb-Bench research tasks, compared with 52.4 for CLI-Any-Browser. The authors also report performance gains on web-search, navigation and multi-source research benchmarks. These measurements reflect specific agents, interfaces and benchmark settings; they do not establish success across general web use.

Data

Table

Measure

DeepWeb-Bench multi-source research tasks; mean turns per task

Mean turns per DeepWeb-Bench task with Web$hell17.6DeepWeb-Bench multi-source research tasks; mean turns per task
Methodology

Values come from the paper's stated benchmark experiments; they are not field measurements or general performance guarantees.

Updated:

Table

Measure

DeepWeb-Bench multi-source research tasks; CLI-Any-Browser comparison

Mean turns per DeepWeb-Bench task with CLI-Any-Browser52.4DeepWeb-Bench multi-source research tasks; CLI-Any-Browser comparison
Methodology

Values come from the paper's stated benchmark experiments; they are not field measurements or general performance guarantees.

Updated:

Structured coverage

Subjects and places

Subject proposals

Geographic references

Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.

  • South Korea · Place role: Coverage · Place type: Country or area · Location precision: Regional
    Evidence
    • Explicit preserved editorial country/region; no city or national centroid inferredarxiv.org ↗
Taxonomy version: IPTC Media Topics 2026-Q2

Continue reading

Technology

Rollout-Time Intervention tests agent training on new tasks

The framework adds multi-agent correction to training rollouts and filters for distribution shift.

arXiv preprint by Alibaba and Hong Kong University of Science and Technology researchers1 min read
Data included
Technology

LongHarness tests evidence management for long-context agents

The framework supports agent answers by updating evidence and conclusions together across long documents.

arXiv preprint by researchers affiliated with Fudan University, Shanghai AI Laboratory, Peking University and other Chinese institutions1 min read
Data included