State Beside the Text: An Afferent Channel for Agents and Game Characters

A near-full context, a dragon at the gate: agents and game characters must act on state the text lacks. An afferent (inward-carrying) channel that only the harness or engine writes feeds it beside every token. Fine-tuned per setting, Qwen3-1.7B names a crossed limit mid-sentence and acts (95.0%) and flags which watched job changed. It picks a message route and says why, missing its strict pass line; forged status text moved the route. At simulated terror, a character breaks off and flees or draws a weapon by role (99.5%). A player's typed alarm drew words, never flight or a weapon (one seed). Choice by water or food at hand was not learned in this run. A matched text-notice model detects as well and met more confirmatory lines, but forged harness-format notices set it off in 19.3% of user-message runs, and the channel model only inside tool responses. Trained on single changes, only the channel model followed a state that changed back. Of 92 pass lines, predictions and stop rules fixed in advance, 33 were not met.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24355790
Primary Topic
Artificial Intelligence in Games
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

State Beside the Text: An Afferent Channel for Agents and Game Characters

Edelmar Schneider
2 links
Zenodo (CERN European Organization for Nuclear Research)
Artificial Intelligence in Games
preprint

State Beside the Text: An Afferent Channel for Agents and Game Characters

Edelmar Schneider
preprint en

Abstract

A near-full context, a dragon at the gate: agents and game characters must act on state the text lacks. An afferent (inward-carrying) channel that only the harness or engine writes feeds it beside every token. Fine-tuned per setting, Qwen3-1.7B names a crossed limit mid-sentence and acts (95.0%) and flags which watched job changed. It picks a message route and says why, missing its strict pass line; forged status text moved the route. At simulated terror, a character breaks off and flees or draws a weapon by role (99.5%). A player's typed alarm drew words, never flight or a weapon (one seed). Choice by water or food at hand was not learned in this run. A matched text-notice model detects as well and met more confirmatory lines, but forged harness-format notices set it off in 19.3% of user-message runs, and the channel model only inside tool responses. Trained on single changes, only the channel model followed a state that changed back. Of 92 pass lines, predictions and stop rules fixed in advance, 33 were not met.

Zenodo (CERN European Organization for Nuclear Research)
Artificial Intelligence in Games
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon