You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
fix: rename semantic drift to lexical drift, update default model names
Jaccard distance measures vocabulary replacement rate, not semantic
similarity. Renamed all labels and comments accordingly. Updated
hardcoded default models to current versions (gpt-5.4-mini,
claude-haiku-4-5-20251001, gemini-2.5-flash).
<plotx="45"autoPlotX="true"yMax="1.0"autoPlotY="true"yAxis="Drift"y="505"xMin="0.0"height="165"legend="true"xMax="10.0"yMin="0.0"width="550"xAxis="Chain Position"display="Semantic Drift by Provider">
648
-
<setup>set-current-plot "Semantic Drift by Provider"</setup>
647
+
<plotx="45"autoPlotX="true"yMax="1.0"autoPlotY="true"yAxis="Drift"y="505"xMin="0.0"height="165"legend="true"xMax="10.0"yMin="0.0"width="550"xAxis="Chain Position"display="Lexical Drift by Provider">
648
+
<setup>set-current-plot "Lexical Drift by Provider"</setup>
A Telephone Game for LLMs. Parallel chains of turtles pass a seed message through repeated paraphrasing by different LLM providers. Each position in the chain asks its provider to rephrase what it received. As the message travels, semantic drift accumulates — and different providers drift at different rates and in different ways.
683
+
A Telephone Game for LLMs. Parallel chains of turtles pass a seed message through repeated paraphrasing by different LLM providers. Each position in the chain asks its provider to rephrase what it received. As the message travels, lexical drift accumulates — vocabulary shifts as each provider substitutes its own word choices, and different providers drift at different rates.
0 commit comments