FIBBLE⁵ Arena for LLMs

The All-Lies Paradox: Why Total Deception Is Paradoxically Solvable

In Fibble⁵, every single tile in every row is a lie — 0 of 5 tiles are truthful. At first glance this seems impossibly hard, but there is a profound paradox at work: when all information is false, you can simply invert it to recover the truth. Total deception is, in a sense, total information.

Here is how inversion works: if a tile shows GREEN, the letter is not correct in that position. If a tile shows GRAY (absent), the letter is guaranteed to be in the word. If a tile shows YELLOW, the letter is not merely in the wrong spot — it is either absent entirely or correct in that exact position. GRAY tiles become the most informative: every GRAY letter is confirmed present in the word.

This creates a fascinating challenge for LLMs. Models that have learned to handle partial deception (1 or 2 lies) through hypothesis tracking must now shift to a completely different strategy: systematic inversion. The cognitive challenge mirrors scenarios where a source is known to always lie — a classic logic puzzle trope. The question is whether LLMs can recognize and exploit this structural property rather than drowning in the apparent complexity of "all clues are wrong."

Compared to partial-lie variants like Fibble (1 lie) or Fibble² (2 lies), Fibble⁵ is paradoxically easier for a solver that recognizes the inversion principle. There is no ambiguity about which tiles are lying — they all are. The benchmark tests whether models can identify and apply this meta-reasoning shortcut, making it a unique probe of strategic reasoning under total deception.