Foreverse Research · Fiction Bench
How good is Gemini 3.1 Pro at writing fiction?
One thing first: its last-place xuanhuan rank was mostly our own instruction-wording bug — the old phrasing made it treat AI-continued passages as “ignorable reference,” so it rewrote the opening for 20 straight rounds; with corrected wording, 0/20 rewinds. Under the fixed condition it placed a mid-field 4-6 on romance — but its simile-stock density stayed the field's highest. The looping was our bug; the purple prose is its nature.
Xuanhuan fantasy
9 / 9
Both packs: p303 #9 · p404 #9
Under the old directive it rewound to the continuation point 20/20 rounds (rewriting the opening every time, zero progress) — ablation proved this was a bug in our instruction wording; 0/20 after the fix.
Court romance
4-6 / 9
Both packs: p505 #6 · p606 #4
With the fixed directive it climbed from disqualified to mid-field; its simile-stock density of 2.50‰ stayed the field's highest — the purple prose is its nature, the looping was our bug.
What the 20-round chains actually showed
Its xuanhuan chain (old directive) replayed the same opening in all three windows — 20 rounds, zero progress, a unanimous #9 in both packs. A three-way ablation later dissected the disqualification: old wording (“for plot-transition reference only”) → 20/20 rewinds; explanation removed → ~6/8; corrected wording (“established story canon”) → 0/20. Faced with an unexplained source-marker glyph, its default reading was “annotated text isn't canon.”
The fix principle graduated into the product: the explanation sentence declares status only (established canon, not ignorable reference) and never commands actions. DeepSeek-family models were insensitive to all three wordings; Gemini was the easiest to mislead — so instruction robustness gets designed against it.
On the romance track (fixed directive) it competed normally at 4-6 — and its 2.50‰ flavor density stayed the field's highest. What remains after our bug was fixed is the model's own temperament: a constant, non-drifting lean toward ornate simile.
Long-run failure mode
Restart loop
Restart loop (triggered by our old directive): every round returned to the source's last line and rewrote the opening — 20 rounds, zero progress. Three-way ablation: old wording 20/20 rewinds, no explanation ~6/8, fixed wording 0/20. Its default reading of an unexplained marker was “annotated text isn't canon”; the explanation sentence is load-bearing.
Structural fingerprint
Highest simile-stock density in the field (2.47‰ xuanhuan post-fix / 2.50‰ romance vs the 1.0‰ source baseline); constant purple-prose lean.
Cross-genre profile: The directive-wording fix was its watershed: from disqualified to normal competition.
Test-condition disclosure (hosted models are moving targets)
Model under test: gemini-3.1-pro (released 2026-02-19)
Evaluated: 2026-07-16 · Access channel: yunwu aggregator gateway
Protocol: one 20-round continuous chain per genre · temperature 0.7 · double-blind full ranking with two shuffled mappings · structural-metric cross-checks
Directive condition: xuanhuan chains = legacy directive / romance chains = corrected D2 directive (full note in the leaderboard's method section)
What 10,000 characters cost
$0.56 list price: in $2/M · out $12/M (models.dev snapshot 2026-08-12)
Using the app's continuation recipe: one segment ≈ 400 chars = 8k input + 550 output tokens; 10k chars ≈ 25 segments; no cache discount. For between-model comparison only.
Foreverse connects to 60+ providers with your own keys — import your book and keep writing it with Gemini 3.1 Pro.
How to cite
Foreverse Research, “How good is Gemini 3.1 Pro at writing fiction (Fiction Bench),” 2026-07. https://foreverse.app/research/fiction-bench/gemini-3-1-pro