What if AI helped you refine a classical Chinese poem instead of simply writing one for you? Most AI poetry tools hand you a finished poem. This one hands you the evidence and makes you do the deciding — because in classical Chinese poetry, the deciding is the craft. Jiuge-Tuiqiao, from @TsinghuaNLP (OpenBMB member), accepted to EMNLP 2026 System Demonstrations.

1️⃣ The user, not the model, holds the pen. A poem is a character matrix M (8×5 for a 5-char lvshi) paired with a binary lock matrix L. Lock a character or a whole line, and generation is hard-constrained to the remaining cells. Init → generate → review & lock → repeat, as a finite state machine, for as many rounds as you want.

2️⃣ Every suggested character comes with its receipts. Four evidence channels open in a side panel: 28,739 collocations mined from 320K Tang-Song jueju and lvshi lines via t-test (co-occurrence ≥3, T > 1.96); up to five famous lines ranked by ascending PPL under Qwen3-32B over 1.8M+ poems from Pre-Qin to Qing, length-binned at 12 and 16 characters; imagery, antithesis and rhyme entries extracted from 11 classical encyclopedias including Hailu Suishi and Baikong Liutie; plus LLM rationales precomputed into an 8,000-entry LFU cache.

3️⃣ Prosody feedback lands 500 ms after you stop typing. The Character Grid Editor recolors in place: green for compliance, yellow for acceptable variants, purple for a Pingshui Yun rhyme mismatch, red for a tonal violation, gold fill for locked tokens. Violations carry an explicit justification rather than a silent autocorrect — a deliberate choice the authors defend on the grounds that silent fixes erode the user's own metrical intuition.

4️⃣ The evidence measurably helps, and the hard parts are reported honestly. On a 50-line cloze task with Qwen-turbo as backbone, context alone gives 12% Hit@1; adding PPL-ranked famous lines lifts it to 60%, and combining famous-line and collocation evidence keeps Hit@5 at 60%, with Hit@1 at 56%. Across 80 generation requests, Format and Rhyme are 100% for all four verse forms and L-Pattern stays ≥91.25% — but global poem-level pattern adherence (incl. nian-dui) drops from 65% on 5-char jueju to 15% on 7-char lvshi. In a preliminary study with 10 poetry enthusiasts and scholars, prosody guidance rated 8.20/9 and overall satisfaction 8.08/9; users favored famous lines and encyclopedia clues (6.90 each) over raw collocation statistics (6.10).

📄 http://arxiv.org/abs/2608.23098

#AI #THUNLP #OpenBMB #LLM #NLP #ClassicalChinesePoetry #HumanAICollaboration #Interpretability