THE GENERATION METHOD
Plan how a discussion grows.
Write what comes next.
ThreadPDG—Thread-Prompt-based Discussion Generation—is a Planner–Writer framework that studies a complete real seed discussion, generates multiple drafts and combines their branches using thread-level statistics.
From a real example
to a new discussion.
- 01PLANNER
Learn how a thread develops.
Read a real seed thread. Plan each branch’s direction and each comment’s role within the reply template.
- 02WRITER
Turn plans into new comments.
Use comment plans, assigned personas and prior generated context to fill the planned discussion.
- 03COMPOSITION
Combine the strongest branches.
Generate multiple drafts and select branches against reference-pool metric targets for the current seed.
What goes into
generation?
A complete real seed thread
The Planner reads the root post, comments and parent–reply relationships. A structural template specifies comment slots, reply structure and length targets. Each draft inherits the seed’s comment count and reply-tree topology. The Planner determines the direction of each branch and what every comment should contribute.
A domain profile and assigned personas
A profile derived from the domain’s separate reference pool guides the discussion patterns. The Writer follows the plans and template, with personas sampled from MatrAIx and previously generated discussion context. This is prompt-based generation; no model-weight update is implied.
A separate reference pool
For branch composition, each seed receives a randomly sampled reference thread from the domain’s reference pool. That one thread supplies the targets for all metrics, preserving relationships among the target metric values. The selection objective uses metric weights and reference-pool interquartile ranges.
Paper · §4, Appendix F ↗Several drafts.
One composed thread.
The main experiments allow up to K = 3 Planner–Writer drafts per seed. A branch contains a parent comment and all its sub-comments. The final thread can take different branches from different drafts of that same seed.
The search minimizes a weighted sum of squared, IQR-normalized differences between the composed thread’s metrics and its sampled reference targets. Selection uses the current seed’s drafts and target plus fixed reference-pool statistics, without other held-out test data.
Paper · §4.2, §5.1 ↗Progress with
a clear boundary.
ThreadPDG is a case study in using ThreadBench to improve discussion realism. It currently depends on corresponding real seed discussions. Generating equally realistic threads independently at inference time is future work.
Results vary by model and domain. With GPT-5.4-mini, the method passes 67–92% of metrics; with DeepSeek-V4.1-Flash, 58–100%; and with GPT-4o-mini, 17–42%. Politeness, disagreement and length remain challenges.
Passing statistical tests means a difference was not detected at the chosen threshold. It does not prove distributional equivalence or establish realism beyond the benchmark’s measured properties.
Paper · §5, Appendix L ↗How does ThreadBench relate to other social simulation projects?
ThreadBench evaluates the distributions of generated Reddit discussions against real threads. OASIS is a social media simulation framework; SynthPAI provides synthetic Reddit comments for personal attribute inference. Their generation methods are evaluated alongside ThreadPDG in this paper. SimBench benchmarks human-behavior simulation across structured tasks such as moral and economic choices; it is related work, with a different evaluation scope and no shared score scale.
Does ThreadPDG fine-tune the language model?
No. ThreadPDG is prompt-based generation: the Planner and Writer are prompted with the seed thread, a structural template, a domain profile and sampled personas. No model-weight update is implied.
Can ThreadPDG generate a discussion from only a root post?
Not in the paper’s main setting. ThreadPDG conditions on a complete real seed thread, including its comments and reply relationships, and each draft inherits the seed’s comment count and reply-tree topology. Independent root-post-only generation is future work.
For the ThreadBench protocol’s direct comparisons, see ThreadPDG, OASIS and SynthPAI results. For metric definitions, see the twelve-metric reference.