Study: post-training makes LLMs funnier but homogenizes their jokes across 11 stages
Gold-Bat-3225 · reddit · 2026-09-30
A study tested whether post-training actually makes LLMs funnier, using open models with fully published training stages — Tulu 3 (on Llama 3.1 70B), OLMo 3.1 32B, and Qwen2.5 — tracking 11 stages with 100 joke prompts, 64 human raters, and 2,330 head-to-head judgments.
Key findings:
- Post-training makes models funnier (later models won in 5 of 7 steps), and jokes got 10–20 words shorter after early post-training, reaching punchlines faster
- But joke diversity drops: in 6 of 7 steps, multiple jokes for the same prompt became more similar — asking for eight jokes yields eight versions of the same joke; the biggest drop was Qwen2.5 base→instruct
- Asking the model to plan a line or two before writing cut variety in all 4 models with no reliable funniness gain
- A comedian persona recovered some variety in all 4 models but only made jokes funnier in 2
Method: humans judged base vs. final models; a calibrated model judge compared intermediate stages. Full report: laugh.so/research/humor-tax.
More from Research
- WorkflowEvals: typesafe open collection for evaluating agents on real workflows — _lewtun · 2026-09-30
- Analog Design Bench: agents pass just 8–78% of hours-long chip design tasks — tokenbender · 2026-09-30
- NeurIPS Paper Introduces MISVO to Steer LLMs at Inference Time Without Fine-Tuning — DanielKhashabi · 2026-09-30
- VFig lands NeurIPS 2026: 4B VLM matches GPT-5.2 at converting complex figures to SVG — jmin__cho · 2026-09-30
- UniMate open-sources unified text-to-animation model that drives diverse 3D skeletons — grandorganics · 2026-09-30
- kalomaze: STE works fine under backprop even for binary weights and latents — kalomaze · 2026-09-30