CritPt
Harbor: pendingCritPt — AI evaluation benchmark listed by opencompass.
Verified71%Knowledge & Reasoning
- Samples
- 70
- Released
- 2025
- Scores
- 141
PerSo
PerSo: Investigating Facets of LLMs' Social Intelligence in Real-World Persuasion Scenarios
PerSo evaluates LLMs' cognitive and behavioral social intelligence using real-world Reddit ChangeMyView persuasion dialogues via multiple-choice questions.
How to run
ltb · perso
$ npm i -g lettokenburn-cli$ ltb bench info perso
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
10 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen-397B | 76.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen-35B | 75.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemma-27B | 71.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Claude Haiku 4.5 | 69.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V3.2 | 69.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek-V3.1 | 67.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen-4B | 64.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemma-12B | 62.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V4 Pro | 62.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemma-4B | 60.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
Constructed from Reddit's ChangeMyView community, PerSo uses an LLM-in-the-loop pipeline to generate evaluation tasks from real persuasion threads. It assesses cognitive intelligence (mental state forecasting) and behavioral intelligence (persuasion success prediction).
Scoring involves multiple-choice accuracy for mental states and selection of successful responses. Results show performance scales with model size, with top LLMs surpassing humans on behavioral tasks but lagging on cognitive ones.