CritPt
Harbor: pendingCritPt — AI evaluation benchmark listed by opencompass.
Verified71%Knowledge & Reasoning
- Samples
- 70
- Released
- 2025
- Scores
- 141
SIT-Eval
Situation-Based Interactive Turn-by-Turn Evaluation of Conversational Agents
A framework evaluating conversational agents on 1292 turns using situation-based rules and an LLM-judge to assess pedagogical adherence.
How to run
ltb · sit-eval
$ npm i -g lettokenburn-cli$ ltb bench info sit-eval
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
8 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5 | 0.8 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Claude Sonnet 4.5 (2025-09-29) | 0.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Gemini 2.5 Pro | 0.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Qwen-2-5-7B-SocraticLM | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Llama-3-3-70B-Instruct-AWQ | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Qwen2.5-7B-Instruct | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Llemma-7B-Science-Tutor | 0.3 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Qwen2.5-Math-7B-Instruct | 0.2 | N/A | Independent | medium | Oct 16, 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
SIT-Eval uses an LLM-interlocutor to enact uniformly sampled situations and a context-aware LLM-judge to evaluate agent responses against expert-crafted pedagogical rules. It focuses on turn-by-turn assessment rather than holistic dialogue scoring.
The benchmark validates the judge on human-verified golden dialogues, achieving 84.4% accuracy. Testing 8 LLMs revealed hidden failure modes, with models struggling most in ambiguity clarification and Socratic questioning despite similar aggregate scores.