AgentHorizon
Harbor: pendingA benchmark of 1,373 computer-use tasks with paired instructions to evaluate agentic judges on long-horizon trajectory success verification across three OSs.
- Samples
- 1,373
- Released
- 2026
- Scores
- 10
XiangqiBench
XiangqiBench for Closed-Loop Evaluation of LLM Agents
An executable benchmark with 119 tactical Chinese chess endgames where LLM agents play Red against an engine defender to measure closed-loop conversion success.
How to run
ltb · xiangqibench
$ npm i -g lettokenburn-cli$ ltb bench info xiangqibench
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
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini 3.1 Pro | 38.7 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| GPT-5.5 | 17.6 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Qwen3.7 Max | 13.4 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Claude Opus 4.7 | 10.9 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| DeepSeek V4 Pro | 5.0 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Seed 2.0 Pro | 2.5 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| DeepSeek R1 | 1.7 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Kimi K2.5 | 1.7 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| DeepSeek V3.2 | 0.8 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| GPT-oss 120B | 0.8 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| MiniMax M2.5 | 0.8 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Mistral Large 3 | 0.8 | N/A | Independent | medium | Oct 1, 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
XiangqiBench comprises 119 search-supported tactical endgames from classical treatises. Agents interact via a REPL interface with simulation tools to deliver checkmate against a strong engine defender.
Scoring relies on terminal win outcomes rather than move matching. Results reveal significant gaps between static competence and closed-loop success, with top models winning only a fraction of trials despite high pass@3 scores.