AptMQL-Bench
Harbor: pendingA text-to-MQL benchmark with 21 document databases and 3,186 natural language requests, converted from BIRD using access-pattern schema design.
Auto86%Coding
- Samples
- 3,186
- Released
- 2026
- Scores
- 8
ArgGYM
ArgGYM: A Procedural, Engine-Verified Benchmark for Structured Defeasible Reasoning
A procedural benchmark with a frozen set of 1,440 verified instances evaluating structured defeasible reasoning across twelve tasks using an ASPIC+ argumentation engine.
How to run
ltb · arggym
$ npm i -g lettokenburn-cli$ ltb bench info arggym
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
11 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | 89.7 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Qwen/Qwen3.5-9B | 81.3 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| GPT-oss 120B | 78.1 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| google/gemma-4-31B-it | 77.4 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| google/gemma-4-26B-A4B-it | 75.1 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| google/gemma-4-E4B-it | 69.4 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| gpt-oss-20b | 65.7 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Qwen/Qwen3.5-4B | 63.6 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Gemini 2.5 Pro | 60.8 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Qwen/Qwen3.5-27B | 60.1 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| google/gemma-4-E2B-it | 13.1 | N/A | Independent | medium | Sep 29, 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
Content & Construction: ArgGYM is a reinforcement learning verifiable rewards (RLVR)-compatible environment based on the ASPIC+ framework. It decomposes defeasible reasoning into twelve distinct operations, such as formalization, defeat diagnosis, and theory revision. The benchmark includes procedural generators that create fresh instances while maintaining engine verification.
Scoring & Key Findings: Scoring is grounded in a symbolic argumentation engine computing formal states (e.g., acceptability). A frozen set of 1,440 instances covers fifteen curriculum configurations with varying structural demands and preference orderings. Results show that frontier models differ significantly from open-weight models on specific reasoning profiles.