Grounding-IQA
Harbor: pendinggrounding multimodal language model for IQA
Incomplete29%Multimodal
- Samples
- N/A
- Released
- 2026
- Scores
- 0
Grounding Awareness
Grounding Awareness benchmark
A controlled benchmark of 160 scenarios and 800 test instances evaluating agents' ability to distinguish accepted from unacknowledged proposals in longitudinal interactions.
How to run
ltb · grounding-awareness
$ npm i -g lettokenburn-cli$ ltb bench info grounding-awareness
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
6 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen Max (2025-01-25) | 0.5 | gl | Independent | medium | Oct 13, 2026 | arxiv.org |
| DeepSeek Chat | 0.4 | gl | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen Max (2025-01-25) | 0.2 | langmem | Independent | medium | Oct 13, 2026 | arxiv.org |
| DS | 0.2 | langmem | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT | 0.1 | langmem | Independent | medium | Oct 13, 2026 | arxiv.org |
| Gemini | 0.1 | langmem | Independent | medium | Oct 13, 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
The benchmark uses an agentic framework to generate matched conversation histories differing in proposal acceptance status. It evaluates sensitivity and contextual selectivity across memory writing, retrieval, and action selection.
Scoring focuses on whether agents correctly execute only authorized arrangements. Findings show strong memory performance does not ensure Grounding Awareness, with persistent memory often causing execution of unacknowledged proposals.