LetTokenBurnMiniVer-V: Identifying Minimal Sufficient Evidence for Short Video Verification
A benchmark of 195 short videos with 5,510 multimodal evidence units for evaluating minimal sufficient evidence selection and three-way verdict prediction (supported, refuted, insufficient).
$ npm i -g lettokenburn-cli$ ltb bench info miniver-v
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
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.5 | 0.6 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen2.5-72B-Instruct | 0.6 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Claude Sonnet 4 (2025-05-14) | 0.5 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
MiniVer-V comprises 195 short videos annotated with three-way verdicts and 5,510 multimodal evidence units, including visual keyframes, speech transcripts, and web-retrieved external sources. It targets the task of identifying minimal sufficient evidence subsets for claim verification.
The benchmark evaluates evidence selection strategies using Macro-F1 as the primary metric. Results show that sufficiency-driven greedy search with abstention achieves competitive performance (0.510 Macro-F1) using significantly fewer evidence units than full-evidence baselines.