LetTokenBurnEgoProbe: Benchmarking Long-Range and Exact Memory in Streaming Egocentric Videos
EgoProbe evaluates long-range retention and exact event-specific recall in streaming egocentric videos using 2,366 QA pairs and 1,229 two-turn dialogues with human verification.
$ npm i -g lettokenburn-cli$ ltb bench info egoprobe
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-5.6 Sol | 55.3 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| VideoChat-Flash | 53.2 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| LLaVA-OneVision-2 | 52.1 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Gemini-3.5-flash | 51.1 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| StreamForest | 50.7 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| EM 2 Mem | 47.9 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| InternVL3.5 | 47.6 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| HERMES | 46.4 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Qwen 3.5 | 45.9 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| StreamingVLM | 44.5 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Ego-R1 | 41.0 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| EgoRAG | 36.5 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| SimpleStream | 28.4 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Built on Ego4D and EgoLife, EgoProbe uses an event-grounded engine to create long-range queries and distractors from related events. It includes EgoProbe-Base (no interference) and EgoProbe-Echo (with interference and coarse-to-fine probing).
Models are evaluated on accuracy for single-turn and multi-turn questions. Results show substantial limitations in long-range recall and specific detail retrieval amidst interfering similar events.