How much longer is LoCoMo?
Per-conversation length vs. Multi-Session Chat (MSC).
Hover a bar.
305 turns, 19 sessions Illustrative placement
Each cell is a turn. Humans edited about 15% (46 of 305 here); where they sit in the grid is drawn at random.
Hover a cell.
Generation pipeline
Click a stage, or step through it.
Three ways to handle history
Average conversation ≈ 9,200 tokens.
Two memory forms
The same forms serve as the generation memory and as the retrieval stores.
Five QA categories
Hover a row. Glyphs are conceptual.
Hover a category.
Token F1 calculator
Word overlap: precision and recall combined, F1 = 2PR / (P + R).
A wordier right answer lowers precision, so F1 drops. The authors admit this penalizes paraphrase.
Reported results
Pick a view. Values marked ≈ are approximate figures stated in the episode.
Truncation floor
Context windows in K tokens vs. the average conversation.
Everything is far below the 128K window of GPT-4-turbo. The paper never ran it on the full conversation.
Claims to read carefully
Hover or tap a chip.
Hover a chip.
Where the benchmark went next
Follow-ups are not results from this paper.
Hover a node.