Skip to content

Latest commit

 

History

History
32 lines (22 loc) · 1.05 KB

File metadata and controls

32 lines (22 loc) · 1.05 KB

LoCoMo QA Benchmark — Cogram vs Full Context

Date: 2026-07-18 21:29 UTC
Model (answer + judge): openai/gpt-4o-mini
Conversation: conv-26 (419 turns)
Questions: 30 (categories 1–4 only; category 5 adversarial skipped)

Per-category counts

Category Count
1 8
2 8
3 7
4 7

Results

Condition Accuracy Avg context tokens
full_context 53.3% (16/30) 15,961
cogram 16.7% (5/30) 1,961

Headline

Cogram uses ~8.1× fewer context tokens at 31% of full-context accuracy (16.7% vs 53.3%).

Honest read

Cogram accuracy (16.7%) was substantially lower than full-context (53.3%) despite ~8× token savings. On a single short conversation, statistical concept retrieval misses evidential turns that keyword-sparse questions need; full-context wins trivially when everything fits in window. Cogram's value proposition is accuracy-per-token on corpora too large to stuff.

Raw per-question data: benchmark_qa_raw.json