CSE Roadmap

Core tier · about 4 hours · 11 parts

Part 5 of 11 · 2 min read

What the example found

Faithfulness-check accuracy by kind of claim, and self-consistency recall by sample count and by how consistent the wrong answers are.

How to read the faithfulness table: 22 hand-written claims against a six-sentence context (13 false, 9 true); each column adds one more rule on top of word overlap.

Detector False claims flagged True claims wrongly flagged
Word overlap (threshold 0.7) 54% 22%
Overlap + number check 85% 22%
Overlap + numbers + negation 100% 22%

What to notice: the false-alarm rate on true claims (22%) never moves across all three rows. Adding number and negation rules fixes false negatives on invented claims. It does not touch the paraphrase cases that cause the false alarms.

How to read the by-kind table: each column is a detector variant; each row is one kind of claim, with the share judged correctly.

Kind Overlap + numbers + numbers + negation
Supported, close wording (7) 100% 100% 100%
True paraphrase (2) 0% 0% 0%
Wrong number (4) 0% 100% 100%
Negation flip (3) 33% 33% 100%
Invented content (5) 100% 100% 100%

What to notice: true paraphrases sit at 0% correct in every column. No rule added here fixes paraphrase. The problem is not a missing check. Low word-overlap is simply the wrong signal for a sentence that says the same thing in different words.

How to read the self-consistency tables: 4,000 simulated questions, 70% genuinely known; flagged when sample agreement is below 100%.

k samples 2 3 5 10
Unknown facts flagged 83% 97% 99.7% 100%
Share answered consistently wrong (at k=5) 0% 30% 60%
Unknown facts flagged 100% 68% 39%

What to notice: more samples (first table) steadily close the gap to 100% recall. But no number of samples fixes the second table's problem. Consistent wrong answers are invisible to agreement-based detection, no matter how many times you ask.

{{svg:detectors|A faithfulness check needs a context; a self-consistency check needs samples; neither catches consistent errors against a wrong source.}} {{svg:results|What each check catches on 22 claims, and the share of unknown facts flagged by self-consistency.}}