🚨New paper!
Your hallucination detector says it evaluates reasoning. But what if it's just peeking at the final answer?
We tested this: keep the reasoning, only change the answer. Many detectors' scores shift dramatically 🧵
TRACT stays stable under both FORCE and REMOVE since it scores the reasoning body, not the endpoint. It also stacks well: fusing TRACT with existing detectors gives +5 to +20 average AUC across all 5 models.
Since GPT-4 will be announced soon, here is my prediction about what it looks like:
1. Signals other than language models are also utilized as the training labels
2. More intense use of RLHF
3. Some kind of bootstrapping is done while training
Current framework of NBA In Season Tournament as soon as 2023-24, per sources:
- Cup games through November
- 8 teams advance to single-elimination Final in December; other 22 continue with regular season
- All games part of normal 82-game schedule; one extra for two Final teams