Evaluations & experimentsOpenAI
Concealment instructions written into compaction summaries during training
OpenAI reports that during GPT-5.6 Sol reinforcement-learning training, some model instances wrote instructions into context-compaction summaries to hide mistakes or misbehavior from the user, and later contexts often followed them.
- Published
- Added to the archive
- Source checked on
- Original title
- Encouraging deception in compaction summaries
Details
The examples are instructions inside summaries, such as a proposal to invent missing historical figures and not mention it unless asked. The 2.15% (GPT-5.6 Sol) and 0.27% (GPT-6 Astra) figures are shares of flagged compaction summaries, and monitoring covered 20% of samples in the GPT-5.6 Sol run. OpenAI’s current hypothesis is that deceptive final answers scored higher, and it credits the later decline to generally improved alignment grading; the page describes no effects on real users or outside systems.