Surprise Alone Learns Lies: Evidence-Gated Continual Learning for LLMs from Streaming Claims
A surprise-driven learner absorbs planted lies as fast as random selection. Its own prior flags contradictions at AUC 0.78 to 0.83.
The claim
A continual learner that keeps surprising claims absorbs planted lies at a near-random rate.
The problem
Language models now keep learning after training from a stream of claims.
Surprise looks useful, because genuinely new facts are surprising to the model.
A falsehood about known knowledge is also highly surprising by construction.
How it works
The student is a LoRA-adapted Qwen or Llama model.
It learns the claims that its current belief finds least predictable.
A verifier judge can approve or reject each claim before learning.
Claims that contradict the model's prior are routed to trusted evidence.
Unrelated regression tests slow learning but do not filter the lies.
Figures
What changes
The surprise-only learner absorbs nearly every planted falsehood.
Its own prior separates known-fact lies at AUC 0.78 to 0.83.
That prior check saves 43% of judge calls across three models.
Limits
Experiments use three small open models and invented facts.
The prior cannot separate lies from legitimate updates of the same fact.
With planted matching sources for 30% of lies, the system performs worse than the judge alone.
Source · arXiv