The authors frame trusting and retiring each memory as a sequential statistical test.
Safe LLM memory reuse requires certificates, not simple success counters
Tests on changing knowledge and code show that likelihood-ratio certificates retire stale memories substantially faster while quantifying the cost of safe reuse.
Chinese Tech
Jianghong Zhou · Hao Rong · Chen Lin · Daoan Zhang
ByteDance · Scale AI
Research Digest··3 min read
Zhou and colleagues study reasoning contraction, where an agent replaces a multi-step derivation with a stored template, skill or answer that can be reused in one step.
Why this paper
From ByteDance and Scale AI
In one line
A certificate that weighs failures by likelihood ratio and restarts on changes reduces stale answers 7-13x on QA and 35-73x on code vs vote counters.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§