A small trained module that transforms sparse historical hidden states into memory shifts improves LLM long-context reasoning, with 11.6–29.3 F1 gains on LoCoMo and 10.2–13.0 F1 gains on HotpotQA.
and Schuetze, Hinrich and Tresp, Volker and Ma, Yunpu
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.MA 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
TransMem: Transforming Hidden States into Memory for Large Language Models
A small trained module that transforms sparse historical hidden states into memory shifts improves LLM long-context reasoning, with 11.6–29.3 F1 gains on LoCoMo and 10.2–13.0 F1 gains on HotpotQA.