A tunable 'overthinking' backdoor can be implanted into large reasoning models so that repeating a trigger word N times forces N extra reasoning steps, increasing token use several-fold without hurting accuracy.
Stealthy and persistent unalignment on large language models via backdoor injections
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
BadReasoner: Planting Tunable Overthinking Backdoors into Large Reasoning Models for Fun or Profit
A tunable 'overthinking' backdoor can be implanted into large reasoning models so that repeating a trigger word N times forces N extra reasoning steps, increasing token use several-fold without hurting accuracy.