KARMA adapts reward models from Reddit karma data to align LLMs with conversational pragmatics, finding that context-only rewards outperform karma-predictive ones downstream while reducing factuality across conditions.
Conversation Modeling on Reddit using a Graph-Structured LSTM
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
This paper presents a novel approach for modeling threaded discussions on social media using a graph-structured bidirectional LSTM which represents both hierarchical and temporal conversation structure. In experiments with a task of predicting popularity of comments in Reddit discussions, the proposed model outperforms a node-independent architecture for different sets of input features. Analyses show a benefit to the model over the full course of the discussion, improving detection in both early and late stages. Further, the use of language cues with the bidirectional tree state updates helps with identifying controversial comments.
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
KARMA: Karma-Aligned Reward Model Adaptation
KARMA adapts reward models from Reddit karma data to align LLMs with conversational pragmatics, finding that context-only rewards outperform karma-predictive ones downstream while reducing factuality across conditions.