Reproduce “Negative Before Positive: Asymmetric Valence Processing in Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sohan Venkatesh

The authors use activation patching in three small instruction-tuned open models to locate valence-related processing, comparing positive and negative versions of matched situations against a shared neutral prompt. Negative-outcome signals peak earlier than positive-outcome signals, and a good-news direction shifts all 50 Qwen2.5-1.5B-Instruct neutral prompts toward a more positive score. These findings measure an anchor-token proxy, not emotion.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.