Goodfire maps how LLM emotion representations drift through activation space while reading stories
burny_tech · x · 2026-10-10
- Goodfire (with Stanford) published a new entry in its Neural Geometry series, studying LLM representations over time for the first time: how internal activations trace trajectories in concept space as the model reads short stories.
- Core idea: prior work examined static geometry within a single forward pass; this post tracks dynamics sentence by sentence.
- Case study: how LLMs represent human emotions both geometrically (in activation space) and temporally — comedies rise toward joy, tragedies fall into loss.
- Metaphor: like a hiker traversing curved terrain, internal computation meanders along manifolds reflecting the model's evolving sense of story type.
- Authors include Eric Bigelow, Atticus Geiger and Jack Merullo; full paper on arXiv.
More from Research
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23