ChenLiu-1996 / LM-Dispersion
[𝗜𝗖𝗠𝗟 𝟮𝟬𝟮𝟲] Dispersion loss counteracts embedding condensation and improves generalization in small language models
Why this job: works in the areas we cover, but does not say which job it does
What it does
We observe a geometric phenomenon which we term embedding condensation, where token embeddings collapse into a narrow cone-like subspace in smaller language models. We then design a training objective called dispersion loss to counteract the effect.
Get started
No single install command. Read the steps in the README (opens in new tab)
- Maintenance
- Active
- Last commit 4 months ago
- Stars
- 126
- +3 this week
- Licence
- Custom
- Read before business use
- Downloads for
- None listed
- No release files
- Latest version
- None yet
- No releases
- Built with
- Python
- 100% of the code
Is it looked after?
- Contributors
- Not stated
- Last 90 days
- Issues
- None
- No issues filed
- Pull requests
- 1 merged
- 0 waiting
- Releases
- None
- Never published
- No contributing guide
- No security policy
- No code of conduct
More: languages, topics and project record
Languages
- Python 100%
Topics
- condensation
- cosine-similarity
- dispersion
- dispersive
- embedding
- embedding-condensation
- embedding-vectors
- geometric-learning
- icml
- icml-2026
- large-language-models
- latent-space
Project record
- JobOther AI tools
- StatusListed
- Commits289
- Forks6
- Watching1
- Stars in 28 days+3
- CreatedApr 16, 2025
- First seen by usNov 7, 2026
- CheckedNov 8, 2026