Research map / Large language models / Language, safety and interpretability
Anti-efficient encoding in emergent communication
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a message-length penalty to the Speaker objective (α=0.5) (ablation) | LSTM Speaker-Listener · Power-law inputs, max_len=30, a=40 | E (mean message length): 29.40 -> 7.22 | Helped |
| Add a message-length penalty to the Speaker objective (ablation) | LSTM Speaker-Listener · Power-law inputs, max_len=30, a=40 | Length-frequency relation (ZLA) | Helped |
| Increase the length-penalty coefficient (ablation) | LSTM Speaker-Listener · Communication task, max_len=30, a=1000 | Successful convergence | Hurt |
| Train the Speaker and Listener jointly rather than use an untrained Speaker | Speaker-Listener communication system · Power-law input distribution | Message-length distribution by input frequency | Hurt |
| Use emergent messages instead of monkey-typing messages | Untrained Listener · max_len=30, varying alphabet size | Average pairwise hidden-representation distance | Helped |
| Use emergent messages instead of optimal-code messages | Untrained Listener · max_len=30, varying alphabet size | Average pairwise hidden-representation distance | Helped |
| Use a uniform rather than power-law input distribution (ablation) | LSTM Speaker-Listener · Uniform inputs, max_len=30 | Mean message length | No clear effect |
| Use emergent messages rather than MT messages | LSTM Speaker-Listener · Uniform inputs, max_len=30 | Mean message length: p < 10^{-9} | Mixed |
| Use a=1000 rather than smaller alphabet sizes (ablation) | LSTM Speaker-Listener · Power-law inputs, max_len=30 | Frequency-rank/message-length relation | No clear effect |
Closest papers by meaning
- Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments (ICLR 2024 · similarity 0.85)
- AlphaZero Neural Scaling and Zipf's Law: a Tale of Board Games and Power Laws (NeurIPS 2025 · similarity 0.83)
- Scaling Laws for Gradient Descent and Sign Descent for Linear Bigram Models under Zipf’s Law (NeurIPS 2025 · similarity 0.82)
- Zipfian Whitening (NeurIPS 2024 · similarity 0.81)
- A compressive-expressive communication framework for compositional representations (NeurIPS 2025 · similarity 0.80)
- Fractal Patterns May Illuminate the Success of Next-Token Prediction (NeurIPS 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.