Research map / Large language models / Language, safety and interpretability
Learning is Forgetting; LLM Training As Lossy Compression
Henry Conklin, Tom Hosking, Tan Yi-Chern, Jonathan D. Cohen, Sarah-Jane Leslie, Thomas L. Griffiths, Max Bartolo, Seraphina Goldfarb-Tarrant
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use optimality (expressivity per bit of complexity) as predictor instead of expressivity alone | 47 open-weights LLMs · aggregate of 6 benchmarks | Spearman/Pearson correlation: r=0.08 (expressivity) -> r=0.52 (optimality) | Improved |
| lower token-level complexity | 47 open-weights LLMs · aggregate of 6 benchmarks | correlation with performance: r=-0.38 | Improved |
| more preference information in representations | 47 open-weights LLMs · aggregate of 6 benchmarks | correlation with performance: r=0.76 | Improved |
| optimal compression of C4 as predictor for instruction following | 47 open-weights LLMs · IFEval | correlation with performance: r=0.07, p=0.631 | No clear effect |
| preference information as predictor of instruction following | 47 open-weights LLMs · IFEval | correlation with performance: r=0.39, p=0.004 | Improved |
| smaller model scale (1B vs 7B and 32B OLMo2) (ablation) | OLMo2 pre-training · information plane over pre-training | approach to IB bound | Hurt |
| condition on wider context (token to trigram/quadgram back-off) (ablation) | OLMo2 7B · information plane over pre-training | proximity to IB bound | Improved |
Closest papers by meaning
- Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws (NeurIPS 2025 · similarity 0.91)
- Compressing LLMs: The Truth is Rarely Pure and Never Simple (ICLR 2024 · similarity 0.91)
- Language Modeling Is Compression (ICLR 2024 · similarity 0.91)
- When Less is More: The LLM Scaling Paradox in Context Compression (NeurIPS 2026 · similarity 0.91)
- Language Models as Zero-shot Lossless Gradient Compressors: Towards General Neural Parameter Prior Models (NeurIPS 2024 · similarity 0.91)
- Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression (ICML 2025 · similarity 0.90)
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.
