Research map / Large language models / Alignment and preferences
Evolution and compression in LLMs: on the emergence of human-aligned categorization
Nathaniel Imel, Noga Zaslavsky
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 |
|---|---|---|---|
| scale up and instruction-tune the LLM | LLMs on English color naming · English color naming (WCS, 39 models) | English-alignment and IB-efficiency | Improved |
| instruction-tune a very large model (Llama 3.3 70B) | Llama 3.3 70B · English color naming | IB-efficiency / English alignment | No clear effect |
| present colors as images instead of sRGB text | larger multimodal LLMs · English color naming | alignment and efficiency | No clear effect |
| present colors as images instead of sRGB text | smaller multimodal LLMs · English color naming | alignment and efficiency | Improved |
| present colors as CIELAB triples instead of sRGB | top four LLMs · English color naming | English alignment | Hurt |
| run iterated in-context language learning over generations | Gemini 2.0, Gemma 3 27B, Qwen 2.5 32B, Llama 3.3 70B · IICLL color naming | IB-efficiency | Improved |
| use Gemini 2.0 instead of other LLMs in IICLL | IICLL · IICLL color naming | complexity range reached | Improved |
| use conversation history window of 10 instead of 0 (ablation) | IICLL · IICLL chains | degenerate systems | Improved |
| increase history window from 10 to 20 or 50 (ablation) | IICLL · IICLL chains | results | No clear effect |
| use small instruction-tuned models in IICLL | Gemma 3 4b, Llama 3.2 3B, Qwen 2.5 3B, OLMo 2 7B · IICLL color naming | non-degenerate systems: 19 of 20 chains degenerate | Hurt |
| instruction-tuning stage in training | Olmo 2 32B · English color naming | English-alignment | Improved |
Closest papers by meaning
- Iterated Learning Improves Compositionality in Large Vision-Language Models (CVPR 2024 · similarity 0.86)
- Large Language Models as Model Organisms for Human Associative Learning (NeurIPS 2025 · similarity 0.86)
- A Phase Transition between Positional and Semantic Learning in a Solvable Model of Dot-Product Attention (NeurIPS 2024 · similarity 0.85)
- ColorConceptBench: A Benchmark for Probabilistic Color-Concept Understanding in Text-to-Image Models (NeurIPS 2026 · similarity 0.85)
- Bridging semantics and pragmatics in information-theoretic emergent communication (NeurIPS 2024 · similarity 0.85)
- Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws (NeurIPS 2025 · similarity 0.85)
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.
