Research map / Large language models / Language, safety and interpretability
Linguistic Properties and Model Scale in Brain Encoding: From Small to Compressed Language Models
SUBBA REDDY OOTA, Vijay Rowtula, Satya Sai Srinath Namburi GNVV, Khushbu Pahwa, Anant Khandelwal, Manish Gupta, Tanmoy Chakraborty, Bapi Raju Surampudi
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 Qwen2.5 from 3B to 14B parameters (ablation) | Qwen2.5 · whole-brain fMRI encoding, Moth Radio Hour | best-layer encoding score: 3B - 14B: 0.000 (p = 1.0) | No clear effect |
| scale Qwen2.5 from 1.5B to 3B parameters (ablation) | Qwen2.5 · whole-brain fMRI encoding | best-layer encoding score: +0.073 (p = 0.004) | Improved |
| scale Qwen2.5 from 7B to 3B (ablation) | Qwen2.5 · whole-brain fMRI encoding | best-layer encoding score: 3B - 7B: 0.028 (p = 0.06) | Mixed |
| scale LLaMA-3.2 3B vs 1B in brain-to-text decoding (ablation) | LLaMA-3.2 · brain-to-text decoding | BLEU-1 / WER: BLEU-1 0.1105 -> 0.1198; WER 4.4869 -> 4.2237 | Improved |
| apply AWQ or SmoothQuant quantization to 3B SLM (ablation) | Qwen2.5-3B · normalized brain alignment | normalized brain alignment: 0.924 -> 0.933 (AWQ), 0.930 (SmoothQuant) | No clear effect |
| apply GPTQ quantization to the model (ablation) | Qwen2.5-3B / 7B · normalized brain alignment | normalized brain alignment: 0.924 -> 0.910 (3B) | Hurt |
| apply AWQ quantization to 7B model (ablation) | Qwen2.5-7B · normalized brain alignment | best-layer score: -0.020 (7B minus AWQ), p < 0.001 | Improved |
| apply 10-50% unstructured magnitude pruning to 3B model (ablation) | Qwen2.5-3B · normalized brain alignment | normalized brain alignment: 0.924 -> 0.910 / 0.908 / 0.907 | No clear effect |
| apply 50% pruning to 1.5B model (ablation) | Qwen2.5-1.5B · normalized brain alignment | normalized brain alignment | Hurt |
| compress or shrink to 1-1.5B models (with or without quantization) (ablation) | Qwen2.5 / LLaMA / DeepSeek · language-ROI normalized alignment | normalized brain alignment | Hurt |
| apply GPTQ to LLMs and 3B SLMs and probe linguistic tasks (ablation) | quantized models · FlashHolmes | discourse/reasoning/morphology probing | Hurt |
| compression degrading discourse, syntax and morphology probing | compressed LMs · brain alignment vs FlashHolmes | normalized brain alignment | No clear effect |
Closest papers by meaning
- Scaling and context steer LLMs along the same computational path as the human brain (NeurIPS 2025 · similarity 0.92)
- Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models (NeurIPS 2025 · similarity 0.92)
- fMRI predictors based on language models of increasing complexity recover brain left lateralization (NeurIPS 2024 · similarity 0.91)
- Improving Semantic Understanding in Speech Language Models via Brain-tuning (ICLR 2025 · similarity 0.90)
- Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain (NeurIPS 2025 · similarity 0.89)
- Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement (NeurIPS 2025 · similarity 0.89)
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.
