Research map / Large language models / Language, safety and interpretability
Mixture of Cognitive Reasoners: Modular Reasoning with Brain-Like Specialization
Badr AlKhamissi, C. Nicolò De Sabbata, Greta Tuckute, Zeming Chen, Martin Schrimpf, Antoine Bosselut
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 |
|---|---|---|---|
| add brain-inspired Stage-1/2 specialization curriculum (MiCRo) over unspecialized mixture-of-blocks (ablation) | MoB · GSM8K | accuracy (%): 30.5 -> 34.7 (Llama-1B); 51.6 -> 57.2 (Llama-3B) | Improved |
| MiCRo versus dense model on reasoning benchmarks | Llama-3.2-1B/3B · GSM8K, MMLU, BBH, Minerva | accuracy (%): GSM8K 36.8 -> 34.7 (1B); MMLU 29.7 -> 31.2 (1B); MMLU 48.6 -> 45.4 (3B) | Mixed |
| ablate the Logic expert (ablation) | MiCRo · MATH / GSM8K | accuracy | Hurt |
| ablate the Social expert (ablation) | MiCRo · math benchmarks | accuracy | Improved |
| ablate the Language expert (ablation) | MiCRo · all benchmarks | accuracy | Hurt |
| apply DPO after MiCRo training versus dense | Llama-3.2-1B · DPO 4-benchmark average | average accuracy (%): 25.4 -> 26.8 | Improved |
| MiCRo with medical instruction tuning versus dense | Llama-3.2-1B · Medical average | average accuracy (%): 41.9 -> 42.2 | Improved |
| use MoE (FFN-only experts) instead of mixture-of-blocks (ablation) | MiCRo · specialization emergence | routing specialization | Hurt |
Closest papers by meaning
- Cognitive Mirrors: Exploring the Diverse Functional Roles of Attention Heads in LLM Reasoning (NeurIPS 2025 · similarity 0.89)
- Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.88)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.88)
- Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces (ICLR 2025 · similarity 0.87)
- When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners (NeurIPS 2025 · similarity 0.87)
- Two Experts Are All You Need for Steering Thinking: Reinforcing Cognitive Effort in MoE Reasoning Models Without Additional Training (NeurIPS 2025 · similarity 0.87)
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.
