Research map / Large language models / Agents, code and math
SLM-MUX: Orchestrating Small Language Models for Reasoning
Chenyu Wang, Zishen Wan, Hao Kang, Emma Chen, Zhiqiang Xie, Tushar Krishna, Vijay Janapa Reddi, Yilun Du
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 SLM-MUX (confidence-based selection, no discussion) instead of discussion-based orchestration | SLM orchestration · MATH (Llama 3.1 8B, Mistral 8x7B, Gemma 2 27B) | accuracy (%): 51.6 -> 61.8 (vs LLM-Debate) | Improved |
| use SLM-MUX instead of Mixture-of-Agents / Debate / Verification | SLM orchestration · GPQA | accuracy (%): 36.8 -> 42.1 (vs LLM-Debate) | Improved |
| discussion-based orchestration (MoA, Debate, Verification) with SLMs | Llama 3.1 8B, Mistral 8x7B, Gemma 2 27B · MATH | accuracy (%): best single 56.8 -> 48.4 to 51.6 | Hurt |
| use SLM-MUX instead of self-consistency of best single model | SLM-MUX · GPQA | accuracy (%): 42.4 -> 42.1 | No clear effect |
| model selection search for complementary pairs (ablation) | SLM-MUX · MATH | accuracy (%): 75.5 -> 80.0 | Improved |
| use SLM-MUX over Agent Forest (2 samples) | SLM-MUX · GPQA | accuracy (%): 40.4 -> 46.3 | Improved |
| use SLM-MUX over Agent Forest at best sample budget | SLM-MUX · MATH | accuracy (%): 79.2 -> 79.5 | No clear effect |
| draw more samples per model (2 to 9) (ablation) | SLM-MUX · MATH, GPQA, GSM8K | accuracy (%) | Improved |
| add more participating models beyond 2 (ablation) | SLM-MUX · GPQA | accuracy (%): GPQA test: 47.98 (K=1), 50.21 (K=2), 51.81 (K=3), 45.18 (K=5) | Mixed |
| use random model selection instead of search (ablation) | SLM-MUX · MATH-500 (K=2) | accuracy (%): 66.1 -> 67.2 | Improved |
| tune the Mixture-of-Agents aggregator prompt (ablation) | Mixture-of-Agents · MATH | accuracy (%): 46.2 -> 48.4 (best tuned) | No clear effect |
| apply SLM-MUX to frontier LLMs (GPT-4o + Gemini-2.5-Flash) | SLM-MUX · GPQA | accuracy (%): 51.1 -> 60.1 | Improved |
Closest papers by meaning
- It Helps to Take a Second Opinion: Teaching Smaller LLMs To Deliberate Mutually via Selective Rationale Optimisation (ICLR 2025 · similarity 0.89)
- Select to Think: Unlocking SLM Potential with Local Sufficiency (ICML 2026 · similarity 0.88)
- Reasoning-as-Logic-Units: Scaling Test-Time Reasoning in Large Language Models Through Logic Unit Alignment (ICML 2025 · similarity 0.88)
- Mutual Reasoning Makes Smaller LLMs Stronger Problem-Solver (ICLR 2025 · similarity 0.88)
- GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models (ICLR 2025 · similarity 0.87)
- SGLang: Efficient Execution of Structured Language Model Programs (NeurIPS 2024 · 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.
