Researchers have developed SLIC, a reinforcement fine-tuning framework for small language models (4B-scale) designed to optimize analog integrated circuit design. This framework enables local deployment, avoiding the need to send proprietary data to external LLM services. SLIC demonstrated strong performance on a benchmark of amplifier topologies, achieving 87.1% in-distribution Pass@1 and outperforming the larger DeepSeek-V4.1-Flash model. The system also showed promising results in transferring learned policies to unseen technology nodes and commercial PDKs. AI
IMPACT Enables local deployment of LLMs for specialized engineering tasks, potentially accelerating adoption in industries with sensitive data.
RANK_REASON The cluster describes a new research framework and benchmark for fine-tuning small language models for a specific engineering task, published on arXiv.
- CMOS
- language model
- TSMC
- Reward-Guided Rollout Steering
- SLIC-Random
- SLIC-Softmax
- Small LM Iterating on Circuits
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →