AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

SLIC: Reinforcement Fine-Tuning Small LMs for Multi-Turn Analog Circuit Optimization

Yan Xu, Zhengqi Gao, Ching-Yun Ko, Zhang-Wei Hong, Wenjie Lu, Tao Yu, Xin Zhang, Duane Boning, Ruonan Han

NeurIPS 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SLIC, a reinforcement fine-tuning framework that trains a locally deployable 7B-scale small LM to iteratively optimize analog circuits in a simulator-grounded multi-turn loop.

Problem

Analog IC design is a multi-turn, simulator-grounded task requiring iterative reasoning over coupled performance trade-offs, but proprietary PDKs make hosted LLM workflows hard to deploy.

Approach

Reinforcement fine-tuning of a locally deployable 7B-scale language model for multi-turn analog circuit optimization, so that the model can run on-premises without sending proprietary design data to hosted LLMs.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add domain-validity reward gates (ablation)SLIC-Random · ID, Turn 8Pass@1 (%): 82.8 -> 87.1Improved
add domain-validity reward gates (ablation)SLIC-Random · held-out, Turn 8Pass@1 (%): 63.6 -> 70.6Improved
add success-conditional slack bonus (ablation)SLIC-Random + filter · ID, Turn 8Pass@1 (%): 76.4 -> 84.9Improved
add success-conditional slack bonus (ablation)SLIC-Random + filter · held-out, Turn 8Pass@1 (%): 56.7 -> 68.8Improved
normalize validity and performance advantages separately (ablation)SLIC-Random · ID, Turn 8Pass@1 (%): 79.0 -> 85.2Improved
use random or softmax continuation instead of greedy (ablation)SLIC RGRS · held-out, Turn 8Pass@1 (%): 43.9 -> 70.6 (Random), 72.6 (Softmax)Improved
use softmax instead of random continuation (ablation)SLIC RGRS · held-out, Turn 8Pass@1 (%): 70.6 -> 72.6No clear effect
use online RGRS construction instead of fixed-base TL-GRPO (ablation)SLIC-Random + filter · ID and held-outPass@1 (%): ID 73.8 -> 76.4; OOD 57.6 -> 56.7Mixed
train multi-turn (RGRS) instead of single-turn GRPO (ablation)SLIC · ID, Turn 8Pass@1 (%): 61.6 -> 87.1Improved
use RGRS instead of Tree MT-GRPOSLIC-Random · held-out, Turn 8Pass@1 (%): 58.3 -> 70.6Improved
train with four-turn horizon instead of eight (ablation)SLIC-Random · ID and held-out, Turn 8Pass@1 (%): ID 87.1 -> 89.5; OOD 70.6 -> 69.3No clear effect
keep raw model outputs in history instead of structured history (ablation)SLIC-Random · ID, Turn 8Pass@1 (%): 87.1 -> 76.8Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy