---
title: "SLIC: Reinforcement Fine-Tuning Small... | SI Paper Desk"
url: https://thesimind.si/papers/2611.13452/
description: "A small 4B-scale SI model, trained with reinforcement fine-tuning and simulator feedback, beats a much larger hosted model at tuning analog circuits and can..."
---

Number 9 of 12 · November 12, 2026

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

A small 4B-scale SI model, trained with reinforcement fine-tuning and simulator feedback, beats a much larger hosted model at tuning analog circuits and can run locally on proprietary chip data.

Yan Xu, Zhengqi Gao, Ching-Yun Ko and 6 more · Submitted November 10, 2026 · arXiv 2611.13452v1 · cs.LG

[Read the paper on arXiv](https://arxiv.org/abs/2611.13452) · [The PDF](https://arxiv.org/pdf/2611.13452) · [The day's 12](https://thesimind.si/papers/day/2026-11-12/)

## What they found

The authors built SLIC, a way to train a 4B-scale language model to improve analog circuits over several turns, using feedback from a circuit simulator. Trained only on open source 65 nm CMOS models, SLIC-Random reaches 87.1% Pass@1 on four specifications within eight turns. That is 48.5 percentage points above the 763B DeepSeek-V4.1-Flash Think. On held-out circuit layouts, SLIC-Softmax reaches 72.6% Pass@1. The trained policies also carry over to five unseen technology nodes without more fine-tuning. On a commercial TSMC 65 nm CMOS PDK, SLIC-Random gets 49.9% on held-out topologies. With a 32-turn budget, held-out Pass@1 rises to as much as 83.2%. The authors also release an open benchmark of 41 amplifier topologies.

## Why it matters

Chip makers often cannot send proprietary design kits to hosted SI services. A small model that runs locally avoids that problem. The result suggests that a small model with domain-specific training and simulator feedback can beat a far larger general model on a narrow engineering task. The open benchmark gives others a shared way to test similar work. The abstract does not say how the results compare with human designers or with other circuit tools.

## How it changes the race

It suggests that small, locally run SI models trained on simulator feedback can compete with much larger hosted models in specialized engineering fields. That could lower the cost and the data-sharing barriers for firms working with private design data. It is a result for one narrow domain, not a general capability jump.

## Safety and control

The abstract does not discuss safety or alignment. It does note a privacy benefit: local inference means proprietary design data need not be sent to outside services.

### The abstract, by the authors

Analog integrated circuit (IC) design is a multi-turn, simulator-grounded engineering task that demands iterative reasoning over coupled performance tradeoffs, yet proprietary process design kits (PDKs) make hosted large language model workflows difficult to deploy in practical design settings. We present SLIC (Small LM Iterating on Circuits), a reinforcement fine-tuning framework that trains a 4B-scale, locally deployable language model for closed-loop analog circuit optimization. SLIC pairs structured circuit representations with discretized relative-update actions, domain- and success-gated multiobjective rewards, and Reward-Guided Rollout Steering (RGRS), which turns per-step simulator feedback into signals for both policy learning and online rollout construction in this sparse, highly constrained domain. We further release an open, goal-conditioned benchmark of 41 amplifier topologies with standardized protocols for in-distribution validation, held-out topology evaluation, and cross-technology-node transfer. Trained solely with open source 65 nm CMOS predictive technology models (PTM), SLIC-Random achieves 87.1% in-distribution four-specification Pass@1 within eight turns, surpassing the 763B DeepSeek-V4.1-Flash Think by 48.5 percentage points. Beyond the training distribution, SLIC-Softmax reaches 72.6% Pass@1 on held-out topologies, and the trained policies consistently transfer across five unseen technology nodes without additional fine-tuning. This transfer extends to a commercial TSMC 65 nm CMOS PDK, where SLIC-Random achieves 49.9% held-out-topology Pass@1, demonstrating practical deployment through local inference without transmitting proprietary data to external LLM services. Performance further scales with test-time simulator interactions, reaching up to 83.2% held-out-topology Pass@1 at a 32-turn budget.

Quoted from arXiv:2611.13452. The research and its words belong to its authors.

Desk score **74** out of 100

Why it was picked

- Interpretability (in the title)
- Reasoning
- Evals
- Studies frontier models (DeepSeek-V4)
- Accepted at NeurIPS
- Language models

Topics

- [Reinforcement learning](https://thesimind.si/papers/?tag=reinforcement-learning)
- [Language models](https://thesimind.si/papers/?tag=language-models)
- [Agents](https://thesimind.si/papers/?tag=agents)
- [Efficiency](https://thesimind.si/papers/?tag=efficiency)
- [Benchmarks](https://thesimind.si/papers/?tag=benchmarks)
- [Privacy](https://thesimind.si/papers/?tag=privacy)

**arXiv ID:** 2611.13452v1
**Submitted:** November 10, 2026
**Fields:** cs.LG (Machine Learning)
**Level:** Intermediate
**Comments:** Accepted at NeurIPS 2026. 43 pages, 12 figures. Project page: [https://hangroup.mit.edu/slic/](https://hangroup.mit.edu/slic/)

Fingerprint, SHA-256

`eda450675ce24150f8f24ae77b92593f86951611650aa183af7413af368998d8`

On Solana: [Proof of Mind seal, 66zHq3...ugfh](https://solscan.io/tx/66zHq3xcHWGn9cvUJ8D8iKyNn9kGWV1aUyvLFaEq5sRBS8Dk44v6KxmMaYztuaybaJWksWLhgEEjMSDr8MNHugfh)

### The exact text fingerprinted

{"arxiv":"2611.13452","version":1,"title":"SLIC: Reinforcement Fine-Tuning Small LMs for Multi-Turn Analog Circuit Optimization","authors":["Yan Xu","Zhengqi Gao","Ching-Yun Ko","Zhang-Wei Hong","Wenjie Lu","Tao Yu","Xin Zhang","Duane S. Boning","Ruonan Han"],"day":"2026-11-12","rank":9,"score":74,"status":"written","summary":{"takeaway":"A small 4B-scale SI model, trained with reinforcement fine-tuning and simulator feedback, beats a much larger hosted model at tuning analog circuits and can run locally on proprietary chip data.","found":"The authors built SLIC, a way to train a 4B-scale language model to improve analog circuits over several turns, using feedback from a circuit simulator. Trained only on open source 65 nm CMOS models, SLIC-Random reaches 87.1% Pass@1 on four specifications within eight turns. That is 48.5 percentage points above the 763B DeepSeek-V4.1-Flash Think. On held-out circuit layouts, SLIC-Softmax reaches 72.6% Pass@1. The trained policies also carry over to five unseen technology nodes without more fine-tuning. On a commercial TSMC 65 nm CMOS PDK, SLIC-Random gets 49.9% on held-out topologies. With a 32-turn budget, held-out Pass@1 rises to as much as 83.2%. The authors also release an open benchmark of 41 amplifier topologies.","matters":"Chip makers often cannot send proprietary design kits to hosted SI services. A small model that runs locally avoids that problem. The result suggests that a small model with domain-specific training and simulator feedback can beat a far larger general model on a narrow engineering task. The open benchmark gives others a shared way to test similar work. The abstract does not say how the results compare with human designers or with other circuit tools.","race":"It suggests that small, locally run SI models trained on simulator feedback can compete with much larger hosted models in specialized engineering fields. That could lower the cost and the data-sharing barriers for firms working with private design data. It is a result for one narrow domain, not a general capability jump.","safety":"The abstract does not discuss safety or alignment. It does note a privacy benefit: local inference means proprietary design data need not be sent to outside services.","level":"intermediate","tags":["reinforcement-learning","language-models","agents","efficiency","benchmarks","privacy"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

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

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 16, 2026 · Number 4LuthorSQL: A One-Pass Text-to-SQL Model as System 1 for a Slower LLM AgentA small local 7B model can answer nearly half of database questions in one fast pass and hand the hard ones to a big SI agent, cutting API cost by 33.3% at the same accuracy.EfficiencyAgentsLanguage models](https://thesimind.si/papers/2611.17780/) · [Nov 10, 2026 · Number 2Tail-risk-aware Agent Safety: Constraining Tail Risks Across Multi-step LLM Agent WorkflowsThe authors propose TAS, a way to keep SI agents safe by tracking how risk builds across workflow steps and capping rare, severe failures with one shared budget.SafetyAgentsSecurity](https://thesimind.si/papers/2611.10411/) · [Nov 18, 2026 · Number 2Nemotron 3.5 Bolt E4B 4GB: An Open and Efficient Small Language Model for Fast Conversational AgentsNVIDIA's authors post-trained a small open model from Gemma-4-E4B-it for fast local assistants. They report the top average accuracy on 21 benchmarks and faster decoding with a drafter model.EfficiencyLanguage modelsAgents](https://thesimind.si/papers/2611.20771/)

[Number 8](https://thesimind.si/papers/2611.13577/) · [All 12 for Nov 12, 2026](https://thesimind.si/papers/day/2026-11-12/) · [Number 10](https://thesimind.si/papers/2611.13595/)
