PaperlayerView exact revision on arXiv

Source-linked AI summary

SKILL: Self-correcting Knowledge-guided Iterative Large Language Model Agent for Logic Optimization

Rui Yang

arXiv:2608.14579v1cs.AI

TL;DR

Logic synthesis optimization must handle exponentially growing search spaces and diverse logic topologies beyond static expert-authored flows. SKILL combines multi-agent LLM reasoning, PPO-based tool interaction, and self-correction, achieving a 12.4±2.1% improvement in Power-Delay-Area product over expert-designed flows.

  • Problem

    Static expert-authored logic-synthesis flows struggle to generalize across diverse logic topologies and technology nodes, while optimization search spaces scale exponentially.

  • Method

    SKILL combines three specialized LLMs, a PPO agent, hierarchical action decomposition, and PDA-feedback-driven self-correction for logic synthesis optimization.

  • Results

    12.4±2.1% improvement in Power-Delay-Area product over expert-designed flows.

  • Takeaways & Limitations

    The results illustrate the potential of embedding intelligent agents in professional EDA environments for automated logic design.

  • Takeaways & Limitations

    SKILL relies on proprietary EDA interfaces, incompletely covers emerging process nodes, and is restricted to synthesis-level interactions.

Abstract

from arXiv · show

Logic synthesis optimization poses significant challenges due to exponentially growing search spaces, sparse reward signals, and diverse logic structures. Traditional expert-designed flows lack adaptability, while reinforcement learning (RL) methods often suffer from low sample efficiency and limited interpretability. We introduce SKILL, a Self-correcting Knowledge-guided Iterative Large Language Model Agent that unifies multi-agent LLM reasoning and RL-based environment interaction for automated synthesis optimization. SKILL coordinates three specialized LLMs: GPT-4o for strategic planning, Claude Sonnet 4 for detailed reasoning, and Gemini 2.5 Pro for efficient analysis with a PPO-based RL agent that learns actionable policies through direct interaction with synthesis tools. A novel self-correcting module monitors environment feedback (PDA metrics), detects suboptimal behaviors, and invokes LLM-guided recovery strategies. Evaluations on IWLS, OpenCores, and EPFL benchmarks show SKILL achieves a 12.4 % PDA improvement over expert flows and 86.3% success rate on logic systems up to 500K gates.

1 Introduction

SKILL addresses the limited generalization of static logic-synthesis flows and the interaction burden of reinforcement learning by combining multi-agent LLM reasoning with PPO-based environment interaction. Its self-correcting, hierarchical design uses PDA feedback to diagnose suboptimal decisions and guide corrective plans, with evaluation across three benchmark suites up to 500K gates.

  • Motivation: Traditional expert-authored EDA scripts struggle to generalize across diverse logic topologies and technology nodes as optimization search spaces grow exponentially.These flows operate through tools such as ABC and Yosys.
  • Architecture: SKILL combines three specialized LLMs with a PPO agent to bridge high-level reasoning and low-level synthesis actions.LLMs provide strategy, root-cause analysis, and correction planning, while PPO performs fine-grained tool invocation through reward-driven learning.
  • Self-Correction: The self-correcting optimization loop uses environmental PDA feedback to detect suboptimal decisions, diagnose failures, and trigger LLM-based corrective plans.The mechanism re-engages LLMs when optimization trajectories show performance regressions.
  • Action Decomposition: Hierarchical action decomposition uses two abstraction levels to connect LLM guidance with executable tool actions and improve learning efficiency.This decomposition bridges abstract reasoning and the low-level synthesis action space.
  • Evaluation: 500K gates: SKILL is evaluated on IWLS, OpenCores, and EPFL logic systems, showing consistent improvements over expert and RL-only baselines.The benchmarking claim covers logic systems up to 500K gates.

2 Related Work

Prior EDA optimization methods use RL and Bayesian optimization but remain limited by convergence, sparse PDA feedback, and interpretability. Related LLM work shows gains in logic and software tasks, while SKILL combines specialized agents with RL coordination and closed-loop correction.

  • RL-Based EDA Optimization: RL methods formulate logic synthesis as an MDP in which actions modify logic and rewards reflect quality-of-result metrics, but convergence and generalization remain challenging.DRiLLS established this formulation for logic synthesis within EDA optimization.
  • RL-Based EDA Optimization: Hybrid PPO-based models improve sample efficiency and reward propagation, while Bayesian optimization navigates tool-operation sequences; sparse PDA feedback and limited interpretability persist.EasySO introduced hybrid PPO-based models, and BOiLS applied Bayesian optimization to synthesis-operation sequences.
  • LLMs for EDA: 87% correctness was achieved by RTLCoder for RTL generation from language prompts, outperforming earlier models.Other LLM applications include logic comprehension, code-quality improvement, verification coverage, analog topology synthesis, and parameter estimation.
  • LLMs for EDA: Most LLM-based EDA efforts remain limited to one-shot generation or static analysis rather than continuous toolchain feedback loops.Surveys nevertheless identify natural-language interfaces as a way to automate and accelerate EDA stages.
  • Multi-Agent LLM Systems: SKILL assigns strategy planning to GPT-4o, detailed analysis to Claude, and structural refinement to Gemini, coordinating them through an RL agent in a PDA-informed correction loop.This role-based collaboration follows multi-agent systems such as AutoGen and MetaGPT, which improved software-task success rates.

3 SKILL Framework

SKILL formulates logic optimization as an augmented POMDP that integrates multi-agent LLM guidance, hierarchical reinforcement learning, and self-correcting feedback. Its architecture coordinates specialized LLM roles with PPO-based optimization and autonomous failure recovery.

  • SKILL-POMDP: SKILL extends the conventional logic-optimization POMDP M = ⟨S, A, T, R, O, Ω⟩ to MSKILL = ⟨S, A, T, R, O, Ω, G, C, E⟩.G represents LLM guidance, C self-correcting feedback, and E ensemble coordination.
  • Multi-LLM Collaboration: Three specialized LLMs collaborate through consensus: GPT-4o plans strategically, Claude Sonnet 4 performs detailed analysis, and Gemini 2.5 Pro provides efficient real-time feedback.The ensemble combines multimodal planning, critical-path and bottleneck analysis, pattern matching, rapid proposal evaluation, and strategy adaptation.
  • Hierarchical RL: Modified PPO incorporates multi-LLM guidance as enhanced input features and uses a two-level hierarchical action space aligned with strategic recommendations.The strategic level selects timing-, area-, power-, or balanced objectives, while the tactical level executes transformations, technology mapping, and constraint management.
  • Self-Correction: SKILL autonomously detects optimization failures, analyzes root causes with LLM and algorithmic methods, and selects corrective actions without human intervention.The correction pipeline extracts failure context, ranks causes, generates strategies, selects an action, and updates correction memory.

4 Experimental Setup and Evaluation Framework

The evaluation spans three complementary benchmark suites covering academic research prototypes through industrial production applications, with additional industrial-scale systems under realistic technology constraints. Comparisons include traditional optimization, reinforcement learning, LLM-based methods, multi-objective optimization, and evolutionary strategies.

  • Benchmark Suites: Three complementary benchmark suites span the spectrum from academic research prototypes to industrial production applications.The suites collectively provide systematic evaluation across diverse logic-optimization settings.
  • Benchmark Suites: IWLS includes 23 arithmetic units, 15 control designs, 12 DSP blocks, and 8 processor components across diverse complexities and architectures.
  • Benchmark Suites: OpenCores contains 64 industrial designs spanning communication, memory, cryptographic, processor, and peripheral controllers, ranging from 500 to 50,000 gates.
  • Benchmark Suites: EPFL comprises 10 arithmetic functions, 7 random logic structures, 6 industrial designs, and 3 large-scale benchmarks containing >10M gates.
  • Industrial-Scale Evaluation: Industrial-scale evaluation covers CPU cores with 100K-500K gates, GPU compute units, and network processing systems across 7nm, 14nm, and 28nm technology nodes.The systems are evaluated with realistic constraints.
  • Baseline Comparisons: Baselines include expert-crafted scripts, GA, SA, DRiLLS, EasySO, BOiLS, LSO-former, direct LLM applications, multi-objective methods, and evolutionary strategies.

5 Results and Analysis

Results show that SKILL improves PDA and reliability across benchmarks, logic types, and industrial-scale systems. Ablations and contribution analyses further indicate that multi-LLM collaboration, self-correction, and dynamic role specialization are central to its performance.

  • Overall Performance: 12.4±2.1% PDA improvement surpasses traditional expert scripts at 7.3±1.9% and the LSO-former RL baseline at 10.1±2.3%.SKILL also achieves an 86.3% success rate across diverse logic types and complexity levels.
  • Ablation Study: -3.0% performance degradation follows complete self-correcting-system removal, while GPT-4o removal causes the largest individual loss at -1.6% PDA improvement.Single-LLM configurations achieve only 8.9-9.2% improvement, versus a collaborative gain of +3.2-3.5% improvement.
  • Logic-Specific Performance: 13.2% improvement is strongest for arithmetic logic, attributed to effective pattern recognition by the LLM ensemble.The analysis reports consistent performance advantages across all logic types.
  • Industrial-Scale Validation: 79.8% success on 500K-gate systems exceeds LSO-former’s 65.4%, a 22.1% relative reliability improvement at industrial scales.SKILL maintains more graceful degradation across systems ranging from 1K to 500K gates.
  • Self-Correcting System: 81.7% average detection accuracy, 74.9% correction success, and 3.8-step average recovery time characterize the self-correcting mechanism.These results demonstrate practical effectiveness while indicating realistic performance levels across failure categories.
  • LLM Contribution Patterns: 48.3% of strategic planning, 52.6% of detailed analysis, and 47.1% of real-time adjustments are led by GPT-4o, Claude, and Gemini, respectively.The specialization pattern demonstrates dynamic role adaptation while maintaining collaboration across optimization phases.

6 Discussion and Analysis

SKILL’s discussion emphasizes that multi-agent LLM collaboration benefits from rich EDA feedback, while introducing computational overhead. Its limitations include proprietary interfaces, incomplete process-node coverage, and synthesis-level scope, motivating broader integration and advanced search.

  • Experimental insights: 3.2-3.5% improvement from multi-agent collaboration shows LLMs can interpret and respond to complex simulation signals.The approach embeds LLM agents within professional EDA ecosystems for continuous learning from realistic design feedback.
  • Environmental interaction: 15% computational overhead enables agents to access detailed performance metrics during feedback-rich exploration.During the 48.7-hour optimization process, agents interact with thousands of simulation iterations.
  • Limitations: SKILL relies on proprietary EDA interfaces, incompletely covers emerging process nodes, and is restricted to synthesis-level interactions.These limitations indicate potential for broader environmental integration.
  • Future work: Future work could integrate more sophisticated optimization frameworks and advanced search techniques.These directions address the broader environmental integration potential identified by the current limitations.

7 Conclusion

SKILL enhances logic optimization by deeply integrating collaborative, feedback-driven LLM agents with industrial EDA environments. Its results and workflow demonstrate the value of authentic synthesis, timing, and power-analysis tool integration for automated logic design.

  • Results: 12.4±2.1% improvement in Power-Delay-Area product over expert-designed flows demonstrates SKILL’s optimization benefit.The result is attributed to deep integration between LLMs and industrial EDA environments.
  • Mechanism: SKILL’s agents learn from professional synthesis and analysis tools, share simulation observations, and adapt through continuous performance feedback.This environmental interaction supports collaborative intelligence and more sophisticated decisions across complex tool-chain interactions.
  • Implications: Grounding LLM agents in synthesis, timing, and power analysis shows that real-world tool integration can unlock optimization capabilities beyond algorithmic advances alone.Embedding intelligent agents within professional design environments is presented as a paradigm for advancing automated logic design as system complexity grows.
Loading 2608.14579v1…