StateTune: Transforming LLM-Assisted EDA Flow Tuning into a Stateful, Closed-Loop Process

Kunlong Li; Shangshang Yao; Su Zheng; Lingli Wang

cs.ARAug 19, 2026

StateTune: Transforming LLM-Assisted EDA Flow Tuning into a Stateful, Closed-Loop Process

Authors: Kunlong Li, Shangshang Yao, Su Zheng, Lingli Wang

Organizations: Independent Researcher Shanghai, China · Chinese University of Hong Kong Hong Kong, China · Fudan University Shanghai, China

Abstract

EDA flow parameter tuning is critical for quality-of-results~(QoR), yet the parameter space is large, tightly coupled, and full evaluations are prohibitively expensive. Prior LLM-assisted tuners mainly use the LLM as an external proposer with transient working context; we instead present StateTune, which reformulates LLM-assisted EDA tuning as a closed-loop, state-carrying process. Its optimizer state is a typed, evidence-gated persistent optimization memory that is updated by every evaluation and shared between candidate generation and budget allocation. On top of this optimizer state, an expected hypervolume improvement (EHVI)-guided, runtime-aware promotion policy ranks quick-stage candidates by expected Pareto frontier gain per unit of runtime cost. Evaluated on a Cadence industrial flow across six benchmark blocks (two technology nodes $\times$ three designs), against five baselines including LLM+retrieval-augmented generation (RAG) and preference-based Bayesian optimization (BO) tuners, StateTune achieves the strongest final hypervolume on all six benchmark blocks, showing a stable improvement in frontier quality across the full matrix; it also matches or surpasses the strongest baselines on worst negative slack (WNS), area, and power across the same set. Ablation shows persistent memory is the largest contributor: removing it costs 58.5% of the hypervolume. Dedicated analyses of evidence-gating sensitivity, memory poisoning, cross-design transfer, and three-seed reproducibility (CV, $<$ ,7% on five of six blocks) further validate the memory design.

Figures & tables

Figure 1. Overall workflow of StateTune. The knowledge layer (left) builds and updates persistent optimization memory from design inputs and evaluation feedback via Prior Initialization and an Analysis Agent. The decision layer (center) generates candidate configurations through Knowledge-guided Search and ranks them via EHVI-based Promotion; an Evaluation Engine arbitrates the final selection. The execution layer (right) schedules multi-fidelity evaluations through a Scheduling Agent and returns QoR metrics. The persistent optimization memory connects all three layers.

Method	LLM search	RAG	Memory type	Gate	Multi- fid.	Multi- obj.	Cross- design
PTPT ( 7 )	$\circ$	$\circ$	–	$\circ$	$\circ$	$\bullet$	$\circ$
FlowTuner ( 13 )	$\circ$	$\circ$	–	$\circ$	$\bullet$ a	$\circ$	$\circ$
RankTuner ( 21 )	$\circ$	$\circ$	–	$\circ$	$\circ$	$\bullet$	$\circ$
CROP ( 14 )	$\bullet$	$\bullet$ b	Prompt hist.	$\circ$	$\circ$	$\circ$	$\circ$
ORFS-agent ( 8 )	$\bullet$	$\circ$	Prompt hist.	$\circ$	$\circ$	$\bullet$ c	$\circ$
AstroTune ( 19 )	$\odot$ d	$\bullet$	–	$\circ$	$\bullet$	$\bullet$	$\circ$

Table 1. Capability comparison of EDA flow tuning methods.

\bullet

= supported;

\circ

= not supported;

\odot

= partial.

Figure 2. Structure of the persistent optimization memory $\mathcal{M}_{t}$ . Evaluation results, failure diagnoses, and sensitivity analyses feed into six components. The design prior $P_{t}$ conditions generation; reusable rules $R_{t}$ guide failure avoidance; Pareto status $\Pi_{t}$ and the runtime model $T_{t}$ drive EHVI-based promotion.

Component	Content	Role
Prior $P_{t}$	Parameter ranges, tool knowledge, design characteristics	Conditions proposals
Rules $R_{t}$	Evidence-gated hard rules & soft heuristics	Filters proposals
Failure $F_{t}$	Categorized failures with root-cause analysis	Guides repair
Sensitivity $S_{t}$	Parameter–QoR correlations & threshold effects	Prioritizes knobs
History $H_{t}$	All configs, QoR, terminal status	LLM context & GP fit
Pareto $\Pi_{t}$	Current non-dominated frontier	EHVI targets

Table 2. Persistent memory components.

Type	Content	Evidence
Hard rule	Ban CORE_UTIL=0.78 with CORE_MARGIN=0 : catastrophic WNS ( $-32$ k) from placement congestion.	Multiple failures
Soft heur.	Increase CORE_UTIL & PLACE_MAX_DENSITY ; keep CORE_MARGIN low. WNS corr. +0.24.	Sensitivity
Sensitivity	CTS_POST_OPT corr. $-$ 0.19 with WNS; its heuristics degrade setup timing.	365 runs
Prior upd.	ASPECT_RATIO : 1.2 $\to$ 1.0 (square). Rectangular shapes worse for bus-heavy datapath.	12 full-valid obs.

Table 3. Knowledge artifacts automatically generated by StateTune.

Benchmarks & Flow
Flow	Cadence Genus 21.17 / Innovus 21.18
Benchmarks	JPEG / AES / IBEX on both ASAP7 ( 4 ) (7 nm) and NanGate45 ( 18 ) (45 nm)
Objectives	Maximize WNS; minimize area & power
Fidelities	Quick = CTS (configurable); Full = route
Parameters	19 across 4 stages (Table 5 )
Budget & Hardware

Table 4. Experimental setup and LLM overhead.

Stg	Parameter	Type	Range
FP	CORE_UTIL	C	$[0.55,\,0.78]$
	ASPECT_RATIO	C	$[0.80,\,1.20]$
	CORE_MARGIN	Int	$[0,\,16]$
PL	PLACE_CONG_EFFORT	Cat	low/med/high
	PLACE_TIMING_EFFORT	Cat	med/high
	PLACE_UNIFORM_DENSITY	Cat	true/false

Table 5. Tuned parameters. C=continuous; Cat=categorical; B=binary.

ASAP Benchmarks							NAN Benchmarks
Design	Method	WNS	Area	Power	GP Hits	HV	Design	Method	WNS	Area	Power	GP Hits	HV
ASAP-JPEG	Random	-26.1420	6042	29.9512	0	1.806e5	NAN-JPEG	Random	-0.0080	106638	111.9460	0	1.911e6
	Optuna-TPE	-13.3730	5973	29.9231	2	3.412e5		Optuna-TPE	0.0130	106204	111.3026	1	2.082e6
	BO (qEHVI)	-26.1130	6025	29.7187	1	2.348e5		BO (qEHVI)	-0.0050	106163	111.6088	0	2.064e6
	RankTuner	-29.5560	6256	33.7636	0	9.938e4		RankTuner	0.0090	108508	126.8801	0	7.849e5
	CROP	-10.8240	6924	33.5890	1	1.286e5		CROP	0.0120	106224	111.6372	0	2.030e6

Table 6. Results across the

2{\times}3

benchmark matrix under the shared budget caps. Six methods are compared: BO (qEHVI), Optuna-TPE, Random, RankTuner (ICCAD’24) ( 21 ) , CROP (ICCAD’25) ( 14 ) , and StateTune. Bold = best per benchmark (per column).

Figure 3. Pareto frontier projections for the ASAP-JPEG case ( $-$ Area vs. WNS, $-$ Area vs. $-$ Power, WNS vs. $-$ Power). StateTune (purple triangles) occupies the strongest region among the compared methods.

Figure 4. Anytime hypervolume trajectory for ASAP7-AES, averaged over three independent runs per method with shaded $\pm 1$ standard-deviation bands. StateTune (cyan) separates from all baselines early and maintains a clear advantage throughout the run.

ASAP7			NAN
Benchmark	HV ( $\mu\pm\sigma$ )	CV (%)	Benchmark	HV ( $\mu\pm\sigma$ )	CV (%)
JPEG	$5.33e5\pm 8.21e4$	15.4	JPEG	$2.53e6\pm 1.29e4$	0.51
AES	$3.73e6\pm 3.32e4$	0.89	AES	$4.42e3\pm 6.34e1$	1.43
IBEX	$4.57e5\pm 2.86e4$	6.27	IBEX	$1.66e3\pm 9.00e1$	5.42

Table 7. StateTune reproducibility over three independent runs.

Variant	Final HV (M)	HV Drop (%)	Pareto Hits	First Hit (h)
Full system	10.10	—	6	10.29
w/o Persistence	4.19	58.5	0	–
w/o KA	4.57	54.8	2	2.10
w/o RAG	5.44	46.1	0	–
w/o CoT	6.92	31.5	1	29.62
w/o EHVI	8.32	17.6	4	12.93

Table 8. Ablation on ASAP7 JPEG under 48 thread

\cdot

h budget. Variants ordered by degradation severity.

Panel of Figure 5 — Figure 5. Additional validation on ASAP7 JPEG under 48 thread $\cdot$ h. (a) Final HV and global Pareto hit rate peak at $k{=}12$ . (b) Transferred memory converges faster and reaches higher final HV than no transfer. Thin curves denote individual runs, thick curves denote means, and shaded bands indicate $\pm 1$ standard deviation.

Electronic Design Automation

Electronic Design Automation uses algorithms and software to design, verify, optimize, and manufacture electronic circuits and systems.

See all 215 papers

ATTUNE: Agentic Throughput Tuning for LLM Training
Oct 13, 2026Leo Fan, Hongyang Wang, Wenbo Chen +10Efficient Language Model Training Deep Learning Optimization
Amazon
LLMs in Digital EDA: A perspective on shifting roles from Generation to Orchestration
Aug 27, 2026Matthew Youngman, Cristian Sestito, Themis ProdromakisElectronic Design Automation LLM Agent Orchestration
Centre for Electronics Frontiers, Institute for Integrated Micro and Nano Systems, School of Engineering, The University of Edinburgh, UK
LLM-Guided Runtime Parameter Optimization for Energy-Efficient Model Inference
Apr 29, 2026Katelyn Crumpacker, Dimitrios NikolopoulosLLM Inference Efficiency Energy-Efficient ML
Virginia Polytechnic and State University

Oct 13, 2026J/K move · Enter open · S save

ATTUNE: Agentic Throughput Tuning for LLM Training

Leo Fan, Hongyang Wang, Wenbo Chen, Jingguo Liu+9

Amazon

Efficient Language Model Training Deep Learning Optimization Hyperparameter Optimization Large Language Model-Guided Optimization

Open paper AskPDF arXiv

The throughput of an LLM training job depends on its recipe: how the model is split across GPUs, how memory is traded for compute, and which kernels and communication overlaps it uses. The best recipe changes with every model, accelerator and framework version. Automated tuners search a fixed space of knobs and check only that a job runs, so they miss levers outside that space and can ship a configuration that silently changes training. ATTUNE is a coding agent that tunes any working recipe unattended within a GPU-hour budget. Guided by a skill of search rules, scripts and lessons from past tuning campaigns, it diagnoses each run and changes one lever at a time. A correctness gate outside the agent refuses any argument known to change the training objective. ATTUNE tunes Amazon's production post-training recipes on H100, H200 and B200 clusters, raising throughput by 1.7x on average. It beats the published throughput of four of NVIDIA's own benchmark recipes on NVIDIA's AutoModel framework by 6-17%. On models held out from its development, it improves the strongest public recipes most where they were written for another setting, by up to 2.6x on MiniMax-M2, and least where the vendor tuned them for the same hardware. Given the same GPU-hour budget, it outperforms existing automated tuners, including Bayesian optimization, random search and OptiCo, a recent LLM-agent tuner. Its gain on GLM-4.5 is more than four times the best alternative's. The same coding agent without ATTUNE's skill gains less and fails half its launches. We release ATTUNE at https://github.com/amazon-science/attune and OpenRecipes, a leaderboard of tuned and verified training recipes, at https://openrecipes.ai.

Abstract

Electronic Design Automation

Explore similar work

ATTUNE: Agentic Throughput Tuning for LLM Training

LLMs in Digital EDA: A perspective on shifting roles from Generation to Orchestration

LLM-Guided Runtime Parameter Optimization for Energy-Efficient Model Inference