Skip to content

HeaRT: A Hierarchical Circuit Reasoning Tree-Based Agentic Framework for AMS Design Optimization

🕒 Published (v1): 2025-11-24 20:11 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

HeaRT is a two-phase agentic framework for analog/mixed-signal (AMS) circuit design that builds a hierarchical circuit reasoning tree offline from raw SPICE netlists, then uses query-conditioned tree traversal online to ground LLM reasoning for sizing and topology optimization. It achieves \(\geq13.5\%\) \(F_{1,\text{subcircuits}}\) and \(\geq37.8\%\) \(F_{1,\text{loops}}\) gains over few-shot baselines across five LLM backbones, and \(\geq3\times\) faster convergence on incremental design adaptation tasks.

Problem

Existing AMS automation approaches — Bayesian optimization, RL, and vanilla LLM prompting — share three compounding failures: (1) no hierarchical structural reasoning, causing hallucinations and inconsistent circuit decomposition; (2) no mechanism to distinguish reusable subcircuits from those needing redesign under specification shifts, forcing full re-optimization from scratch ("catastrophic forgetting"); and (3) poor transferability across circuit families due to reliance on task-specific training data or unstructured prompting.

Method

HeaRT operates in two phases:

Offline (one-time): A raw SPICE netlist is converted to a device–net bipartite graph \(G = (V_D, V_N, E)\). DC-connected components are extracted via multi-source BFS from supply rails to identify structurally active devices. Residual AC-connected components are handled separately. An LLM (Claude-4.6-Sonnet) uses CoT + in-context examples to decompose these components into functional subcircuits \(S_{\text{decomp}}\) with metadata. A second LLM pass performs bottom-up knowledge consolidation — reasoning leaf-to-root, identifying local feedback loops at each node, and storing annotated netlists at every abstraction level — producing a hierarchical circuit reasoning tree in structured JSON.

Online (agentic retrieval): Given a user query \(Q\), a single LLM pass assigns query-conditioned relevance weights to all tree edges. BFS from root enqueues children only if the Branch Cut criterion (Eq. 4) is not met — halting when children are uniformly weak (\(\max w < \tau_{\text{stop}}\)) or uniformly strong (\(\max w - \min w < \epsilon\)). The resulting priority regions guide downstream tasks. For topology optimization, candidate subcircuits are matched to a topology knowledge database (organized by circuit family, ranked by metric-wise performance using ISSCC/JSSC grounding) via Weisfeiler–Lehman graph fingerprints, and ranked by an aggregate \(\text{FoM}_{\text{rank}}\) (Eq. 5) subject to constraint-set filtering (Eq. 6).

Design effort reuse under specification shifts is quantified by \(R_{\text{effort}} = \frac{1}{M}\sum_{j=1}^{M}\left[1 - \min\!\left(1,\log_{10}(x_j/x_j^{(0)})\right)\right] \in [0,1]\).

Key Contributions

  • Offline/online two-phase harness separating expensive tree construction from lightweight real-time retrieval.
  • Graph-guided functional subcircuit decomposition algorithm (Algorithm 1) operating directly on flattened SPICE without handcrafted templates.
  • Bottom-up knowledge consolidation producing a multi-level JSON reasoning tree with annotated netlists at each node.
  • Query-conditioned BFS traversal with branch-cut pruning for token-efficient, focused retrieval.
  • Topology knowledge database with metric-wise rankings (WL-fingerprint indexed) enabling objective-driven topology selection.
  • \(R_{\text{effort}}\) metric for quantifying design effort preservation across specification shifts.
  • 40-circuit AMS benchmark spanning Simple/Medium/Hard complexity tiers with expert-annotated ground truth (to be open-sourced).

Results

  • \(F_{1,\text{subcircuits}}\): HeaRT with Claude-4.6-Sonnet reaches 0.997/0.988/0.955 (Simple/Medium/Hard) vs. 0.762/0.709/0.684 for few-shot text baseline — \(\geq13.5\%\) improvement across all backbones.
  • \(F_{1,\text{loops}}\): HeaRT with Claude-4.6-Sonnet reaches 0.984/0.962/0.937 vs. 0.741/0.675/0.614 — \(\geq37.8\%\) improvement across all backbones.
  • Weakest backbone (LLaMA-3.3-70B): HeaRT still outperforms few-shot text (e.g., \(F_{1,\text{subcircuits}}\): 0.724 vs. 0.571 on Simple).
  • Incremental design adaptation: \(\geq3\times\) faster convergence on both topology reconfiguration and sizing scenarios within a 200-simulation budget.
  • Schematic image input provides marginal gains over text-only on simpler circuits but underperforms HeaRT at all complexity tiers.

Limitations

  • Offline tree construction uses Claude-4.6-Sonnet exclusively — cost and latency of this one-time phase are not reported.
  • Benchmark is limited to 40 circuits; generalization to industrial-scale designs with hundreds of subcircuits is undemonstrated.
  • \(R_{\text{effort}}\) evaluation only reported for fixed-architecture (sizing-only) scenario, not topology reconfiguration.
  • Branch Cut hyperparameters (\(\tau_{\text{stop}}\), \(\epsilon\)) are not ablated; sensitivity is unknown.
  • Evaluation uses TSMC 180nm only; portability to advanced nodes not demonstrated despite claiming commercial PDK support.
  • Paper text is truncated; full quantitative results for downstream optimization FoM are not visible.

Relevance to Harnesses / Meta-Harnesses

HeaRT exemplifies the offline-build / online-execute harness pattern: expensive structured knowledge extraction runs once to build a reusable artifact (the reasoning tree), while lightweight agentic inference uses that artifact at query time — exactly the separation of concerns central to meta-harness design. The topology knowledge database with LLM-maintained rankings is a domain-specific knowledge store that functions as a persistent memory layer, analogous to retrieval stores in RAG-style harnesses. The branch-cut traversal policy is a harness-level routing mechanism that controls which LLM calls are made and which subtrees are skipped, demonstrating how a harness can enforce token efficiency without exposing that logic to the individual agent. For researchers building harnesses over structured technical artifacts (code, specs, configurations), HeaRT's graph-to-tree pipeline offers a concrete template for transforming flat structured text into hierarchical, query-indexable context.