Ares: Adaptive Reasoning-Effort Steering for PPA- and Cost-Aware RTL Optimization with LLM Agents

Research Question and Approach

Automating the optimization of Register-Transfer-Level (RTL) designs using LLM agents is a growing field, but existing approaches often suffer from inefficient resource allocation. Agents typically use a fixed level of reasoning effort for every iteration, which wastes compute on simple edits or under-thinks complex ones. Furthermore, prior research often attributes performance gains to complex, engineered long-term memory structures without accounting for the dollar cost of the LLM calls required to generate those results. This paper introduces Ares, an agent that optimizes RTL designs by tracking the normalized dollar cost of every LLM call and dynamically steering the reasoning effort based on the agent's progress.

Methodology

Ares introduces three primary innovations. First, it implements a normalized cost metric that allows for fair comparisons of optimization quality (measured by a Figure of Merit, or FoM) against the actual dollar spend. Second, it evaluates the necessity of complex long-term memory by comparing an engineered memory (using structured, deduplicated, and abstracted rules) against a simple concatenation of past experience and a memoryless baseline. Third, it implements an adaptive reasoning-effort policy. This policy uses a patience counter that tracks stalled or failed optimization attempts; when the counter reaches a threshold, the agent escalates to a higher reasoning effort level. The parameters for this counter were fit once on a training set of 21 designs.

The study reveals that the construction of long-term memory matters significantly less than previously assumed; at equal cost, the engineered memory provides no dependable gain over a plain concatenation of experience. In contrast, the adaptive reasoning-effort policy proves highly effective. By allocating deeper reasoning only when progress at lower effort levels stalls, Ares achieves a 23–27% lower FoM on unseen test designs compared to the best fixed-effort strategy at the same cost. Additionally, Ares reaches a 25% deeper FoM than the state-of-the-art Dr. RTL while consuming only 12% of the tokens.