Waveform debug is where hardware verification cycles go to die. A senior RTL engineer opens GTKWave, loads a VCD, and manually correlates signal edges across 50 wires to find the root cause of a one-cycle handshake violation. WaveformQA, published at ICLAD 2026, is the first systematic benchmark of whether LLMs can take that workload. The headline finding is not impressive: frontier models do well on simple single-signal queries and fall apart on complex temporal or multi-step questions. But the failure mode is worth reading carefully.
The binding constraint is not reasoning quality. It is context window. VCD files for non-trivial simulations are enormous; the LLM runs out of window before it can correlate events across the waveform. The paper's most actionable result is that switching from raw VCD to an event-time JSON representation (a compact encoding that surfaces only state changes with timestamps) substantially improves accuracy on complex queries. That is not a model capability finding. It is a representation finding: the bottleneck is information density, and the right format can move the constraint from "model can't reason over this" to "model can reason over a larger slice of the waveform per token budget."
The 360 questions span eight categories including multi-signal correlation and event ordering, with ground truths generated programmatically from open-source implementations. That grounding in real hardware behavior matters. The benchmark does not measure LLM performance on toy HDL; it measures performance on the kind of waveform data teams actually deal with. The gap this quantifies is not a 6-month gap. Multi-signal temporal correlation in long simulation traces is exactly the class of problem where context window limits and positional encoding degradation interact badly. LLM-native waveform debug is 2-3 years out at current trajectory. The near-term payoff is in tools that pre-process VCDs into event-time JSON before feeding them to the model, which any team can build today.