§ 2.0Core Concepts

IterCAD is governed by a "Look and Loop" philosophy. Look leverages multi-view engineering drawings with dimensional constraints as a persistent reference to isolate defects. Loop establishes a multi-dimensional refinement mechanism that integrates compiler, execution, and visual feedback rather than blind regeneration. [click a term to focus it]

§ 3.0The Two-Stage Training Pipeline

A static, open-loop policy can't learn to debug itself from imitation alone. IterCAD bootstraps procedural reasoning first, then hands the policy to closed-loop reinforcement learning to teach genuine self-correction. [click a stage for its measured outcome]

Stage 1 Imitation

Progressive
Cold-Start SFT

20K expert trajectories establish the basic mapping from engineering specifications to executable code. A further 8K on-policy trajectories, corrected by a teacher model against the base policy's own execution failures, instill iterative debugging behavior before RL ever begins.

⇄refine

Stage 2 Reinforcement

GSPO +
Geometry Rewards

Sequence-level policy optimization with group-relative advantages and a length-normalized importance ratio, restricted to 2K hard Drawing-to-Code samples the task that most rewards closed-loop geometric feedback over imitation alone.

{{ stageReadoutLabel }}{{ stageReadoutBody }}

Reward & advantage GSPO with GVPM R(y) = R_CD(y) + λf · R_fmt(y) + λp · R_prog(y) // λf = 0.5, λp = 0.2 geometric fidelity, format compliance, multi-turn progress Â_g = (r_g − mean({r_g'})) / std({r_g'}) // group size G = 8, length-normalized importance ratio prevents long trajectories from dominating f = min(f_exec, f_stall) // GVPM masks tokens past the viable-prefix boundary; masked trajectories clamp Â_g ≥ 0

§ 4.0Benchmark Results

Agentic Workflow on IterCAD-Draw closed-loop sandbox feedback, up to 5 turns. IterCAD (built on a 4B backbone) converts the multi-turn budget into accuracy that an order-of-magnitude-larger competitor can't match. [click any column header to sort]

IterCAD invalid ratio vs 62.30% backbone baseline

AUC-TR best of all evaluated models

Mean Chamfer Distance ×10³ lowest in field

Average turns most parts solved in a single pass

IterCAD-Draw Agentic Workflow (closed-loop sandbox)

Model	IR%	AUC-TR	Mean CD	Med. CD	Avg Turn
GPT-5	4.70	0.50	12.18	1.07	2.44
Gemini-3-flash-lite	12.70	0.56	5.79	0.09	3.12
GLM-4.6v	9.70	0.43	13.72	3.27	2.59
Qwen3.5-35B-A3B	37.50	0.36	12.51	0.10	3.50
Qwen3.5-4B (backbone)	62.30	0.21	13.04	0.09	4.39
IterCAD (ours)	0.30	0.61	5.09	0.35	2.48

IterCAD-Edit instruction-based local editing

Model	IR%	AUC-TR	Mean CD	Med. CD	Avg Turn
GPT-5	0.50	0.79	2.14	0.05	2.31
Gemini-3-flash-lite	11.00	0.51	4.27	0.06	4.23
Qwen3.5-35B-A3B	11.00	0.55	7.49	0.14	3.84
Qwen3.5-4B (backbone)	63.00	0.18	9.96	1.25	4.49
IterCAD (ours)	1.00	0.54	7.52	0.77	2.34

§ 5.0What Each Stage Buys

Starting from the bare Qwen3.5-4B backbone, every added component compounds. The full recipe SFT, on-policy refinement, GSPO, and GVPM is what closes the gap from a 62.30% invalid ratio to 0.30%. [click a step to expand it]

Cold-Start SFT on Expert Trajectories+

The first imitation pass alone drops invalid ratio from 62.30% to 7.50% and lifts AUC-TR from 0.21 to 0.52 establishing the basic specification-to-program mapping before any closed-loop signal is introduced.

On-Policy Refinement Data+

Adding teacher-corrected trajectories that target the base policy's own execution failures pushes invalid ratio down to 0.80% proof that imitation alone, without geometry-aware RL, still leaves Mean CD high at 12.44.

GSPO Reinforcement Learning+

Closed-loop reward optimization lowers Mean CD to 8.00 and lifts AUC-TR to 0.58 but without further correction, average turns collapse to 2.18 as the policy learns premature termination is locally rewarding.

Geometry-Viable Prefix Masking+

Masking gradient contributions from stalled or error-cascading suffixes fixes the credit-assignment problem directly: invalid ratio falls to 0.30%, AUC-TR reaches 0.61, Mean CD drops sharply to 5.09 and the agent persists through productive multi-turn correction instead of bailing early.

§ 6.0What IterCAD Unlocks

Open-loop, one-shot CAD generation can't recover from its own mistakes by definition. IterCAD's closed-loop architecture generalizes well beyond its own benchmark confirming the gains aren't an artifact of one evaluation protocol.

Drawings as Spatial Anchors+

Multi-view engineering drawings with dimensional constraints persist across every turn as a physical reference, letting the agent isolate which specific feature is wrong rather than regenerating the whole part from scratch.

One Agent, Three Modalities+

A single unified session reconstructs a base plate from a drawing, then applies sequential natural-language edits increase thickness, add fillets, undo an operation, add a chamfered boss while preserving unaffected geometry throughout.

Generalizes Beyond Its Own Benchmark+

On the external Text2CAD benchmark, IterCAD posts a 0.64% invalid ratio against general-purpose models exceeding 50%. On CADPrompt, it beats the domain-specific CAD-Judge baseline by an order of magnitude in median Chamfer Distance.

~/conclusion

$ query: what does IterCAD change // CAD generation stops being open-loop and one-shot. // the drawing becomes a persistent reference the agent checks itself against. $ query: what does this cost // two training stages instead of one, plus a sandbox the agent can fail inside safely. // GVPM exists specifically so failure in turn 4 doesn't punish a correct turn 1. $ query: what is the actual result // 0.30% invalid ratio, AUC-TR 0.61, 2.48 average turns. // a 4B-parameter model out-debugging baselines an order of magnitude larger.

CAD GENERATION
ISN'T
ONE-SHOT.
IT'S A
LOOP.

CLOSING THE LOOP ON CAD GENERATION

CLOSING
THE LOOP
ON CAD
GENERATION