Efficient Parallel Samplers for Recurrent-Depth Models

ICML 2026 Reproduction Challenge • Independent Artifact Evidence Presentation

Paper: h7WBYYJF1Q arXiv: 2510.14961v1 Attempt: 534db42c-5b16-4f00-9a7d-a47056fc9dd4 Mode: Deterministic CPU Audit

Claim 1: Wavefront Sampler

PARTIAL

"The sampler decodes new tokens every forward pass while refining latent states for those tokens in parallel through recurrent depth (Section 3.1)."

Findings: Released sampler AST (recpre/raven_modeling_minimal.py) confirms generate_diffusion_style applies inner_recurrence steps per outer iteration across active states, decodes logits, and appends new tokens. However, decoding occurs after the inner loop per outer step rather than after every individual inner step.

Claim 2: Expressiveness Theorem

UNAVAILABLE

"The paper proves the sampler is strictly more expressive than baseline autoregressive generation under the same time budget on modern hardware (Theorem 4.2)."

Findings: Source TeX audit reveals a citation mismatch: Theorem 4.2 is the prefilling result, while same-runtime decoding expressiveness is Theorem 4.4. No complete independently checkable proof of the decoding theorem is included in the v1 source.

⚡ Wavefront Mechanism Schedule Simulation

Outer Step	Active Positions Before	Recurrence Updates	Decoded Position	Candidate Positions	Retained Positions	Active Width / Bound
Step 1	`[0]`	+4 updates	Position 0	`[1]`	`[1]`	2 / 8
Step 2	`[0, 1]`	+4 updates	Position 1	`[2]`	`[2]`	3 / 8
Step 3	`[0, 1, 2]`	+4 updates	Position 2	`[3]`	`[3]`	4 / 8
Step 4	`[0, 1, 2, 3]`	+4 updates	Position 3	`[4]`	`[4]`	5 / 8
Step 5	`[0, 1, 2, 3, 4]`	+4 updates	Position 4	`[5]`	`[5]`	6 / 8
Step 6	`[0, 1, 2, 3, 4, 5]`	+4 updates	Position 5	`[6]`	`[6]`	7 / 8
Step 7	`[0, 1, 2, 3, 4, 5, 6]`	+4 updates	Position 6	`[7]`	`[7]`	8 / 8
Step 8	`[0, 1, 2, 3, 4, 5, 6, 7]`	+4 updates	Position 7	`[8]`	`[]`	8 / 8

📜 Provenance & Pinned Input Hashes

Provenance Token:

arxiv:2510.14961v1+pdf-sha256:74e7985abe41ee2a75914a65e3778a15353fb0c0964d6ea34e7bfeb1f18312c8+source-sha256:60a795d123a2d2d642971834b6e0cba6dda80b5dfcd539f78d01639582d9c41d+github:seal-rg/recurrent-pretraining@1ea7220ec7eb42d13e89db0663df254d0bcdc28e+git-blob:recpre/raven_modeling_minimal.py@0e83a0766644df9113a8923f43350c6a1b5a182c

Input & Claim Hashes:

• Claim 1 SHA-256: d0da87ee16f7485d3dff369e7465f66299c55ac003a54e1cf8c00b3a0ad8b265

• Claim 2 SHA-256: 2e15221c8b5516b0ab705e29a3d7c5d924ed5f0187c970a0caf60a1402757804

• arxiv_submission.tex SHA-256: cdc058830d1e51f631e4fb8d1f2de0b79de91670fd4111646fe624f8c258d3b8

• raven_modeling_minimal.py SHA-256: 18fcacd53fb5696a76c0d3bda44480f2f3900aa9659c137a08962c593a9a9e42

• raven_modeling_minimal.py Git Blob: 0e83a0766644df9113a8923f43350c6a1b5a182c

• LICENSE (Apache-2.0) SHA-256: bc6c264d8ba4450599cf95c4699c6b82142f32ca1ecd91011c17b50a5a36a2f5

• ATTRIBUTION.md (CC-BY-4.0) SHA-256: 79775b50c72988b90eae75ef87e9d4df9dbd0bfceefaed60b398656a88d8a735

📥 Download Evidence Bundles

📄 manifest.json 📄 claim-1-wavefront.json 📄 claim-2-theorem-audit.json 📄 results.json 📝 REPORT.md

⚠️ Scope & Limitations

No Model Execution: 3.5B Huginn-0125 weights were not loaded or run.
No GPU Benchmarking: Modern GPU hardware timing and reported ~5x speedups were not measured.
Audit Only: Evidence statuses reflect static AST, schedule, and TeX audits.