README.md

metadata

title: Recurrent-Depth Parallel Sampler Reproduction
sdk: static
app_file: index.html
tags:
  - icml2026-repro
  - paper-h7WBYYJF1Q

Reproduction: Efficient Parallel Samplers for Recurrent-Depth Models

This directory contains the reproduction project for: "Efficient Parallel Samplers for Recurrent-Depth Models and Their Connection to Diffusion Language Models" (OpenReview ID: h7WBYYJF1Q, arXiv: 2510.14961v1).

Attempt ID: 534db42c-5b16-4f00-9a7d-a47056fc9dd4

Target Claims & Evidence Status

1. Claim 1 (Wavefront Sampler Mechanism):

   • Text: "The sampler decodes new tokens every forward pass while refining latent states for those tokens in parallel through recurrent depth (Section 3.1)."
   • Status: partial
   • Findings: Source AST inspection of recpre/raven_modeling_minimal.py confirms generate() dispatches to generate_diffusion_style(). The control flow executes inner_recurrence steps per outer iteration across active latent states, decodes logits for the active wavefront, appends new positions, and bounds active width via max_wavefront. However, decoding occurs once after the inner-recurrence loop per outer step rather than after each single inner step.

2. Claim 2 (Expressiveness Theorem):

   • Text: "The paper proves the sampler is strictly more expressive than baseline autoregressive generation under the same time budget on modern hardware (Theorem 4.2)."
   • Status: unavailable
   • Findings: Citation audit of arxiv_submission.tex reveals a citation mismatch: Theorem 4.2 addresses prefilling depth vs width scaling, while the same-runtime decoding result is Theorem 4.4 (conditional on $r > 1$, KV sharing, $W \le L_*$). The released source does not contain an independently checkable proof for strict hardware-dependent expressiveness.

Running Evidence Generation & Tests

Generate evidence artifacts:

uv run --project submissions/efficient-parallel-samplers-for-recurrent-depth-models \
  python -m recurrent_sampler_repro.evidence \
  --project-root submissions/efficient-parallel-samplers-for-recurrent-depth-models

Run unit tests:

uv run --project submissions/efficient-parallel-samplers-for-recurrent-depth-models \
  python -m pytest \
  submissions/efficient-parallel-samplers-for-recurrent-depth-models/tests

Structure

• src/recurrent_sampler_repro/: Core verification logic, AST parser, schedule simulator, theorem auditor, and evidence generator.
• tests/: Complete test suite covering claim bindings, provenance, source audit, schedule invariants, theorem audit, output determinism, and static Space assets.
• vendor/: Pinned immutable upstream source files.
• evidence/: Generated deterministic evidence JSON bundles and report.
• space/: Static Hugging Face Space application assets.