| --- |
| title: Top-W Geometry-Aware Decoding Reproduction |
| emoji: 🧠 |
| colorFrom: blue |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: 5.16.0 |
| python_version: "3.10" |
| app_file: app.py |
| pinned: false |
| tags: |
| - icml2026-repro |
| - paper-HSuU4xBmAv |
| --- |
| |
| # Reproduction for "Geometry-Aware Decoding with Wasserstein-Regularized Truncation and Mass Penalties for Large Language Models" |
|
|
| Paper ID: `HSuU4xBmAv` |
| Attempt ID: `c1b6dd10-f227-4d24-89a0-17fb00ec9147` |
|
|
| ## Target Claims |
| 1. Top-W decoding selects token subsets by optimizing a Wasserstein-entropy-mass objective using embedding-induced geometry (Section 3, Algorithm 1). |
| 2. The method instantiates a practical alternating decoder with an exact subset-update step inside a candidate-pool loop (Section 4.2). |
| 3. Top-W is evaluated against Min-p, Top-p, and Top-H on GSM8K across multiple temperatures and models (Table 1). |
|
|
| ## Logbook |
|
|
| The claim-by-claim evidence lives in `pages/*.md` and is rendered by the |
| Space app. Claims 1 and 2 are numerical audits of the mechanism: the |
| exact prefix-form S-step (Theorem 3.4a) is checked against brute-force |
| enumeration of all 1023 subsets per instance and against the |
| byte-exact vendored official implementation |
| (`evidence/inputs/upstream/`, pinned in |
| `evidence/inputs/upstream_manifest.json`). Claim 3 (GSM8K benchmark |
| table) is **not reproduced** — no language model was run. |
|
|
| ## Reproduce |
|
|
| ```bash |
| uv run --project . python generate_evidence.py |
| uv run --project . python -m pytest tests -q |
| ``` |
|
|
| CPU-only, zero metered API cost, fixed seeds; `evidence/bundle.json` is |
| the machine-readable output backing every number on the pages. |
|
|
