| # Reproduction: "Beyond Confidence: Adaptive and Coherent Decoding for Diffusion Language Models" |
|
|
| Independent reproduction of ICML 2026 paper #6883 (OpenReview `b0O96emqNj`, arXiv |
| [2512.02044](https://arxiv.org/abs/2512.02044)). **No official code was released**, so |
| everything here is written from the equations in Section 3. |
|
|
| ## Layout |
|
|
| | path | what | |
| |---|---| |
| | `scripts/ccd_decode.py` | CCD + CCD-DS decoder for Dream (Eq. 1, 6, 16, 17, 18, 20) | |
| | `scripts/run_eval.py` | Trip Plan / HumanEval driver, Dream's official hyperparameters | |
| | `scripts/check_propositions.py` | Claim 2: exact numerical audit of Prop. 1 and Prop. 2 | |
| | `scripts/check_budget_bound.py` | the `k <= V/(d+1)` speedup ceiling + simulation | |
| | `scripts/smoke_test.py` | decoder mechanics on a tiny random Dream (CPU) | |
| | `scripts/test_scorers.py` | metric validation against gold answers (CPU) | |
| | `scripts/make_figures.py` | logbook figures + raw CSVs | |
| | `scripts/job_*.sh` | the exact HF Jobs run scripts | |
| | `data/trip_planning.json` | Trip Plan benchmark (from DreamLM/Dream `eval/data/`) | |
| | `outputs/` | all results, per-example scores/steps/responses, figures | |
|
|
| ## Reproducing |
|
|
| ```bash |
| pip install torch "transformers==4.46.2" "huggingface_hub<1.0" "datasets<4" accelerate |
| # free, no GPU: mechanism + metric gates + both analytical claims |
| python scripts/smoke_test.py |
| python scripts/test_scorers.py |
| python scripts/check_propositions.py |
| python scripts/check_budget_bound.py |
| # GPU (~24GB): one benchmark config |
| python scripts/run_eval.py --task trip --method ccd_ds --limit 64 --out outputs/x.json |
| ``` |
|
|
| ## Determinism / seeds |
|
|
| | component | status | |
| |---|---| |
| | `check_propositions.py`, `check_budget_bound.py` | Seeded (`np.random.default_rng(0)`). Bit-identical across runs — verified 3x. | |
| | Example selection (`load_trip`) | Seeded, fixed at `seed=0`, independent of `--seed`, so every arm scores the same stratified sample. | |
| | Decoding at **T=0** (Trip Plan, HumanEval, all ablations) | Deterministic: `_pick_token` takes the argmax, no RNG involved. | |
| | Decoding at **T>0** (the Claim 6 sweep, T=0.1/0.4/0.7/1.0) | Samples via `Categorical.sample()`. `run_eval.py --seed N` (default 0) now seeds `random`/`numpy`/`torch`, and the seed is recorded in each result JSON. | |
|
|
| **Honest caveat:** `--seed` was added *after* the runs in `outputs/` were produced. The T=0 |
| results (the large majority, including every number on the poster's ceiling, Trip Plan, |
| HumanEval and ablation cards) are argmax and reproduce exactly regardless. The eight |
| temperature-sweep runs at T>0 (`c6_*_t0.1/0.4/0.7/1.0`) predate the seeding and therefore will |
| **not** reproduce bit-exactly; re-running them under `--seed 0` is the first thing to do with a |
| fresh GPU budget. At n=16 those points are our weakest evidence anyway, and the poster says so. |
|
|
| ## Hyperparameters |
|
|
| Taken from the Dream authors' own eval scripts, which is what the paper says it does |
| ("we follow the base models' default settings without tuning"): |
|
|
| * Trip Plan: `steps=256, max_new_tokens=256, temperature=0, top_p=1, alg=entropy`, 2-shot |
| (`DreamLM/Dream eval/eval_dream_gen_planning.sh`) |
| * HumanEval: `steps=768, max_new_tokens=768, temperature=0.1, top_p=0.9, alg=entropy`, 0-shot |
| chat template (`DreamLM/Dream eval_instruct/eval.sh`) |
|
|
| CCD: `V=4`, `d=3` for Dream (paper Sec. 4.2). |
|
|
| ## Gotchas found |
|
|
| * The Trip Plan file is **ordered by difficulty** (first 200 examples are all |
| `num_cities=3`). Prefix sampling inflates the score ~55% vs ~15%. Use the |
| stratified sampler in `load_trip`. |
| * HumanEval responses are the function **body only** (the prompt's `gen_prefix` |
| already contains the signature), so the `\ndef` stop sequence must be applied to |
| the continuation, not to prompt+continuation. |
|
|
