Beyond Confidence: Adaptive and Coherent Decoding for Diffusion Language Models [ A Reproduction ]

The idea reproduces. The speedup it is packaged with is structurally impossible at the paper's own hyperparameters.

Independent reproduction of ICML 2026 paper #6883 (OpenReview b0O96emqNj · arXiv 2512.02044) · no official code released — reimplemented from the equations · Dream-7B-Instruct · 1×L40S, ~7 GPU-h, ~$13

code, per-example
outputs & job scripts

Verdict CCD's quality gain reproduces (+4.69 vs +4.65 claimed). Its advertised 3.48× speedup cannot exist at the stated V=4, d=3 — a ceiling k ≤ V/(d+1), derived from the paper's own Eqs. (16)–(17), caps decoding at exactly 1 token per step. Widening the buffer to V≈16 buys back the reported speed, and spends the quality gain to do it. The two are each reachable alone, never together.

1What was reproduced

CCD replaces a diffusion LM's single-step confidence with a marginal p̄ averaged over a sliding buffer of the last d iterations, keeping the top-V tokens (Eq. 6/16/17). CCD-DS spends that signal as an adaptive unmasking budget (Eq. 20).

No code was released. We reimplemented from the equations as a drop-in replacement for Dream's _sample, copying its confidence metric, logit shift and budget schedule verbatim — the arms differ only in which distribution a token is drawn from.

2The harness is faithful

Three independent baselines land on the paper's numbers — this is what licenses every verdict below.

baseline	paper	ours
Trip Plan	15.10	15.62
HumanEval	52.66	53.12
Trip City=3	58	55.0

Hyperparameters are Dream's own eval scripts — we tuned nothing.

3Bonus: Dream's own recipe is broken

Dream's published HumanEval setting (T=0.1, top_p=0.9) emits 768/768 EOS under Dream's own diffusion_generate — yet this paper and the Dream paper both report 52.66 from it.

Cause: T<1 scales logits ×1/T and top_p<1 keeps only the argmax, so negative entropy is exactly 0 at 242/256 positions and the ranking ties. Only the conjunction is fatal — we ran HumanEval at T=0.

4Scope, cost & what's next

~7 GPU-h on one L40S, ~$13, against ~250 GPU-h (~$450) to run all of Table 1 at full size — spent on the claims that discriminate.

Honest limits. Scores ran at n=64 (Trip, HumanEval), n=40 (ablation), n=16 (temperature) — underpowered, so we report McNemar paired tests, not point estimates, and call Claim 3's score inconclusive rather than refuted. Next: full HumanEval (n=164, ~$6) would take CCD's 3–0 to p≈0.008.

5Theory: Prop. 1 ✓, Prop. 2 ✗

Prop. 1 — H(x|s) = H(x|c,s) + I(x;c|s) — is a correct identity, exact to 2.2e-15 over 2000 seeded random joints.

Prop. 2 does not follow. Lemma 1 is an upper bound; Eq. (12) swaps its RHS for a trajectory-averaged MI that is strictly smaller (0.191 vs 0.669 bits in our counterexample) while keeping the ≤.

Worse, its RHS is sampler-independent — it depends on the data and the trained model, not on how tokens are picked. The only lever is G/T, so CCD-DS's fewer steps make the bound ~3.4× looser — the opposite of governing it.

6A ceiling on the speedup

Let S_u be the top-V set at step u. Eq. (17) intersects the last d+1 of them; every token decoded in that window came from S_t−d and never returns, which in steady state forces:

|I^c_t| ≤ V − Σk ⇒ k ≤ V/(d+1)

Speedup is mean tokens/step, so V/(d+1) caps it. At the paper's V=4, d=3: 4/4 = 1.0× — no speedup is possible. The cap follows from position bookkeeping alone; no model or benchmark enters.

7Measured: the budget starves

The stability heuristic — the part the paper describes tuning — passes on 96.1% of candidates. It is not the bottleneck; the intersection is.

8The ceiling predicts, 8/8

It holds on the d axis too: d=1→1.95 (predicted 2.00), d=2→1.31 (1.33), d=3→1.01 (1.00). At V=1–2 CCD-DS degenerates exactly to baseline: 256.0 steps, 99.6% fallback, identical score.

9V≈16 reconciles the paper

The bound says 3.48× needs V = 3.48×4 ≈ 14. We ran V=16 — nothing else changed — and two independent reported quantities reappear:

	paper	V=4	V=16
CCD Δ	+1.83	−1.56	+1.56 ✓
DS steps	75.2	252.1	70.3 ✓
DS speedup	3.48×	1.02×	3.64× ✓

On HumanEval too: the bound says 3.04× needs V≥12.2; V=12 gives 2.89×. The published V=4 appears to fit neither benchmark.

10Claim 3 — Trip Plan

arm	paper	ours	speedup
baseline	15.10	15.62	1.00×
+ CCD	16.93	14.06	1.00×
+ CCD-DS	19.01	14.06	1.02×
CCD-DS V=16	—	12.50	3.64×

Speedup refuted: 1.02× vs 3.48× claimed (k=1.015 against the bound's 1.00). The score is inconclusive, not refuted — only 1 of 64 examples changed (McNemar p=1.000), though CCD drove 77% of steps and rewrote 69% of the plans.

11Claim 4 — CCD reproduces

The paper's headline quality claim lands to within 0.04 points:

HumanEval, n=64	paper	ours
baseline	52.66	48.44
+ CCD Δ	+4.65	+4.69 ✓
+ CCD-DS speedup	3.04×	1.01× ✗

Unanimous but underpowered: 3 of 64 improved, none worsened, p=0.250. CCD drove 96% of steps and rewrote 33% of programs.

12Claim 5 — buffer ablation

The reported "95.54 steps at buffer 4" sits between our V=8 (132.8) and V=16 (68.6) — V≈11 again.

13Claim 6 — temperature

pass@1 vs sampling temperature, baseline vs CCD-DS

At T=0, 50.0 vs 43.75. Caveat: n=16, and T=0.1/0.4/0.7 sit inside the collapse regime — our weakest test.