- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes C3 (Candidate-specific Context Coupling) for preference optimization of discrete diffusion models. C3 couples the reveal progress, reveal counts, and Gumbel ordering keys across preference candidates while preserving each candidate's marginal distribution over selected contexts, with low-confidence priorities determining which tokens serve as conditioning. The paper proves Kendall-optimal ordering and minimization of a preference-discrepancy upper bound under stated assumptions, and evaluates the method on DNA sequence design, protein inverse folding, and language-model alignment.
Critical strengths
- The separation of candidate marginals from their coupling is principled. The shared-Gumbel proof gives compatible full permutations attaining all pairwise Kendall minima, and the loss-bound argument is clearly scoped.
- Matched ablations distinguish the priority choice, shared budgets, and shared orders. Fixed-checkpoint loss, variance, and complete-gradient diagnostics provide evidence beyond aggregate generation scores.
- The results span three domains and include useful complementary metrics. Protein joint success improves from SDPO's 75.5% to 79.7%, and the paper explicitly reports the accompanying structural trade-off.
Critical weaknesses
- Most DNA and protein baseline results are imported from prior work, whereas C3 uses its own training and scoring protocol. In addition, the protein component ablations cover nine epochs while the main result includes two continuation epochs (Appendices C.2 and D.1). These choices complicate attributing the headline improvement to context coupling alone.
- Several of the incremental gains need stronger uncertainty analysis. Protein order sharing improves joint success from 77.48 +/- 0.37% to 78.51 +/- 1.04% over three seeds, and the language improvements over the best baselines are small and single-run, including 0.49 points in LC-WR and 0.61 points on HumanEval (Table 3).
- The extra completed-candidate confidence evaluation is specified, but its wall-clock and memory overhead are not quantified. Equal-compute comparisons would clarify the practical value of the incremental gains. Moreover, the theory optimizes an upper bound at fixed priorities rather than the actual evolving training objective, so empirical validation remains essential.
The marginal-preserving construction, the sound fixed-priority analysis, and the informative component experiments are valuable; better baseline matching, uncertainty estimates, and cost reporting would strengthen the empirical case.
Questions for authors
- How were the baseline training budgets, scoring rules, and selection protocols matched? Do the component gains persist under the full continuation schedule used for the main protein result?
- Can you report paired uncertainty for the order-sharing gain, and the stability of the small language gains across training seeds?
- What training-time and peak-memory overhead does the confidence pass add, and how do the results compare at equal total compute?
2 comments
w3&q3. Quantify the confidence pass’s wall-clock and peak-memory overhead and compare performance at equal total compute. The fixed-priority upper-bound analysis also needs empirical validation for training with evolving priorities.
Thank you for suggesting a direct assessment of the confidence-pass overhead and performance under a matched training budget. We have measured CUDA-synchronized training-update wall time and peak memory usage. For the single-confidence-pass implementation in Algorithm 1, DNA update time after warm-up increases from 157.3 to 193.2 ms (+22.8%). Peak allocated memory increases by only 0.38 MiB, while peak reserved memory remains 3532 MiB. The confidence pass requires no backpropagation.
Separately, using the main-result continuation implementation, we completed a DNA comparison under a matched training-time budget, allowing the confidence-free control to spend the saved time on additional updates:
| Method | Updates | Cumulative update wall time (s) | Pred-Activity ↑ | ATAC-Acc (%) ↑ |
|---|---|---|---|---|
| Uniform, without confidence | 110–114 | 21.36 | 6.330 (0.062) | 97.97 (0.63) |
| C3 | 50 | 21.21 | 6.411 (0.041) | 98.23 (0.33) |
Values are means (sample SDs) across three training seeds. C3 achieves higher mean activity and accessibility even when the control receives more than twice as many updates. The mean gains are 0.081 in Pred-Activity and 0.26 percentage points in ATAC-Acc.
Regarding the theory, training directly optimizes the sampled preference loss, as specified in Algorithm 1. Priorities are recomputed from the current model before each update and held fixed within that update (Section 4.4). Proposition 1 gives an exact conditional ordering guarantee; Corollary 1 connects it to an upper bound on the additional preference loss.
The paper already tests the actual loss and gradient beyond this bound. Section 5.4 and Appendix D.3, Table 6 show that full sharing reduces comparison loss from 0.4328 to 0.3494; order sharing alone reduces it from 0.3607 to 0.3494. Appendix D.5, Table 7 compares complete gradients averaged over 48 context draws with the context-averaged reference gradient: mean cosine similarity across the five candidate collections increases from 0.419 to 0.665, with improvement in every collection. These measurements directly evaluate the preference loss and mean update direction.
Thank you for your careful review and for recognizing the principled coupling construction, the clearly scoped theoretical analysis, and the complementary evidence from component diagnostics and experiments in three domains. Following your suggestions, we have added full-schedule protein controls, paired analyses across training seeds, and training-time and peak-memory measurements.
Thank you for suggesting that we align the ablations with the main training schedule and clarify the scoring and selection protocols. The policy–reference log-probability scores defined in Section 3 are used for preference training, while evaluation follows the same benchmark metrics and success criteria as SDPO (Appendices C.1–C.2). Tables 1–2 identify the imported baseline results, which retain their published training budgets and selection procedures.
To isolate component contributions, all six protein configurations now follow the main result’s 9+2-epoch schedule across three training seeds. They share the same initialization, data, loss, and stage-specific learning-rate and scoring settings, together with the same decoder, sample counts, and evaluation procedure, without reward-based filtering or reranking. Revised Appendix D.1 explicitly states the full schedule and matched settings; updated Table 4 reports the complete results.
Low-confidence selection increases Pred-ddG from 0.858 to 1.644 and Success Rate by 13.22 percentage points over uniform selection, and exceeds high-confidence selection by 28.24 points in Success Rate. Halving the priority strength retains 79.38% Success Rate, close to 79.73% at the default strength. Order sharing alone, with low-confidence selection and the sampling of t (the expected visible fraction) and visible-token counts unchanged, improves Success Rate by 2.71 points and the fraction with scRMSD < 2 Å by 4.33 points. Full context sharing improves Success Rate by 2.98 points and entropy by 7.68 bits over fully independent contexts. Thus, the gains from low-confidence selection and context coupling persist under the full training schedule.
We appreciate your suggestion to examine paired uncertainty and the stability of the language gains across training seeds. We have added a paired uncertainty analysis of the full 9+2-epoch results in revised Appendix D.1, Table 4, extending the original nine-epoch comparison cited in your review. This comparison changes only whether the ordering random values are shared, retaining the same low-confidence selection rule and shared sampling of t and visible-token counts.
The three seed-wise gains are +3.06, +2.44, and +2.64 points: order sharing improves Success Rate in every seed, and its paired interval lies entirely above zero.
For language, we have completed new paired experiments with training seeds 42, 43, and 44 against Diffu-GRPO, the strongest baseline on LC-WR and HumanEval in Table 3. Within each benchmark, both methods use the same generation and scoring settings across seeds. These supplementary evaluations are reported separately from the original single-run Table 3.
LC-WR improves in all three training seeds; HumanEval improves in two seeds and ties in one. Their mean gains are 0.817 and 2.439 points, respectively, with paired 95% intervals of [−1.162, 2.796] and [−3.620, 8.498]; both intervals include zero.
± denotes sample SD across training seeds; 95% CIs are paired t-intervals computed from within-seed differences.